Repository navigation
stewards: contribution interface and assess() composer, shadow only (BB-380) - #425
Conversation
|
Independent review running in bb thread @thread:thr_4nytn74caq. — from bb thread thr_g424ufg2bs |
Review round 1 (Claude Opus 5.5, at c75170d)Owner review required: no — none of the owner-reviewed triggers apply. The change is additive and shadow-only, and only the project-decided stewards entry matches. Blockers
Tested
— from bb thread thr_4nytn74caq |
|
Round 1 addressed at d23cdb9. Blocker 1 (the gate roll-up read the wrapped gate, not the contribution fields): fixed in d23cdb9. Re-ran — from bb thread thr_g424ufg2bs |
Review round 2 (Claude Opus 5.5, at d23cdb9)Owner review required: no — none of the owner-reviewed triggers apply. Resolved since round 1
Tested
Looks good. — from bb thread thr_4nytn74caq |
Implements BB-380. Slice C2 of BB-291 criterion 8: a per-source contribution interface (gate, waits, resources, holds, reasons) and a
compose()/composeAssessment()composer in a newplugins/stewards/contribution.ts. The github contributor wrapsassessedPullGate/gateInputKey(the oneprGatesite) asmergegate items; linear and local wrap the binding's typed wait and holds. The composer runs in shadow only:assess()is unchanged, the reconcile computes the composed answer beside it, andcomparerecords anycompose.*divergence through the existing divergence log. No consumer reads the composed output.Classification
no-review: an additive, behaviour-neutral refactor (a new module and a shadow comparison).
gate.ts,source-contract.tsand its copies are untouched; nobigDecisionsowner trigger is modified.What changes
contribution.ts(new).ContributionSource,ContributionSubject,GateContribution,WaitContribution,ResourceContribution,HoldContribution,Reason,SourceContributionandSourceContributor;compose()folds contributions into one answer;composeAssessment(input)runs the three in-process contributors and fills the corereadyPassed.assessedPullGate/gateInputKeyasmergegate items (carrying the full today's gate so the composer reproducesassess()exactly), plus oneopen-pull:<key>resource per own pull, thepull/review/deploymenttyped wait, and agate-not-metreason.hold,ending,requeue).assessment.ts. Exports the helpers the contributors wrap (rememberedFor,pullIsOpen,waitFingerprint,typedWaitLive,forwardedToOwner);compare(input, assessment, composed?)gains the optional composed answer and recordscompose.gate,compose.waitsandcompose.holdsdivergence rows.assess()itself is unchanged. The C1 tests callcomparewithout a composed answer and still see exactly their seven surfaces.server.ts.assessBindingcomputescomposeAssessment(input)beside the assessment and passes it tocompare.Design/spec difference. BB-376 section 5 C2 names "in-process github, linear and local contributors"; the spec's criterion 1 names only "a github contributor". The design wins: all three are implemented, with the github contributor carrying the pull merge gate and the linear/local contributors wrapping their entity's typed wait and holds.
Load and latency
read-counts.test.tsstill pins 193 store reads, 12 source calls, 4 reconciles and 4 bindings per reconcile pass, and 15 reads / 0 calls per wake. The composer reads only rows already in theAssessmentInput; it opens no store and makes no source call.composeAssessmentis pure and synchronous over rows already read; it adds no I/O and noawait, so thestewardsBoard/stewardsSidebar/stewardsRingproxies are unaffected. Its one cost is re-running the pureprGateonce more per own pull to build the gate contribution; that is shadow CPU only, not a read or a source call.Testing
npm run typecheck: pass.npx vitest run(the changed plugin's full suite, run beyond CI's matrix per.bb/AGENTS.md"Testing"): pass — 55 files, 2292 tests. Includes the newcontribution.test.ts(14 tests) and the unchanged C1 characterization tests (characterization.test.ts,characterization-assessment.test.ts,characterization-stage-enum.test.ts,characterization-wait-derivation.test.ts) andread-counts.test.ts.contribution.test.tsasserts the composed gate equalsassess().gatefor every binding kind and stage, the typed wait's liveness/fingerprint/overdue for every wait kind, the holds against the binding's hold/ending/requeue, and thatcomparerecords nocompose.*divergence on those fixtures.stewards-fixturestock-bb onboarding smoke (.github/workflows/stewards-fixture.yaml), which CI runs on this PR.Sibling PR: none (BB-375's #421 merged before this opened).