From f60db1f0d6029aa59b42c651938e8875a1132ee2 Mon Sep 17 00:00:00 2001 From: zy95-12 <135135297+zy95-12@users.noreply.github.com> Date: Fri, 6 Nov 2026 09:55:31 +0800 Subject: [PATCH 1/3] Bind Chief next delivery to its validated invocation snapshot --- PROJECT.md | 2 +- ROADMAP.md | 2 +- changelog.d/2560.fixed.md | 1 + core/src/host/project-chief-control.js | 49 ++- core/src/host/project-chief-protocol.js | 4 +- core/src/host/project-delivery-options.js | 2 +- core/test/project-chief-control.test.js | 94 ++++++ core/test/project-delivery.test.js | 11 +- ...2026-11-06-chief-continuation-reference.md | 13 + docs/development-service-api.md | 18 +- docs/openapi/flybot.openapi.json | 288 +++++++++++++++++- docs/subsystems/backend-service.md | 2 + docs/subsystems/project-chief-control.md | 14 +- docs/subsystems/project-delivery.md | 6 +- .../files/src/host/project-chief-control.js | 49 ++- .../files/src/host/project-chief-protocol.js | 4 +- .../src/host/project-delivery-options.js | 2 +- .../files/test/project-chief-control.test.js | 94 ++++++ .../files/test/project-delivery.test.js | 11 +- scripts/api-contracts.mjs | 3 + 20 files changed, 614 insertions(+), 55 deletions(-) create mode 100644 changelog.d/2560.fixed.md create mode 100644 decisions/2026-11-06-chief-continuation-reference.md diff --git a/PROJECT.md b/PROJECT.md index 0765d62b..78fe8197 100644 --- a/PROJECT.md +++ b/PROJECT.md @@ -143,6 +143,6 @@ Detailed dated implementation checkpoints are [preserved verbatim](docs/history/ #2519: SDK Reviewer receives the complete hash-verified host log inline, without scratch access. Preserve exact candidate/current authority, artifact and execution-context budgets; persist only provenance metadata. Host output is data, never Review or CI approval. -#2526: Chief may settle a merged Attempt under its pinned requirements after Task revisions; retain both versions and keep Task open. Re-read changes during judgement. Next delivery uses current requirements. Chief decides; Personal Agent help is the last resort. +#2526/#2560: Settle pinned Attempt scope after Task revisions; Task stays open. Chief chooses current-goal next delivery with a validated snapshot reference. Evidence/identity/rights/exit remain gates; human help is Chief's last resort. #2526: same-scope repeated Chief action refusals reuse task-attention warnings and the configured no-progress threshold. Current exact failures reach Chief; changed delivery facts retire the old warning. Warnings never choose roles or escalate to Personal Agent automatically. diff --git a/ROADMAP.md b/ROADMAP.md index 79a46663..4c7d53cb 100644 --- a/ROADMAP.md +++ b/ROADMAP.md @@ -103,4 +103,4 @@ privilege or mission scope and complete no outstanding operational acceptance. #2519 removes the host-log scratch dependency through sealed inline evidence. Verify full input, configured context refusal and read-only SDK behavior separately from live OpenToday acceptance. -#2526 separates pinned-scope settlement, current-goal next delivery and whole-Task closure. Verify commit failure/replay, stale facts, unknown exit, quota admission and Task-role continuity; no automatic Developer or human route. +#2526 separates pinned-scope settlement, current-goal next delivery via validated references (#2560), and whole-Task closure. Verify commit failure/replay, stale facts, unknown exit, quota admission and Task-role continuity; no automatic Developer or human route. diff --git a/changelog.d/2560.fixed.md b/changelog.d/2560.fixed.md new file mode 100644 index 00000000..9ebec64e --- /dev/null +++ b/changelog.d/2560.fixed.md @@ -0,0 +1 @@ +Chief can start the next delivery by referencing its observed continuation facts, without copying workspace and SHA fields; ambiguous targets, stale facts and explicitly incorrect fields still require correction, and actual bindings remain auditable. diff --git a/core/src/host/project-chief-control.js b/core/src/host/project-chief-control.js index 5ec6989b..d01e78dd 100644 --- a/core/src/host/project-chief-control.js +++ b/core/src/host/project-chief-control.js @@ -85,7 +85,15 @@ export function createProjectChiefControl(rt) { function view(projectId, actor) { rt.requireProjectAccess(actor, projectId) const policy = policyOf(projectId), logical = policy?.scheduleId ? rt.state.projectScheduledTasks?.[policy.scheduleId] : null - return { warnings: rt.projectWarnings?.summaries(projectId) ?? [], compensation: compensationSchedules(rt.state, projectId, policy?.scheduleId), health: inspectionHealth(rt.state, projectId), policy: structuredClone(policy ?? { enabled: false }), template: structuredClone(template), schedule: logical ? { id: logical.id, enabled: logical.enabled, status: logical.status, config: structuredClone(logical.config), deliveries: Object.values(logical.deliveries).slice(-5).map(row => structuredClone(row)) } : null, runs: Object.values(rt.state.projectExecutions ?? {}).filter(run => run.projectId === projectId && run.kind === 'chief-control').slice(-30).map(({ capability: _capability, ...row }) => structuredClone(row)) } + const runs = Object.values(rt.state.projectExecutions ?? {}).filter(run => run.projectId === projectId && run.kind === 'chief-control').slice(-30).map(({ capability: _capability, ...row }) => { + const copy = structuredClone(row) + for (const applied of copy.appliedDecisions ?? []) { + const audit = rt.state.projectChiefDecisions?.[applied.id] + if (audit?.projectId === projectId && audit.runId === copy.id && audit.taskId === applied.taskId && audit.continuationBinding) applied.continuationBinding = structuredClone(audit.continuationBinding) + } + return copy + }) + return { warnings: rt.projectWarnings?.summaries(projectId) ?? [], compensation: compensationSchedules(rt.state, projectId, policy?.scheduleId), health: inspectionHealth(rt.state, projectId), policy: structuredClone(policy ?? { enabled: false }), template: structuredClone(template), schedule: logical ? { id: logical.id, enabled: logical.enabled, status: logical.status, config: structuredClone(logical.config), deliveries: Object.values(logical.deliveries).slice(-5).map(row => structuredClone(row)) } : null, runs } } async function configure(body, actor) { rt.requireProjectAccess(actor, body.projectId, 'configure') @@ -173,15 +181,26 @@ export function createProjectChiefControl(rt) { })) } async function continuationChoice(decision, saved, roles, actor, projectId) { - if (decision.continuation === undefined) return null - const choice = decision.continuation, facts = saved?.repositoryRecovery?.continuationOptions - const target = roles.find(row => row.id === decision.targetBotId) - if (!['route', 'return', 'start-next'].includes(decision.action) || (target?.policyRole ?? target?.role) !== 'developer' || decision.repositoryInputs !== undefined) throw chiefProtocolError('continuation', 'an explicit Developer route/return, separate from repositoryInputs admission', decision.taskId) - if (!facts || facts.status === 'unavailable' || !(decision.action === 'start-next' ? ['next'] : ['continue', 'latest']).includes(choice?.mode) || choice.previousAttemptId !== facts.attemptId || choice.generation !== facts.generation || !isDeepStrictEqual(choice.expectedWorkspace, facts.snapshot) || ['latest', 'next'].includes(choice.mode) && choice.expectedBaseSha !== facts.latestBaseSha) throw chiefProtocolError('continuation', 'the exact current continuationOptions snapshot, generation and selected latestBaseSha', decision.taskId) + const next = decision.action === 'start-next' + if (!next && decision.continuation === undefined) return null + const facts = saved?.repositoryRecovery?.continuationOptions + const implementations = roles.filter(row => (row.policyRole ?? row.role) === 'developer') + const target = next && decision.targetBotId === undefined && implementations.length === 1 ? implementations[0] : roles.find(row => row.id === decision.targetBotId) + if (!['route', 'return', 'start-next'].includes(decision.action) || (target?.policyRole ?? target?.role) !== 'developer' || decision.repositoryInputs !== undefined) throw chiefProtocolError('targetBotId', 'one active implementation role; choose explicitly when more than one is available, separate from repositoryInputs', decision.taskId) + let choice = decision.continuation + if (next && decision.continuationRef !== undefined && (typeof decision.continuationRef !== 'string' || decision.continuationRef !== facts?.expectedScope)) throw chiefProtocolError('continuationRef', 'the expectedScope of the continuationOptions read in this invocation', decision.taskId) + const source = choice === undefined ? decision.continuationRef === undefined ? 'chief-snapshot' : 'continuation-ref' : 'explicit' + if (next && choice === undefined) { + if (!facts || facts.status === 'unavailable' || facts.attemptId !== saved.attemptId || !Array.isArray(facts.modes) || !facts.modes.includes('next') || decision.continuationRef === undefined && (facts.modes.length !== 1 || facts.modes[0] !== 'next') || typeof facts.expectedScope !== 'string' || facts.expectedScope !== saved.deliverySnapshot?.expectedScope || saved.repositoryRecovery?.nextActions?.actions?.find(row => row.action === 'start-next')?.status !== 'ready') throw chiefProtocolError('continuation', 'current ready next-delivery facts with one unambiguous mode, or an explicit selection/reference', decision.taskId) + choice = { mode: 'next', previousAttemptId: facts.attemptId, generation: facts.generation, expectedWorkspace: structuredClone(facts.snapshot), expectedBaseSha: facts.latestBaseSha } + } + if (!facts || facts.status === 'unavailable' || !(next ? ['next'] : ['continue', 'latest']).includes(choice?.mode) || choice.previousAttemptId !== facts.attemptId || choice.generation !== facts.generation || !isDeepStrictEqual(choice.expectedWorkspace, facts.snapshot) || ['latest', 'next'].includes(choice.mode) && choice.expectedBaseSha !== facts.latestBaseSha) throw chiefProtocolError('continuation', 'the exact current continuationOptions snapshot, generation and selected latestBaseSha', decision.taskId) const response = await rt.projectDelivery.endpoint('project-attempt-continuation-state', { projectId, attemptId: facts.attemptId }, actor) const { nextActions: _nextActions, ...current } = response if (!isDeepStrictEqual(current, facts)) throw chiefProtocolError('continuation', 'fresh facts after the repository baseline or workspace changed', decision.taskId) - return target + const continuation = { mode: choice.mode, previousAttemptId: choice.previousAttemptId, generation: choice.generation, expectedWorkspace: structuredClone(choice.expectedWorkspace) } + if (['latest', 'next'].includes(choice.mode)) continuation.expectedBaseSha = choice.expectedBaseSha + return { target, continuation, source } } async function applyDecision(job, decision) { const { run } = job @@ -193,11 +212,15 @@ export function createProjectChiefControl(rt) { if ((task.collaboration?.generation ?? 0) !== saved.generation || task.deliveryAttemptId !== saved.attemptId && (task.deliveryAttemptId ?? null) !== saved.attemptId || (rt.state.projectAttempts?.[task.deliveryAttemptId]?.candidateSha ?? null) !== saved.candidateSha) throw refuse('Task progressed after the Chief snapshot; decision is stale') if (saved.deliverySnapshot && (saved.deliverySnapshot.status !== 'current' || saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task))) throw chiefProtocolError('snapshot', 'fresh Task, Attempt, requirements and evidence facts before action', task.id) if (digest(rt.state.projectAttempts?.[task.deliveryAttemptId]?.evidence?.merge ?? null) !== digest(saved.evidence?.merge ?? null)) throw chiefProtocolError('snapshot', 'fresh facts after merge evidence changed', task.id) - const continuationTarget = await continuationChoice(decision, saved, rt.projectAgents.deliveryDefinitions(run.projectId).roles, actor, run.projectId) + const continuationBinding = await continuationChoice(decision, saved, rt.projectAgents.deliveryDefinitions(run.projectId).roles, actor, run.projectId) + if (continuationBinding) await validateActor(actor) + if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw chiefProtocolError('targetBotId', 'fresh role catalog after continuation validation', task.id) + if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw chiefProtocolError('snapshot', 'fresh Task facts after continuation validation', task.id) const reason = typeof decision.reason === 'string' ? decision.reason.trim() : '' if (!reason || reason.length > 4000 || !['close', 'route', 'return', 'request-tests', 'needs-human', 'continue', 'admit', 'revise', 'cancel', 'reopen', 'resolve-dependency', 'retry-notification', 'reconcile-exit', 'stop-execution', 'accept-attempt', 'start-next'].includes(decision.action)) throw refuse('Chief decision needs a supported action and bounded reason') if (decision.action === 'needs-human') { assistanceRequest(decision); if (decision.recipient === 'operations' && !operationsContact(rt, task.projectId, policyOf(task.projectId).operationsContact)) throw refuse('No operations contact is configured') } const audit = { id: `chief-decision-${randomUUID()}`, runId: run.id, projectId: run.projectId, taskId: task.id, initiatedBy: run.initiatedBy, executorRole: 'chief', action: decision.action, reason, snapshotSha256: run.snapshotSha256, candidateSha: saved.candidateSha, expectedScope: saved.deliverySnapshot?.expectedScope ?? null, criteria: structuredClone(decision.criteria ?? []), inputReceiptSummary: inputDeclaration(decision, saved), status: 'requested', invocation: job.invocation, at: Date.now() } + if (continuationBinding) audit.continuationBinding = { targetBotId: continuationBinding.target.id, source: continuationBinding.source, expectedScope: saved.deliverySnapshot?.expectedScope ?? null, continuation: structuredClone(continuationBinding.continuation) } rt.state.projectChiefDecisions ??= {}; setEntry(rt.state.projectChiefDecisions, audit.id, audit) actor = rt.projectInputs?.chiefActor(actor, task.id, saved, run, decision.inputReceipts) ?? actor try { @@ -238,8 +261,8 @@ export function createProjectChiefControl(rt) { } else if (decision.action === 'accept-attempt') { await rt.projectDelivery.chiefAccept({ taskId: task.id, expectedSha: saved.candidateSha, decision: { ...audit, delivery: decision.delivery } }, actor) } else if (decision.action === 'start-next') { - if (!continuationTarget || decision.continuation?.mode !== 'next') throw chiefProtocolError('continuation', 'a fresh next-delivery snapshot and explicit implementation target', task.id) - const result = await rt.projectScheduler.request({ taskId: task.id, requestId: audit.id, targetRole: continuationTarget.role, continuation: decision.continuation }, actor) + if (!continuationBinding || continuationBinding.continuation.mode !== 'next') throw chiefProtocolError('continuation', 'a fresh next-delivery snapshot and explicit implementation target', task.id) + const result = await rt.projectScheduler.request({ taskId: task.id, requestId: audit.id, targetRole: continuationBinding.target.role, continuation: continuationBinding.continuation }, actor) audit.successorAttemptId = result.attempt?.id ?? null; audit.queueId = result.queue?.id ?? null } else if (decision.action === 'close') { rt.projectTaskLifecycle?.requireCompletedDependencies(task) @@ -251,8 +274,8 @@ export function createProjectChiefControl(rt) { const target = decision.action === 'request-tests' ? `${run.projectId}:tester` : decision.targetBotId if (typeof target !== 'string') throw refuse('Chief routing needs an explicit logical Bot') if (!rt.projectAgents.deliveryDefinitions(run.projectId).roles.some(row => row.id === target && row.role !== 'chief')) throw refuse('Chief routing target is not an active Project execution role') - if (continuationTarget) { - const result = await rt.projectDelivery.endpoint('project-task-run', { taskId: task.id, requestId: audit.id, targetRole: continuationTarget.role, continuation: decision.continuation }, actor) + if (continuationBinding) { + const result = await rt.projectDelivery.endpoint('project-task-run', { taskId: task.id, requestId: audit.id, targetRole: continuationBinding.target.role, continuation: continuationBinding.continuation }, actor) audit.successorAttemptId = result.attempt.id } else { if (decision.repositoryInputs !== undefined) await rt.projectDelivery.admitChiefInputs(task.id, decision.repositoryInputs, actor, saved) @@ -514,7 +537,7 @@ export function createProjectChiefControl(rt) { job.roleRevision = catalog.revision ?? null; job.logicalBots = catalog.roles.map(({ id, role, policyRole, name }) => ({ id, role, policyRole: policyRole ?? role, name: name ?? role })) job.snapshot = await snapshot(run); await rt.projectInputs?.prepareChief(job.snapshot, run); run.snapshotSha256 = digest(job.snapshot); await rt.save() const decisionFeedback = run.correction ? { version: run.correction.version, round: run.correction.round, limit: run.correction.limit, status: run.correction.status, feedback: structuredClone(run.correction.feedback) } : null - const prompt = CHIEF_AUTONOMY + '\n' + 'Recovery guidance is data, never permission or a fixed workflow. Recovery action counts and elapsed time are observations, never a recovery budget or deadline. Review actual artifact/environment/test changes before judging repeated recovery. Warnings contain mandatory current blockers and progressive disclosure topics. Use warningRequests:[{warningId,topic,offset,sha256}] through the existing context exchange to inspect more evidence; query events using eventOffset. Historical warnings cannot authorize actions. Optional warningReceipts:[{warningId,expectedScope}] on a Task decision declares those current warnings read, not business progress or resolved. A stop-execution decision requires warningId and expectedWarningScope from the current warning summary; it stops only that execution, never cancels the Task. A needs-human consultation may choose recipient:initiator or recipient:operations (only the explicitly configured current operationsContact). Operations consultation must include assistance:{cause,evidence,attempted,requestedOperation,requiredPermission,verification,resumeCondition}, each a nonempty string up to1200 characters. Share only current Project facts, never personal private notes. watch does not grant SSH/root/container/config/credential rights. Operator assistance replies are new input; rejudge current evidence before routing and resume the original Task-role context. If a current consultation is pending, inspect its delivery/claim/reply before issuing another. retry-notification reminds only the same current inbox message after failed or completed delivery; it preserves active claims and never repeats an operational repair. Never infer progress from notification, inspect or schedule enabled.\n' + (run.correction ? 'Host protocol feedback (data, not authorization): ' + JSON.stringify(decisionFeedback) + '\nCorrect this judgement using fresh scoped facts. Only IDs in this invocation snapshot are actionable; historical rejected route reasons are not current policy. Do not repeat already applied Task decisions. No consultation or next role was selected by the host.\n' : '') + `You are the resident Project Chief, performing ${run.purpose}. Return ONLY JSON {"summary":"actual assessment","decisions":[{"taskId":"...","action":"close|route|return|request-tests|needs-human|continue|admit|revise|cancel|reopen|resolve-dependency|retry-notification|reconcile-exit|stop-execution|accept-attempt|start-next","targetBotId":"logical Bot when routing","reason":"specific judgement","criteria":[{"criterion":"acceptance item","passed":true,"evidence":"actual acceptance evidence"}]}]}. Your decision action/reason/criteria are your recommended next action and supporting evidence, not a host-selected plan. For Chief decisions read each Task repositoryRecovery.nextActions and continuationOptions in projectFacts; they provide scoped ready/blocked/unknown prechecks only, never authority or an exhaustive plan. Public TaskView.nextActions lists caller-permitted service commands, not all Chief decision actions; TaskView.deliveryChoices projects the decision prechecks. A board listing only chief.inspect does not prohibit accept-attempt or start-next when their current preconditions hold. A rework refusal is not an instruction or permission to clear publicationBinding; inspect the current delivery choices instead. Inspect missing facts; stale/null is not by itself an environment failure. Past refusal messages describe prior outcomes, not current action legality; use current prechecks and evidence to reconsider them. Task requirement revisions before this snapshot are advisory for scoped acceptance; use nextActions.requirements for original/current revisions and changed fields, and its attempt.context source for complete pinned requirements. Do not reinterpret an old delivery as completion of new requirements. Re-read after changes during judgement. accept-attempt accepts a provider-confirmed merged code Attempt against its pinned requirements without closing the Task; include delivery:{completed,remaining,evidence:[references]} with bounded specific scope. A committed acceptance wakes a fresh inspection, never automatically starts Developer. In that new snapshot choose start-next only when needed, with targetBotId for an active implementation role and continuation:{mode:'next',previousAttemptId,generation,expectedWorkspace,expectedBaseSha} copied from fresh continuationOptions. This pins verified target HEAD containing the merged result and creates a new PR identity while retaining Task-role context/history. Never rework or inherit the merged PR binding. close remains available only when all Task criteria/dependencies/inputs are satisfied. Personal Agent assistance is the last resort when actual information/authority or a human decision is missing; failure counts/time never automatically escalate. A conversation/turn completion is not Task completion. Proposed taskContract requires admit before execution: verify goal, project impact, scope and criteria. Prefer admit without repeating already adequate requirements. For admit/revise, every provided requirements.title, requirements.criteria, requirements.impact and requirements.scope must be a bounded nonempty string. In particular requirements.criteria is a string; the separate outer decision.criteria is an evidence array and must never be copied into requirements.criteria. Optional requirements.repository is an associated owner/name or null for a repository-free decision. Uncertain intent requires needs-human. Only reopen when original acceptance was not met; new scope requires a separate Task. Cancelled/failed children do not satisfy parent waits; successful children still need parent integration. All Task mutations remain scoped Chief decisions. To resolve a failed/cancelled child condition without faking child completion, use resolve-dependency with dependency:{taskId,version:the exact current dependency snapshot version,evidence:[actual references],replacementTaskId:optional accepted alternative,replacementVersion:its task outcome version}. This is explicit Chief judgement, not host test success or parent completion. Leave unresolved conditions blocked when evidence is insufficient. Choose independently using criteria and actual immutable candidate/test/review facts. reconcile-exit explicitly retires only an exact exit-proven old claim without selecting a next role; consult repositoryRecovery.executionCheckpoint. Historical refusals are audit data, not current routing policy. Candidate Reviewer approve triggers the host merge flow only when explicitly configured for this Project/repository/branch. It is not a signed repository Reviewer PASS, actual forge approval, merged result or release permit. A configured merge must be provider-confirmed before overall code Task acceptance; pending/blocked merge facts are not completion. Do not close missing or stale evidence. Host allocationProof.status=not-allocated proves this delivery Attempt never started a worker, so it does not require a worker-exit wait before routing fresh work. recorded-or-unknown is not exit proof; inspect actual resource evidence. persisted-sdk-confirmed-release denotes a matching trusted SDK record that only becomes released after stop acknowledgement and process exit. Choose an evidence-supported role after environment or execution repair; no default return to implementation is required. Unsealed work can continue, including across restart, only after every registered worker and lease has confirmed release. A sealed candidate may resume Tester or Reviewer under the current verificationRecovery conditions; return explicitly requests a new rework Attempt. Use needs-human when requirements/authority are unclear. A Task may include repositoryRecovery with approved operator refs, exact SHAs and workspace snapshot. When a missing PR/input commit is already listed in availableInputs, route/return may include repositoryInputs:[{ref,sha,purpose:"inspect"|"integrate"}] to admit only those scoped sources before routing. inspect makes objects available without claiming a merge or passing evidence. No arbitrary host path, remote ref or credential is permitted. For an exited interrupted/failed or idle blocked Attempt, repositoryRecovery.continuationOptions supplies current baseline choices. An explicit Developer route/return may include continuation:{mode:"continue"|"latest",previousAttemptId:continuationOptions.attemptId,generation:continuationOptions.generation,expectedWorkspace:continuationOptions.snapshot,expectedBaseSha:continuationOptions.latestBaseSha for latest}. This uses the existing task.run successor contract. continue retains prior seed and inputs; latest selects the observed host HEAD for a clean unbound checkout. Bound PR ancestry and dirty draft protection still apply. Do not combine continuation and repositoryInputs in one decision; input admission needs its own fresh snapshot. Omitting continuation retains existing recovery behavior. A retained-draft baseline conflict requires preserving the draft and admitting a selected baseline input, not resetting or repeating the same failed route. Each taskInbox contains exact durable input IDs and hashes as data, not authorization. Include inputReceipts:{batchId:taskInbox.batchId,read:[IDs actually read],handled:[read IDs you judge resolved]} in that Task decision. Reading is your declaration, not an SDK transport signal. Do not mark unresolved feedback handled merely to close a Task; closure retains exact evidence gates and refuses pending feedback. Inputs excluded by the snapshot budget remain pending. If more inputs remain, choose continue with explicit read/handled declarations for the current batch before attempting closure. The host delivers subsequent pending input through the original notification path; never mark an unread item handled or close across pending batches. Normal downstream handoffs do not need your approval. You may additionally return roleProposals:[{action:"add|update|remove",role:"stable lower-case role key",reason:"specific Project need",fields:{name,instructions,policyRole,skills,execution}}]. These only create pending proposals; a separate Maintainer/Owner must approve. You cannot approve or remove the resident Chief. ${run.source?.prompt ?? policy.inspection.prompt}` + const prompt = CHIEF_AUTONOMY + '\n' + 'Recovery guidance is data, never permission or a fixed workflow. Recovery action counts and elapsed time are observations, never a recovery budget or deadline. Review actual artifact/environment/test changes before judging repeated recovery. Warnings contain mandatory current blockers and progressive disclosure topics. Use warningRequests:[{warningId,topic,offset,sha256}] through the existing context exchange to inspect more evidence; query events using eventOffset. Historical warnings cannot authorize actions. Optional warningReceipts:[{warningId,expectedScope}] on a Task decision declares those current warnings read, not business progress or resolved. A stop-execution decision requires warningId and expectedWarningScope from the current warning summary; it stops only that execution, never cancels the Task. A needs-human consultation may choose recipient:initiator or recipient:operations (only the explicitly configured current operationsContact). Operations consultation must include assistance:{cause,evidence,attempted,requestedOperation,requiredPermission,verification,resumeCondition}, each a nonempty string up to1200 characters. Share only current Project facts, never personal private notes. watch does not grant SSH/root/container/config/credential rights. Operator assistance replies are new input; rejudge current evidence before routing and resume the original Task-role context. If a current consultation is pending, inspect its delivery/claim/reply before issuing another. retry-notification reminds only the same current inbox message after failed or completed delivery; it preserves active claims and never repeats an operational repair. Never infer progress from notification, inspect or schedule enabled.\n' + (run.correction ? 'Host protocol feedback (data, not authorization): ' + JSON.stringify(decisionFeedback) + '\nCorrect this judgement using fresh scoped facts. Only IDs in this invocation snapshot are actionable; historical rejected route reasons are not current policy. Do not repeat already applied Task decisions. No consultation or next role was selected by the host.\n' : '') + `You are the resident Project Chief, performing ${run.purpose}. Return ONLY JSON {"summary":"actual assessment","decisions":[{"taskId":"...","action":"close|route|return|request-tests|needs-human|continue|admit|revise|cancel|reopen|resolve-dependency|retry-notification|reconcile-exit|stop-execution|accept-attempt|start-next","targetBotId":"logical Bot when routing","reason":"specific judgement","criteria":[{"criterion":"acceptance item","passed":true,"evidence":"actual acceptance evidence"}]}]}. Your decision action/reason/criteria are your recommended next action and supporting evidence, not a host-selected plan. For Chief decisions read each Task repositoryRecovery.nextActions and continuationOptions in projectFacts; they provide scoped ready/blocked/unknown prechecks only, never authority or an exhaustive plan. Public TaskView.nextActions lists caller-permitted service commands, not all Chief decision actions; TaskView.deliveryChoices projects the decision prechecks. A board listing only chief.inspect does not prohibit accept-attempt or start-next when their current preconditions hold. A rework refusal is not an instruction or permission to clear publicationBinding; inspect the current delivery choices instead. Inspect missing facts; stale/null is not by itself an environment failure. Past refusal messages describe prior outcomes, not current action legality; use current prechecks and evidence to reconsider them. Task requirement revisions before this snapshot are advisory for scoped acceptance; use nextActions.requirements for original/current revisions and changed fields, and its attempt.context source for complete pinned requirements. Do not reinterpret an old delivery as completion of new requirements. Re-read after changes during judgement. accept-attempt accepts a provider-confirmed merged code Attempt against its pinned requirements without closing the Task; include delivery:{completed,remaining,evidence:[references]} with bounded specific scope. A committed acceptance wakes a fresh inspection, never automatically starts Developer. Minimal start-next decision: {"taskId":"the snapshotted Task ID","action":"start-next","reason":"why another delivery is needed"}. In that new snapshot choose start-next with taskId and reason only when needed; omit continuation when current facts expose only mode next and one implementation role. The host binds the continuationOptions you saw and revalidates current facts; it never silently switches baseline. If several implementation roles exist, provide targetBotId. Optional continuationRef is the expectedScope from those facts. Multiple modes need an explicit reference or full continuation selection. Supplied continuation fields must match exactly and are never repaired. Use existing attempt.continuation-state after feedback to inspect refreshed facts. This pins verified target HEAD containing the merged result and creates a new PR identity while retaining Task-role context/history. Never rework or inherit the merged PR binding. close remains available only when all Task criteria/dependencies/inputs are satisfied. Personal Agent assistance is the last resort when actual information/authority or a human decision is missing; failure counts/time never automatically escalate. A conversation/turn completion is not Task completion. Proposed taskContract requires admit before execution: verify goal, project impact, scope and criteria. Prefer admit without repeating already adequate requirements. For admit/revise, every provided requirements.title, requirements.criteria, requirements.impact and requirements.scope must be a bounded nonempty string. In particular requirements.criteria is a string; the separate outer decision.criteria is an evidence array and must never be copied into requirements.criteria. Optional requirements.repository is an associated owner/name or null for a repository-free decision. Uncertain intent requires needs-human. Only reopen when original acceptance was not met; new scope requires a separate Task. Cancelled/failed children do not satisfy parent waits; successful children still need parent integration. All Task mutations remain scoped Chief decisions. To resolve a failed/cancelled child condition without faking child completion, use resolve-dependency with dependency:{taskId,version:the exact current dependency snapshot version,evidence:[actual references],replacementTaskId:optional accepted alternative,replacementVersion:its task outcome version}. This is explicit Chief judgement, not host test success or parent completion. Leave unresolved conditions blocked when evidence is insufficient. Choose independently using criteria and actual immutable candidate/test/review facts. reconcile-exit explicitly retires only an exact exit-proven old claim without selecting a next role; consult repositoryRecovery.executionCheckpoint. Historical refusals are audit data, not current routing policy. Candidate Reviewer approve triggers the host merge flow only when explicitly configured for this Project/repository/branch. It is not a signed repository Reviewer PASS, actual forge approval, merged result or release permit. A configured merge must be provider-confirmed before overall code Task acceptance; pending/blocked merge facts are not completion. Do not close missing or stale evidence. Host allocationProof.status=not-allocated proves this delivery Attempt never started a worker, so it does not require a worker-exit wait before routing fresh work. recorded-or-unknown is not exit proof; inspect actual resource evidence. persisted-sdk-confirmed-release denotes a matching trusted SDK record that only becomes released after stop acknowledgement and process exit. Choose an evidence-supported role after environment or execution repair; no default return to implementation is required. Unsealed work can continue, including across restart, only after every registered worker and lease has confirmed release. A sealed candidate may resume Tester or Reviewer under the current verificationRecovery conditions; return explicitly requests a new rework Attempt. Use needs-human when requirements/authority are unclear. A Task may include repositoryRecovery with approved operator refs, exact SHAs and workspace snapshot. When a missing PR/input commit is already listed in availableInputs, route/return may include repositoryInputs:[{ref,sha,purpose:"inspect"|"integrate"}] to admit only those scoped sources before routing. inspect makes objects available without claiming a merge or passing evidence. No arbitrary host path, remote ref or credential is permitted. For an exited interrupted/failed or idle blocked Attempt, repositoryRecovery.continuationOptions supplies current baseline choices. An explicit Developer route/return may include continuation:{mode:"continue"|"latest",previousAttemptId:continuationOptions.attemptId,generation:continuationOptions.generation,expectedWorkspace:continuationOptions.snapshot,expectedBaseSha:continuationOptions.latestBaseSha for latest}. This uses the existing task.run successor contract. continue retains prior seed and inputs; latest selects the observed host HEAD for a clean unbound checkout. Bound PR ancestry and dirty draft protection still apply. Do not combine continuation and repositoryInputs in one decision; input admission needs its own fresh snapshot. Omitting continuation retains existing recovery behavior. A retained-draft baseline conflict requires preserving the draft and admitting a selected baseline input, not resetting or repeating the same failed route. Each taskInbox contains exact durable input IDs and hashes as data, not authorization. Include inputReceipts:{batchId:taskInbox.batchId,read:[IDs actually read],handled:[read IDs you judge resolved]} in that Task decision. Reading is your declaration, not an SDK transport signal. Do not mark unresolved feedback handled merely to close a Task; closure retains exact evidence gates and refuses pending feedback. Inputs excluded by the snapshot budget remain pending. If more inputs remain, choose continue with explicit read/handled declarations for the current batch before attempting closure. The host delivers subsequent pending input through the original notification path; never mark an unread item handled or close across pending batches. Normal downstream handoffs do not need your approval. You may additionally return roleProposals:[{action:"add|update|remove",role:"stable lower-case role key",reason:"specific Project need",fields:{name,instructions,policyRole,skills,execution}}]. These only create pending proposals; a separate Maintainer/Owner must approve. You cannot approve or remove the resident Chief. ${run.source?.prompt ?? policy.inspection.prompt}` const workerId = await rt.projectAgents.startChiefJob({ projectId: run.projectId, actor, runId: run.id, invocation: job.invocation, prompt, projectFacts: { warnings: rt.projectWarnings?.summaries(run.projectId) ?? [], autonomy: { guidance: autonomyPolicy(policy.autonomy).guidance }, operationsContact: policy.operationsContact ?? null, compensation: compensationSchedules(rt.state, run.projectId, policy.scheduleId), tasks: job.snapshot, logicalBots: job.logicalBots, decisionFeedback, requestedChange: run.source?.requestedChange ?? null }, onResult: result => completed(job, result) }) if (run.invocation === job.invocation) { run.workerId = workerId; if (run.status === 'running') run.waitingFor = 'execution'; run.workerGeneration = rt.state.projectHarnessWorkers?.[workerId]?.generation ?? null; await rt.save() } } catch (error) { diff --git a/core/src/host/project-chief-protocol.js b/core/src/host/project-chief-protocol.js index ff765970..435a6167 100644 --- a/core/src/host/project-chief-protocol.js +++ b/core/src/host/project-chief-protocol.js @@ -19,9 +19,11 @@ export function parseChiefResponse(text, roles, { onInvalid } = {}) { seen.add(taskId) if (!actions.has(decision.action)) throw chiefProtocolError('action', 'a supported Chief action', taskId) if (typeof decision.reason !== 'string' || !decision.reason.trim() || decision.reason.length > 4000) throw chiefProtocolError('reason', 'bounded nonempty text up to4000 characters', taskId) - if (['route', 'return', 'start-next'].includes(decision.action)) { + if (['route', 'return'].includes(decision.action) || decision.action === 'start-next' && decision.targetBotId !== undefined) { if (typeof decision.targetBotId !== 'string' || !decision.targetBotId.trim() || !roles.some(row => row.id === decision.targetBotId && row.role !== 'chief')) throw chiefProtocolError('targetBotId', 'one exact active logical Bot ID from logicalBots, excluding Chief', taskId) } + if (decision.mode !== undefined && (decision.action !== 'start-next' || decision.mode !== 'next')) throw chiefProtocolError('mode', 'next only for start-next', taskId) + if (decision.continuationRef !== undefined && (decision.action !== 'start-next' || typeof decision.continuationRef !== 'string' || !decision.continuationRef)) throw chiefProtocolError('continuationRef', 'a nonempty scoped reference only for start-next', taskId) if (decision.criteria !== undefined) { if (!Array.isArray(decision.criteria)) throw chiefProtocolError('criteria', 'an evidence array, separate from requirements.criteria text', taskId) for (const [index, item] of decision.criteria.entries()) { diff --git a/core/src/host/project-delivery-options.js b/core/src/host/project-delivery-options.js index 30791c38..439107ec 100644 --- a/core/src/host/project-delivery-options.js +++ b/core/src/host/project-delivery-options.js @@ -40,7 +40,7 @@ export function deliveryChoices(state, task, observation) { const acceptance = option('accept-attempt', () => requireAttemptAcceptance(task, a), 'Chief must judge delivered scope and remaining work; execution rechecks all evidence') if (acceptance.status === 'ready' && (observation?.status === 'unknown' || !merged)) Object.assign(acceptance, { status: observation?.status === 'unknown' ? 'unknown' : 'blocked', reason: observation?.status === 'unknown' ? 'Query the current provider merge result; no merge outcome is inferred' : 'Provider-confirmed merge is required for scoped acceptance' }) return { requirements, expectedScope: scope, observedAt: Date.now(), admission: admissions(state, task).slice(-4), observation: observation ?? null, recommendation: previous?.taskId === task.id && previous.status === 'applied' ? { decisionId: previous.id, action: previous.action, reason: previous.reason, criteria: previous.criteria, expectedScope: previous.expectedScope, alreadyApplied: true } : null, recommendationSource: 'Chief decision action/reason/evidence in the original inference turn', actions: [acceptance, - option('start-next', () => requireNextAttempt(task, a), 'Read fresh continuation-state; pin verified target HEAD, choose an implementation role and use task.run mode next'), + option('start-next', () => requireNextAttempt(task, a), 'Chief may reference observed continuationOptions with start-next; choose ambiguous roles/modes explicitly. Execution revalidates the target; external task.run keeps explicit fields'), merged ? { action: 'return', status: 'blocked', reason: 'A provider-confirmed merged PR cannot be reworked or reused for publication' } : { action: 'return', status: observation?.status === 'unknown' ? 'unknown' : 'ready', reason: 'Existing route/rework, exit, binding and current authority gates still apply' }, { action: 'close', status: 'unknown', reason: 'Chief must establish every Task criterion, dependency and input; merged or accepted Attempt alone does not complete the Task' }, { action: 'continue', status: 'ready', reason: 'Inspect current scoped facts or defer judgement without claiming completion' }, diff --git a/core/test/project-chief-control.test.js b/core/test/project-chief-control.test.js index 4275be7d..a63d22b8 100644 --- a/core/test/project-chief-control.test.js +++ b/core/test/project-chief-control.test.js @@ -1,6 +1,8 @@ import assert from 'node:assert/strict' +import { createHash } from 'node:crypto' import { test } from 'node:test' import { createProjectChiefControl, inspectionFacts, allocationProof, inspectionResources } from '../src/host/project-chief-control.js' +const createHashForTest = value => createHash('sha256').update(JSON.stringify(value)).digest('hex') const thresholds = { maxHandoffsWithoutProgress: 3, longExecutionMs: 10000, missingHeartbeatMs: 5000, longWaitMs: 8000 } test('Chief inspection counts verified artifact progress rather than message or transfer churn', () => { const task = { id: 't', updatedAt: 1, deliveryAttemptId: 'a', collaboration: { status: 'queued', generation: 7, history: Array.from({ length: 5 }, (_, index) => ({ evidence: [], reason: 'Progress message ' + index, createdAt: index + 1 })) } } @@ -879,3 +881,95 @@ test('requirement changes after scoped acceptance judgement request fresh facts assert.equal(run.status, 'done'); assert.equal(f.accepted(), 0) await f.rt.projectEffects.close() }) + +async function referenceFixture({ modes = ['next'], targetCount = 1, policyRole = 'developer', read } = {}) { + const { deliveryScope } = await import('../src/host/project-delivery-options.js') + let facts, calls = [], catalog, reads = 0 + const f = await controlFixture({ prepare(rt, task) { + const role = policyRole === 'developer' ? 'developer' : 'implementer' + catalog = { revision: 1, roles: [{ id: 'p:chief', role: 'chief' }, ...Array.from({ length: targetCount }, (_, index) => ({ id: 'p:' + role + (index ? index : ''), role: role + (index ? index : ''), policyRole: 'developer' }))] } + rt.projectAgents.deliveryDefinitions = () => catalog + const scope = deliveryScope(rt.state, task) + facts = { attemptId: 'a', generation: 1, snapshot: { workspaceId: 'w', sha256: 'workspace' }, latestBaseSha: 'a'.repeat(40), modes, expectedScope: scope } + rt.projectDelivery.recoveryFacts = async () => { facts.expectedScope = deliveryScope(rt.state, task); return { continuationOptions: structuredClone(facts), nextActions: { actions: [{ action: 'start-next', status: 'ready' }] } } } + rt.projectDelivery.endpoint = async () => { ++reads; return read ? read(facts, rt, task, catalog, reads) : structuredClone(facts) } + rt.projectScheduler = { wake() {}, request: async (args) => { calls.push(structuredClone(args)); return { queue: { id: 'queue-reference' } } } } + } }) + return { ...f, facts, calls, catalog } +} + +test('next snapshot reference binds current fields and the unique custom implementation role without mutating the decision', async () => { + const f = await referenceFixture({ policyRole: 'custom' }) + const decision = { action: 'start-next', mode: 'next', continuationRef: f.facts.expectedScope } + const original = structuredClone(decision), result = f.result(decision) + await f.request.onResult(result); await f.request.onResult(result) + assert.deepEqual(decision, original); assert.equal(f.calls.length, 1, JSON.stringify(f.rt.state.projectExecutions)) + assert.equal(f.calls[0].targetRole, 'implementer') + const audit = Object.values(f.rt.state.projectChiefDecisions).find(row => row.status === 'applied') + assert.equal(audit.continuationBinding.source, 'continuation-ref') + assert.equal(audit.continuationBinding.targetBotId, 'p:implementer') + assert.equal(audit.queueId, 'queue-reference'); assert.equal(audit.successorAttemptId, null) + assert.equal(f.calls[0].continuation.expectedBaseSha, f.facts.latestBaseSha) + assert.deepEqual(f.calls[0].continuation.expectedWorkspace, f.facts.snapshot) + const intent = createHashForTest(JSON.parse(result.text).decisions[0]) + assert.equal(f.rt.state.projectExecutions[f.request.runId].appliedDecisions[0].intent, intent) + const visible = f.control.view('p', { memberId: 'owner' }).runs.find(run => run.id === f.request.runId).appliedDecisions[0].continuationBinding + assert.deepEqual(visible, audit.continuationBinding) + visible.continuation.expectedBaseSha = 'changed-view-only'; assert.equal(audit.continuationBinding.continuation.expectedBaseSha, f.facts.latestBaseSha) + audit.projectId = 'foreign' + assert.equal(f.control.view('p', { memberId: 'owner' }).runs.find(run => run.id === f.request.runId).appliedDecisions[0].continuationBinding, undefined) +}) + +for (const kind of ['ambiguous-target', 'no-target', 'ambiguous-mode', 'unavailable', 'wrong-ref', 'bad-fields', 'bad-mode', 'not-ready', 'foreign-attempt', 'target-changed', 'workspace-changed', 'local-changed', 'roles-changed', 'rights-revoked']) test('next reference refuses ' + kind + ' without dispatch or inferred recovery', async () => { + const f = await referenceFixture({ targetCount: kind === 'ambiguous-target' ? 2 : kind === 'no-target' ? 0 : 1, modes: kind === 'ambiguous-mode' ? ['next', 'latest'] : ['next'], read: async (facts, rt, task, catalog, reads) => { + if (kind === 'target-changed') return { ...structuredClone(facts), latestBaseSha: 'b'.repeat(40) } + if (kind === 'workspace-changed') return { ...structuredClone(facts), snapshot: { workspaceId: 'other' } } + if (kind === 'local-changed' && reads === 2) task.criteria = 'Changed during validation' + if (kind === 'roles-changed' && reads === 2) catalog.revision++ + if (kind === 'rights-revoked' && reads === 2) rt.requireProjectAccess = () => { throw Error('Current role denied') } + return structuredClone(facts) + } }) + const saved = f.request.projectFacts.tasks[0] + if (kind === 'unavailable') saved.repositoryRecovery.continuationOptions.status = 'unavailable' + if (kind === 'not-ready') saved.repositoryRecovery.nextActions.actions[0].status = 'blocked' + if (kind === 'foreign-attempt') saved.repositoryRecovery.continuationOptions.attemptId = 'other' + const selection = kind === 'wrong-ref' ? { continuationRef: 'wrong' } : kind === 'bad-fields' ? { continuation: { mode: 'next', previousAttemptId: 'a', generation: 1, expectedWorkspace: {}, expectedBaseSha: f.facts.latestBaseSha } } : kind === 'bad-mode' ? { mode: 'latest' } : {} + await f.request.onResult(f.result({ action: 'start-next', ...selection })) + assert.equal(f.calls.length, 0) + assert.equal(f.task.chiefQuestion, undefined) + const run = f.rt.state.projectExecutions[f.request.runId] + assert.notEqual(run.status, 'done') + const expected = kind === 'roles-changed' ? /fresh role catalog/ : kind === 'rights-revoked' ? /Current role denied/ : kind === 'local-changed' ? /fresh Task facts/ : ['target-changed', 'workspace-changed'].includes(kind) ? /fresh facts after/ : kind === 'wrong-ref' ? /expectedScope/ : kind === 'bad-mode' ? /next only/ : kind.includes('target') ? /active implementation role/ : /continuation/ + assert.match(run.failure, expected) +}) + +test('explicit target resolves multiple implementation roles, and explicit scope reference resolves multiple modes', async () => { + const f = await referenceFixture({ targetCount: 2, modes: ['next', 'latest'] }) + await f.request.onResult(f.result({ action: 'start-next', targetBotId: 'p:developer1', continuationRef: f.facts.expectedScope })) + assert.equal(f.calls.length, 1, JSON.stringify(f.rt.state.projectExecutions)); assert.equal(f.calls[0].targetRole, 'developer1') +}) + +for (const targetBotId of [null, '', 'p:unknown', 'p:chief', 3]) test('explicit invalid next target remains a protocol error: ' + JSON.stringify(targetBotId), async () => { + const f = await referenceFixture() + await f.request.onResult(f.result({ action: 'start-next', targetBotId })) + assert.equal(f.calls.length, 0) + assert.match(f.rt.state.projectExecutions[f.request.runId].failure, /targetBotId/) +}) +test('an incomplete explicit continuation is refused rather than silently filled from the snapshot', async () => { + const f = await referenceFixture() + await f.request.onResult(f.result({ action: 'start-next', continuation: { mode: 'next' } })) + assert.equal(f.calls.length, 0); assert.match(f.rt.state.projectExecutions[f.request.runId].failure, /exact current continuationOptions/) +}) +for (const ref of [null, {}, '']) test('invalid reference type is not treated as omission: ' + JSON.stringify(ref), async () => { + const f = await referenceFixture() + await f.request.onResult(f.result({ action: 'start-next', continuationRef: ref })) + assert.equal(f.calls.length, 0); assert.match(f.rt.state.projectExecutions[f.request.runId].failure, /continuationRef/) +}) + +for (const [field, value] of [['mode', 'next'], ['continuationRef', 'scope']]) test('next-only reference field cannot be silently used on another action: ' + field, async () => { + const f = await referenceFixture() + await f.request.onResult(f.result({ action: 'continue', [field]: value })) + assert.equal(f.calls.length, 0) + assert.equal(f.rt.state.projectExecutions[f.request.runId].status, 'correction-pending') + assert.match(f.rt.state.projectExecutions[f.request.runId].failure, new RegExp(field)) +}) diff --git a/core/test/project-delivery.test.js b/core/test/project-delivery.test.js index 620f407f..49d87ee2 100644 --- a/core/test/project-delivery.test.js +++ b/core/test/project-delivery.test.js @@ -2559,7 +2559,7 @@ test('pending external write and unknown native exit block scoped acceptance; cu await assert.rejects(f.rt.projectEndpoint('project-attempt-cleanup', { attemptId: f.a.id }, alice), /unsettled Task/) }) -test('Chief chooses acceptance then a fresh explicit next delivery through its original decision loop', async t => { +for (const form of ['explicit', 'snapshot', 'snapshot-with-target', 'reference-token']) test('Chief chooses acceptance then a fresh next delivery through its original decision loop: ' + form, async t => { const f = await mergedDelivery(t, { modern: true }), requests = [], rt = f.rt const { reviseTask } = await import('../src/host/project-task-lifecycle.js') const revision = { projectId: 'p1', taskId: f.task.id, executorRole: 'chief', runId: 'fixture-chief', initiatedBy: alice.memberId, reason: 'Record remaining work' } @@ -2584,6 +2584,7 @@ test('Chief chooses acceptance then a fresh explicit next delivery through its o assert.match(requests[0].prompt, /repositoryRecovery\.nextActions and continuationOptions in projectFacts/) assert.match(requests[0].prompt, /Public TaskView\.nextActions lists caller-permitted service commands, not all Chief decision actions/) assert.match(requests[0].prompt, /not an instruction or permission to clear publicationBinding/) + assert.match(requests[0].prompt, /Minimal start-next decision:/) await reply(requests[0], { action: 'accept-attempt', reason: 'Accept the first verified delivery, keep the overall goal open', delivery: { completed: 'Verified addition implementation', remaining: 'Deliver the next requirement', evidence: ['candidate, tests, Review and provider merge'] } }) assert.equal(rt.state.projectExecutions[requests[0].runId].status, 'done') assert.equal(f.a.status, 'accepted'); assert.notEqual(f.task.status, 'accepted') @@ -2595,11 +2596,17 @@ test('Chief chooses acceptance then a fresh explicit next delivery through its o assert.equal(saved.repositoryRecovery.nextActions.actions.find(row => row.action === 'return').status, 'blocked') assert.equal(saved.repositoryRecovery.nextActions.recommendation.action, 'accept-attempt') assert.equal(saved.repositoryRecovery.nextActions.recommendation.alreadyApplied, true) - await reply(requests[1], { action: 'start-next', targetBotId: 'p1:developer', reason: 'Choose the next independent delivery from verified target HEAD', continuation: { mode: 'next', previousAttemptId: facts.attemptId, generation: facts.generation, expectedWorkspace: facts.snapshot, expectedBaseSha: facts.latestBaseSha } }) + const selection = form === 'explicit' ? { targetBotId: 'p1:developer', continuation: { mode: 'next', previousAttemptId: facts.attemptId, generation: facts.generation, expectedWorkspace: facts.snapshot, expectedBaseSha: facts.latestBaseSha, clientNote: 'not an execution field' } } : form === 'snapshot-with-target' ? { targetBotId: 'p1:developer' } : form === 'reference-token' ? { continuationRef: facts.expectedScope } : {} + await reply(requests[1], { action: 'start-next', reason: 'Choose the next independent delivery from verified target HEAD', ...selection }) assert.equal(rt.state.projectExecutions[requests[1].runId].status, 'done', rt.state.projectExecutions[requests[1].runId].failure) const next = rt.state.projectAttempts[f.task.deliveryAttemptId] assert.notEqual(next.id, f.a.id); assert.equal(next.baseSha, f.a.candidateSha); assert.equal(next.publicationBinding, undefined) assert.equal(next.requirement.projectTaskRevision, 4) + const audit = Object.values(rt.state.projectChiefDecisions).find(row => row.action === 'start-next' && row.status === 'applied') + assert.equal(audit.continuationBinding.targetBotId, 'p1:developer') + assert.equal(audit.continuationBinding.continuation.expectedBaseSha, facts.latestBaseSha) + assert.deepEqual(audit.continuationBinding.continuation.expectedWorkspace, facts.snapshot) + assert.equal(audit.continuationBinding.continuation.clientNote, undefined) assert.notEqual(f.task.status, 'accepted'); assert.equal(f.task.chiefQuestion, undefined) }) diff --git a/decisions/2026-11-06-chief-continuation-reference.md b/decisions/2026-11-06-chief-continuation-reference.md new file mode 100644 index 00000000..3ff0d648 --- /dev/null +++ b/decisions/2026-11-06-chief-continuation-reference.md @@ -0,0 +1,13 @@ +# Reference observed continuation facts for Chief start-next + +Date: 2026-11-06 (Asia/Shanghai). +Class: Owner-authorized contract clarification for #2560. +Status: Accepted for delivery/ai-centric-development-service source implementation. + +Owner reported repeated mechanical snapshot echo errors after durable acceptance and approved snapshot reference first, existing read query tools second. Fresh base80cf5795 real Git/Runtime two-turn regressions reproduce missing target/continuation refusal despite a unique eligible next delivery. + +Chief chooses whether to start next and any ambiguous implementation target. Omission of machine fields is an explicit reference to that authenticated invocation's immutable snapshot, not model repair or selection of latest facts. A unique approved implementation capability can be resolved; route/return do not get a generic Developer fallback. Optional reference must match saved expectedScope; explicit malformed fields still refuse. Read current continuation-state and revalidate actor/catalog/Task scope; stale facts require another judgement. Keep original settlement/exit/provider ancestry/quota/replay gates. + +Reuse original Store audits, Effects, scheduler and existing query endpoints. Persist actual binding separately from unchanged model intent, expose only same Project/Run/Task audit data. No new planner/store/timer/draft transaction or automatic human escalation. This preserves #2327's semantic validation boundary. Dispatch/inbox stale races and mandatory document slimming remain separately scoped. Native/SDK path, deterministic fixtures, real vendor model and onsite acceptance are distinct evidence; source delivery authorizes no live deployment, credential operations or paid inference. + +Written-by: Codex diff --git a/docs/development-service-api.md b/docs/development-service-api.md index efbaeca2..01b12625 100644 --- a/docs/development-service-api.md +++ b/docs/development-service-api.md @@ -301,8 +301,8 @@ settlement；原 Task 不关闭。`task.accept` 的公开语义仍是完整 Task 字段绕过 Chief 裁决。若整个目标已完成，Chief 仍可选择 `close`。 提交后通过原队列通知 Chief 重新观察；不自动启动 Dev。Chief 可选 `start-next`， -显式给出实现角色 `targetBotId` 及新快照中的 next continuation。配置授权的调用方 -也可沿原接口查询/请求： +按当前快照引用选择（#2560），有歧义时显式选择角色/continuation。配置授权的 +外部调用方也可沿原显式接口查询/请求： 1. `attempt.continuation-state {attemptId}` 返回 `modes:["next"]`、generation、snapshot、 latestBaseSha、target 和 nextActions。target SHA 来自获准 provider 的实际目标分支， @@ -334,3 +334,17 @@ Chief 应读取当前交付选项再判断，Personal Agent 的文字输入也 Task 需求修订号与 Attempt 领取时的修订号允许不同。`accept-attempt` 验收原承诺的已合成果，保留 `delivery.completed/remaining/evidence`，不自动认可当前新需求完成。`deliveryChoices.requirements` / Chief 的 `repositoryRecovery.nextActions.requirements` 返回 `attemptRevision`、`taskRevision`、`changedFields` 和这些字段的 `priorValues`；当前值在 Task requirements 中。原始完整快照由已有 `attempt.context` 的 `section=requirement` 分页读取，不把整段历史反复加入上下文。 Settlement 的 `requirementsRevision` 是 Attempt 原版本；可选 `taskRequirementsRevision` 是验收时当前 Task 版本。旧 settlement 无需迁移。下一轮 Attempt 复制当前 Task 需求。Chief 读后需求再变化时，宿主通过既有 stale-snapshot 纠正反馈要求重新判断；候选/测试/Review/PR SHA、权限、代次、退出与幂等校验保持严格，整单 close 仍独立验证当前需求。 + +### Chief 下一轮交付引用（#2560） + +在当前 Chief 调用中，若 `continuationOptions.modes=["next"]`、`start-next` 预检 ready 且只有一个当前获准实施角色，可输出： + +```json +{"summary":"本轮已验收，继续剩余目标","decisions":[{"taskId":"当前快照中的 Task ID","action":"start-next","reason":"开始下一轮独立交付"}]} +``` + +不必复述 workspace、generation、SHA。宿主绑定的是 **Chief 看过的那份快照**，再通过现有 `attempt.continuation-state` 和当前权限/角色/Task 闸复验；变化就反馈刷新，不静默换成新基线。多个实施角色须显式给 `targetBotId`；可选 `mode` 只能是 `next`，可选 `continuationRef` 是该份 facts 的 `expectedScope`，供显式引用。多 mode 时必须有明确引用或完整选择；不是让宿主代替 Chief 决定路线。 + +旧的完整 `continuation` 仍支持；一旦提供，其字段就必须完整且准确，宿主不会修正错误字段或替你补半个 JSON。`route/return` 的显式目标约定不变。外部程序直接调用 `task.run` 仍沿用原显式 continuation 参数，可先读 `attempt.continuation-state`；本次不新增准备记录或公共写接口。 + +`chief.state.runs[].appliedDecisions[].continuationBinding` 返回实际绑定的目标、来源、scope 与 continuation，便于审计。引用不是凭证，不绕过 settlement、配额、退出或 provider target 校验，也不自动关闭 Task、重用已合 PR 或升级人工。 diff --git a/docs/openapi/flybot.openapi.json b/docs/openapi/flybot.openapi.json index ac479a8b..a92822bc 100644 --- a/docs/openapi/flybot.openapi.json +++ b/docs/openapi/flybot.openapi.json @@ -6584,6 +6584,102 @@ "reauthenticationRequired" ] }, + "ChiefContinuationBinding": { + "type": "object", + "properties": { + "targetBotId": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "source": { + "type": "string", + "enum": [ + "chief-snapshot", + "continuation-ref", + "explicit" + ] + }, + "expectedScope": { + "anyOf": [ + { + "type": "string", + "pattern": "^[a-f0-9]{64}$" + }, + { + "type": "null" + } + ] + }, + "continuation": { + "type": "object", + "properties": { + "mode": { + "type": "string", + "enum": [ + "continue", + "latest", + "next" + ] + }, + "expectedBaseSha": { + "type": "string", + "pattern": "^[a-f0-9]{40}$" + }, + "previousAttemptId": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "generation": { + "type": "integer", + "minimum": 0 + }, + "expectedWorkspace": { + "$ref": "#/components/schemas/JsonValue" + } + }, + "additionalProperties": false, + "required": [ + "mode", + "previousAttemptId", + "generation", + "expectedWorkspace" + ], + "allOf": [ + { + "if": { + "properties": { + "mode": { + "enum": [ + "latest", + "next" + ] + } + }, + "required": [ + "mode" + ] + }, + "then": { + "required": [ + "expectedBaseSha" + ] + } + } + ] + } + }, + "additionalProperties": false, + "required": [ + "targetBotId", + "source", + "expectedScope", + "continuation" + ] + }, "DeliveryChoices": { "type": "object", "properties": { @@ -12258,6 +12354,40 @@ "recoveryIssue": { "$ref": "#/components/schemas/ChiefRecoveryIssue" }, + "appliedDecisions": { + "type": "array", + "items": { + "type": "object", + "properties": { + "taskId": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "id": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "intent": { + "type": "string", + "pattern": "^[a-f0-9]{64}$" + }, + "continuationBinding": { + "$ref": "#/components/schemas/ChiefContinuationBinding" + } + }, + "additionalProperties": false, + "required": [ + "taskId", + "id", + "intent" + ] + }, + "maxItems": 30 + }, "lastResult": { "type": "object", "properties": { @@ -18483,6 +18613,40 @@ "recoveryIssue": { "$ref": "#/components/schemas/ChiefRecoveryIssue" }, + "appliedDecisions": { + "type": "array", + "items": { + "type": "object", + "properties": { + "taskId": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "id": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "intent": { + "type": "string", + "pattern": "^[a-f0-9]{64}$" + }, + "continuationBinding": { + "$ref": "#/components/schemas/ChiefContinuationBinding" + } + }, + "additionalProperties": false, + "required": [ + "taskId", + "id", + "intent" + ] + }, + "maxItems": 30 + }, "lastResult": { "type": "object", "properties": { @@ -24170,6 +24334,40 @@ "recoveryIssue": { "$ref": "#/components/schemas/ChiefRecoveryIssue" }, + "appliedDecisions": { + "type": "array", + "items": { + "type": "object", + "properties": { + "taskId": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "id": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "intent": { + "type": "string", + "pattern": "^[a-f0-9]{64}$" + }, + "continuationBinding": { + "$ref": "#/components/schemas/ChiefContinuationBinding" + } + }, + "additionalProperties": false, + "required": [ + "taskId", + "id", + "intent" + ] + }, + "maxItems": 30 + }, "lastResult": { "type": "object", "properties": { @@ -32050,6 +32248,40 @@ "recoveryIssue": { "$ref": "#/components/schemas/ChiefRecoveryIssue" }, + "appliedDecisions": { + "type": "array", + "items": { + "type": "object", + "properties": { + "taskId": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "id": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "intent": { + "type": "string", + "pattern": "^[a-f0-9]{64}$" + }, + "continuationBinding": { + "$ref": "#/components/schemas/ChiefContinuationBinding" + } + }, + "additionalProperties": false, + "required": [ + "taskId", + "id", + "intent" + ] + }, + "maxItems": 30 + }, "lastResult": { "type": "object", "properties": { @@ -32483,6 +32715,40 @@ "recoveryIssue": { "$ref": "#/components/schemas/ChiefRecoveryIssue" }, + "appliedDecisions": { + "type": "array", + "items": { + "type": "object", + "properties": { + "taskId": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "id": { + "type": "string", + "minLength": 1, + "maxLength": 200, + "description": "Opaque server-issued ID; never infer authority from it." + }, + "intent": { + "type": "string", + "pattern": "^[a-f0-9]{64}$" + }, + "continuationBinding": { + "$ref": "#/components/schemas/ChiefContinuationBinding" + } + }, + "additionalProperties": false, + "required": [ + "taskId", + "id", + "intent" + ] + }, + "maxItems": 30 + }, "lastResult": { "type": "object", "properties": { @@ -35865,10 +36131,10 @@ "unifiedCommandReceipt": false }, "invokesModel": false, - "notes": "All current authorization, Project/Task bindings and domain fences are rechecked server-side.", + "notes": "undefined Applied continuation decisions disclose the actually bound target/snapshot under appliedDecisions[].continuationBinding; it is audit data, not an authorization token. Chief start-next can reference its own ready snapshot instead of echoing machine fields. External task.run keeps its explicit continuation request contract.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "7fdff3f88afff250f3c3588d587d2ea2b3c4db4267383101a462397d1fcbe8d1" + "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" } }, { @@ -39070,7 +39336,7 @@ "notes": "All current authorization, Project/Task bindings and domain fences are rechecked server-side.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "7fdff3f88afff250f3c3588d587d2ea2b3c4db4267383101a462397d1fcbe8d1" + "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" } }, { @@ -39175,7 +39441,7 @@ "notes": "operationsContact selects an explicit assistance recipient, never host privileges; null clears it, omission retains it. autonomy budgets/guidance inform Chief judgement, not a fixed escalation workflow. Use existing schedule.configure targeting project-chief for hourly compensation; chief.state.compensation reports original nextCheckAt. 30-minute inspection is a configurable schedule default, not a framework deadline. correctionBudget defaults to 2 (0–4); protocol correction reuses the Chief Session. Exhaustion is incomplete judgement, never automatic consultation or default routing. inspection.lifecycle configures queue/execution deadlines, bounded recovery delay/count and health overdue threshold; timeout is not exit proof.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "7fdff3f88afff250f3c3588d587d2ea2b3c4db4267383101a462397d1fcbe8d1" + "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" } }, { @@ -50055,7 +50321,7 @@ "notes": "All current authorization, Project/Task bindings and domain fences are rechecked server-side.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "7fdff3f88afff250f3c3588d587d2ea2b3c4db4267383101a462397d1fcbe8d1" + "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" }, "aliasOf": "development.command.chief.inspect" }, @@ -50158,7 +50424,7 @@ "notes": "operationsContact selects an explicit assistance recipient, never host privileges; null clears it, omission retains it. autonomy budgets/guidance inform Chief judgement, not a fixed escalation workflow. Use existing schedule.configure targeting project-chief for hourly compensation; chief.state.compensation reports original nextCheckAt. 30-minute inspection is a configurable schedule default, not a framework deadline. correctionBudget defaults to 2 (0–4); protocol correction reuses the Chief Session. Exhaustion is incomplete judgement, never automatic consultation or default routing. inspection.lifecycle configures queue/execution deadlines, bounded recovery delay/count and health overdue threshold; timeout is not exit proof.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "7fdff3f88afff250f3c3588d587d2ea2b3c4db4267383101a462397d1fcbe8d1" + "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" }, "aliasOf": "development.command.chief.configure" }, @@ -50231,10 +50497,10 @@ "unifiedCommandReceipt": false }, "invokesModel": false, - "notes": "All current authorization, Project/Task bindings and domain fences are rechecked server-side.", + "notes": "undefined Applied continuation decisions disclose the actually bound target/snapshot under appliedDecisions[].continuationBinding; it is audit data, not an authorization token. Chief start-next can reference its own ready snapshot instead of echoing machine fields. External task.run keeps its explicit continuation request contract.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "7fdff3f88afff250f3c3588d587d2ea2b3c4db4267383101a462397d1fcbe8d1" + "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" }, "aliasOf": "development.query.chief.state" }, @@ -51115,9 +51381,9 @@ "core/src/host/project-assistant.js": "4fa4309d4aa7abf78b6ba4e55ba754f320384f931e0e4a5c6620c859e1b50676", "core/src/host/project-board.js": "dfb0b4bc05ae42335632e3edbba883a03c5da2fb17899b23e8226320ea08aec4", "core/src/host/project-chief-assistance.js": "261f3767cf7fd4466f057c625b32e74b9c0a90c53c0fca3fcc835ac0f11c567b", - "core/src/host/project-chief-control.js": "7fdff3f88afff250f3c3588d587d2ea2b3c4db4267383101a462397d1fcbe8d1", + "core/src/host/project-chief-control.js": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78", "core/src/host/project-chief-lifecycle.js": "ed85bb1be90198910b20ffd605895812dfa848fce1049f7bd27b32dbfae3a328", - "core/src/host/project-chief-protocol.js": "75d97be3ff530e8cbd5bdf9c9a70b9c671cc7db9bd7ee2f926143a7d8f801ee9", + "core/src/host/project-chief-protocol.js": "8e0ad56f025e08dac5318ce48c17c39754f001d5aa40b0beddbee0d66c9a179f", "core/src/host/project-chief-recovery.js": "c80a9d791aab06d4b343e8740a25a471ddf93e0af52b59dd9b7e6fc835c0d824", "core/src/host/project-chief.js": "a3f01a053177fd8f10d4b778d79ac98c25aee70aa3c4d967dc7e1da2aea2cf6e", "core/src/host/project-child-inputs.js": "db1f534c4222dba1b500d308b1cd07f0df50a02eb228f520b150a3b79d6f126d", @@ -51129,7 +51395,7 @@ "core/src/host/project-context-memory.js": "933893b3d6b7b2c7a79365991c1303fb03a665d9529f16ccf57429a753ad4cd0", "core/src/host/project-credentials.js": "18fe61057fb920664fc5077d81eee0957f7eaf64821dd10d2d2a3cd463f5fdea", "core/src/host/project-delegations.js": "d4d86192551994a3419e0d2ca0f64bc4d50a75960970d2b79ed9f97b82940903", - "core/src/host/project-delivery-options.js": "591aff7f7a8f6b6938b0fcb02592444c5398d960be5f824cf80c31a76d4d38e4", + "core/src/host/project-delivery-options.js": "026d37e7d1a86b60667a702b6b02d4f79df060b4035942c45559c3fd716c73bb", "core/src/host/project-delivery-template.js": "f149d09786ac6db41a7ebeccb7620cce62b57f334d3ddd5e3c66ee673e5b66a8", "core/src/host/project-delivery.js": "17d976ca26336000ebed6dbf328aed4a016754b87be410a1c5aa8f5bf34ea500", "core/src/host/project-development-service.js": "8d3ee6533b55c52fd5989a49db1d575effac9622f92070c32b39aa25c3e89691", diff --git a/docs/subsystems/backend-service.md b/docs/subsystems/backend-service.md index 529f9965..b215cf6f 100644 --- a/docs/subsystems/backend-service.md +++ b/docs/subsystems/backend-service.md @@ -214,3 +214,5 @@ The original task.accept endpoint retains whole-Task semantics. See the developm API reference for current prerequisites and fresh target/generation/workspace fields. #2526 follow-up keeps service commands unchanged. DeliveryChoices.requirements exposes original/current revisions, changed fields and their priorValues, plus the existing attempt.context requirement source; current values remain in Task requirements. Scoped settlement records the pinned requirementsRevision and optional current taskRequirementsRevision. This read-only advisory data does not grant acceptance or waive evidence/permission/exit checks. Whole-Task close and current-goal next delivery remain separate. Old settlement records need no migration. + +#2560 adds optional ChiefContinuationBinding on chief.state runs[].appliedDecisions: actual targetBotId, source (chief-snapshot/continuation-ref/explicit), expectedScope and bound continuation. The read projection joins only the same Project/Run/Task audit and cannot mutate durable state. This describes executed binding, not a reusable authorization token. Existing continuation-state/context remain the read-only sources; external task.run fields stay explicit and compatible. Chief shorthand is bound within its authenticated invocation, with fresh validation before the original scheduler request. diff --git a/docs/subsystems/project-chief-control.md b/docs/subsystems/project-chief-control.md index b79da50f..f337d3c2 100644 --- a/docs/subsystems/project-chief-control.md +++ b/docs/subsystems/project-chief-control.md @@ -381,9 +381,9 @@ completion. The normal Chief decision action/reason/criteria supplies its recomm The read view may reference that already-applied decision, explicitly marked as history; no recommendation is selected by the host or executed from that read projection. accept-attempt additionally supplies delivery:{completed,remaining,evidence:[references]}. -After committed acceptance a new scoped inspection lets Chief choose start-next with -an explicit active implementation Bot and fresh continuation:{mode:'next',...}. Other -lawful choices remain available. close still needs every overall criterion satisfied. +After committed acceptance a new scoped inspection lets Chief choose start-next by +validated snapshot reference (#2560), or an explicit active implementation Bot/full +continuation. Other lawful choices remain; close still needs every overall criterion. Chief must first query facts, correct or change its approach autonomously. The addressed Personal Agent, then its human if necessary, is the last resort for genuinely missing @@ -413,3 +413,11 @@ Scoped acceptance judges the immutable Attempt requirement, never equality with Requirements that change after a Chief snapshot use existing protocol correction with stale-snapshot feedback, so Chief can reconsider the original intent. Earlier legitimate Task revisions do not permanently block historical settlement. Neither outcome satisfies new Task criteria or chooses a downstream role/Personal Agent automatically. start-next copies current requirements; whole-Task closure remains independently gated. Scoped parent settlement uses the pinned parent requirement revision to locate its original child-input admission. Actual child evidence, committed admission Effects, accepted sources and parent integration remain required; whole-Task verification/admission still uses current requirements. + +## Next delivery by snapshot reference (#2560) + +A Chief decision {taskId,action:'start-next',reason} references that invocation's observed continuationOptions when only next is available and the active approved implementation role is unique. This is semantic intent, not repair of invalid model fields; route/return still require explicit logical targets. With multiple implementation roles Chief must select targetBotId. Optional mode must be next. Optional continuationRef equals the observed expectedScope and can disambiguate a mode set containing next. A complete explicit continuation remains supported; any supplied malformed/conflicting snapshot fields are refused, never filled in silently. + +The host derives machine fields from the saved immutable snapshot, reads the original continuation-state endpoint, compares it to those saved facts, then revalidates current actor, role catalog and Task scope. It never binds to a newer target/workspace implicitly. Original settlement, confirmed exit, target ancestry, quota and publication identity checks remain in the existing scheduler/delivery boundary. No new planner, draft store, timer, tool executor or generic field repair is added. Existing read queries support inspection after feedback. + +Audit.continuationBinding records actual target, source, expectedScope and continuation. The raw model decision is not mutated; existing intent hashes/replay keys remain stable. chief.state decorates appliedDecisions with matching project/run/task audit bindings, read-only and detached from Store. No independent authority is conveyed. Native and SDK Chief use the same parser/apply path and actual job prompt. No live deployment or new paid model validation follows from these source contracts. diff --git a/docs/subsystems/project-delivery.md b/docs/subsystems/project-delivery.md index 2a82710f..353c1ac2 100644 --- a/docs/subsystems/project-delivery.md +++ b/docs/subsystems/project-delivery.md @@ -750,9 +750,9 @@ inputs, unknown native/SDK exit and unresolved external writes retain their gate The Effect notifies the original Chief inspection queue for a fresh decision; it never selects a downstream role. Old merge handoff backfill stops at this settlement. -The original continuation-state/task.run contract adds explicit mode next. It requires -a committed settlement, fresh generation/workspace and expectedBaseSha, live configure -rights and normal scheduler quota/resource admission. Read the actual provider target +Chief start-next may reference its observed snapshot when next/implementation are unique. +The existing task.run still requires committed settlement, pinned generation/workspace/base +SHA, current configure rights and normal scheduler quota/resource admission. Read the actual provider target HEAD and verify its local Git objects include the accepted merge. Missing objects and non-containing target histories have distinct diagnostics. A moved target, dirty draft, new input or repository mismatch refuses, preserving history. The successor has a new diff --git a/patches/ds-bot/files/src/host/project-chief-control.js b/patches/ds-bot/files/src/host/project-chief-control.js index 5ec6989b..d01e78dd 100644 --- a/patches/ds-bot/files/src/host/project-chief-control.js +++ b/patches/ds-bot/files/src/host/project-chief-control.js @@ -85,7 +85,15 @@ export function createProjectChiefControl(rt) { function view(projectId, actor) { rt.requireProjectAccess(actor, projectId) const policy = policyOf(projectId), logical = policy?.scheduleId ? rt.state.projectScheduledTasks?.[policy.scheduleId] : null - return { warnings: rt.projectWarnings?.summaries(projectId) ?? [], compensation: compensationSchedules(rt.state, projectId, policy?.scheduleId), health: inspectionHealth(rt.state, projectId), policy: structuredClone(policy ?? { enabled: false }), template: structuredClone(template), schedule: logical ? { id: logical.id, enabled: logical.enabled, status: logical.status, config: structuredClone(logical.config), deliveries: Object.values(logical.deliveries).slice(-5).map(row => structuredClone(row)) } : null, runs: Object.values(rt.state.projectExecutions ?? {}).filter(run => run.projectId === projectId && run.kind === 'chief-control').slice(-30).map(({ capability: _capability, ...row }) => structuredClone(row)) } + const runs = Object.values(rt.state.projectExecutions ?? {}).filter(run => run.projectId === projectId && run.kind === 'chief-control').slice(-30).map(({ capability: _capability, ...row }) => { + const copy = structuredClone(row) + for (const applied of copy.appliedDecisions ?? []) { + const audit = rt.state.projectChiefDecisions?.[applied.id] + if (audit?.projectId === projectId && audit.runId === copy.id && audit.taskId === applied.taskId && audit.continuationBinding) applied.continuationBinding = structuredClone(audit.continuationBinding) + } + return copy + }) + return { warnings: rt.projectWarnings?.summaries(projectId) ?? [], compensation: compensationSchedules(rt.state, projectId, policy?.scheduleId), health: inspectionHealth(rt.state, projectId), policy: structuredClone(policy ?? { enabled: false }), template: structuredClone(template), schedule: logical ? { id: logical.id, enabled: logical.enabled, status: logical.status, config: structuredClone(logical.config), deliveries: Object.values(logical.deliveries).slice(-5).map(row => structuredClone(row)) } : null, runs } } async function configure(body, actor) { rt.requireProjectAccess(actor, body.projectId, 'configure') @@ -173,15 +181,26 @@ export function createProjectChiefControl(rt) { })) } async function continuationChoice(decision, saved, roles, actor, projectId) { - if (decision.continuation === undefined) return null - const choice = decision.continuation, facts = saved?.repositoryRecovery?.continuationOptions - const target = roles.find(row => row.id === decision.targetBotId) - if (!['route', 'return', 'start-next'].includes(decision.action) || (target?.policyRole ?? target?.role) !== 'developer' || decision.repositoryInputs !== undefined) throw chiefProtocolError('continuation', 'an explicit Developer route/return, separate from repositoryInputs admission', decision.taskId) - if (!facts || facts.status === 'unavailable' || !(decision.action === 'start-next' ? ['next'] : ['continue', 'latest']).includes(choice?.mode) || choice.previousAttemptId !== facts.attemptId || choice.generation !== facts.generation || !isDeepStrictEqual(choice.expectedWorkspace, facts.snapshot) || ['latest', 'next'].includes(choice.mode) && choice.expectedBaseSha !== facts.latestBaseSha) throw chiefProtocolError('continuation', 'the exact current continuationOptions snapshot, generation and selected latestBaseSha', decision.taskId) + const next = decision.action === 'start-next' + if (!next && decision.continuation === undefined) return null + const facts = saved?.repositoryRecovery?.continuationOptions + const implementations = roles.filter(row => (row.policyRole ?? row.role) === 'developer') + const target = next && decision.targetBotId === undefined && implementations.length === 1 ? implementations[0] : roles.find(row => row.id === decision.targetBotId) + if (!['route', 'return', 'start-next'].includes(decision.action) || (target?.policyRole ?? target?.role) !== 'developer' || decision.repositoryInputs !== undefined) throw chiefProtocolError('targetBotId', 'one active implementation role; choose explicitly when more than one is available, separate from repositoryInputs', decision.taskId) + let choice = decision.continuation + if (next && decision.continuationRef !== undefined && (typeof decision.continuationRef !== 'string' || decision.continuationRef !== facts?.expectedScope)) throw chiefProtocolError('continuationRef', 'the expectedScope of the continuationOptions read in this invocation', decision.taskId) + const source = choice === undefined ? decision.continuationRef === undefined ? 'chief-snapshot' : 'continuation-ref' : 'explicit' + if (next && choice === undefined) { + if (!facts || facts.status === 'unavailable' || facts.attemptId !== saved.attemptId || !Array.isArray(facts.modes) || !facts.modes.includes('next') || decision.continuationRef === undefined && (facts.modes.length !== 1 || facts.modes[0] !== 'next') || typeof facts.expectedScope !== 'string' || facts.expectedScope !== saved.deliverySnapshot?.expectedScope || saved.repositoryRecovery?.nextActions?.actions?.find(row => row.action === 'start-next')?.status !== 'ready') throw chiefProtocolError('continuation', 'current ready next-delivery facts with one unambiguous mode, or an explicit selection/reference', decision.taskId) + choice = { mode: 'next', previousAttemptId: facts.attemptId, generation: facts.generation, expectedWorkspace: structuredClone(facts.snapshot), expectedBaseSha: facts.latestBaseSha } + } + if (!facts || facts.status === 'unavailable' || !(next ? ['next'] : ['continue', 'latest']).includes(choice?.mode) || choice.previousAttemptId !== facts.attemptId || choice.generation !== facts.generation || !isDeepStrictEqual(choice.expectedWorkspace, facts.snapshot) || ['latest', 'next'].includes(choice.mode) && choice.expectedBaseSha !== facts.latestBaseSha) throw chiefProtocolError('continuation', 'the exact current continuationOptions snapshot, generation and selected latestBaseSha', decision.taskId) const response = await rt.projectDelivery.endpoint('project-attempt-continuation-state', { projectId, attemptId: facts.attemptId }, actor) const { nextActions: _nextActions, ...current } = response if (!isDeepStrictEqual(current, facts)) throw chiefProtocolError('continuation', 'fresh facts after the repository baseline or workspace changed', decision.taskId) - return target + const continuation = { mode: choice.mode, previousAttemptId: choice.previousAttemptId, generation: choice.generation, expectedWorkspace: structuredClone(choice.expectedWorkspace) } + if (['latest', 'next'].includes(choice.mode)) continuation.expectedBaseSha = choice.expectedBaseSha + return { target, continuation, source } } async function applyDecision(job, decision) { const { run } = job @@ -193,11 +212,15 @@ export function createProjectChiefControl(rt) { if ((task.collaboration?.generation ?? 0) !== saved.generation || task.deliveryAttemptId !== saved.attemptId && (task.deliveryAttemptId ?? null) !== saved.attemptId || (rt.state.projectAttempts?.[task.deliveryAttemptId]?.candidateSha ?? null) !== saved.candidateSha) throw refuse('Task progressed after the Chief snapshot; decision is stale') if (saved.deliverySnapshot && (saved.deliverySnapshot.status !== 'current' || saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task))) throw chiefProtocolError('snapshot', 'fresh Task, Attempt, requirements and evidence facts before action', task.id) if (digest(rt.state.projectAttempts?.[task.deliveryAttemptId]?.evidence?.merge ?? null) !== digest(saved.evidence?.merge ?? null)) throw chiefProtocolError('snapshot', 'fresh facts after merge evidence changed', task.id) - const continuationTarget = await continuationChoice(decision, saved, rt.projectAgents.deliveryDefinitions(run.projectId).roles, actor, run.projectId) + const continuationBinding = await continuationChoice(decision, saved, rt.projectAgents.deliveryDefinitions(run.projectId).roles, actor, run.projectId) + if (continuationBinding) await validateActor(actor) + if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw chiefProtocolError('targetBotId', 'fresh role catalog after continuation validation', task.id) + if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw chiefProtocolError('snapshot', 'fresh Task facts after continuation validation', task.id) const reason = typeof decision.reason === 'string' ? decision.reason.trim() : '' if (!reason || reason.length > 4000 || !['close', 'route', 'return', 'request-tests', 'needs-human', 'continue', 'admit', 'revise', 'cancel', 'reopen', 'resolve-dependency', 'retry-notification', 'reconcile-exit', 'stop-execution', 'accept-attempt', 'start-next'].includes(decision.action)) throw refuse('Chief decision needs a supported action and bounded reason') if (decision.action === 'needs-human') { assistanceRequest(decision); if (decision.recipient === 'operations' && !operationsContact(rt, task.projectId, policyOf(task.projectId).operationsContact)) throw refuse('No operations contact is configured') } const audit = { id: `chief-decision-${randomUUID()}`, runId: run.id, projectId: run.projectId, taskId: task.id, initiatedBy: run.initiatedBy, executorRole: 'chief', action: decision.action, reason, snapshotSha256: run.snapshotSha256, candidateSha: saved.candidateSha, expectedScope: saved.deliverySnapshot?.expectedScope ?? null, criteria: structuredClone(decision.criteria ?? []), inputReceiptSummary: inputDeclaration(decision, saved), status: 'requested', invocation: job.invocation, at: Date.now() } + if (continuationBinding) audit.continuationBinding = { targetBotId: continuationBinding.target.id, source: continuationBinding.source, expectedScope: saved.deliverySnapshot?.expectedScope ?? null, continuation: structuredClone(continuationBinding.continuation) } rt.state.projectChiefDecisions ??= {}; setEntry(rt.state.projectChiefDecisions, audit.id, audit) actor = rt.projectInputs?.chiefActor(actor, task.id, saved, run, decision.inputReceipts) ?? actor try { @@ -238,8 +261,8 @@ export function createProjectChiefControl(rt) { } else if (decision.action === 'accept-attempt') { await rt.projectDelivery.chiefAccept({ taskId: task.id, expectedSha: saved.candidateSha, decision: { ...audit, delivery: decision.delivery } }, actor) } else if (decision.action === 'start-next') { - if (!continuationTarget || decision.continuation?.mode !== 'next') throw chiefProtocolError('continuation', 'a fresh next-delivery snapshot and explicit implementation target', task.id) - const result = await rt.projectScheduler.request({ taskId: task.id, requestId: audit.id, targetRole: continuationTarget.role, continuation: decision.continuation }, actor) + if (!continuationBinding || continuationBinding.continuation.mode !== 'next') throw chiefProtocolError('continuation', 'a fresh next-delivery snapshot and explicit implementation target', task.id) + const result = await rt.projectScheduler.request({ taskId: task.id, requestId: audit.id, targetRole: continuationBinding.target.role, continuation: continuationBinding.continuation }, actor) audit.successorAttemptId = result.attempt?.id ?? null; audit.queueId = result.queue?.id ?? null } else if (decision.action === 'close') { rt.projectTaskLifecycle?.requireCompletedDependencies(task) @@ -251,8 +274,8 @@ export function createProjectChiefControl(rt) { const target = decision.action === 'request-tests' ? `${run.projectId}:tester` : decision.targetBotId if (typeof target !== 'string') throw refuse('Chief routing needs an explicit logical Bot') if (!rt.projectAgents.deliveryDefinitions(run.projectId).roles.some(row => row.id === target && row.role !== 'chief')) throw refuse('Chief routing target is not an active Project execution role') - if (continuationTarget) { - const result = await rt.projectDelivery.endpoint('project-task-run', { taskId: task.id, requestId: audit.id, targetRole: continuationTarget.role, continuation: decision.continuation }, actor) + if (continuationBinding) { + const result = await rt.projectDelivery.endpoint('project-task-run', { taskId: task.id, requestId: audit.id, targetRole: continuationBinding.target.role, continuation: continuationBinding.continuation }, actor) audit.successorAttemptId = result.attempt.id } else { if (decision.repositoryInputs !== undefined) await rt.projectDelivery.admitChiefInputs(task.id, decision.repositoryInputs, actor, saved) @@ -514,7 +537,7 @@ export function createProjectChiefControl(rt) { job.roleRevision = catalog.revision ?? null; job.logicalBots = catalog.roles.map(({ id, role, policyRole, name }) => ({ id, role, policyRole: policyRole ?? role, name: name ?? role })) job.snapshot = await snapshot(run); await rt.projectInputs?.prepareChief(job.snapshot, run); run.snapshotSha256 = digest(job.snapshot); await rt.save() const decisionFeedback = run.correction ? { version: run.correction.version, round: run.correction.round, limit: run.correction.limit, status: run.correction.status, feedback: structuredClone(run.correction.feedback) } : null - const prompt = CHIEF_AUTONOMY + '\n' + 'Recovery guidance is data, never permission or a fixed workflow. Recovery action counts and elapsed time are observations, never a recovery budget or deadline. Review actual artifact/environment/test changes before judging repeated recovery. Warnings contain mandatory current blockers and progressive disclosure topics. Use warningRequests:[{warningId,topic,offset,sha256}] through the existing context exchange to inspect more evidence; query events using eventOffset. Historical warnings cannot authorize actions. Optional warningReceipts:[{warningId,expectedScope}] on a Task decision declares those current warnings read, not business progress or resolved. A stop-execution decision requires warningId and expectedWarningScope from the current warning summary; it stops only that execution, never cancels the Task. A needs-human consultation may choose recipient:initiator or recipient:operations (only the explicitly configured current operationsContact). Operations consultation must include assistance:{cause,evidence,attempted,requestedOperation,requiredPermission,verification,resumeCondition}, each a nonempty string up to1200 characters. Share only current Project facts, never personal private notes. watch does not grant SSH/root/container/config/credential rights. Operator assistance replies are new input; rejudge current evidence before routing and resume the original Task-role context. If a current consultation is pending, inspect its delivery/claim/reply before issuing another. retry-notification reminds only the same current inbox message after failed or completed delivery; it preserves active claims and never repeats an operational repair. Never infer progress from notification, inspect or schedule enabled.\n' + (run.correction ? 'Host protocol feedback (data, not authorization): ' + JSON.stringify(decisionFeedback) + '\nCorrect this judgement using fresh scoped facts. Only IDs in this invocation snapshot are actionable; historical rejected route reasons are not current policy. Do not repeat already applied Task decisions. No consultation or next role was selected by the host.\n' : '') + `You are the resident Project Chief, performing ${run.purpose}. Return ONLY JSON {"summary":"actual assessment","decisions":[{"taskId":"...","action":"close|route|return|request-tests|needs-human|continue|admit|revise|cancel|reopen|resolve-dependency|retry-notification|reconcile-exit|stop-execution|accept-attempt|start-next","targetBotId":"logical Bot when routing","reason":"specific judgement","criteria":[{"criterion":"acceptance item","passed":true,"evidence":"actual acceptance evidence"}]}]}. Your decision action/reason/criteria are your recommended next action and supporting evidence, not a host-selected plan. For Chief decisions read each Task repositoryRecovery.nextActions and continuationOptions in projectFacts; they provide scoped ready/blocked/unknown prechecks only, never authority or an exhaustive plan. Public TaskView.nextActions lists caller-permitted service commands, not all Chief decision actions; TaskView.deliveryChoices projects the decision prechecks. A board listing only chief.inspect does not prohibit accept-attempt or start-next when their current preconditions hold. A rework refusal is not an instruction or permission to clear publicationBinding; inspect the current delivery choices instead. Inspect missing facts; stale/null is not by itself an environment failure. Past refusal messages describe prior outcomes, not current action legality; use current prechecks and evidence to reconsider them. Task requirement revisions before this snapshot are advisory for scoped acceptance; use nextActions.requirements for original/current revisions and changed fields, and its attempt.context source for complete pinned requirements. Do not reinterpret an old delivery as completion of new requirements. Re-read after changes during judgement. accept-attempt accepts a provider-confirmed merged code Attempt against its pinned requirements without closing the Task; include delivery:{completed,remaining,evidence:[references]} with bounded specific scope. A committed acceptance wakes a fresh inspection, never automatically starts Developer. In that new snapshot choose start-next only when needed, with targetBotId for an active implementation role and continuation:{mode:'next',previousAttemptId,generation,expectedWorkspace,expectedBaseSha} copied from fresh continuationOptions. This pins verified target HEAD containing the merged result and creates a new PR identity while retaining Task-role context/history. Never rework or inherit the merged PR binding. close remains available only when all Task criteria/dependencies/inputs are satisfied. Personal Agent assistance is the last resort when actual information/authority or a human decision is missing; failure counts/time never automatically escalate. A conversation/turn completion is not Task completion. Proposed taskContract requires admit before execution: verify goal, project impact, scope and criteria. Prefer admit without repeating already adequate requirements. For admit/revise, every provided requirements.title, requirements.criteria, requirements.impact and requirements.scope must be a bounded nonempty string. In particular requirements.criteria is a string; the separate outer decision.criteria is an evidence array and must never be copied into requirements.criteria. Optional requirements.repository is an associated owner/name or null for a repository-free decision. Uncertain intent requires needs-human. Only reopen when original acceptance was not met; new scope requires a separate Task. Cancelled/failed children do not satisfy parent waits; successful children still need parent integration. All Task mutations remain scoped Chief decisions. To resolve a failed/cancelled child condition without faking child completion, use resolve-dependency with dependency:{taskId,version:the exact current dependency snapshot version,evidence:[actual references],replacementTaskId:optional accepted alternative,replacementVersion:its task outcome version}. This is explicit Chief judgement, not host test success or parent completion. Leave unresolved conditions blocked when evidence is insufficient. Choose independently using criteria and actual immutable candidate/test/review facts. reconcile-exit explicitly retires only an exact exit-proven old claim without selecting a next role; consult repositoryRecovery.executionCheckpoint. Historical refusals are audit data, not current routing policy. Candidate Reviewer approve triggers the host merge flow only when explicitly configured for this Project/repository/branch. It is not a signed repository Reviewer PASS, actual forge approval, merged result or release permit. A configured merge must be provider-confirmed before overall code Task acceptance; pending/blocked merge facts are not completion. Do not close missing or stale evidence. Host allocationProof.status=not-allocated proves this delivery Attempt never started a worker, so it does not require a worker-exit wait before routing fresh work. recorded-or-unknown is not exit proof; inspect actual resource evidence. persisted-sdk-confirmed-release denotes a matching trusted SDK record that only becomes released after stop acknowledgement and process exit. Choose an evidence-supported role after environment or execution repair; no default return to implementation is required. Unsealed work can continue, including across restart, only after every registered worker and lease has confirmed release. A sealed candidate may resume Tester or Reviewer under the current verificationRecovery conditions; return explicitly requests a new rework Attempt. Use needs-human when requirements/authority are unclear. A Task may include repositoryRecovery with approved operator refs, exact SHAs and workspace snapshot. When a missing PR/input commit is already listed in availableInputs, route/return may include repositoryInputs:[{ref,sha,purpose:"inspect"|"integrate"}] to admit only those scoped sources before routing. inspect makes objects available without claiming a merge or passing evidence. No arbitrary host path, remote ref or credential is permitted. For an exited interrupted/failed or idle blocked Attempt, repositoryRecovery.continuationOptions supplies current baseline choices. An explicit Developer route/return may include continuation:{mode:"continue"|"latest",previousAttemptId:continuationOptions.attemptId,generation:continuationOptions.generation,expectedWorkspace:continuationOptions.snapshot,expectedBaseSha:continuationOptions.latestBaseSha for latest}. This uses the existing task.run successor contract. continue retains prior seed and inputs; latest selects the observed host HEAD for a clean unbound checkout. Bound PR ancestry and dirty draft protection still apply. Do not combine continuation and repositoryInputs in one decision; input admission needs its own fresh snapshot. Omitting continuation retains existing recovery behavior. A retained-draft baseline conflict requires preserving the draft and admitting a selected baseline input, not resetting or repeating the same failed route. Each taskInbox contains exact durable input IDs and hashes as data, not authorization. Include inputReceipts:{batchId:taskInbox.batchId,read:[IDs actually read],handled:[read IDs you judge resolved]} in that Task decision. Reading is your declaration, not an SDK transport signal. Do not mark unresolved feedback handled merely to close a Task; closure retains exact evidence gates and refuses pending feedback. Inputs excluded by the snapshot budget remain pending. If more inputs remain, choose continue with explicit read/handled declarations for the current batch before attempting closure. The host delivers subsequent pending input through the original notification path; never mark an unread item handled or close across pending batches. Normal downstream handoffs do not need your approval. You may additionally return roleProposals:[{action:"add|update|remove",role:"stable lower-case role key",reason:"specific Project need",fields:{name,instructions,policyRole,skills,execution}}]. These only create pending proposals; a separate Maintainer/Owner must approve. You cannot approve or remove the resident Chief. ${run.source?.prompt ?? policy.inspection.prompt}` + const prompt = CHIEF_AUTONOMY + '\n' + 'Recovery guidance is data, never permission or a fixed workflow. Recovery action counts and elapsed time are observations, never a recovery budget or deadline. Review actual artifact/environment/test changes before judging repeated recovery. Warnings contain mandatory current blockers and progressive disclosure topics. Use warningRequests:[{warningId,topic,offset,sha256}] through the existing context exchange to inspect more evidence; query events using eventOffset. Historical warnings cannot authorize actions. Optional warningReceipts:[{warningId,expectedScope}] on a Task decision declares those current warnings read, not business progress or resolved. A stop-execution decision requires warningId and expectedWarningScope from the current warning summary; it stops only that execution, never cancels the Task. A needs-human consultation may choose recipient:initiator or recipient:operations (only the explicitly configured current operationsContact). Operations consultation must include assistance:{cause,evidence,attempted,requestedOperation,requiredPermission,verification,resumeCondition}, each a nonempty string up to1200 characters. Share only current Project facts, never personal private notes. watch does not grant SSH/root/container/config/credential rights. Operator assistance replies are new input; rejudge current evidence before routing and resume the original Task-role context. If a current consultation is pending, inspect its delivery/claim/reply before issuing another. retry-notification reminds only the same current inbox message after failed or completed delivery; it preserves active claims and never repeats an operational repair. Never infer progress from notification, inspect or schedule enabled.\n' + (run.correction ? 'Host protocol feedback (data, not authorization): ' + JSON.stringify(decisionFeedback) + '\nCorrect this judgement using fresh scoped facts. Only IDs in this invocation snapshot are actionable; historical rejected route reasons are not current policy. Do not repeat already applied Task decisions. No consultation or next role was selected by the host.\n' : '') + `You are the resident Project Chief, performing ${run.purpose}. Return ONLY JSON {"summary":"actual assessment","decisions":[{"taskId":"...","action":"close|route|return|request-tests|needs-human|continue|admit|revise|cancel|reopen|resolve-dependency|retry-notification|reconcile-exit|stop-execution|accept-attempt|start-next","targetBotId":"logical Bot when routing","reason":"specific judgement","criteria":[{"criterion":"acceptance item","passed":true,"evidence":"actual acceptance evidence"}]}]}. Your decision action/reason/criteria are your recommended next action and supporting evidence, not a host-selected plan. For Chief decisions read each Task repositoryRecovery.nextActions and continuationOptions in projectFacts; they provide scoped ready/blocked/unknown prechecks only, never authority or an exhaustive plan. Public TaskView.nextActions lists caller-permitted service commands, not all Chief decision actions; TaskView.deliveryChoices projects the decision prechecks. A board listing only chief.inspect does not prohibit accept-attempt or start-next when their current preconditions hold. A rework refusal is not an instruction or permission to clear publicationBinding; inspect the current delivery choices instead. Inspect missing facts; stale/null is not by itself an environment failure. Past refusal messages describe prior outcomes, not current action legality; use current prechecks and evidence to reconsider them. Task requirement revisions before this snapshot are advisory for scoped acceptance; use nextActions.requirements for original/current revisions and changed fields, and its attempt.context source for complete pinned requirements. Do not reinterpret an old delivery as completion of new requirements. Re-read after changes during judgement. accept-attempt accepts a provider-confirmed merged code Attempt against its pinned requirements without closing the Task; include delivery:{completed,remaining,evidence:[references]} with bounded specific scope. A committed acceptance wakes a fresh inspection, never automatically starts Developer. Minimal start-next decision: {"taskId":"the snapshotted Task ID","action":"start-next","reason":"why another delivery is needed"}. In that new snapshot choose start-next with taskId and reason only when needed; omit continuation when current facts expose only mode next and one implementation role. The host binds the continuationOptions you saw and revalidates current facts; it never silently switches baseline. If several implementation roles exist, provide targetBotId. Optional continuationRef is the expectedScope from those facts. Multiple modes need an explicit reference or full continuation selection. Supplied continuation fields must match exactly and are never repaired. Use existing attempt.continuation-state after feedback to inspect refreshed facts. This pins verified target HEAD containing the merged result and creates a new PR identity while retaining Task-role context/history. Never rework or inherit the merged PR binding. close remains available only when all Task criteria/dependencies/inputs are satisfied. Personal Agent assistance is the last resort when actual information/authority or a human decision is missing; failure counts/time never automatically escalate. A conversation/turn completion is not Task completion. Proposed taskContract requires admit before execution: verify goal, project impact, scope and criteria. Prefer admit without repeating already adequate requirements. For admit/revise, every provided requirements.title, requirements.criteria, requirements.impact and requirements.scope must be a bounded nonempty string. In particular requirements.criteria is a string; the separate outer decision.criteria is an evidence array and must never be copied into requirements.criteria. Optional requirements.repository is an associated owner/name or null for a repository-free decision. Uncertain intent requires needs-human. Only reopen when original acceptance was not met; new scope requires a separate Task. Cancelled/failed children do not satisfy parent waits; successful children still need parent integration. All Task mutations remain scoped Chief decisions. To resolve a failed/cancelled child condition without faking child completion, use resolve-dependency with dependency:{taskId,version:the exact current dependency snapshot version,evidence:[actual references],replacementTaskId:optional accepted alternative,replacementVersion:its task outcome version}. This is explicit Chief judgement, not host test success or parent completion. Leave unresolved conditions blocked when evidence is insufficient. Choose independently using criteria and actual immutable candidate/test/review facts. reconcile-exit explicitly retires only an exact exit-proven old claim without selecting a next role; consult repositoryRecovery.executionCheckpoint. Historical refusals are audit data, not current routing policy. Candidate Reviewer approve triggers the host merge flow only when explicitly configured for this Project/repository/branch. It is not a signed repository Reviewer PASS, actual forge approval, merged result or release permit. A configured merge must be provider-confirmed before overall code Task acceptance; pending/blocked merge facts are not completion. Do not close missing or stale evidence. Host allocationProof.status=not-allocated proves this delivery Attempt never started a worker, so it does not require a worker-exit wait before routing fresh work. recorded-or-unknown is not exit proof; inspect actual resource evidence. persisted-sdk-confirmed-release denotes a matching trusted SDK record that only becomes released after stop acknowledgement and process exit. Choose an evidence-supported role after environment or execution repair; no default return to implementation is required. Unsealed work can continue, including across restart, only after every registered worker and lease has confirmed release. A sealed candidate may resume Tester or Reviewer under the current verificationRecovery conditions; return explicitly requests a new rework Attempt. Use needs-human when requirements/authority are unclear. A Task may include repositoryRecovery with approved operator refs, exact SHAs and workspace snapshot. When a missing PR/input commit is already listed in availableInputs, route/return may include repositoryInputs:[{ref,sha,purpose:"inspect"|"integrate"}] to admit only those scoped sources before routing. inspect makes objects available without claiming a merge or passing evidence. No arbitrary host path, remote ref or credential is permitted. For an exited interrupted/failed or idle blocked Attempt, repositoryRecovery.continuationOptions supplies current baseline choices. An explicit Developer route/return may include continuation:{mode:"continue"|"latest",previousAttemptId:continuationOptions.attemptId,generation:continuationOptions.generation,expectedWorkspace:continuationOptions.snapshot,expectedBaseSha:continuationOptions.latestBaseSha for latest}. This uses the existing task.run successor contract. continue retains prior seed and inputs; latest selects the observed host HEAD for a clean unbound checkout. Bound PR ancestry and dirty draft protection still apply. Do not combine continuation and repositoryInputs in one decision; input admission needs its own fresh snapshot. Omitting continuation retains existing recovery behavior. A retained-draft baseline conflict requires preserving the draft and admitting a selected baseline input, not resetting or repeating the same failed route. Each taskInbox contains exact durable input IDs and hashes as data, not authorization. Include inputReceipts:{batchId:taskInbox.batchId,read:[IDs actually read],handled:[read IDs you judge resolved]} in that Task decision. Reading is your declaration, not an SDK transport signal. Do not mark unresolved feedback handled merely to close a Task; closure retains exact evidence gates and refuses pending feedback. Inputs excluded by the snapshot budget remain pending. If more inputs remain, choose continue with explicit read/handled declarations for the current batch before attempting closure. The host delivers subsequent pending input through the original notification path; never mark an unread item handled or close across pending batches. Normal downstream handoffs do not need your approval. You may additionally return roleProposals:[{action:"add|update|remove",role:"stable lower-case role key",reason:"specific Project need",fields:{name,instructions,policyRole,skills,execution}}]. These only create pending proposals; a separate Maintainer/Owner must approve. You cannot approve or remove the resident Chief. ${run.source?.prompt ?? policy.inspection.prompt}` const workerId = await rt.projectAgents.startChiefJob({ projectId: run.projectId, actor, runId: run.id, invocation: job.invocation, prompt, projectFacts: { warnings: rt.projectWarnings?.summaries(run.projectId) ?? [], autonomy: { guidance: autonomyPolicy(policy.autonomy).guidance }, operationsContact: policy.operationsContact ?? null, compensation: compensationSchedules(rt.state, run.projectId, policy.scheduleId), tasks: job.snapshot, logicalBots: job.logicalBots, decisionFeedback, requestedChange: run.source?.requestedChange ?? null }, onResult: result => completed(job, result) }) if (run.invocation === job.invocation) { run.workerId = workerId; if (run.status === 'running') run.waitingFor = 'execution'; run.workerGeneration = rt.state.projectHarnessWorkers?.[workerId]?.generation ?? null; await rt.save() } } catch (error) { diff --git a/patches/ds-bot/files/src/host/project-chief-protocol.js b/patches/ds-bot/files/src/host/project-chief-protocol.js index ff765970..435a6167 100644 --- a/patches/ds-bot/files/src/host/project-chief-protocol.js +++ b/patches/ds-bot/files/src/host/project-chief-protocol.js @@ -19,9 +19,11 @@ export function parseChiefResponse(text, roles, { onInvalid } = {}) { seen.add(taskId) if (!actions.has(decision.action)) throw chiefProtocolError('action', 'a supported Chief action', taskId) if (typeof decision.reason !== 'string' || !decision.reason.trim() || decision.reason.length > 4000) throw chiefProtocolError('reason', 'bounded nonempty text up to4000 characters', taskId) - if (['route', 'return', 'start-next'].includes(decision.action)) { + if (['route', 'return'].includes(decision.action) || decision.action === 'start-next' && decision.targetBotId !== undefined) { if (typeof decision.targetBotId !== 'string' || !decision.targetBotId.trim() || !roles.some(row => row.id === decision.targetBotId && row.role !== 'chief')) throw chiefProtocolError('targetBotId', 'one exact active logical Bot ID from logicalBots, excluding Chief', taskId) } + if (decision.mode !== undefined && (decision.action !== 'start-next' || decision.mode !== 'next')) throw chiefProtocolError('mode', 'next only for start-next', taskId) + if (decision.continuationRef !== undefined && (decision.action !== 'start-next' || typeof decision.continuationRef !== 'string' || !decision.continuationRef)) throw chiefProtocolError('continuationRef', 'a nonempty scoped reference only for start-next', taskId) if (decision.criteria !== undefined) { if (!Array.isArray(decision.criteria)) throw chiefProtocolError('criteria', 'an evidence array, separate from requirements.criteria text', taskId) for (const [index, item] of decision.criteria.entries()) { diff --git a/patches/ds-bot/files/src/host/project-delivery-options.js b/patches/ds-bot/files/src/host/project-delivery-options.js index 30791c38..439107ec 100644 --- a/patches/ds-bot/files/src/host/project-delivery-options.js +++ b/patches/ds-bot/files/src/host/project-delivery-options.js @@ -40,7 +40,7 @@ export function deliveryChoices(state, task, observation) { const acceptance = option('accept-attempt', () => requireAttemptAcceptance(task, a), 'Chief must judge delivered scope and remaining work; execution rechecks all evidence') if (acceptance.status === 'ready' && (observation?.status === 'unknown' || !merged)) Object.assign(acceptance, { status: observation?.status === 'unknown' ? 'unknown' : 'blocked', reason: observation?.status === 'unknown' ? 'Query the current provider merge result; no merge outcome is inferred' : 'Provider-confirmed merge is required for scoped acceptance' }) return { requirements, expectedScope: scope, observedAt: Date.now(), admission: admissions(state, task).slice(-4), observation: observation ?? null, recommendation: previous?.taskId === task.id && previous.status === 'applied' ? { decisionId: previous.id, action: previous.action, reason: previous.reason, criteria: previous.criteria, expectedScope: previous.expectedScope, alreadyApplied: true } : null, recommendationSource: 'Chief decision action/reason/evidence in the original inference turn', actions: [acceptance, - option('start-next', () => requireNextAttempt(task, a), 'Read fresh continuation-state; pin verified target HEAD, choose an implementation role and use task.run mode next'), + option('start-next', () => requireNextAttempt(task, a), 'Chief may reference observed continuationOptions with start-next; choose ambiguous roles/modes explicitly. Execution revalidates the target; external task.run keeps explicit fields'), merged ? { action: 'return', status: 'blocked', reason: 'A provider-confirmed merged PR cannot be reworked or reused for publication' } : { action: 'return', status: observation?.status === 'unknown' ? 'unknown' : 'ready', reason: 'Existing route/rework, exit, binding and current authority gates still apply' }, { action: 'close', status: 'unknown', reason: 'Chief must establish every Task criterion, dependency and input; merged or accepted Attempt alone does not complete the Task' }, { action: 'continue', status: 'ready', reason: 'Inspect current scoped facts or defer judgement without claiming completion' }, diff --git a/patches/ds-bot/files/test/project-chief-control.test.js b/patches/ds-bot/files/test/project-chief-control.test.js index 4275be7d..a63d22b8 100644 --- a/patches/ds-bot/files/test/project-chief-control.test.js +++ b/patches/ds-bot/files/test/project-chief-control.test.js @@ -1,6 +1,8 @@ import assert from 'node:assert/strict' +import { createHash } from 'node:crypto' import { test } from 'node:test' import { createProjectChiefControl, inspectionFacts, allocationProof, inspectionResources } from '../src/host/project-chief-control.js' +const createHashForTest = value => createHash('sha256').update(JSON.stringify(value)).digest('hex') const thresholds = { maxHandoffsWithoutProgress: 3, longExecutionMs: 10000, missingHeartbeatMs: 5000, longWaitMs: 8000 } test('Chief inspection counts verified artifact progress rather than message or transfer churn', () => { const task = { id: 't', updatedAt: 1, deliveryAttemptId: 'a', collaboration: { status: 'queued', generation: 7, history: Array.from({ length: 5 }, (_, index) => ({ evidence: [], reason: 'Progress message ' + index, createdAt: index + 1 })) } } @@ -879,3 +881,95 @@ test('requirement changes after scoped acceptance judgement request fresh facts assert.equal(run.status, 'done'); assert.equal(f.accepted(), 0) await f.rt.projectEffects.close() }) + +async function referenceFixture({ modes = ['next'], targetCount = 1, policyRole = 'developer', read } = {}) { + const { deliveryScope } = await import('../src/host/project-delivery-options.js') + let facts, calls = [], catalog, reads = 0 + const f = await controlFixture({ prepare(rt, task) { + const role = policyRole === 'developer' ? 'developer' : 'implementer' + catalog = { revision: 1, roles: [{ id: 'p:chief', role: 'chief' }, ...Array.from({ length: targetCount }, (_, index) => ({ id: 'p:' + role + (index ? index : ''), role: role + (index ? index : ''), policyRole: 'developer' }))] } + rt.projectAgents.deliveryDefinitions = () => catalog + const scope = deliveryScope(rt.state, task) + facts = { attemptId: 'a', generation: 1, snapshot: { workspaceId: 'w', sha256: 'workspace' }, latestBaseSha: 'a'.repeat(40), modes, expectedScope: scope } + rt.projectDelivery.recoveryFacts = async () => { facts.expectedScope = deliveryScope(rt.state, task); return { continuationOptions: structuredClone(facts), nextActions: { actions: [{ action: 'start-next', status: 'ready' }] } } } + rt.projectDelivery.endpoint = async () => { ++reads; return read ? read(facts, rt, task, catalog, reads) : structuredClone(facts) } + rt.projectScheduler = { wake() {}, request: async (args) => { calls.push(structuredClone(args)); return { queue: { id: 'queue-reference' } } } } + } }) + return { ...f, facts, calls, catalog } +} + +test('next snapshot reference binds current fields and the unique custom implementation role without mutating the decision', async () => { + const f = await referenceFixture({ policyRole: 'custom' }) + const decision = { action: 'start-next', mode: 'next', continuationRef: f.facts.expectedScope } + const original = structuredClone(decision), result = f.result(decision) + await f.request.onResult(result); await f.request.onResult(result) + assert.deepEqual(decision, original); assert.equal(f.calls.length, 1, JSON.stringify(f.rt.state.projectExecutions)) + assert.equal(f.calls[0].targetRole, 'implementer') + const audit = Object.values(f.rt.state.projectChiefDecisions).find(row => row.status === 'applied') + assert.equal(audit.continuationBinding.source, 'continuation-ref') + assert.equal(audit.continuationBinding.targetBotId, 'p:implementer') + assert.equal(audit.queueId, 'queue-reference'); assert.equal(audit.successorAttemptId, null) + assert.equal(f.calls[0].continuation.expectedBaseSha, f.facts.latestBaseSha) + assert.deepEqual(f.calls[0].continuation.expectedWorkspace, f.facts.snapshot) + const intent = createHashForTest(JSON.parse(result.text).decisions[0]) + assert.equal(f.rt.state.projectExecutions[f.request.runId].appliedDecisions[0].intent, intent) + const visible = f.control.view('p', { memberId: 'owner' }).runs.find(run => run.id === f.request.runId).appliedDecisions[0].continuationBinding + assert.deepEqual(visible, audit.continuationBinding) + visible.continuation.expectedBaseSha = 'changed-view-only'; assert.equal(audit.continuationBinding.continuation.expectedBaseSha, f.facts.latestBaseSha) + audit.projectId = 'foreign' + assert.equal(f.control.view('p', { memberId: 'owner' }).runs.find(run => run.id === f.request.runId).appliedDecisions[0].continuationBinding, undefined) +}) + +for (const kind of ['ambiguous-target', 'no-target', 'ambiguous-mode', 'unavailable', 'wrong-ref', 'bad-fields', 'bad-mode', 'not-ready', 'foreign-attempt', 'target-changed', 'workspace-changed', 'local-changed', 'roles-changed', 'rights-revoked']) test('next reference refuses ' + kind + ' without dispatch or inferred recovery', async () => { + const f = await referenceFixture({ targetCount: kind === 'ambiguous-target' ? 2 : kind === 'no-target' ? 0 : 1, modes: kind === 'ambiguous-mode' ? ['next', 'latest'] : ['next'], read: async (facts, rt, task, catalog, reads) => { + if (kind === 'target-changed') return { ...structuredClone(facts), latestBaseSha: 'b'.repeat(40) } + if (kind === 'workspace-changed') return { ...structuredClone(facts), snapshot: { workspaceId: 'other' } } + if (kind === 'local-changed' && reads === 2) task.criteria = 'Changed during validation' + if (kind === 'roles-changed' && reads === 2) catalog.revision++ + if (kind === 'rights-revoked' && reads === 2) rt.requireProjectAccess = () => { throw Error('Current role denied') } + return structuredClone(facts) + } }) + const saved = f.request.projectFacts.tasks[0] + if (kind === 'unavailable') saved.repositoryRecovery.continuationOptions.status = 'unavailable' + if (kind === 'not-ready') saved.repositoryRecovery.nextActions.actions[0].status = 'blocked' + if (kind === 'foreign-attempt') saved.repositoryRecovery.continuationOptions.attemptId = 'other' + const selection = kind === 'wrong-ref' ? { continuationRef: 'wrong' } : kind === 'bad-fields' ? { continuation: { mode: 'next', previousAttemptId: 'a', generation: 1, expectedWorkspace: {}, expectedBaseSha: f.facts.latestBaseSha } } : kind === 'bad-mode' ? { mode: 'latest' } : {} + await f.request.onResult(f.result({ action: 'start-next', ...selection })) + assert.equal(f.calls.length, 0) + assert.equal(f.task.chiefQuestion, undefined) + const run = f.rt.state.projectExecutions[f.request.runId] + assert.notEqual(run.status, 'done') + const expected = kind === 'roles-changed' ? /fresh role catalog/ : kind === 'rights-revoked' ? /Current role denied/ : kind === 'local-changed' ? /fresh Task facts/ : ['target-changed', 'workspace-changed'].includes(kind) ? /fresh facts after/ : kind === 'wrong-ref' ? /expectedScope/ : kind === 'bad-mode' ? /next only/ : kind.includes('target') ? /active implementation role/ : /continuation/ + assert.match(run.failure, expected) +}) + +test('explicit target resolves multiple implementation roles, and explicit scope reference resolves multiple modes', async () => { + const f = await referenceFixture({ targetCount: 2, modes: ['next', 'latest'] }) + await f.request.onResult(f.result({ action: 'start-next', targetBotId: 'p:developer1', continuationRef: f.facts.expectedScope })) + assert.equal(f.calls.length, 1, JSON.stringify(f.rt.state.projectExecutions)); assert.equal(f.calls[0].targetRole, 'developer1') +}) + +for (const targetBotId of [null, '', 'p:unknown', 'p:chief', 3]) test('explicit invalid next target remains a protocol error: ' + JSON.stringify(targetBotId), async () => { + const f = await referenceFixture() + await f.request.onResult(f.result({ action: 'start-next', targetBotId })) + assert.equal(f.calls.length, 0) + assert.match(f.rt.state.projectExecutions[f.request.runId].failure, /targetBotId/) +}) +test('an incomplete explicit continuation is refused rather than silently filled from the snapshot', async () => { + const f = await referenceFixture() + await f.request.onResult(f.result({ action: 'start-next', continuation: { mode: 'next' } })) + assert.equal(f.calls.length, 0); assert.match(f.rt.state.projectExecutions[f.request.runId].failure, /exact current continuationOptions/) +}) +for (const ref of [null, {}, '']) test('invalid reference type is not treated as omission: ' + JSON.stringify(ref), async () => { + const f = await referenceFixture() + await f.request.onResult(f.result({ action: 'start-next', continuationRef: ref })) + assert.equal(f.calls.length, 0); assert.match(f.rt.state.projectExecutions[f.request.runId].failure, /continuationRef/) +}) + +for (const [field, value] of [['mode', 'next'], ['continuationRef', 'scope']]) test('next-only reference field cannot be silently used on another action: ' + field, async () => { + const f = await referenceFixture() + await f.request.onResult(f.result({ action: 'continue', [field]: value })) + assert.equal(f.calls.length, 0) + assert.equal(f.rt.state.projectExecutions[f.request.runId].status, 'correction-pending') + assert.match(f.rt.state.projectExecutions[f.request.runId].failure, new RegExp(field)) +}) diff --git a/patches/ds-bot/files/test/project-delivery.test.js b/patches/ds-bot/files/test/project-delivery.test.js index 620f407f..49d87ee2 100644 --- a/patches/ds-bot/files/test/project-delivery.test.js +++ b/patches/ds-bot/files/test/project-delivery.test.js @@ -2559,7 +2559,7 @@ test('pending external write and unknown native exit block scoped acceptance; cu await assert.rejects(f.rt.projectEndpoint('project-attempt-cleanup', { attemptId: f.a.id }, alice), /unsettled Task/) }) -test('Chief chooses acceptance then a fresh explicit next delivery through its original decision loop', async t => { +for (const form of ['explicit', 'snapshot', 'snapshot-with-target', 'reference-token']) test('Chief chooses acceptance then a fresh next delivery through its original decision loop: ' + form, async t => { const f = await mergedDelivery(t, { modern: true }), requests = [], rt = f.rt const { reviseTask } = await import('../src/host/project-task-lifecycle.js') const revision = { projectId: 'p1', taskId: f.task.id, executorRole: 'chief', runId: 'fixture-chief', initiatedBy: alice.memberId, reason: 'Record remaining work' } @@ -2584,6 +2584,7 @@ test('Chief chooses acceptance then a fresh explicit next delivery through its o assert.match(requests[0].prompt, /repositoryRecovery\.nextActions and continuationOptions in projectFacts/) assert.match(requests[0].prompt, /Public TaskView\.nextActions lists caller-permitted service commands, not all Chief decision actions/) assert.match(requests[0].prompt, /not an instruction or permission to clear publicationBinding/) + assert.match(requests[0].prompt, /Minimal start-next decision:/) await reply(requests[0], { action: 'accept-attempt', reason: 'Accept the first verified delivery, keep the overall goal open', delivery: { completed: 'Verified addition implementation', remaining: 'Deliver the next requirement', evidence: ['candidate, tests, Review and provider merge'] } }) assert.equal(rt.state.projectExecutions[requests[0].runId].status, 'done') assert.equal(f.a.status, 'accepted'); assert.notEqual(f.task.status, 'accepted') @@ -2595,11 +2596,17 @@ test('Chief chooses acceptance then a fresh explicit next delivery through its o assert.equal(saved.repositoryRecovery.nextActions.actions.find(row => row.action === 'return').status, 'blocked') assert.equal(saved.repositoryRecovery.nextActions.recommendation.action, 'accept-attempt') assert.equal(saved.repositoryRecovery.nextActions.recommendation.alreadyApplied, true) - await reply(requests[1], { action: 'start-next', targetBotId: 'p1:developer', reason: 'Choose the next independent delivery from verified target HEAD', continuation: { mode: 'next', previousAttemptId: facts.attemptId, generation: facts.generation, expectedWorkspace: facts.snapshot, expectedBaseSha: facts.latestBaseSha } }) + const selection = form === 'explicit' ? { targetBotId: 'p1:developer', continuation: { mode: 'next', previousAttemptId: facts.attemptId, generation: facts.generation, expectedWorkspace: facts.snapshot, expectedBaseSha: facts.latestBaseSha, clientNote: 'not an execution field' } } : form === 'snapshot-with-target' ? { targetBotId: 'p1:developer' } : form === 'reference-token' ? { continuationRef: facts.expectedScope } : {} + await reply(requests[1], { action: 'start-next', reason: 'Choose the next independent delivery from verified target HEAD', ...selection }) assert.equal(rt.state.projectExecutions[requests[1].runId].status, 'done', rt.state.projectExecutions[requests[1].runId].failure) const next = rt.state.projectAttempts[f.task.deliveryAttemptId] assert.notEqual(next.id, f.a.id); assert.equal(next.baseSha, f.a.candidateSha); assert.equal(next.publicationBinding, undefined) assert.equal(next.requirement.projectTaskRevision, 4) + const audit = Object.values(rt.state.projectChiefDecisions).find(row => row.action === 'start-next' && row.status === 'applied') + assert.equal(audit.continuationBinding.targetBotId, 'p1:developer') + assert.equal(audit.continuationBinding.continuation.expectedBaseSha, facts.latestBaseSha) + assert.deepEqual(audit.continuationBinding.continuation.expectedWorkspace, facts.snapshot) + assert.equal(audit.continuationBinding.continuation.clientNote, undefined) assert.notEqual(f.task.status, 'accepted'); assert.equal(f.task.chiefQuestion, undefined) }) diff --git a/scripts/api-contracts.mjs b/scripts/api-contracts.mjs index c554e350..773827bd 100644 --- a/scripts/api-contracts.mjs +++ b/scripts/api-contracts.mjs @@ -426,6 +426,9 @@ native.get('project-terminal-open').notes += ' Returns the actual HOME/configDir commands.get('task.run').request.properties.continuation = obj({ mode: choice('continue', 'latest', 'next'), expectedBaseSha: sha, previousAttemptId: id, generation: int(), expectedWorkspace: json }, ['mode', 'previousAttemptId', 'generation', 'expectedWorkspace']) commands.get('task.run').request.properties.continuation.allOf = [{ if: { properties: { mode: { enum: ['latest', 'next'] } }, required: ['mode'] }, then: { required: ['expectedBaseSha'] } }] +schemas.ChiefContinuationBinding = obj({ targetBotId: id, source: choice('chief-snapshot', 'continuation-ref', 'explicit'), expectedScope: nullable(hash), continuation: commands.get('task.run').request.properties.continuation }, ['targetBotId', 'source', 'expectedScope', 'continuation']) +queries.get('chief.state').response.properties.runs.items.properties.appliedDecisions = array(obj({ taskId: id, id, intent: hash, continuationBinding: ref('ChiefContinuationBinding') }, ['taskId', 'id', 'intent']), 30) +queries.get('chief.state').notes += ' Applied continuation decisions disclose the actually bound target/snapshot under appliedDecisions[].continuationBinding; it is audit data, not an authorization token. Chief start-next can reference its own ready snapshot instead of echoing machine fields. External task.run keeps its explicit continuation request contract.' schemas.DeliveryChoices = obj({ expectedScope: str(), observedAt: int(), admission: array(json), observation: nullable(json), recommendation: nullable(obj({ decisionId: id, action: str(), reason: str(), criteria: array(json), expectedScope: nullable(str()), alreadyApplied: lit(true) })), recommendationSource: str(), actions: array(obj({ action: str(), status: choice('ready', 'blocked', 'unknown'), reason: str() }, ['action', 'status', 'reason'])), note: str() }, ['expectedScope', 'observedAt', 'actions', 'note']) schemas.DeliveryChoices.properties.requirements = nullable(obj({ basis: lit('attempt.requirement'), attemptRevision: nullable(int()), taskRevision: nullable(int()), changedFields: array(choice('revision', 'title', 'criteria', 'impact', 'scope')), priorValues: map(str(16000)), source: obj({ query: lit('attempt.context'), attemptId: id, section: lit('requirement') }, ['query', 'attemptId', 'section']), notice: str() }, ['basis', 'attemptRevision', 'taskRevision', 'changedFields', 'priorValues', 'source', 'notice'])) schemas.TaskView.properties.deliveryChoices = nullable(ref('DeliveryChoices')) From ef2f5855c75394c881cede7a3a35c21b471197e8 Mon Sep 17 00:00:00 2001 From: zy95-12 <135135297+zy95-12@users.noreply.github.com> Date: Fri, 6 Nov 2026 10:01:10 +0800 Subject: [PATCH 2/3] Classify changed continuation facts through existing stale snapshot feedback --- core/src/host/project-chief-control.js | 6 +++--- core/test/project-chief-control.test.js | 1 + docs/openapi/flybot.openapi.json | 14 +++++++------- .../ds-bot/files/src/host/project-chief-control.js | 6 +++--- .../files/test/project-chief-control.test.js | 1 + 5 files changed, 15 insertions(+), 13 deletions(-) diff --git a/core/src/host/project-chief-control.js b/core/src/host/project-chief-control.js index d01e78dd..16c36b33 100644 --- a/core/src/host/project-chief-control.js +++ b/core/src/host/project-chief-control.js @@ -197,7 +197,7 @@ export function createProjectChiefControl(rt) { if (!facts || facts.status === 'unavailable' || !(next ? ['next'] : ['continue', 'latest']).includes(choice?.mode) || choice.previousAttemptId !== facts.attemptId || choice.generation !== facts.generation || !isDeepStrictEqual(choice.expectedWorkspace, facts.snapshot) || ['latest', 'next'].includes(choice.mode) && choice.expectedBaseSha !== facts.latestBaseSha) throw chiefProtocolError('continuation', 'the exact current continuationOptions snapshot, generation and selected latestBaseSha', decision.taskId) const response = await rt.projectDelivery.endpoint('project-attempt-continuation-state', { projectId, attemptId: facts.attemptId }, actor) const { nextActions: _nextActions, ...current } = response - if (!isDeepStrictEqual(current, facts)) throw chiefProtocolError('continuation', 'fresh facts after the repository baseline or workspace changed', decision.taskId) + if (!isDeepStrictEqual(current, facts)) throw chiefProtocolError('continuation', 'fresh facts after snapshot changed in the repository baseline or workspace', decision.taskId) const continuation = { mode: choice.mode, previousAttemptId: choice.previousAttemptId, generation: choice.generation, expectedWorkspace: structuredClone(choice.expectedWorkspace) } if (['latest', 'next'].includes(choice.mode)) continuation.expectedBaseSha = choice.expectedBaseSha return { target, continuation, source } @@ -214,8 +214,8 @@ export function createProjectChiefControl(rt) { if (digest(rt.state.projectAttempts?.[task.deliveryAttemptId]?.evidence?.merge ?? null) !== digest(saved.evidence?.merge ?? null)) throw chiefProtocolError('snapshot', 'fresh facts after merge evidence changed', task.id) const continuationBinding = await continuationChoice(decision, saved, rt.projectAgents.deliveryDefinitions(run.projectId).roles, actor, run.projectId) if (continuationBinding) await validateActor(actor) - if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw chiefProtocolError('targetBotId', 'fresh role catalog after continuation validation', task.id) - if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw chiefProtocolError('snapshot', 'fresh Task facts after continuation validation', task.id) + if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw chiefProtocolError('targetBotId', 'fresh role catalog after snapshot changed during continuation validation', task.id) + if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw chiefProtocolError('snapshot', 'fresh Task facts after scope changed during continuation validation', task.id) const reason = typeof decision.reason === 'string' ? decision.reason.trim() : '' if (!reason || reason.length > 4000 || !['close', 'route', 'return', 'request-tests', 'needs-human', 'continue', 'admit', 'revise', 'cancel', 'reopen', 'resolve-dependency', 'retry-notification', 'reconcile-exit', 'stop-execution', 'accept-attempt', 'start-next'].includes(decision.action)) throw refuse('Chief decision needs a supported action and bounded reason') if (decision.action === 'needs-human') { assistanceRequest(decision); if (decision.recipient === 'operations' && !operationsContact(rt, task.projectId, policyOf(task.projectId).operationsContact)) throw refuse('No operations contact is configured') } diff --git a/core/test/project-chief-control.test.js b/core/test/project-chief-control.test.js index a63d22b8..152b5626 100644 --- a/core/test/project-chief-control.test.js +++ b/core/test/project-chief-control.test.js @@ -941,6 +941,7 @@ for (const kind of ['ambiguous-target', 'no-target', 'ambiguous-mode', 'unavaila assert.notEqual(run.status, 'done') const expected = kind === 'roles-changed' ? /fresh role catalog/ : kind === 'rights-revoked' ? /Current role denied/ : kind === 'local-changed' ? /fresh Task facts/ : ['target-changed', 'workspace-changed'].includes(kind) ? /fresh facts after/ : kind === 'wrong-ref' ? /expectedScope/ : kind === 'bad-mode' ? /next only/ : kind.includes('target') ? /active implementation role/ : /continuation/ assert.match(run.failure, expected) + if (['target-changed', 'workspace-changed', 'local-changed', 'roles-changed'].includes(kind)) assert.equal(run.correction.feedback.at(-1).recovery.kind, 'stale-snapshot') }) test('explicit target resolves multiple implementation roles, and explicit scope reference resolves multiple modes', async () => { diff --git a/docs/openapi/flybot.openapi.json b/docs/openapi/flybot.openapi.json index a92822bc..21749e60 100644 --- a/docs/openapi/flybot.openapi.json +++ b/docs/openapi/flybot.openapi.json @@ -36134,7 +36134,7 @@ "notes": "undefined Applied continuation decisions disclose the actually bound target/snapshot under appliedDecisions[].continuationBinding; it is audit data, not an authorization token. Chief start-next can reference its own ready snapshot instead of echoing machine fields. External task.run keeps its explicit continuation request contract.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" + "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" } }, { @@ -39336,7 +39336,7 @@ "notes": "All current authorization, Project/Task bindings and domain fences are rechecked server-side.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" + "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" } }, { @@ -39441,7 +39441,7 @@ "notes": "operationsContact selects an explicit assistance recipient, never host privileges; null clears it, omission retains it. autonomy budgets/guidance inform Chief judgement, not a fixed escalation workflow. Use existing schedule.configure targeting project-chief for hourly compensation; chief.state.compensation reports original nextCheckAt. 30-minute inspection is a configurable schedule default, not a framework deadline. correctionBudget defaults to 2 (0–4); protocol correction reuses the Chief Session. Exhaustion is incomplete judgement, never automatic consultation or default routing. inspection.lifecycle configures queue/execution deadlines, bounded recovery delay/count and health overdue threshold; timeout is not exit proof.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" + "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" } }, { @@ -50321,7 +50321,7 @@ "notes": "All current authorization, Project/Task bindings and domain fences are rechecked server-side.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" + "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" }, "aliasOf": "development.command.chief.inspect" }, @@ -50424,7 +50424,7 @@ "notes": "operationsContact selects an explicit assistance recipient, never host privileges; null clears it, omission retains it. autonomy budgets/guidance inform Chief judgement, not a fixed escalation workflow. Use existing schedule.configure targeting project-chief for hourly compensation; chief.state.compensation reports original nextCheckAt. 30-minute inspection is a configurable schedule default, not a framework deadline. correctionBudget defaults to 2 (0–4); protocol correction reuses the Chief Session. Exhaustion is incomplete judgement, never automatic consultation or default routing. inspection.lifecycle configures queue/execution deadlines, bounded recovery delay/count and health overdue threshold; timeout is not exit proof.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" + "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" }, "aliasOf": "development.command.chief.configure" }, @@ -50500,7 +50500,7 @@ "notes": "undefined Applied continuation decisions disclose the actually bound target/snapshot under appliedDecisions[].continuationBinding; it is audit data, not an authorization token. Chief start-next can reference its own ready snapshot instead of echoing machine fields. External task.run keeps its explicit continuation request contract.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78" + "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" }, "aliasOf": "development.query.chief.state" }, @@ -51381,7 +51381,7 @@ "core/src/host/project-assistant.js": "4fa4309d4aa7abf78b6ba4e55ba754f320384f931e0e4a5c6620c859e1b50676", "core/src/host/project-board.js": "dfb0b4bc05ae42335632e3edbba883a03c5da2fb17899b23e8226320ea08aec4", "core/src/host/project-chief-assistance.js": "261f3767cf7fd4466f057c625b32e74b9c0a90c53c0fca3fcc835ac0f11c567b", - "core/src/host/project-chief-control.js": "ee797c7e85be6c48db958077b48d11511c5a52994e4c8a059abf69f515f8be78", + "core/src/host/project-chief-control.js": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67", "core/src/host/project-chief-lifecycle.js": "ed85bb1be90198910b20ffd605895812dfa848fce1049f7bd27b32dbfae3a328", "core/src/host/project-chief-protocol.js": "8e0ad56f025e08dac5318ce48c17c39754f001d5aa40b0beddbee0d66c9a179f", "core/src/host/project-chief-recovery.js": "c80a9d791aab06d4b343e8740a25a471ddf93e0af52b59dd9b7e6fc835c0d824", diff --git a/patches/ds-bot/files/src/host/project-chief-control.js b/patches/ds-bot/files/src/host/project-chief-control.js index d01e78dd..16c36b33 100644 --- a/patches/ds-bot/files/src/host/project-chief-control.js +++ b/patches/ds-bot/files/src/host/project-chief-control.js @@ -197,7 +197,7 @@ export function createProjectChiefControl(rt) { if (!facts || facts.status === 'unavailable' || !(next ? ['next'] : ['continue', 'latest']).includes(choice?.mode) || choice.previousAttemptId !== facts.attemptId || choice.generation !== facts.generation || !isDeepStrictEqual(choice.expectedWorkspace, facts.snapshot) || ['latest', 'next'].includes(choice.mode) && choice.expectedBaseSha !== facts.latestBaseSha) throw chiefProtocolError('continuation', 'the exact current continuationOptions snapshot, generation and selected latestBaseSha', decision.taskId) const response = await rt.projectDelivery.endpoint('project-attempt-continuation-state', { projectId, attemptId: facts.attemptId }, actor) const { nextActions: _nextActions, ...current } = response - if (!isDeepStrictEqual(current, facts)) throw chiefProtocolError('continuation', 'fresh facts after the repository baseline or workspace changed', decision.taskId) + if (!isDeepStrictEqual(current, facts)) throw chiefProtocolError('continuation', 'fresh facts after snapshot changed in the repository baseline or workspace', decision.taskId) const continuation = { mode: choice.mode, previousAttemptId: choice.previousAttemptId, generation: choice.generation, expectedWorkspace: structuredClone(choice.expectedWorkspace) } if (['latest', 'next'].includes(choice.mode)) continuation.expectedBaseSha = choice.expectedBaseSha return { target, continuation, source } @@ -214,8 +214,8 @@ export function createProjectChiefControl(rt) { if (digest(rt.state.projectAttempts?.[task.deliveryAttemptId]?.evidence?.merge ?? null) !== digest(saved.evidence?.merge ?? null)) throw chiefProtocolError('snapshot', 'fresh facts after merge evidence changed', task.id) const continuationBinding = await continuationChoice(decision, saved, rt.projectAgents.deliveryDefinitions(run.projectId).roles, actor, run.projectId) if (continuationBinding) await validateActor(actor) - if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw chiefProtocolError('targetBotId', 'fresh role catalog after continuation validation', task.id) - if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw chiefProtocolError('snapshot', 'fresh Task facts after continuation validation', task.id) + if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw chiefProtocolError('targetBotId', 'fresh role catalog after snapshot changed during continuation validation', task.id) + if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw chiefProtocolError('snapshot', 'fresh Task facts after scope changed during continuation validation', task.id) const reason = typeof decision.reason === 'string' ? decision.reason.trim() : '' if (!reason || reason.length > 4000 || !['close', 'route', 'return', 'request-tests', 'needs-human', 'continue', 'admit', 'revise', 'cancel', 'reopen', 'resolve-dependency', 'retry-notification', 'reconcile-exit', 'stop-execution', 'accept-attempt', 'start-next'].includes(decision.action)) throw refuse('Chief decision needs a supported action and bounded reason') if (decision.action === 'needs-human') { assistanceRequest(decision); if (decision.recipient === 'operations' && !operationsContact(rt, task.projectId, policyOf(task.projectId).operationsContact)) throw refuse('No operations contact is configured') } diff --git a/patches/ds-bot/files/test/project-chief-control.test.js b/patches/ds-bot/files/test/project-chief-control.test.js index a63d22b8..152b5626 100644 --- a/patches/ds-bot/files/test/project-chief-control.test.js +++ b/patches/ds-bot/files/test/project-chief-control.test.js @@ -941,6 +941,7 @@ for (const kind of ['ambiguous-target', 'no-target', 'ambiguous-mode', 'unavaila assert.notEqual(run.status, 'done') const expected = kind === 'roles-changed' ? /fresh role catalog/ : kind === 'rights-revoked' ? /Current role denied/ : kind === 'local-changed' ? /fresh Task facts/ : ['target-changed', 'workspace-changed'].includes(kind) ? /fresh facts after/ : kind === 'wrong-ref' ? /expectedScope/ : kind === 'bad-mode' ? /next only/ : kind.includes('target') ? /active implementation role/ : /continuation/ assert.match(run.failure, expected) + if (['target-changed', 'workspace-changed', 'local-changed', 'roles-changed'].includes(kind)) assert.equal(run.correction.feedback.at(-1).recovery.kind, 'stale-snapshot') }) test('explicit target resolves multiple implementation roles, and explicit scope reference resolves multiple modes', async () => { From 248bf92f3fd1011e22c8a9ff47889aab0dc09ba7 Mon Sep 17 00:00:00 2001 From: zy95-12 <135135297+zy95-12@users.noreply.github.com> Date: Fri, 6 Nov 2026 10:06:22 +0800 Subject: [PATCH 3/3] Reuse structured snapshot feedback for continuation proof changes --- core/src/host/project-chief-control.js | 11 ++++++----- core/test/project-chief-control.test.js | 2 +- docs/openapi/flybot.openapi.json | 14 +++++++------- docs/subsystems/project-chief-control.md | 2 ++ .../ds-bot/files/src/host/project-chief-control.js | 11 ++++++----- .../files/test/project-chief-control.test.js | 2 +- 6 files changed, 23 insertions(+), 19 deletions(-) diff --git a/core/src/host/project-chief-control.js b/core/src/host/project-chief-control.js index 16c36b33..85a60834 100644 --- a/core/src/host/project-chief-control.js +++ b/core/src/host/project-chief-control.js @@ -19,6 +19,7 @@ import { stageRoleSummary } from './project-context-memory.js' const template = JSON.parse(readFileSync(new URL('./project-chief-schedule.json', import.meta.url), 'utf8')) const refuse = message => Object.assign(new Error(message), { code: 'bot/chief' }) const digest = value => createHash('sha256').update(JSON.stringify(value)).digest('hex') +const staleSnapshot = (taskId, changedFields) => Object.assign(chiefProtocolError('snapshot', `fresh facts after ${changedFields.join(', ')} changed`, taskId), { recovery: { version: 'chief-recovery/v1', kind: 'stale-snapshot', changedFields, retryable: false, conditions: ['fresh scoped snapshot; never rebase old input receipts'], actions: ['Chief reassesses current facts'] } }) const inputDeclaration = (decision, saved) => ({ provided: decision.inputReceipts !== undefined, batchId: decision.inputReceipts?.batchId ?? null, read: [...(decision.inputReceipts?.read ?? [])], handled: [...(decision.inputReceipts?.handled ?? [])], remaining: saved.taskInbox?.remaining ?? 0 }) const terminal = new Set(['accepted', 'cancelled']) export function inspectionResources(state, task) { @@ -197,7 +198,7 @@ export function createProjectChiefControl(rt) { if (!facts || facts.status === 'unavailable' || !(next ? ['next'] : ['continue', 'latest']).includes(choice?.mode) || choice.previousAttemptId !== facts.attemptId || choice.generation !== facts.generation || !isDeepStrictEqual(choice.expectedWorkspace, facts.snapshot) || ['latest', 'next'].includes(choice.mode) && choice.expectedBaseSha !== facts.latestBaseSha) throw chiefProtocolError('continuation', 'the exact current continuationOptions snapshot, generation and selected latestBaseSha', decision.taskId) const response = await rt.projectDelivery.endpoint('project-attempt-continuation-state', { projectId, attemptId: facts.attemptId }, actor) const { nextActions: _nextActions, ...current } = response - if (!isDeepStrictEqual(current, facts)) throw chiefProtocolError('continuation', 'fresh facts after snapshot changed in the repository baseline or workspace', decision.taskId) + if (!isDeepStrictEqual(current, facts)) throw staleSnapshot(decision.taskId, ['repository baseline or workspace']) const continuation = { mode: choice.mode, previousAttemptId: choice.previousAttemptId, generation: choice.generation, expectedWorkspace: structuredClone(choice.expectedWorkspace) } if (['latest', 'next'].includes(choice.mode)) continuation.expectedBaseSha = choice.expectedBaseSha return { target, continuation, source } @@ -214,8 +215,8 @@ export function createProjectChiefControl(rt) { if (digest(rt.state.projectAttempts?.[task.deliveryAttemptId]?.evidence?.merge ?? null) !== digest(saved.evidence?.merge ?? null)) throw chiefProtocolError('snapshot', 'fresh facts after merge evidence changed', task.id) const continuationBinding = await continuationChoice(decision, saved, rt.projectAgents.deliveryDefinitions(run.projectId).roles, actor, run.projectId) if (continuationBinding) await validateActor(actor) - if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw chiefProtocolError('targetBotId', 'fresh role catalog after snapshot changed during continuation validation', task.id) - if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw chiefProtocolError('snapshot', 'fresh Task facts after scope changed during continuation validation', task.id) + if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw staleSnapshot(task.id, ['role catalog']) + if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw staleSnapshot(task.id, ['Task scope']) const reason = typeof decision.reason === 'string' ? decision.reason.trim() : '' if (!reason || reason.length > 4000 || !['close', 'route', 'return', 'request-tests', 'needs-human', 'continue', 'admit', 'revise', 'cancel', 'reopen', 'resolve-dependency', 'retry-notification', 'reconcile-exit', 'stop-execution', 'accept-attempt', 'start-next'].includes(decision.action)) throw refuse('Chief decision needs a supported action and bounded reason') if (decision.action === 'needs-human') { assistanceRequest(decision); if (decision.recipient === 'operations' && !operationsContact(rt, task.projectId, policyOf(task.projectId).operationsContact)) throw refuse('No operations contact is configured') } @@ -348,7 +349,7 @@ export function createProjectChiefControl(rt) { if (digest(rt.state.projectAttempts?.[current.deliveryAttemptId]?.evidence?.merge ?? null) !== digest(saved.evidence?.merge ?? null)) changed.push('merge evidence') if (current.title !== saved.title || (current.criteria ?? '') !== saved.criteria || current.projectTask && current.projectTask.requirements.revision !== saved.taskContract?.requirements.revision) changed.push('requirements') } - if (changed.length) throw Object.assign(chiefProtocolError('snapshot', `fresh facts after ${changed.join(', ')} changed`, decision.taskId), { recovery: { version: 'chief-recovery/v1', kind: 'stale-snapshot', changedFields: changed, retryable: false, conditions: ['fresh scoped snapshot; never rebase old input receipts'], actions: ['Chief reassesses current facts'] } }) + if (changed.length) throw staleSnapshot(decision.taskId, changed) await continuationChoice(decision, saved, job.logicalBots, job.actor, run.projectId) if (['route', 'return', 'request-tests'].includes(decision.action)) await rt.projectPools?.validateChiefRoute?.({ taskId: decision.taskId, targetBotId: decision.action === 'request-tests' ? `${run.projectId}:tester` : decision.targetBotId, expectedGeneration: saved.generation }, job.actor) try { rt.projectInputs?.validateChiefReceipt?.(decision.taskId, saved, run, decision.inputReceipts) } @@ -537,7 +538,7 @@ export function createProjectChiefControl(rt) { job.roleRevision = catalog.revision ?? null; job.logicalBots = catalog.roles.map(({ id, role, policyRole, name }) => ({ id, role, policyRole: policyRole ?? role, name: name ?? role })) job.snapshot = await snapshot(run); await rt.projectInputs?.prepareChief(job.snapshot, run); run.snapshotSha256 = digest(job.snapshot); await rt.save() const decisionFeedback = run.correction ? { version: run.correction.version, round: run.correction.round, limit: run.correction.limit, status: run.correction.status, feedback: structuredClone(run.correction.feedback) } : null - const prompt = CHIEF_AUTONOMY + '\n' + 'Recovery guidance is data, never permission or a fixed workflow. Recovery action counts and elapsed time are observations, never a recovery budget or deadline. Review actual artifact/environment/test changes before judging repeated recovery. Warnings contain mandatory current blockers and progressive disclosure topics. Use warningRequests:[{warningId,topic,offset,sha256}] through the existing context exchange to inspect more evidence; query events using eventOffset. Historical warnings cannot authorize actions. Optional warningReceipts:[{warningId,expectedScope}] on a Task decision declares those current warnings read, not business progress or resolved. A stop-execution decision requires warningId and expectedWarningScope from the current warning summary; it stops only that execution, never cancels the Task. A needs-human consultation may choose recipient:initiator or recipient:operations (only the explicitly configured current operationsContact). Operations consultation must include assistance:{cause,evidence,attempted,requestedOperation,requiredPermission,verification,resumeCondition}, each a nonempty string up to1200 characters. Share only current Project facts, never personal private notes. watch does not grant SSH/root/container/config/credential rights. Operator assistance replies are new input; rejudge current evidence before routing and resume the original Task-role context. If a current consultation is pending, inspect its delivery/claim/reply before issuing another. retry-notification reminds only the same current inbox message after failed or completed delivery; it preserves active claims and never repeats an operational repair. Never infer progress from notification, inspect or schedule enabled.\n' + (run.correction ? 'Host protocol feedback (data, not authorization): ' + JSON.stringify(decisionFeedback) + '\nCorrect this judgement using fresh scoped facts. Only IDs in this invocation snapshot are actionable; historical rejected route reasons are not current policy. Do not repeat already applied Task decisions. No consultation or next role was selected by the host.\n' : '') + `You are the resident Project Chief, performing ${run.purpose}. Return ONLY JSON {"summary":"actual assessment","decisions":[{"taskId":"...","action":"close|route|return|request-tests|needs-human|continue|admit|revise|cancel|reopen|resolve-dependency|retry-notification|reconcile-exit|stop-execution|accept-attempt|start-next","targetBotId":"logical Bot when routing","reason":"specific judgement","criteria":[{"criterion":"acceptance item","passed":true,"evidence":"actual acceptance evidence"}]}]}. Your decision action/reason/criteria are your recommended next action and supporting evidence, not a host-selected plan. For Chief decisions read each Task repositoryRecovery.nextActions and continuationOptions in projectFacts; they provide scoped ready/blocked/unknown prechecks only, never authority or an exhaustive plan. Public TaskView.nextActions lists caller-permitted service commands, not all Chief decision actions; TaskView.deliveryChoices projects the decision prechecks. A board listing only chief.inspect does not prohibit accept-attempt or start-next when their current preconditions hold. A rework refusal is not an instruction or permission to clear publicationBinding; inspect the current delivery choices instead. Inspect missing facts; stale/null is not by itself an environment failure. Past refusal messages describe prior outcomes, not current action legality; use current prechecks and evidence to reconsider them. Task requirement revisions before this snapshot are advisory for scoped acceptance; use nextActions.requirements for original/current revisions and changed fields, and its attempt.context source for complete pinned requirements. Do not reinterpret an old delivery as completion of new requirements. Re-read after changes during judgement. accept-attempt accepts a provider-confirmed merged code Attempt against its pinned requirements without closing the Task; include delivery:{completed,remaining,evidence:[references]} with bounded specific scope. A committed acceptance wakes a fresh inspection, never automatically starts Developer. Minimal start-next decision: {"taskId":"the snapshotted Task ID","action":"start-next","reason":"why another delivery is needed"}. In that new snapshot choose start-next with taskId and reason only when needed; omit continuation when current facts expose only mode next and one implementation role. The host binds the continuationOptions you saw and revalidates current facts; it never silently switches baseline. If several implementation roles exist, provide targetBotId. Optional continuationRef is the expectedScope from those facts. Multiple modes need an explicit reference or full continuation selection. Supplied continuation fields must match exactly and are never repaired. Use existing attempt.continuation-state after feedback to inspect refreshed facts. This pins verified target HEAD containing the merged result and creates a new PR identity while retaining Task-role context/history. Never rework or inherit the merged PR binding. close remains available only when all Task criteria/dependencies/inputs are satisfied. Personal Agent assistance is the last resort when actual information/authority or a human decision is missing; failure counts/time never automatically escalate. A conversation/turn completion is not Task completion. Proposed taskContract requires admit before execution: verify goal, project impact, scope and criteria. Prefer admit without repeating already adequate requirements. For admit/revise, every provided requirements.title, requirements.criteria, requirements.impact and requirements.scope must be a bounded nonempty string. In particular requirements.criteria is a string; the separate outer decision.criteria is an evidence array and must never be copied into requirements.criteria. Optional requirements.repository is an associated owner/name or null for a repository-free decision. Uncertain intent requires needs-human. Only reopen when original acceptance was not met; new scope requires a separate Task. Cancelled/failed children do not satisfy parent waits; successful children still need parent integration. All Task mutations remain scoped Chief decisions. To resolve a failed/cancelled child condition without faking child completion, use resolve-dependency with dependency:{taskId,version:the exact current dependency snapshot version,evidence:[actual references],replacementTaskId:optional accepted alternative,replacementVersion:its task outcome version}. This is explicit Chief judgement, not host test success or parent completion. Leave unresolved conditions blocked when evidence is insufficient. Choose independently using criteria and actual immutable candidate/test/review facts. reconcile-exit explicitly retires only an exact exit-proven old claim without selecting a next role; consult repositoryRecovery.executionCheckpoint. Historical refusals are audit data, not current routing policy. Candidate Reviewer approve triggers the host merge flow only when explicitly configured for this Project/repository/branch. It is not a signed repository Reviewer PASS, actual forge approval, merged result or release permit. A configured merge must be provider-confirmed before overall code Task acceptance; pending/blocked merge facts are not completion. Do not close missing or stale evidence. Host allocationProof.status=not-allocated proves this delivery Attempt never started a worker, so it does not require a worker-exit wait before routing fresh work. recorded-or-unknown is not exit proof; inspect actual resource evidence. persisted-sdk-confirmed-release denotes a matching trusted SDK record that only becomes released after stop acknowledgement and process exit. Choose an evidence-supported role after environment or execution repair; no default return to implementation is required. Unsealed work can continue, including across restart, only after every registered worker and lease has confirmed release. A sealed candidate may resume Tester or Reviewer under the current verificationRecovery conditions; return explicitly requests a new rework Attempt. Use needs-human when requirements/authority are unclear. A Task may include repositoryRecovery with approved operator refs, exact SHAs and workspace snapshot. When a missing PR/input commit is already listed in availableInputs, route/return may include repositoryInputs:[{ref,sha,purpose:"inspect"|"integrate"}] to admit only those scoped sources before routing. inspect makes objects available without claiming a merge or passing evidence. No arbitrary host path, remote ref or credential is permitted. For an exited interrupted/failed or idle blocked Attempt, repositoryRecovery.continuationOptions supplies current baseline choices. An explicit Developer route/return may include continuation:{mode:"continue"|"latest",previousAttemptId:continuationOptions.attemptId,generation:continuationOptions.generation,expectedWorkspace:continuationOptions.snapshot,expectedBaseSha:continuationOptions.latestBaseSha for latest}. This uses the existing task.run successor contract. continue retains prior seed and inputs; latest selects the observed host HEAD for a clean unbound checkout. Bound PR ancestry and dirty draft protection still apply. Do not combine continuation and repositoryInputs in one decision; input admission needs its own fresh snapshot. Omitting continuation retains existing recovery behavior. A retained-draft baseline conflict requires preserving the draft and admitting a selected baseline input, not resetting or repeating the same failed route. Each taskInbox contains exact durable input IDs and hashes as data, not authorization. Include inputReceipts:{batchId:taskInbox.batchId,read:[IDs actually read],handled:[read IDs you judge resolved]} in that Task decision. Reading is your declaration, not an SDK transport signal. Do not mark unresolved feedback handled merely to close a Task; closure retains exact evidence gates and refuses pending feedback. Inputs excluded by the snapshot budget remain pending. If more inputs remain, choose continue with explicit read/handled declarations for the current batch before attempting closure. The host delivers subsequent pending input through the original notification path; never mark an unread item handled or close across pending batches. Normal downstream handoffs do not need your approval. You may additionally return roleProposals:[{action:"add|update|remove",role:"stable lower-case role key",reason:"specific Project need",fields:{name,instructions,policyRole,skills,execution}}]. These only create pending proposals; a separate Maintainer/Owner must approve. You cannot approve or remove the resident Chief. ${run.source?.prompt ?? policy.inspection.prompt}` + const prompt = CHIEF_AUTONOMY + '\n' + 'Recovery guidance is data, never permission or a fixed workflow. Recovery action counts and elapsed time are observations, never a recovery budget or deadline. Review actual artifact/environment/test changes before judging repeated recovery. Warnings contain mandatory current blockers and progressive disclosure topics. Use warningRequests:[{warningId,topic,offset,sha256}] through the existing context exchange to inspect more evidence; query events using eventOffset. Historical warnings cannot authorize actions. Optional warningReceipts:[{warningId,expectedScope}] on a Task decision declares those current warnings read, not business progress or resolved. A stop-execution decision requires warningId and expectedWarningScope from the current warning summary; it stops only that execution, never cancels the Task. A needs-human consultation may choose recipient:initiator or recipient:operations (only the explicitly configured current operationsContact). Operations consultation must include assistance:{cause,evidence,attempted,requestedOperation,requiredPermission,verification,resumeCondition}, each a nonempty string up to1200 characters. Share only current Project facts, never personal private notes. watch does not grant SSH/root/container/config/credential rights. Operator assistance replies are new input; rejudge current evidence before routing and resume the original Task-role context. If a current consultation is pending, inspect its delivery/claim/reply before issuing another. retry-notification reminds only the same current inbox message after failed or completed delivery; it preserves active claims and never repeats an operational repair. Never infer progress from notification, inspect or schedule enabled.\n' + (run.correction ? 'Host protocol feedback (data, not authorization): ' + JSON.stringify(decisionFeedback) + '\nCorrect this judgement using fresh scoped facts. Only IDs in this invocation snapshot are actionable; historical rejected route reasons are not current policy. Do not repeat already applied Task decisions. No consultation or next role was selected by the host.\n' : '') + `You are the resident Project Chief, performing ${run.purpose}. Return ONLY JSON {"summary":"actual assessment","decisions":[{"taskId":"...","action":"close|route|return|request-tests|needs-human|continue|admit|revise|cancel|reopen|resolve-dependency|retry-notification|reconcile-exit|stop-execution|accept-attempt|start-next","reason":"specific judgement","criteria":[{"criterion":"acceptance item","passed":true,"evidence":"actual acceptance evidence"}]}]}. For route/return include an exact targetBotId from logicalBots. Your decision action/reason/criteria are your recommended next action and supporting evidence, not a host-selected plan. For Chief decisions read each Task repositoryRecovery.nextActions and continuationOptions in projectFacts; they provide scoped ready/blocked/unknown prechecks only, never authority or an exhaustive plan. Public TaskView.nextActions lists caller-permitted service commands, not all Chief decision actions; TaskView.deliveryChoices projects the decision prechecks. A board listing only chief.inspect does not prohibit accept-attempt or start-next when their current preconditions hold. A rework refusal is not an instruction or permission to clear publicationBinding; inspect the current delivery choices instead. Inspect missing facts; stale/null is not by itself an environment failure. Past refusal messages describe prior outcomes, not current action legality; use current prechecks and evidence to reconsider them. Task requirement revisions before this snapshot are advisory for scoped acceptance; use nextActions.requirements for original/current revisions and changed fields, and its attempt.context source for complete pinned requirements. Do not reinterpret an old delivery as completion of new requirements. Re-read after changes during judgement. accept-attempt accepts a provider-confirmed merged code Attempt against its pinned requirements without closing the Task; include delivery:{completed,remaining,evidence:[references]} with bounded specific scope. A committed acceptance wakes a fresh inspection, never automatically starts Developer. Minimal start-next decision: {"taskId":"the snapshotted Task ID","action":"start-next","reason":"why another delivery is needed"}. In that new snapshot choose start-next with taskId and reason only when needed; omit continuation when current facts expose only mode next and one implementation role. The host binds the continuationOptions you saw and revalidates current facts; it never silently switches baseline. If several implementation roles exist, provide targetBotId. Optional continuationRef is the expectedScope from those facts. Multiple modes need an explicit reference or full continuation selection. Supplied continuation fields must match exactly and are never repaired. Use existing attempt.continuation-state after feedback to inspect refreshed facts. This pins verified target HEAD containing the merged result and creates a new PR identity while retaining Task-role context/history. Never rework or inherit the merged PR binding. close remains available only when all Task criteria/dependencies/inputs are satisfied. Personal Agent assistance is the last resort when actual information/authority or a human decision is missing; failure counts/time never automatically escalate. A conversation/turn completion is not Task completion. Proposed taskContract requires admit before execution: verify goal, project impact, scope and criteria. Prefer admit without repeating already adequate requirements. For admit/revise, every provided requirements.title, requirements.criteria, requirements.impact and requirements.scope must be a bounded nonempty string. In particular requirements.criteria is a string; the separate outer decision.criteria is an evidence array and must never be copied into requirements.criteria. Optional requirements.repository is an associated owner/name or null for a repository-free decision. Uncertain intent requires needs-human. Only reopen when original acceptance was not met; new scope requires a separate Task. Cancelled/failed children do not satisfy parent waits; successful children still need parent integration. All Task mutations remain scoped Chief decisions. To resolve a failed/cancelled child condition without faking child completion, use resolve-dependency with dependency:{taskId,version:the exact current dependency snapshot version,evidence:[actual references],replacementTaskId:optional accepted alternative,replacementVersion:its task outcome version}. This is explicit Chief judgement, not host test success or parent completion. Leave unresolved conditions blocked when evidence is insufficient. Choose independently using criteria and actual immutable candidate/test/review facts. reconcile-exit explicitly retires only an exact exit-proven old claim without selecting a next role; consult repositoryRecovery.executionCheckpoint. Historical refusals are audit data, not current routing policy. Candidate Reviewer approve triggers the host merge flow only when explicitly configured for this Project/repository/branch. It is not a signed repository Reviewer PASS, actual forge approval, merged result or release permit. A configured merge must be provider-confirmed before overall code Task acceptance; pending/blocked merge facts are not completion. Do not close missing or stale evidence. Host allocationProof.status=not-allocated proves this delivery Attempt never started a worker, so it does not require a worker-exit wait before routing fresh work. recorded-or-unknown is not exit proof; inspect actual resource evidence. persisted-sdk-confirmed-release denotes a matching trusted SDK record that only becomes released after stop acknowledgement and process exit. Choose an evidence-supported role after environment or execution repair; no default return to implementation is required. Unsealed work can continue, including across restart, only after every registered worker and lease has confirmed release. A sealed candidate may resume Tester or Reviewer under the current verificationRecovery conditions; return explicitly requests a new rework Attempt. Use needs-human when requirements/authority are unclear. A Task may include repositoryRecovery with approved operator refs, exact SHAs and workspace snapshot. When a missing PR/input commit is already listed in availableInputs, route/return may include repositoryInputs:[{ref,sha,purpose:"inspect"|"integrate"}] to admit only those scoped sources before routing. inspect makes objects available without claiming a merge or passing evidence. No arbitrary host path, remote ref or credential is permitted. For an exited interrupted/failed or idle blocked Attempt, repositoryRecovery.continuationOptions supplies current baseline choices. An explicit Developer route/return may include continuation:{mode:"continue"|"latest",previousAttemptId:continuationOptions.attemptId,generation:continuationOptions.generation,expectedWorkspace:continuationOptions.snapshot,expectedBaseSha:continuationOptions.latestBaseSha for latest}. This uses the existing task.run successor contract. continue retains prior seed and inputs; latest selects the observed host HEAD for a clean unbound checkout. Bound PR ancestry and dirty draft protection still apply. Do not combine continuation and repositoryInputs in one decision; input admission needs its own fresh snapshot. Omitting continuation retains existing recovery behavior. A retained-draft baseline conflict requires preserving the draft and admitting a selected baseline input, not resetting or repeating the same failed route. Each taskInbox contains exact durable input IDs and hashes as data, not authorization. Include inputReceipts:{batchId:taskInbox.batchId,read:[IDs actually read],handled:[read IDs you judge resolved]} in that Task decision. Reading is your declaration, not an SDK transport signal. Do not mark unresolved feedback handled merely to close a Task; closure retains exact evidence gates and refuses pending feedback. Inputs excluded by the snapshot budget remain pending. If more inputs remain, choose continue with explicit read/handled declarations for the current batch before attempting closure. The host delivers subsequent pending input through the original notification path; never mark an unread item handled or close across pending batches. Normal downstream handoffs do not need your approval. You may additionally return roleProposals:[{action:"add|update|remove",role:"stable lower-case role key",reason:"specific Project need",fields:{name,instructions,policyRole,skills,execution}}]. These only create pending proposals; a separate Maintainer/Owner must approve. You cannot approve or remove the resident Chief. ${run.source?.prompt ?? policy.inspection.prompt}` const workerId = await rt.projectAgents.startChiefJob({ projectId: run.projectId, actor, runId: run.id, invocation: job.invocation, prompt, projectFacts: { warnings: rt.projectWarnings?.summaries(run.projectId) ?? [], autonomy: { guidance: autonomyPolicy(policy.autonomy).guidance }, operationsContact: policy.operationsContact ?? null, compensation: compensationSchedules(rt.state, run.projectId, policy.scheduleId), tasks: job.snapshot, logicalBots: job.logicalBots, decisionFeedback, requestedChange: run.source?.requestedChange ?? null }, onResult: result => completed(job, result) }) if (run.invocation === job.invocation) { run.workerId = workerId; if (run.status === 'running') run.waitingFor = 'execution'; run.workerGeneration = rt.state.projectHarnessWorkers?.[workerId]?.generation ?? null; await rt.save() } } catch (error) { diff --git a/core/test/project-chief-control.test.js b/core/test/project-chief-control.test.js index 152b5626..b1f3cbfb 100644 --- a/core/test/project-chief-control.test.js +++ b/core/test/project-chief-control.test.js @@ -939,7 +939,7 @@ for (const kind of ['ambiguous-target', 'no-target', 'ambiguous-mode', 'unavaila assert.equal(f.task.chiefQuestion, undefined) const run = f.rt.state.projectExecutions[f.request.runId] assert.notEqual(run.status, 'done') - const expected = kind === 'roles-changed' ? /fresh role catalog/ : kind === 'rights-revoked' ? /Current role denied/ : kind === 'local-changed' ? /fresh Task facts/ : ['target-changed', 'workspace-changed'].includes(kind) ? /fresh facts after/ : kind === 'wrong-ref' ? /expectedScope/ : kind === 'bad-mode' ? /next only/ : kind.includes('target') ? /active implementation role/ : /continuation/ + const expected = kind === 'roles-changed' ? /role catalog changed/ : kind === 'rights-revoked' ? /Current role denied/ : kind === 'local-changed' ? /Task scope changed/ : ['target-changed', 'workspace-changed'].includes(kind) ? /fresh facts after/ : kind === 'wrong-ref' ? /expectedScope/ : kind === 'bad-mode' ? /next only/ : kind.includes('target') ? /active implementation role/ : /continuation/ assert.match(run.failure, expected) if (['target-changed', 'workspace-changed', 'local-changed', 'roles-changed'].includes(kind)) assert.equal(run.correction.feedback.at(-1).recovery.kind, 'stale-snapshot') }) diff --git a/docs/openapi/flybot.openapi.json b/docs/openapi/flybot.openapi.json index 21749e60..edf8617c 100644 --- a/docs/openapi/flybot.openapi.json +++ b/docs/openapi/flybot.openapi.json @@ -36134,7 +36134,7 @@ "notes": "undefined Applied continuation decisions disclose the actually bound target/snapshot under appliedDecisions[].continuationBinding; it is audit data, not an authorization token. Chief start-next can reference its own ready snapshot instead of echoing machine fields. External task.run keeps its explicit continuation request contract.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" + "sha256": "418b841a8597ed5b3c511a180aad75d10df0682113507b88b19fb1d4749ae2b3" } }, { @@ -39336,7 +39336,7 @@ "notes": "All current authorization, Project/Task bindings and domain fences are rechecked server-side.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" + "sha256": "418b841a8597ed5b3c511a180aad75d10df0682113507b88b19fb1d4749ae2b3" } }, { @@ -39441,7 +39441,7 @@ "notes": "operationsContact selects an explicit assistance recipient, never host privileges; null clears it, omission retains it. autonomy budgets/guidance inform Chief judgement, not a fixed escalation workflow. Use existing schedule.configure targeting project-chief for hourly compensation; chief.state.compensation reports original nextCheckAt. 30-minute inspection is a configurable schedule default, not a framework deadline. correctionBudget defaults to 2 (0–4); protocol correction reuses the Chief Session. Exhaustion is incomplete judgement, never automatic consultation or default routing. inspection.lifecycle configures queue/execution deadlines, bounded recovery delay/count and health overdue threshold; timeout is not exit proof.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" + "sha256": "418b841a8597ed5b3c511a180aad75d10df0682113507b88b19fb1d4749ae2b3" } }, { @@ -50321,7 +50321,7 @@ "notes": "All current authorization, Project/Task bindings and domain fences are rechecked server-side.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" + "sha256": "418b841a8597ed5b3c511a180aad75d10df0682113507b88b19fb1d4749ae2b3" }, "aliasOf": "development.command.chief.inspect" }, @@ -50424,7 +50424,7 @@ "notes": "operationsContact selects an explicit assistance recipient, never host privileges; null clears it, omission retains it. autonomy budgets/guidance inform Chief judgement, not a fixed escalation workflow. Use existing schedule.configure targeting project-chief for hourly compensation; chief.state.compensation reports original nextCheckAt. 30-minute inspection is a configurable schedule default, not a framework deadline. correctionBudget defaults to 2 (0–4); protocol correction reuses the Chief Session. Exhaustion is incomplete judgement, never automatic consultation or default routing. inspection.lifecycle configures queue/execution deadlines, bounded recovery delay/count and health overdue threshold; timeout is not exit proof.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" + "sha256": "418b841a8597ed5b3c511a180aad75d10df0682113507b88b19fb1d4749ae2b3" }, "aliasOf": "development.command.chief.configure" }, @@ -50500,7 +50500,7 @@ "notes": "undefined Applied continuation decisions disclose the actually bound target/snapshot under appliedDecisions[].continuationBinding; it is audit data, not an authorization token. Chief start-next can reference its own ready snapshot instead of echoing machine fields. External task.run keeps its explicit continuation request contract.", "implementation": { "path": "core/src/host/project-chief-control.js", - "sha256": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67" + "sha256": "418b841a8597ed5b3c511a180aad75d10df0682113507b88b19fb1d4749ae2b3" }, "aliasOf": "development.query.chief.state" }, @@ -51381,7 +51381,7 @@ "core/src/host/project-assistant.js": "4fa4309d4aa7abf78b6ba4e55ba754f320384f931e0e4a5c6620c859e1b50676", "core/src/host/project-board.js": "dfb0b4bc05ae42335632e3edbba883a03c5da2fb17899b23e8226320ea08aec4", "core/src/host/project-chief-assistance.js": "261f3767cf7fd4466f057c625b32e74b9c0a90c53c0fca3fcc835ac0f11c567b", - "core/src/host/project-chief-control.js": "3fa23cb8e165157d8f03311d96fa19f3a47da9139dcd6eb00b78ebd20fc27f67", + "core/src/host/project-chief-control.js": "418b841a8597ed5b3c511a180aad75d10df0682113507b88b19fb1d4749ae2b3", "core/src/host/project-chief-lifecycle.js": "ed85bb1be90198910b20ffd605895812dfa848fce1049f7bd27b32dbfae3a328", "core/src/host/project-chief-protocol.js": "8e0ad56f025e08dac5318ce48c17c39754f001d5aa40b0beddbee0d66c9a179f", "core/src/host/project-chief-recovery.js": "c80a9d791aab06d4b343e8740a25a471ddf93e0af52b59dd9b7e6fc835c0d824", diff --git a/docs/subsystems/project-chief-control.md b/docs/subsystems/project-chief-control.md index f337d3c2..edcc2fd6 100644 --- a/docs/subsystems/project-chief-control.md +++ b/docs/subsystems/project-chief-control.md @@ -421,3 +421,5 @@ A Chief decision {taskId,action:'start-next',reason} references that invocation' The host derives machine fields from the saved immutable snapshot, reads the original continuation-state endpoint, compares it to those saved facts, then revalidates current actor, role catalog and Task scope. It never binds to a newer target/workspace implicitly. Original settlement, confirmed exit, target ancestry, quota and publication identity checks remain in the existing scheduler/delivery boundary. No new planner, draft store, timer, tool executor or generic field repair is added. Existing read queries support inspection after feedback. Audit.continuationBinding records actual target, source, expectedScope and continuation. The raw model decision is not mutated; existing intent hashes/replay keys remain stable. chief.state decorates appliedDecisions with matching project/run/task audit bindings, read-only and detached from Store. No independent authority is conveyed. Native and SDK Chief use the same parser/apply path and actual job prompt. No live deployment or new paid model validation follows from these source contracts. + +Continuation proof changes and post-read Task/catalog changes reuse the same structured stale-snapshot feedback helper as ordinary preflight. No dispatch/inbox recovery policy is changed. diff --git a/patches/ds-bot/files/src/host/project-chief-control.js b/patches/ds-bot/files/src/host/project-chief-control.js index 16c36b33..85a60834 100644 --- a/patches/ds-bot/files/src/host/project-chief-control.js +++ b/patches/ds-bot/files/src/host/project-chief-control.js @@ -19,6 +19,7 @@ import { stageRoleSummary } from './project-context-memory.js' const template = JSON.parse(readFileSync(new URL('./project-chief-schedule.json', import.meta.url), 'utf8')) const refuse = message => Object.assign(new Error(message), { code: 'bot/chief' }) const digest = value => createHash('sha256').update(JSON.stringify(value)).digest('hex') +const staleSnapshot = (taskId, changedFields) => Object.assign(chiefProtocolError('snapshot', `fresh facts after ${changedFields.join(', ')} changed`, taskId), { recovery: { version: 'chief-recovery/v1', kind: 'stale-snapshot', changedFields, retryable: false, conditions: ['fresh scoped snapshot; never rebase old input receipts'], actions: ['Chief reassesses current facts'] } }) const inputDeclaration = (decision, saved) => ({ provided: decision.inputReceipts !== undefined, batchId: decision.inputReceipts?.batchId ?? null, read: [...(decision.inputReceipts?.read ?? [])], handled: [...(decision.inputReceipts?.handled ?? [])], remaining: saved.taskInbox?.remaining ?? 0 }) const terminal = new Set(['accepted', 'cancelled']) export function inspectionResources(state, task) { @@ -197,7 +198,7 @@ export function createProjectChiefControl(rt) { if (!facts || facts.status === 'unavailable' || !(next ? ['next'] : ['continue', 'latest']).includes(choice?.mode) || choice.previousAttemptId !== facts.attemptId || choice.generation !== facts.generation || !isDeepStrictEqual(choice.expectedWorkspace, facts.snapshot) || ['latest', 'next'].includes(choice.mode) && choice.expectedBaseSha !== facts.latestBaseSha) throw chiefProtocolError('continuation', 'the exact current continuationOptions snapshot, generation and selected latestBaseSha', decision.taskId) const response = await rt.projectDelivery.endpoint('project-attempt-continuation-state', { projectId, attemptId: facts.attemptId }, actor) const { nextActions: _nextActions, ...current } = response - if (!isDeepStrictEqual(current, facts)) throw chiefProtocolError('continuation', 'fresh facts after snapshot changed in the repository baseline or workspace', decision.taskId) + if (!isDeepStrictEqual(current, facts)) throw staleSnapshot(decision.taskId, ['repository baseline or workspace']) const continuation = { mode: choice.mode, previousAttemptId: choice.previousAttemptId, generation: choice.generation, expectedWorkspace: structuredClone(choice.expectedWorkspace) } if (['latest', 'next'].includes(choice.mode)) continuation.expectedBaseSha = choice.expectedBaseSha return { target, continuation, source } @@ -214,8 +215,8 @@ export function createProjectChiefControl(rt) { if (digest(rt.state.projectAttempts?.[task.deliveryAttemptId]?.evidence?.merge ?? null) !== digest(saved.evidence?.merge ?? null)) throw chiefProtocolError('snapshot', 'fresh facts after merge evidence changed', task.id) const continuationBinding = await continuationChoice(decision, saved, rt.projectAgents.deliveryDefinitions(run.projectId).roles, actor, run.projectId) if (continuationBinding) await validateActor(actor) - if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw chiefProtocolError('targetBotId', 'fresh role catalog after snapshot changed during continuation validation', task.id) - if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw chiefProtocolError('snapshot', 'fresh Task facts after scope changed during continuation validation', task.id) + if (continuationBinding && (rt.projectAgents.deliveryDefinitions(run.projectId).revision ?? null) !== (job.roleRevision ?? null)) throw staleSnapshot(task.id, ['role catalog']) + if (continuationBinding && saved.deliverySnapshot && saved.deliverySnapshot.expectedScope !== deliveryScope(rt.state, task)) throw staleSnapshot(task.id, ['Task scope']) const reason = typeof decision.reason === 'string' ? decision.reason.trim() : '' if (!reason || reason.length > 4000 || !['close', 'route', 'return', 'request-tests', 'needs-human', 'continue', 'admit', 'revise', 'cancel', 'reopen', 'resolve-dependency', 'retry-notification', 'reconcile-exit', 'stop-execution', 'accept-attempt', 'start-next'].includes(decision.action)) throw refuse('Chief decision needs a supported action and bounded reason') if (decision.action === 'needs-human') { assistanceRequest(decision); if (decision.recipient === 'operations' && !operationsContact(rt, task.projectId, policyOf(task.projectId).operationsContact)) throw refuse('No operations contact is configured') } @@ -348,7 +349,7 @@ export function createProjectChiefControl(rt) { if (digest(rt.state.projectAttempts?.[current.deliveryAttemptId]?.evidence?.merge ?? null) !== digest(saved.evidence?.merge ?? null)) changed.push('merge evidence') if (current.title !== saved.title || (current.criteria ?? '') !== saved.criteria || current.projectTask && current.projectTask.requirements.revision !== saved.taskContract?.requirements.revision) changed.push('requirements') } - if (changed.length) throw Object.assign(chiefProtocolError('snapshot', `fresh facts after ${changed.join(', ')} changed`, decision.taskId), { recovery: { version: 'chief-recovery/v1', kind: 'stale-snapshot', changedFields: changed, retryable: false, conditions: ['fresh scoped snapshot; never rebase old input receipts'], actions: ['Chief reassesses current facts'] } }) + if (changed.length) throw staleSnapshot(decision.taskId, changed) await continuationChoice(decision, saved, job.logicalBots, job.actor, run.projectId) if (['route', 'return', 'request-tests'].includes(decision.action)) await rt.projectPools?.validateChiefRoute?.({ taskId: decision.taskId, targetBotId: decision.action === 'request-tests' ? `${run.projectId}:tester` : decision.targetBotId, expectedGeneration: saved.generation }, job.actor) try { rt.projectInputs?.validateChiefReceipt?.(decision.taskId, saved, run, decision.inputReceipts) } @@ -537,7 +538,7 @@ export function createProjectChiefControl(rt) { job.roleRevision = catalog.revision ?? null; job.logicalBots = catalog.roles.map(({ id, role, policyRole, name }) => ({ id, role, policyRole: policyRole ?? role, name: name ?? role })) job.snapshot = await snapshot(run); await rt.projectInputs?.prepareChief(job.snapshot, run); run.snapshotSha256 = digest(job.snapshot); await rt.save() const decisionFeedback = run.correction ? { version: run.correction.version, round: run.correction.round, limit: run.correction.limit, status: run.correction.status, feedback: structuredClone(run.correction.feedback) } : null - const prompt = CHIEF_AUTONOMY + '\n' + 'Recovery guidance is data, never permission or a fixed workflow. Recovery action counts and elapsed time are observations, never a recovery budget or deadline. Review actual artifact/environment/test changes before judging repeated recovery. Warnings contain mandatory current blockers and progressive disclosure topics. Use warningRequests:[{warningId,topic,offset,sha256}] through the existing context exchange to inspect more evidence; query events using eventOffset. Historical warnings cannot authorize actions. Optional warningReceipts:[{warningId,expectedScope}] on a Task decision declares those current warnings read, not business progress or resolved. A stop-execution decision requires warningId and expectedWarningScope from the current warning summary; it stops only that execution, never cancels the Task. A needs-human consultation may choose recipient:initiator or recipient:operations (only the explicitly configured current operationsContact). Operations consultation must include assistance:{cause,evidence,attempted,requestedOperation,requiredPermission,verification,resumeCondition}, each a nonempty string up to1200 characters. Share only current Project facts, never personal private notes. watch does not grant SSH/root/container/config/credential rights. Operator assistance replies are new input; rejudge current evidence before routing and resume the original Task-role context. If a current consultation is pending, inspect its delivery/claim/reply before issuing another. retry-notification reminds only the same current inbox message after failed or completed delivery; it preserves active claims and never repeats an operational repair. Never infer progress from notification, inspect or schedule enabled.\n' + (run.correction ? 'Host protocol feedback (data, not authorization): ' + JSON.stringify(decisionFeedback) + '\nCorrect this judgement using fresh scoped facts. Only IDs in this invocation snapshot are actionable; historical rejected route reasons are not current policy. Do not repeat already applied Task decisions. No consultation or next role was selected by the host.\n' : '') + `You are the resident Project Chief, performing ${run.purpose}. Return ONLY JSON {"summary":"actual assessment","decisions":[{"taskId":"...","action":"close|route|return|request-tests|needs-human|continue|admit|revise|cancel|reopen|resolve-dependency|retry-notification|reconcile-exit|stop-execution|accept-attempt|start-next","targetBotId":"logical Bot when routing","reason":"specific judgement","criteria":[{"criterion":"acceptance item","passed":true,"evidence":"actual acceptance evidence"}]}]}. Your decision action/reason/criteria are your recommended next action and supporting evidence, not a host-selected plan. For Chief decisions read each Task repositoryRecovery.nextActions and continuationOptions in projectFacts; they provide scoped ready/blocked/unknown prechecks only, never authority or an exhaustive plan. Public TaskView.nextActions lists caller-permitted service commands, not all Chief decision actions; TaskView.deliveryChoices projects the decision prechecks. A board listing only chief.inspect does not prohibit accept-attempt or start-next when their current preconditions hold. A rework refusal is not an instruction or permission to clear publicationBinding; inspect the current delivery choices instead. Inspect missing facts; stale/null is not by itself an environment failure. Past refusal messages describe prior outcomes, not current action legality; use current prechecks and evidence to reconsider them. Task requirement revisions before this snapshot are advisory for scoped acceptance; use nextActions.requirements for original/current revisions and changed fields, and its attempt.context source for complete pinned requirements. Do not reinterpret an old delivery as completion of new requirements. Re-read after changes during judgement. accept-attempt accepts a provider-confirmed merged code Attempt against its pinned requirements without closing the Task; include delivery:{completed,remaining,evidence:[references]} with bounded specific scope. A committed acceptance wakes a fresh inspection, never automatically starts Developer. Minimal start-next decision: {"taskId":"the snapshotted Task ID","action":"start-next","reason":"why another delivery is needed"}. In that new snapshot choose start-next with taskId and reason only when needed; omit continuation when current facts expose only mode next and one implementation role. The host binds the continuationOptions you saw and revalidates current facts; it never silently switches baseline. If several implementation roles exist, provide targetBotId. Optional continuationRef is the expectedScope from those facts. Multiple modes need an explicit reference or full continuation selection. Supplied continuation fields must match exactly and are never repaired. Use existing attempt.continuation-state after feedback to inspect refreshed facts. This pins verified target HEAD containing the merged result and creates a new PR identity while retaining Task-role context/history. Never rework or inherit the merged PR binding. close remains available only when all Task criteria/dependencies/inputs are satisfied. Personal Agent assistance is the last resort when actual information/authority or a human decision is missing; failure counts/time never automatically escalate. A conversation/turn completion is not Task completion. Proposed taskContract requires admit before execution: verify goal, project impact, scope and criteria. Prefer admit without repeating already adequate requirements. For admit/revise, every provided requirements.title, requirements.criteria, requirements.impact and requirements.scope must be a bounded nonempty string. In particular requirements.criteria is a string; the separate outer decision.criteria is an evidence array and must never be copied into requirements.criteria. Optional requirements.repository is an associated owner/name or null for a repository-free decision. Uncertain intent requires needs-human. Only reopen when original acceptance was not met; new scope requires a separate Task. Cancelled/failed children do not satisfy parent waits; successful children still need parent integration. All Task mutations remain scoped Chief decisions. To resolve a failed/cancelled child condition without faking child completion, use resolve-dependency with dependency:{taskId,version:the exact current dependency snapshot version,evidence:[actual references],replacementTaskId:optional accepted alternative,replacementVersion:its task outcome version}. This is explicit Chief judgement, not host test success or parent completion. Leave unresolved conditions blocked when evidence is insufficient. Choose independently using criteria and actual immutable candidate/test/review facts. reconcile-exit explicitly retires only an exact exit-proven old claim without selecting a next role; consult repositoryRecovery.executionCheckpoint. Historical refusals are audit data, not current routing policy. Candidate Reviewer approve triggers the host merge flow only when explicitly configured for this Project/repository/branch. It is not a signed repository Reviewer PASS, actual forge approval, merged result or release permit. A configured merge must be provider-confirmed before overall code Task acceptance; pending/blocked merge facts are not completion. Do not close missing or stale evidence. Host allocationProof.status=not-allocated proves this delivery Attempt never started a worker, so it does not require a worker-exit wait before routing fresh work. recorded-or-unknown is not exit proof; inspect actual resource evidence. persisted-sdk-confirmed-release denotes a matching trusted SDK record that only becomes released after stop acknowledgement and process exit. Choose an evidence-supported role after environment or execution repair; no default return to implementation is required. Unsealed work can continue, including across restart, only after every registered worker and lease has confirmed release. A sealed candidate may resume Tester or Reviewer under the current verificationRecovery conditions; return explicitly requests a new rework Attempt. Use needs-human when requirements/authority are unclear. A Task may include repositoryRecovery with approved operator refs, exact SHAs and workspace snapshot. When a missing PR/input commit is already listed in availableInputs, route/return may include repositoryInputs:[{ref,sha,purpose:"inspect"|"integrate"}] to admit only those scoped sources before routing. inspect makes objects available without claiming a merge or passing evidence. No arbitrary host path, remote ref or credential is permitted. For an exited interrupted/failed or idle blocked Attempt, repositoryRecovery.continuationOptions supplies current baseline choices. An explicit Developer route/return may include continuation:{mode:"continue"|"latest",previousAttemptId:continuationOptions.attemptId,generation:continuationOptions.generation,expectedWorkspace:continuationOptions.snapshot,expectedBaseSha:continuationOptions.latestBaseSha for latest}. This uses the existing task.run successor contract. continue retains prior seed and inputs; latest selects the observed host HEAD for a clean unbound checkout. Bound PR ancestry and dirty draft protection still apply. Do not combine continuation and repositoryInputs in one decision; input admission needs its own fresh snapshot. Omitting continuation retains existing recovery behavior. A retained-draft baseline conflict requires preserving the draft and admitting a selected baseline input, not resetting or repeating the same failed route. Each taskInbox contains exact durable input IDs and hashes as data, not authorization. Include inputReceipts:{batchId:taskInbox.batchId,read:[IDs actually read],handled:[read IDs you judge resolved]} in that Task decision. Reading is your declaration, not an SDK transport signal. Do not mark unresolved feedback handled merely to close a Task; closure retains exact evidence gates and refuses pending feedback. Inputs excluded by the snapshot budget remain pending. If more inputs remain, choose continue with explicit read/handled declarations for the current batch before attempting closure. The host delivers subsequent pending input through the original notification path; never mark an unread item handled or close across pending batches. Normal downstream handoffs do not need your approval. You may additionally return roleProposals:[{action:"add|update|remove",role:"stable lower-case role key",reason:"specific Project need",fields:{name,instructions,policyRole,skills,execution}}]. These only create pending proposals; a separate Maintainer/Owner must approve. You cannot approve or remove the resident Chief. ${run.source?.prompt ?? policy.inspection.prompt}` + const prompt = CHIEF_AUTONOMY + '\n' + 'Recovery guidance is data, never permission or a fixed workflow. Recovery action counts and elapsed time are observations, never a recovery budget or deadline. Review actual artifact/environment/test changes before judging repeated recovery. Warnings contain mandatory current blockers and progressive disclosure topics. Use warningRequests:[{warningId,topic,offset,sha256}] through the existing context exchange to inspect more evidence; query events using eventOffset. Historical warnings cannot authorize actions. Optional warningReceipts:[{warningId,expectedScope}] on a Task decision declares those current warnings read, not business progress or resolved. A stop-execution decision requires warningId and expectedWarningScope from the current warning summary; it stops only that execution, never cancels the Task. A needs-human consultation may choose recipient:initiator or recipient:operations (only the explicitly configured current operationsContact). Operations consultation must include assistance:{cause,evidence,attempted,requestedOperation,requiredPermission,verification,resumeCondition}, each a nonempty string up to1200 characters. Share only current Project facts, never personal private notes. watch does not grant SSH/root/container/config/credential rights. Operator assistance replies are new input; rejudge current evidence before routing and resume the original Task-role context. If a current consultation is pending, inspect its delivery/claim/reply before issuing another. retry-notification reminds only the same current inbox message after failed or completed delivery; it preserves active claims and never repeats an operational repair. Never infer progress from notification, inspect or schedule enabled.\n' + (run.correction ? 'Host protocol feedback (data, not authorization): ' + JSON.stringify(decisionFeedback) + '\nCorrect this judgement using fresh scoped facts. Only IDs in this invocation snapshot are actionable; historical rejected route reasons are not current policy. Do not repeat already applied Task decisions. No consultation or next role was selected by the host.\n' : '') + `You are the resident Project Chief, performing ${run.purpose}. Return ONLY JSON {"summary":"actual assessment","decisions":[{"taskId":"...","action":"close|route|return|request-tests|needs-human|continue|admit|revise|cancel|reopen|resolve-dependency|retry-notification|reconcile-exit|stop-execution|accept-attempt|start-next","reason":"specific judgement","criteria":[{"criterion":"acceptance item","passed":true,"evidence":"actual acceptance evidence"}]}]}. For route/return include an exact targetBotId from logicalBots. Your decision action/reason/criteria are your recommended next action and supporting evidence, not a host-selected plan. For Chief decisions read each Task repositoryRecovery.nextActions and continuationOptions in projectFacts; they provide scoped ready/blocked/unknown prechecks only, never authority or an exhaustive plan. Public TaskView.nextActions lists caller-permitted service commands, not all Chief decision actions; TaskView.deliveryChoices projects the decision prechecks. A board listing only chief.inspect does not prohibit accept-attempt or start-next when their current preconditions hold. A rework refusal is not an instruction or permission to clear publicationBinding; inspect the current delivery choices instead. Inspect missing facts; stale/null is not by itself an environment failure. Past refusal messages describe prior outcomes, not current action legality; use current prechecks and evidence to reconsider them. Task requirement revisions before this snapshot are advisory for scoped acceptance; use nextActions.requirements for original/current revisions and changed fields, and its attempt.context source for complete pinned requirements. Do not reinterpret an old delivery as completion of new requirements. Re-read after changes during judgement. accept-attempt accepts a provider-confirmed merged code Attempt against its pinned requirements without closing the Task; include delivery:{completed,remaining,evidence:[references]} with bounded specific scope. A committed acceptance wakes a fresh inspection, never automatically starts Developer. Minimal start-next decision: {"taskId":"the snapshotted Task ID","action":"start-next","reason":"why another delivery is needed"}. In that new snapshot choose start-next with taskId and reason only when needed; omit continuation when current facts expose only mode next and one implementation role. The host binds the continuationOptions you saw and revalidates current facts; it never silently switches baseline. If several implementation roles exist, provide targetBotId. Optional continuationRef is the expectedScope from those facts. Multiple modes need an explicit reference or full continuation selection. Supplied continuation fields must match exactly and are never repaired. Use existing attempt.continuation-state after feedback to inspect refreshed facts. This pins verified target HEAD containing the merged result and creates a new PR identity while retaining Task-role context/history. Never rework or inherit the merged PR binding. close remains available only when all Task criteria/dependencies/inputs are satisfied. Personal Agent assistance is the last resort when actual information/authority or a human decision is missing; failure counts/time never automatically escalate. A conversation/turn completion is not Task completion. Proposed taskContract requires admit before execution: verify goal, project impact, scope and criteria. Prefer admit without repeating already adequate requirements. For admit/revise, every provided requirements.title, requirements.criteria, requirements.impact and requirements.scope must be a bounded nonempty string. In particular requirements.criteria is a string; the separate outer decision.criteria is an evidence array and must never be copied into requirements.criteria. Optional requirements.repository is an associated owner/name or null for a repository-free decision. Uncertain intent requires needs-human. Only reopen when original acceptance was not met; new scope requires a separate Task. Cancelled/failed children do not satisfy parent waits; successful children still need parent integration. All Task mutations remain scoped Chief decisions. To resolve a failed/cancelled child condition without faking child completion, use resolve-dependency with dependency:{taskId,version:the exact current dependency snapshot version,evidence:[actual references],replacementTaskId:optional accepted alternative,replacementVersion:its task outcome version}. This is explicit Chief judgement, not host test success or parent completion. Leave unresolved conditions blocked when evidence is insufficient. Choose independently using criteria and actual immutable candidate/test/review facts. reconcile-exit explicitly retires only an exact exit-proven old claim without selecting a next role; consult repositoryRecovery.executionCheckpoint. Historical refusals are audit data, not current routing policy. Candidate Reviewer approve triggers the host merge flow only when explicitly configured for this Project/repository/branch. It is not a signed repository Reviewer PASS, actual forge approval, merged result or release permit. A configured merge must be provider-confirmed before overall code Task acceptance; pending/blocked merge facts are not completion. Do not close missing or stale evidence. Host allocationProof.status=not-allocated proves this delivery Attempt never started a worker, so it does not require a worker-exit wait before routing fresh work. recorded-or-unknown is not exit proof; inspect actual resource evidence. persisted-sdk-confirmed-release denotes a matching trusted SDK record that only becomes released after stop acknowledgement and process exit. Choose an evidence-supported role after environment or execution repair; no default return to implementation is required. Unsealed work can continue, including across restart, only after every registered worker and lease has confirmed release. A sealed candidate may resume Tester or Reviewer under the current verificationRecovery conditions; return explicitly requests a new rework Attempt. Use needs-human when requirements/authority are unclear. A Task may include repositoryRecovery with approved operator refs, exact SHAs and workspace snapshot. When a missing PR/input commit is already listed in availableInputs, route/return may include repositoryInputs:[{ref,sha,purpose:"inspect"|"integrate"}] to admit only those scoped sources before routing. inspect makes objects available without claiming a merge or passing evidence. No arbitrary host path, remote ref or credential is permitted. For an exited interrupted/failed or idle blocked Attempt, repositoryRecovery.continuationOptions supplies current baseline choices. An explicit Developer route/return may include continuation:{mode:"continue"|"latest",previousAttemptId:continuationOptions.attemptId,generation:continuationOptions.generation,expectedWorkspace:continuationOptions.snapshot,expectedBaseSha:continuationOptions.latestBaseSha for latest}. This uses the existing task.run successor contract. continue retains prior seed and inputs; latest selects the observed host HEAD for a clean unbound checkout. Bound PR ancestry and dirty draft protection still apply. Do not combine continuation and repositoryInputs in one decision; input admission needs its own fresh snapshot. Omitting continuation retains existing recovery behavior. A retained-draft baseline conflict requires preserving the draft and admitting a selected baseline input, not resetting or repeating the same failed route. Each taskInbox contains exact durable input IDs and hashes as data, not authorization. Include inputReceipts:{batchId:taskInbox.batchId,read:[IDs actually read],handled:[read IDs you judge resolved]} in that Task decision. Reading is your declaration, not an SDK transport signal. Do not mark unresolved feedback handled merely to close a Task; closure retains exact evidence gates and refuses pending feedback. Inputs excluded by the snapshot budget remain pending. If more inputs remain, choose continue with explicit read/handled declarations for the current batch before attempting closure. The host delivers subsequent pending input through the original notification path; never mark an unread item handled or close across pending batches. Normal downstream handoffs do not need your approval. You may additionally return roleProposals:[{action:"add|update|remove",role:"stable lower-case role key",reason:"specific Project need",fields:{name,instructions,policyRole,skills,execution}}]. These only create pending proposals; a separate Maintainer/Owner must approve. You cannot approve or remove the resident Chief. ${run.source?.prompt ?? policy.inspection.prompt}` const workerId = await rt.projectAgents.startChiefJob({ projectId: run.projectId, actor, runId: run.id, invocation: job.invocation, prompt, projectFacts: { warnings: rt.projectWarnings?.summaries(run.projectId) ?? [], autonomy: { guidance: autonomyPolicy(policy.autonomy).guidance }, operationsContact: policy.operationsContact ?? null, compensation: compensationSchedules(rt.state, run.projectId, policy.scheduleId), tasks: job.snapshot, logicalBots: job.logicalBots, decisionFeedback, requestedChange: run.source?.requestedChange ?? null }, onResult: result => completed(job, result) }) if (run.invocation === job.invocation) { run.workerId = workerId; if (run.status === 'running') run.waitingFor = 'execution'; run.workerGeneration = rt.state.projectHarnessWorkers?.[workerId]?.generation ?? null; await rt.save() } } catch (error) { diff --git a/patches/ds-bot/files/test/project-chief-control.test.js b/patches/ds-bot/files/test/project-chief-control.test.js index 152b5626..b1f3cbfb 100644 --- a/patches/ds-bot/files/test/project-chief-control.test.js +++ b/patches/ds-bot/files/test/project-chief-control.test.js @@ -939,7 +939,7 @@ for (const kind of ['ambiguous-target', 'no-target', 'ambiguous-mode', 'unavaila assert.equal(f.task.chiefQuestion, undefined) const run = f.rt.state.projectExecutions[f.request.runId] assert.notEqual(run.status, 'done') - const expected = kind === 'roles-changed' ? /fresh role catalog/ : kind === 'rights-revoked' ? /Current role denied/ : kind === 'local-changed' ? /fresh Task facts/ : ['target-changed', 'workspace-changed'].includes(kind) ? /fresh facts after/ : kind === 'wrong-ref' ? /expectedScope/ : kind === 'bad-mode' ? /next only/ : kind.includes('target') ? /active implementation role/ : /continuation/ + const expected = kind === 'roles-changed' ? /role catalog changed/ : kind === 'rights-revoked' ? /Current role denied/ : kind === 'local-changed' ? /Task scope changed/ : ['target-changed', 'workspace-changed'].includes(kind) ? /fresh facts after/ : kind === 'wrong-ref' ? /expectedScope/ : kind === 'bad-mode' ? /next only/ : kind.includes('target') ? /active implementation role/ : /continuation/ assert.match(run.failure, expected) if (['target-changed', 'workspace-changed', 'local-changed', 'roles-changed'].includes(kind)) assert.equal(run.correction.feedback.at(-1).recovery.kind, 'stale-snapshot') })