[Tracker] Docs and plans steward

Goal

Every document and plan in the repository tells the truth about current main. Every plan is executing, explicitly blocked, or retired with a reason. Val can plan the next session from docs/status.md alone.

How this issue is used

Start a steward session with the first prompt below, and resume one with the second. Both work in Claude Code and Codex.
Each cycle ends with a checkpoint comment here: the main SHA reviewed, PRs opened and merged, plan state changes with evidence, drift fixed and left, questions for Val, and the next action.
The steward asks Val here before it restructures docs/status.md, deletes or moves a doc, changes a spec's decisions or reorders "Next". It sets that item aside until Val answers.
Docs-only reconciliation PRs merge through the guard script once CI is green. Code PRs wait for Val's go-ahead. The steward never tags, publishes, uploads to TestFlight or turns on flags.

Long-running goal prompt — start

Act as the docs-and-plans steward for TypeSafeAI/modex. The goal: every document and plan in the repository tells the truth about current main; every plan is executing, explicitly blocked, or retired with a reason; and Val can plan the next session from docs/status.md alone. Work in cycles until the stop condition at the end holds, leaving a checkpoint after each cycle.

GROUND RULES (they apply whatever your harness loads)
- Read AGENTS.md and ~/.claude/CLAUDE.md first. Their signing, attribution, merge-guard, worktree and ~/.coven rules bind you here.
- Never edit the primary checkout. Cut each cycle's branch with `scripts/worktree.sh new docs-steward-<yyyymmdd>` (slug, no dots). Install with pnpm only. Clean up with `scripts/worktree.sh release`/`remove` per docs/worktree-hygiene.md, never by force.
- Sign every commit with `git commit -S -F <message-file>`, then confirm a good signature with `git log -1 --show-signature`. Stop if it is missing.
- Credit no model, tool or vendor: no Co-Authored-By for Claude, Codex or similar, and no "Generated with" footer. A non-familiar session ends commits with only `Coven-Harness: <harness> / <model id>`.
- Merge only with `~/.claude/scripts/gh-merge-when-green.sh <pr> --squash`.
- Never touch ~/.coven. Never resume, rebase or delete a worktree, branch or PR you did not create.
- Other Claude and Codex sessions land work here at the same time. Before each cycle, list open PRs (`gh pr list`) and worktrees (`scripts/worktree.sh audit`). Leave alone any doc section an open PR is changing (`gh pr diff <n> --name-only`). Release review docs and the release entry under "In flight" belong to whoever is cutting that release.
- Change a status only on evidence. Tick a box, mark something shipped or close a gap only when you can cite a merged PR, a release tag, a review doc, a CI run id or file:line on main. A previous session's summary is not evidence. Leave unknowns marked unknown.

WHAT YOU STEWARD (classify every file before editing)
1. Living map: docs/status.md (Shipped, per-release sections, On main unreleased, In flight, Known rough edges, Next, Conventions).
2. Reference docs: docs/*.md feature pages and README.md. They describe behaviour on main, so check them against code and the Shipped table.
3. Specs: docs/specs/*.md. Each status line ("Approved; not implemented", "direction accepted") must match what has landed.
4. Plans and ledgers: docs/superpowers/plans/*.md, docs/mac-store-preview-plan.md and any other checkbox ledger. Each is in exactly one state: active, blocked (on what or whom), delivered (by which PR or release), superseded (by what) or abandoned (why).
5. Evidence: docs/reviews/ is dated history. Never rewrite it. Fix broken links only, and put corrections in status.md or a new dated review.
6. GitHub trackers and issues, such as #273 and its children: plans that live outside the tree. Their checklists must match the issues that have closed and the PRs that have merged.

KNOWN DRIFT AS OF 2026-10-18 (re-check each item first; it may already be fixed)
- README.md calls automatic updates "in-flight" and says published builds need a manual install, but v0.0.12 shipped self-update. It also says app-owned sign-in is "under review".
- docs/specs/2026-10-11-modex-routines-spec.md says nothing has shipped, but Phases 0 and 1 are on main.
- docs/specs/2026-10-12-sticky-auto-routing-spec.md says "not implemented", but v0.0.11 shipped Sticky Auto routing.
- In docs/superpowers/plans/, 2026-10-11-desktop-roadmap.md still has open boxes for Canvas, signed updates and native installers, which shipped in v0.0.11, v0.0.12 and v0.1.0. Check the open boxes in 2026-10-12-sticky-auto-routing.md (70), 2026-10-09-browser-extensions.md and 2026-10-08-workos-oauth.md the same way.
- Tracker #273's checklist has no boxes ticked, but #274 (CF-1) has closed and #282, #284 and #285 have merged.

EACH CYCLE
1. Refresh. Fetch, then read what changed since your last checkpoint: main's log, merged PRs, closed issues, new tags and releases, open PRs. That delta is your work queue.
2. Reconcile. For each change, grep for its feature names, PR numbers, versions and file paths to find every doc that should now say something different. Verify claims against code on main, not against other docs. Check that relative links and anchors resolve. Move resolved items out of "Known rough edges". Edit precisely and in each file's existing voice. Keep status.md a map, not a changelog: link to a review instead of copying its evidence.
3. Disposition plans. Record each plan's and spec's state with evidence. For a delivered plan, add one line at the top ("Delivered in vX by #N; the boxes below are historical") rather than ticking boxes you cannot prove. For a superseded or abandoned plan, say what replaced it or why it stopped. Delete no plans.
4. Advance one plan. Pick the single most valuable unblocked next step across status.md's "In flight" and "Next", the active plans and the trackers' dependency orders. Prefer steps that unblock other work, then finishing partial work, then starting new work. If the step is small and no one owns it, do it in its own worktree through the normal flow: read the plan and spec, write the test first, make the minimal change, run AGENTS.md's pre-push checks, make signed commits, and open a PR that updates the plan ledger and status.md in the same change. If the step is large, owned by another session or waiting on a decision, do not start it. Instead, write an exact next-step note on the plan or its issue.
5. Land. A docs-only reconciliation PR may merge through the guard script once CI is green. Open each code PR and get it green, then stop for Val's go-ahead before merging it. Never tag, publish a release, upload to TestFlight, approve the Release environment, or turn on a feature flag or approval gate.
6. Checkpoint. Comment on tracker #290 (TypeSafeAI/modex). Include: the date and the main SHA you reviewed; PRs opened and merged; each plan's change of state, with evidence; drift you fixed; drift you left, and why; questions for Val; and the first action for the next cycle.

JUDGEMENT
- Ask Val in the tracker, and set the item aside, before you restructure or split status.md, delete or move a doc, change a decision a spec records, re-scope or reorder "Next", or change release, signing or credential docs beyond factual corrections.
- Correct an existing doc instead of writing a new one. The tracker is your only process document.
- Many PRs edit status.md. Keep your PRs small, rebase onto main just before you push, and expect a CI rerun whenever main moves.
- Stop condition: no unreconciled drift remains, and the next plan step is blocked on a person or an external gate, or is waiting for Val's merge. Write the checkpoint and stop. Never report something complete that you did not verify in this cycle.

Follow-up long-running goal prompt — resume

Resume the Modex docs-and-plans steward goal. Work from evidence, not from the previous session's summary. Read AGENTS.md, ~/.claude/CLAUDE.md, the latest checkpoint comment on tracker #290 (TypeSafeAI/modex), and everything on main since that checkpoint's SHA: commits, merged and closed PRs, closed issues, tags. Check whether each worktree and PR the checkpoint names still exists and is yours. Resume one only after `scripts/worktree.sh claim`/`check` proves you own it; otherwise start a fresh worktree. Close out half-done work first. For an unmerged docs PR: rebase, wait for CI, merge with the guard. Leave code PRs that are waiting for Val as they are. Finish any plan step left halfway. Then run the normal cycle: refresh, reconcile, disposition the plans, advance one unblocked plan step, land, checkpoint. Every ground rule of the original goal still applies:
- signed commits, each confirmed with --show-signature
- no harness attribution (the Coven-Harness trailer only)
- merges through the guard script only
- nothing under ~/.coven is touched
- no tags, releases, publication or flags
- evidence behind every change of status
- restructuring and scope are Val's decisions
Stop with a checkpoint when the stop condition holds.

Close the two finished In flight entries and add the GitHub App to Next

Move fixed rough edges into a dated review

Move status.md's release sections into their release reviews

Trim the Mac App Store entry to its current state

Record what shipped across plans, specs and reference docs (steward c…

Checkpoint: cycle 1, 2026-10-18

Reviewed: main at d3711c6 (#287), then f6976df (#289). Worktree docs-steward-20261018, owned by this session.

PRs: #292 (docs only), merged through the guard as 04b7ec7 with all 9 checks green.

Trackers: in #273, ticked #274 (CF-1, closed). The other children are still open.

Plan states

Plan or spec	State	Evidence
workos-oauth	delivered	#141, v0.0.9
browser-extensions	delivered	#152, v0.0.11
knowledge-agents, pages-agents	delivered	#150, v0.0.10
desktop-roadmap	delivered except Cloud tasks (out of scope)	#164, #181/#200/#201, #173/#186
worktree-hygiene	delivered	#171
release-candidates	delivered	#186, v0.1.0
sticky-auto-routing (plan + spec)	delivered	#180, #183, v0.0.11
images-plugins	delivered	#236, v0.1.0
mac-store-preview-plan	partial; no owner, no open PR	#119; parity matrix, real-CLI and fresh-user acceptance open
routines spec	Phase 1 and most of Phase 0 on main, unreleased	#206, #253, #257, #267; gate decision open, Phase 2 waits on it
approval-rules spec	parts 1–2 shipped, gate off	#54, #107, #185
tracker #273 (CF-1…8)	active, owned by other sessions	#289 landed; open: #283, #286, #288, #291, #293, #294

Drift fixed (#292)

Plan and spec status lines, README (5 places), 15 reference docs, and in status.md the site fallback, the In flight pointer, Next 3, and the node-pty and demo-thread lines. Each change was checked against code on main, a status.md release section or a review. The PR body lists them.

Drift left, and why

Left for the open PRs: status.md "On main, unreleased" and docs/conversation-events.md, which Render transcript Markdown with CommonMark and GitHub's extensions (CF-4 parser slice) #283, Make Claude Stop truthful on claude -p: native interrupt, bounded exit, withdrawn prompts (CF-2) #286, Read a completed Codex web search as done, so it folds again (CF-4) #288, Keep failed Claude turns owned through Stop (CF-2 repair) #291, Answer each approval and question once, and show Codex questions as cards (CF-3) #293 and Prove the conversation path end to end and record what fidelity still needs (CF-8, first part) #294 are editing. Make Claude Stop truthful on claude -p: native interrupt, bounded exit, withdrawn prompts (CF-2) #286 now also edits In flight and Next 3 right next to Record what shipped across plans, specs and reference docs (steward cycle 1) #292's changes, so I left it a note on how to resolve the conflict.
Not proven wrong, left as is:
- modex-signin.md names a fixed callback port 43821, but the code listens on an ephemeral port. The WorkOS redirect settings would settle it.
- auto-routing.md says to install jev with npm link and pins jev-latest → jev-1.13.0 with its price. Both are external facts.
- auto-routing.md calls tier 2 of 0–3 "the middle tier".
- analytics.md gives fixture counts that no test asserts.
- gemini-feasibility.md uses npm install --prefix for a probe outside the workspace.
- ios-companion.md names Xcode 26.6.
- pages-maintenance.md says Codex tools last "until the app-server restarts".
- notes.md gives a benchmark figure.
- The test counts in update-recovery.md are dated 2026-10-12.
Wording for Val: mac-app-store.md gate 2 reads as unimplemented. The host transport and MAS shell exist, but production acceptance is open. The question is whether to reword it as "implemented; acceptance open".

Plan step advanced

None. Every unblocked step has an owner or waits on a person:

the CF slices belong to sessions with live worktrees;
v0.2.0 belongs to release-v0-2-0;
Next 1 needs Windows and Linux machines;
Next 2 waits on the approval gate decision;
the Store gates need a fresh macOS user and Apple.

Questions for Val

Finished In flight entries. Answered: removed both, and the GitHub App item is now Next 4, ahead of the out-of-scope Cloud tasks (Record what shipped across plans, specs and reference docs (steward cycle 1) #292).
Fixed incidents in Known rough edges. Answered: six fixed incidents moved word for word to docs/reviews/2026-10-18-resolved-rough-edges.md, each under its fixing PR (Record what shipped across plans, specs and reference docs (steward cycle 1) #292). Besides the four named here, these include the native-picker and owner-removal test fixes, which had nothing left open. The demo-thread and sizing entries stay because their causes are unproven.
status.md size. Answered: the v0.0.6–v0.1.0 release sections moved unchanged into a "Status summary" section of each release review. The Shipped rows link them, a new convention says where a release's details go, and status.md went from 1,060 to 388 lines, then to 352 after the Store trim (Record what shipped across plans, specs and reference docs (steward cycle 1) #292).
Mac App Store entry. Answered: cut from 52 lines to 15, giving the current state with links to the receipts. The old text moved unchanged to the October 11 signed-acceptance review (Record what shipped across plans, specs and reference docs (steward cycle 1) #292). Still waiting on Val: the receipt says the attribution of one GUI Connect grant from that run is awaiting Val's response.

Noticed, not acted on

status.md ends with a separate "Local Mac Store development preview (unreleased)" section that overlaps the trimmed Store entry. It could fold into the entry or into mac-store-preview-plan.md. Val's call.

Next cycle's first action

Refresh from 04b7ec7 (#292), then:

update the [Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273 checklist and the In flight pointer as Render transcript Markdown with CommonMark and GitHub's extensions (CF-4 parser slice) #283, Make Claude Stop truthful on claude -p: native interrupt, bounded exit, withdrawn prompts (CF-2) #286, Read a completed Codex web search as done, so it folds again (CF-4) #288, Keep failed Claude turns owned through Stop (CF-2 repair) #291, Answer each approval and question once, and show Codex questions as cards (CF-3) #293 and Prove the conversation path end to end and record what fidelity still needs (CF-8, first part) #294 land;
once v0.2.0 publishes, change "unreleased" to v0.2.0 in the routines spec and the README's "Not (yet) here", and check the site fallback;
act on any answers above.

Mark mac-app-store gate 2 as implemented, acceptance open

Fold the Store preview section into In flight; mark Store gate 2 impl…

Follow-up: cycle 1 decisions, 2026-10-18

#296 (docs only) merged through the guard as db0bbfd, with all 9 checks green. It acts on two more of Val's decisions:

Store section folded. status.md's "Local Mac Store development preview (unreleased)" section is folded into the Mac App Store edition entry under In flight. That adds the preview (Add an opt-in sandboxed Mac Store preview with a local host #119), the recovery behavior from Keep Companion reachable and preserve Store host recovery #166, and two open items: graceful host shutdown and the full parity matrix. The section moved unchanged to the "Status summary" of docs/reviews/2026-10-11-store-recovery.md. status.md is now 329 lines.
Gate 2 reworded. mac-app-store.md gate 2 now reads "Implemented; acceptance open" and links the readiness matrix for the signed acceptance that is still pending.

Still waiting on Val: where the grant from the GUI Connect run in the October 11 signed-acceptance review came from.

Next cycle: refresh from db0bbfd.

Point In flight at the tracker instead of listing CF PRs

Record the GUI Connect grant as Val's manual testing

Settle open docs claims, point In flight at the tracker, record the S…

Checkpoint: cycle 2, 2026-10-18

Reviewed: main at db0bbfd (#296), with no commits since the last checkpoint. #297 and #301 (CF-8) landed during the cycle. Worktree docs-steward-cycle2, owned by this session.

PRs: #300 (docs only), merged through the guard as 28c6723 with all 9 checks green. It went in on the third try, after #297 and #301 landed during CI.

Trackers: #273's children #275–#281 are all still open, so its checklist needs no change.

Unverified claims from cycle 1, now settled

Claim	Result
`modex-signin.md`: fixed callback port 43821	Only the default registered with WorkOS. Modex uses a system-assigned port (`listen(0)` since #141). Clarified.
`auto-routing.md`: "the middle tier" for unknown models	Tier 2 of 0–3 (`catalog.ts`). Reworded.
`analytics.md`: fixture counts	All correct when `sample.ts` is run (471 / 147 / 45 / 11 / 212). Unchanged.
`pages-maintenance.md`: Codex keeps its tools until the app-server restarts	Correct: the config is sent only when a thread is first loaded. Unchanged.
`update-recovery.md`: test counts	From the 2026-10-12 run. Dated.
`gemini-feasibility.md` `npm --prefix`; `ios-companion.md` Xcode 26.6; `notes.md` benchmark; `auto-routing.md` `jev` install and model pin	Still unverified. Each needs an external system, a CI log or a rerun of a measurement. These are research or dated facts, not instructions for the current build.

Plan step advanced

None. The unblocked work is still owned or gated:

the CF slices belong to live sessions;
v0.2.0 belongs to its release session;
Next 1 needs Windows and Linux machines;
Next 2 waits on the approval gate decision;
Next 4 needs a GitHub App owner;
the Store gates need a fresh macOS user and Apple.

Questions for Val

None open.

Next cycle's first action

Refresh from 28c6723 (#300), then:

after Make Claude Stop truthful on claude -p: native interrupt, bounded exit, withdrawn prompts (CF-2) #286 lands, check that Next 3 is gone;
once v0.2.0 publishes, change "unreleased" to v0.2.0 in the routines spec and the README, and check that the release has a Shipped row linking its review and no status.md section.

Claim: post-v0.2.0 docs cycle, 2026-10-18

v0.2.0 was published at 23:14:37Z (by github-actions[bot], tag on 01b1b8a). Nothing in the tree records it yet, and the release-v0-2-0 worktree has been idle for 18 hours. This session is taking the docs side, in one PR. It follows the layout Val chose:

Shipped: add a v0.2.0 row linking its review.
Review: move the items shipped in v0.2.0 unchanged out of "On main, unreleased", into a "Status summary" in the v0.2.0 review.
Unreleased: keep only work after 01b1b8a there, which is the CF slices.
In flight: trim the v0.2.0 entry to the gates no receipt records yet: the independent download checks, the v0.1.0 → v0.2.0 banner upgrade on the release Mac, and the Companion 0.2.0 (7) TestFlight upload.
Wording: change "unreleased" to v0.2.0 in the routines spec and the README.

Not taken: the website's static fallback (apps/site/release.json) and any publication evidence. Those belong to the release owner or Val.

If the release session is already doing this, say so here and I'll stop.

Checkpoint: cycle 13, 2026-10-18

Reviewed: main from a4039b2 to 81798d7 (#293, #306, #307).

Correction to cycle 12

Cycle 12 said every relative link and anchor resolves. That check ran on an empty export: git archive failed on a path that doesn't exist, and I had silenced the error. I re-ran it on the full tree, and it finds 0 broken links across 109 Markdown files, on 57e7607 and on each PR head. The claim was true, but it wasn't verified when I posted it.

Ownership

None of the 26 live worktrees was the steward's:

every one is claimed by another session in .git/modex-maintenance/owners/, except usage-credits-failure, which its own session released;
the four earlier docs-steward-* records are all released, with no live directory.

I resumed none. I cut docs-steward-approval-spec and docs-steward-canvas-comment with new, and check passed in both.

Landed

PR	Merge	What
#306 (docs)	`3560064`	Approval rules spec and status.md "Still open". The first guard run refused it because `Upload TypeScript coverage` timed out at GitHub's processing step (`processing_timeout`, job 117817438712). Re-running that job passed, and the guard merged.
#307 (comments only)	`81798d7`	Merged through the guard on Val's go-ahead. All 11 checks passed.

Drift fixed

docs/status.md:109: "Still open" no longer lists "the Markdown parser (Render transcript Markdown with CommonMark and GitHub's extensions (CF-4 parser slice) #283)". Answer each approval and question once, and show Codex questions as cards (CF-3) #293 merged as 57e7607 at the same head it had before my comment there, so it kept the item. The line now names "CF-4's remaining Markdown checks ([P1][CF-4] Render typed conversation turns with safe Markdown, code controls, and stable streaming UX #277)" (Record #185's downgrades in the approval spec; drop the merged parser from Still open #306).
docs/specs/2026-10-01-modex-jev-approval-rules-spec.md: it described the gate as it was before Ask before approval rules allow chained or clipped commands #185. "Rule evaluation" and "Known limits" now name the compound and incomplete downgrades (gate.ts:155-158, digest.ts:103-111) (Record #185's downgrades in the approval spec; drop the merged parser from Still open #306).
canvas.ts:20 and shared/inline.ts:2: both comments described the transcript's Markdown as it was before Render transcript Markdown with CommonMark and GitHub's extensions (CF-4 parser slice) #283 (Correct two comments that still describe the transcript's old Markdown #307).

Since cycle 12, besides my PRs

Answer each approval and question once, and show Codex questions as cards (CF-3) #293 (CF-3) merged. Its claims hold on main: the phone's 409 (companion.ts:356), item/tool/requestUserInput (codex.ts:599) and the withdrawn state (human-requests.ts:55).
[P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276 is still open, so tracker [Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273 is unchanged.
v0.2.0 was published at 23:14:37Z. Another steward session claimed the post-release docs here and opened Say read-only routines ship from v0.2.0 in the README and the routines spec #310 for cycle 3's release follow-ups. I'm leaving those to it.

Pre-landing check

I ran git merge-tree for #288, #299, #304, #308, #309, #310 and #311 against 81798d7.

All seven merge cleanly.
Say read-only routines ship from v0.2.0 in the README and the routines spec #310's only change under In flight is its own v0.2.0 entry. Every other status.md addition lands under "On main, unreleased".
Carry each request's answer, who gave it and its delivery through the journal, replay and the phone (CF-6, request half) #308 rewrites "Still open" again: it drops "CF-6's request half" and keeps Record #185's downgrades in the approval spec; drop the merged parser from Still open #306's CF-4 wording.

Plans

Approval rules spec: a dated progress note under the enablement list. Ask before approval rules allow chained or clipped commands #185 closed the clipped-command false allow. The quoted-command false allow is open and the corpus hasn't been run live since, so item 1 is unmet.
No other plan changed state.

Plan step

Most valuable step: the approval gate's quoted-command false allow, because Routines Phase 2 (Next 2) waits for the gate decision.

Not started, because:

it needs a design that tells a quoted mention from an operation. Masking quoted text alone would also hide sh -c '…' from Jev.
it needs a live Jev run of the corpus plus a holdout, which spends usage.

The exact next step is now in the spec (#306). No open PR or worktree carries it.

Drift left

None known.

Question for Val

The gate's item 1 has no owner. If you want Routines Phase 2 to move, that's the step to assign.

Next cycle's first action

Refresh from 81798d7.
After Say read-only routines ship from v0.2.0 in the README and the routines spec #310 lands: confirm the v0.2.0 Shipped row is there, and that "On main, unreleased" holds only work after 01b1b8a.

Both steward worktrees are being removed through scripts/worktree.sh remove.

Record v0.2.0's publication in status, the release review, the README…

Checkpoint: cycle 14, 2026-10-18

Reviewed: main from 81798d7 to 396b17c (#308, #312, #313, #314, #310).

Since cycle 13

Carry each request's answer, who gave it and its delivery through the journal, replay and the phone (CF-6, request half) #308 (CF-6, request half), Keep each plan clarification and its answer in the task record and Notes (CF-7, clarification half) #312 (CF-7, clarification half) and Prove the request variants in the fidelity path: withdrawal, two clients at once, reconnect, a clarifying question (CF-8) #313 (CF-8, request variants) each recorded their slice under "On main, unreleased" and shortened CF's "Still open" line themselves. No drift.
Record v0.2.0 publication and advance the website fallback #314 recorded v0.2.0's publication: the Shipped row, "What shipped" and the download checks in the release review, the In flight gates, and the website fallback moved to 0.2.0.
Say read-only routines ship from v0.2.0 in the README and the routines spec #310 overlapped Record v0.2.0 publication and advance the website fallback #314. Both recorded v0.2.0 in status.md and the release review, and Record v0.2.0 publication and advance the website fallback #314 landed first at 02:36Z, so Say read-only routines ship from v0.2.0 in the README and the routines spec #310 conflicted. On Val's instruction I merged main into Say read-only routines ship from v0.2.0 in the README and the routines spec #310 and took main's side of both conflicts (868c54f, signed). Record v0.2.0 publication and advance the website fallback #314 is the more current version, because it records the download checks that Say read-only routines ship from v0.2.0 in the README and the routines spec #310 still listed as open. Say read-only routines ship from v0.2.0 in the README and the routines spec #310 kept only the two lines Record v0.2.0 publication and advance the website fallback #314 had missed, where the README and the routines spec still called read-only routines unreleased. I retitled it to match. The owner session merged it at 03:29Z as 396b17c, while my guard's merge call was waiting out GitHub's GraphQL rate limit. Record note: 396b17c's commit message still describes Say read-only routines ship from v0.2.0 in the README and the routines spec #310's original scope, the status.md moves, the review's "Status summary" and an In flight entry. Its diff is only README.md and the routines spec (+3 −3). That status and review work is in e25dd9d (Record v0.2.0 publication and advance the website fallback #314).

Confirmed on `396b17c`

Shipped: the v0.2.0 row links its release and review.
On main, unreleased: all 16 entries are newer than the tag. None of their titles appears in status.md at 01b1b8a. They are the CF slices and Render transcript Markdown with CommonMark and GitHub's extensions (CF-4 parser slice) #283's renderer.
In flight: the v0.2.0 entry lists the two gates that need Val at the Mac: the v0.1.0 → v0.2.0 banner upgrade on the release Mac, and the TestFlight upload of Companion 0.2.0 (7).
Routines: Next 2, the README and the routines spec now all say read-only routines shipped in v0.2.0.
Trackers: on [Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273, [P1][CF-1] Preserve native turn/item identity, message phases, and execution outcomes end to end #274 and [P1][CF-2] Harden Claude Stop, late-message draining, and cancelled-request cleanup on claude -p #275 are ticked and closed. [P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276 to [P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281 are open, and none of Carry each request's answer, who gave it and its delivery through the journal, replay and the phone (CF-6, request half) #308, Keep each plan clarification and its answer in the task record and Notes (CF-7, clarification half) #312 or Prove the request variants in the fidelity path: withdrawal, two clients at once, reconnect, a clarifying question (CF-8) #313 claims to close one, so the checkboxes match.
Links: 0 broken across all 110 Markdown files, on a full-tree export of 396b17c.
Open PRs: Read a completed Codex web search as done, so it folds again (CF-4) #288, Show and copy the lines a Codex memory citation points to (CF-5) #299, Move a project to a new workspace from the chat sidebar #304, Harden the event journal against failed writes, crashes and oversized frames (CF-6, journal mechanics) #309, Give the macOS CI job 45 minutes instead of 30 #315 (macOS CI timeout), Make CF-3 answer delivery receipts truthful #316 (CF-3 delivery receipts) and Update macOS over the air: download in the background, install on quit #317 (over-the-air macOS updates) merge cleanly onto main, and Keep a long command's full log and read it back a page at a time (CF-5) #311 merges cleanly onto its base.

Plans

No status changed. The step that matters most is still the approval gate's quoted-command false allow, which blocks Routines Phase 2. It needs a design (masking quoted text alone would also hide sh -c '…' from Jev), a live Jev run and an owner. The approval spec records the exact next step. Every other Next item waits on machines, accounts or a GitHub App owner.

Process note

Three sessions worked on v0.2.0's docs at once: this steward, the docs-steward-v020-shipped steward (which claimed the work here) and the session behind #314 (which did not claim it here). Separately, two of us pushed identical merges of main into #310 three seconds apart. The non-force push refused the second, so nothing was lost. Before starting docs work for a release, check this issue for claims.

Questions for Val

Approval gate fix: nobody owns it. If Routines Phase 2 should move, that's the work to assign.

Next action

Refresh from 396b17c. The v0.2.0 gates (the banner upgrade and TestFlight build 7) are Val's at the Mac. Re-check #273 as CF-4 and CF-5 land.

Checkpoint: cycle 15, 2026-10-19

Reviewed: main from 396b17c to 58b468d (#299, #315, #288, #304, #311, #309, #328, #331).

Landed by this session

PR	Merge	What
#315 (CI)	`cb33db5`	`unit + e2e (macOS)` gets 45 minutes instead of 30. Merged through the guard on Val's go-ahead.
#309 (code)	`c345779`	CF-6 journal hardening, merged through the guard on Val's go-ahead. It passed every check four times. #288 and #304 each landed while its CI was running, and the guard refused it as BEHIND each time.
#331 (docs)	`58b468d`	The fidelity group's "Still open" line now names what remains for each of CF-3 to CF-8. #328 landed during its run, so CF-5's comparison item was dropped before merge.

For each of the three, the squash tree matches the branch.

Since cycle 14, besides mine

Each of these recorded its own status entry and reference doc, so there was no drift:

PR	Docs it updated
#299 (CF-5 memory citations, patch comparison)	`conversation-evidence.md`
#288 (CF-4 web-search fold)	`conversation-events.md`
#304 (move to a new workspace)	`notes.md`
#311 (CF-5 retained log)	`conversation-evidence.md`
#328 (CF-5 removal-only and notebook comparison)	`conversation-evidence.md`

#315 changes only CI. No doc mentions the time limit.

Drift fixed

docs/status.md "Still open" (Name what is still open for each fidelity child in status.md (steward cycle 15) #331). It named only CF-4, live providers and VoiceOver. CF-3, CF-5, CF-6 and CF-7 had dropped off as their first slices landed, while [P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276 to [P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281 all stay open.

Plans and trackers

No plan or spec changed state.
On [Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273, [P1][CF-1] Preserve native turn/item identity, message phases, and execution outcomes end to end #274 and [P1][CF-2] Harden Claude Stop, late-message draining, and cancelled-request cleanup on claude -p #275 are ticked and closed, and [P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276 to [P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281 are open, so the checkboxes match.
New: tracker [Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319 (Code Review workspace) and its children [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 to [CR-8][P1] Verify Code Review fidelity, accessibility, recovery and end-to-end safety #327, opened 05:39Z. No doc mentions it yet (see the questions below).
CF-6 residue is in review as Read the thread again when the Mac refuses an answer from the phone (CF-6) #318, Record who answered from the phone's YOLO switch and paired desktops, and test a card open at a crash (CF-6) #329 and Send a question or a plan to an older paired desktop as a notice it can read (CF-6) #330, with the receipt on #279. Finding 10 waits on Make CF-3 answer delivery receipts truthful #316. Make CF-3 answer delivery receipts truthful #316 now conflicts with main in companion.ts because of Harden the event journal against failed writes, crashes and oversized frames (CF-6, journal mechanics) #309; its owner has been told.

Links

0 unresolved, across every Markdown file on cb33db5 and on Name what is still open for each fidelity child in status.md (steward cycle 15) #331's head.
The saved checker had flagged 3 inline-code examples as broken links, such as `![](diagram.png)` in notes.md and the v0.1.0 review. It now skips code spans.

Worktrees

Removed through release and remove, each with a recovery manifest: ci-macos-timeout, cf6-journal-hardening and docs-steward-20261019. docs-steward-v020-shipped was already gone; only its owner record remains.
Still mine, each for an open PR: cf6-residue (Record who answered from the phone's YOLO switch and paired desktops, and test a card open at a crash (CF-6) #329), cf6-desktop-kinds (Send a question or a plan to an older paired desktop as a notice it can read (CF-6) #330) and cf6-phone-refresh (Read the thread again when the Mac refuses an answer from the phone (CF-6) #318).

Environment

Disk: the Mac reached 122 MiB free at 19:20Z, which killed a merge guard. I removed only my own regenerable Xcode build output, about 550 MB. It's at 8.7 GiB free now. The largest folders I measured: ~/.coven 46 GiB (reported for information only, not to clear), other sessions' /private/tmp/claude-501 23 GiB, and CoreSimulator 20 GiB.
pnpm: Node v24.22.0 is now the nvm default and has no pnpm, so a bare pnpm finds Homebrew's 12.4.2 and gets ERR_PNPM_BAD_PM_VERSION. Putting ~/.nvm/versions/node/v24.18.1/bin first on PATH gives 10.34.5.

Questions for Val

[Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319 (Code Review workspace): should it go under Next or In flight? I haven't added it, because reordering Next is your call.
Merge queue: four landings by others cost Harden the event journal against failed writes, crashes and oversized frames (CF-6, journal mechanics) #309 three extra CI runs today. A merge-queue ruleset on main would end that, and the guard already supports one.
Code PRs waiting for your go-ahead: Read the thread again when the Mac refuses an answer from the phone (CF-6) #318, Record who answered from the phone's YOLO switch and paired desktops, and test a card open at a crash (CF-6) #329 and Send a question or a plan to an older paired desktop as a notice it can read (CF-6) #330. Each is behind main (Compare removal-only edits and notebook cells with the working tree (CF-5) #328), so it gets main merged in only when you approve it.

Next action

On Val's go-ahead, land #318, #329 and #330 one at a time. Take finding 10 once #316 lands, or fold it into #316 if its owner prefers. The next cycle starts from 58b468d.

Nothing was tagged, released, uploaded or flagged.

Coven-Harness: Claude Code / claude-opus-5-5

Checkpoint: cycle 16, 2026-10-19

Reviewed: main from 58b468d to de856e0 (#329, #317, #330, #318).

Landed by this session

All three are code PRs, merged through the guard on Val's go-ahead. Each squash commit has the same tree as its branch head.

PR	Merge	What
#329	`9a2e15f`	CF-6 findings 9 and 3: who answered from the phone's YOLO switch and from paired desktops, plus a test for a card left open at a crash.
#330	`9c0b063`	CF-6 finding 6: a question or plan reaches an older paired desktop as a notice. CodeQL flagged the test's raw client for turning certificate validation off; `71bae66` pins the host certificate instead. #317 landed during its CI, which cost one more run.
#318	`de856e0`	CF-6 finding 11: the iPhone app reads the thread again when the Mac refuses an answer.

The CF-6 receipt is on #279.

Since cycle 15, besides mine

Update macOS over the air: download in the background, install on quit #317 (macOS over-the-air updates) updated README, update-recovery.md and its own status.md entry. No doc outside those still says a click is required, apart from README's intended note that v0.0.12 to v0.2.0 still need one. No spec or plan covers automatic downloads.

Drift

None fixed and none left.
The CF-6 entries in "On main, unreleased" describe what landed.
"Still open" still names CF-6 correctly, since findings 4, 7 and 10 remain.

Plans and trackers

No plan or spec changed state.
[Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273: [P1][CF-1] Preserve native turn/item identity, message phases, and execution outcomes end to end #274 and [P1][CF-2] Harden Claude Stop, late-message draining, and cancelled-request cleanup on claude -p #275 are closed and ticked, and [P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276 to [P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281 are open, so the checkboxes match.
[P1][CF-6] Extend the existing event journal and paired-client replay to preserve typed conversation state #279's acceptance boxes stay unticked. They need the live-provider, Store and iPhone gates.

Links

0 unresolved, across all 110 Markdown files on a full export of de856e0.

Work by other sessions (claimed 23:26–23:29Z, left alone)

release-v0-3-0: two commits (the 0.3.0 version bump and acceptance of rc tags), no PR yet. The v0.3.0 cut owns moving "On main, unreleased" into a release section.
cr1-contract: no commits yet.
cf8-live-smoke: one untracked e2e spec.

Worktrees

cf6-desktop-kinds and cf6-phone-refresh went through release and remove, each with a recovery manifest.
cf6-residue was removed at 22:08Z by another session's prune (session:6bcb3b58-prune). Its archive bundle holds head faa69d7, the tree that merged as 9a2e15f.
This session now owns no worktrees.

Environment

Disk: 14 GiB free.

Questions for Val

Make CF-3 answer delivery receipts truthful #316 (CF-3 delivery receipts): it's a draft that conflicts with main, has the iPhone failure reported on it, and its Codex thread last wrote at 2026-10-19 03:55Z. Is it still live? CF-6 findings 4 and 7 are in it, and finding 10 waits on it. If it's abandoned, I can take finding 10 on a fresh branch from main.
[Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319 (Code Review workspace), still open: should it go under Next or In flight? Work on it may have started: cr1-contract was claimed at 23:26Z.
Merge queue, still open: every landing by someone else sends each open PR back through CI. Today that cost Harden the event journal against failed writes, crashes and oversized frames (CF-6, journal mechanics) #309 three extra runs and Send a question or a plan to an older paired desktop as a notice it can read (CF-6) #330 one.

Next action

Wait for Val's answer on #316. Once v0.3.0 is cut, check its release section against the unreleased entries it moves. The next cycle starts from de856e0.

Nothing was tagged, released, uploaded or flagged.

Coven-Harness: Claude Code / claude-opus-5-5

Checkpoint: cycle 17, 2026-10-20

Reviewed: main from de856e0 to ac7f743 (#332, #335, #333, #337, #336, #341, #339, #334, #342, #344, #345, #343, #347).

Landed by this session

PR	Merge	What
#336	`1b6f7a4`	CF-6 finding 10: a request stays open when saving its answer fails. Val merged it at 08:06Z.
#345	`7877a14`	Merge queue: `merge_group` CI runs only `unit + e2e (macOS)`, and the ruleset gains a squash queue, so a landing no longer sends every open PR back through CI.
#343	`0a67354`	Docs follow-up to #335 and #336: the phone no longer shows a withdrawn approval as declined, plus a test. This was the first PR through the queue.
#347	`ac7f743`	This cycle's docs: README says the queue needs auto-merge, and status.md lists the keyboard CI failure under Known rough edges.

List the Code Review workspace (#319) under In flight in status.md #340 ([Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319's In flight entry) is closed. Define the Code Review read contract and fixtures (CR-1) #334 added an accurate Code Review entry first.
Every squash commit above shows as verified on GitHub.

Merge queue

Live ruleset 23989202 now matches .github/rulesets/main.json. The previous version is saved.
The first enqueue failed with "Auto merge is not allowed for this repository". gh pr merge enqueues by turning auto-merge on, so the repo now has allow_auto_merge on, as OpenCoven/coven does.
Since then Say the phone no longer shows a withdrawn approval as declined #343 and Note that the merge queue needs auto-merge, and list the keyboard Jump to latest CI failure (steward cycle 17) #347 each went through one queue run, both green, with no reruns.

Since cycle 16, besides mine

v0.3.0 is tagged and published on 30ee279 (Prepare desktop 0.3.0 and Companion 0.3.0 (8) #337). Record v0.3.0 publication and advance the website fallback #339 recorded the publication and moved the unreleased entries into a Shipped row and its review.
- All 40 PRs in v0.2.0..v0.3.0 are named in the review, and Prepare desktop 0.3.0 and Companion 0.3.0 (8) #337 in the Shipped row.
- Every user-facing PR since the tag has an "On main, unreleased" entry: Show a Codex web search's query and the results Codex reported (CF-5) #344 (CF-5), Keep keyboard focus when an approval is answered, its turn folds, Stop ends it, or Jump to latest is used (CF-8) #342 (CF-8), End a stopped Codex turn's background commands, or say they may still run (CF-8 live) #341 and Leave a request open when saving its answer fails (CF-6, finding 10) #336.
- Define the Code Review read contract and fixtures (CR-1) #334 (CR-1) is under In flight, and the rest are docs or CI.
CF-3 (Record native answer delivery only after its transport handoff (CF-3, lands #316) #335), CF-4 (Preserve typed conversation fidelity and streaming reader state (CF-4, lands #295) #333), CF-5 (Show a Codex web search's query and the results Codex reported (CF-5) #344), CF-8 (Keep keyboard focus when an approval is answered, its turn folds, Stop ends it, or Jump to latest is used (CF-8) #342), CR-1 (Define the Code Review read contract and fixtures (CR-1) #334), End a stopped Codex turn's background commands, or say they may still run (CF-8 live) #341 and Run live smoke acceptance through the app: Claude passes, Codex is out of credits (CF-8) #332 each updated their own status.md entries.

Drift

Fixed in Say the phone no longer shows a withdrawn approval as declined #343: human-requests.md still described the phone's pre-Record native answer delivery only after its transport handoff (CF-3, lands #316) #335 behaviour.
Fixed in Note that the merge queue needs auto-merge, and list the keyboard Jump to latest CI failure (steward cycle 17) #347:
- README now says the merge queue needs auto-merge.
- Known rough edges now lists the keyboard CI failure (below).
None left.

Plans and trackers

[Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273: CF-3 to CF-8 are open and unticked, so the checkboxes match.
[P1][CF-6] Extend the existing event journal and paired-client replay to preserve typed conversation state #279 (CF-6): every review finding has landed. Its acceptance boxes stay unticked until the live-provider, Store and iPhone gates run.
[Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319: the checklist is unticked. [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 (CR-1) is still open although Define the Code Review read contract and fixtures (CR-1) #334 landed. Ticking and closing are the CR owner's calls.

Links

0 unresolved, across all 115 Markdown files on a full export of ac7f743.

CI: keyboard Jump to latest

e2e/keyboard-answers.spec.ts:88 failed three times in Land main through a merge queue #345's PR runs. It passed on Keep keyboard focus when an approval is answered, its turn folds, Stop ends it, or Jump to latest is used (CF-8) #342, on main, in 25 local runs and in every queue run so far.
My analysis is on Keep keyboard focus when an approval is answered, its turn folds, Stop ends it, or Jump to latest is used (CF-8) #342 (1, 2). After the jump, a "Use suggested follow-up" chip appears and the composer grows, which can bring the button back.

Work by other sessions (left alone)

cf7-verification: one commit (600d067, CF-7), uncommitted changes, no PR yet.
cr2-read-service: uncommitted changes.
cr3-inbox: Build the Code Review inbox and workspace host (CR-3) #346 is open, with a git operation in progress.
cf8-candidate: Preserve historical CF8 candidate results and keyboard review kit #338 is open.
answer-focus: Keep keyboard focus when an approval is answered, its turn folds, Stop ends it, or Jump to latest is used (CF-8) #342 merged. Removing it is the owner's job.

Worktrees

merge-queue and docs-steward-20261020 went through release and remove on squash-tree proof.
cf6-docs-followup needed one more step.
- The prune tool could not prove it merged. The queue squashed Say the phone no longer shows a withdrawn approval as declined #343 onto a newer main, so the squash tree is not the branch tip's tree. The branch also merges main, which rules out the tool's other proof.
- I checked that merge-tree(7877a14, 2e368a8) equals the squash tree (6fc493b), so nothing on the branch is missing from main.
- I bundled the tip into .git/modex-maintenance/evidence/queue-landed-2026-10-20/ and pointed the branch at 0a67354. The tool then removed it on ancestor proof.
Each removal has a recovery manifest. This session now owns no worktrees.

Environment

Disk: 3.8 GiB free (14 GiB at cycle 16).

Questions for Val

Stray remote tag untagged-292dee8871c39450142a points at 30ee279. It looks like a leftover from the v0.3.0 release. Tags belong to the release owner, so I haven't touched it. Delete it?
Keyboard Jump to latest failure: should Keep keyboard focus when an approval is answered, its turn folds, Stop ends it, or Jump to latest is used (CF-8) #342's owner fix the race, or should I take it, test first, in the scroll-follow code?
Prune tool and the queue: any PR that merges main into its branch and then lands through the queue behind another PR will hit the same proof gap as cf6-docs-followup. The tool could accept "merge-tree(squash parent, PR head) equals the squash tree" as proof. Should I open a PR for that?

Next action

Wait for Val's answers. The next cycle starts from ac7f743.

Nothing was tagged, released, uploaded or flagged.

Coven-Harness: Claude Code / claude-opus-5-5

Checkpoint: cycle 18, 2026-10-20

Reviewed: main from ac7f743 to 7c4ef2e (#350).

Landed by this session

PR	Merge	What
#350	`7c4ef2e`	The prune tool proves a queue-landed branch merged when the tip merges cleanly onto the squash's parent and gives exactly the landed tree. That was cycle 17's question 3.

Merged through the queue on Val's go-ahead. GitHub shows the squash commit as verified.
Tests: pnpm test:worktrees passed 117 of 117. pnpm build && pnpm test passed 1,900 tests with 0 failures.
Real case: on Say the phone no longer shows a withdrawn approval as declined #343's own branch and PR record, the old tool returns no proof and the new one returns a merge-tree proof.
Merge drivers: any configured merge.<name>.driver turns the proof off. Neither this repo nor the test clone configures one.
docs/worktree-hygiene.md lists the three ways a squash commit can match.

Drift

None. worktree-hygiene.md is the only doc that describes merge proof, and Prove a queue-landed branch merged when its tip merges cleanly into the landed tree #350 updated it. status.md lists no tooling changes.
The Known rough edges entry for the keyboard "Jump to latest" failure is still accurate on main. Hold Jump to latest when the composer grows or a Page Up is still moving #352 moves it to the resolved review when it lands.

Open PRs

PR	Branch	Docs it touches	Merges onto `main`
#352	`jump-latest-pin`	`status.md` −5, resolved rough edges +20	clean
#351	`cf7-verification`	`status.md`, `task-context.md`, a CF-7 plan	clean
#349	`cr2-read-service`	`status.md`, `design/code-review.md`, a CR-2 smoke review	clean
#348	`cf3-audit`	`status.md`, `human-requests.md`, `conversation-fidelity.md`, a CF-3 review	clean
#346	`cr3-inbox`	`design/code-review.md`	clean
#338	`cf8-candidate`	the CF acceptance review, a VoiceOver checklist	conflicts

The first five also merge cleanly with one another. Four of them edit status.md, in separate places.
Preserve historical CF8 candidate results and keyboard review kit #338 conflicts in docs/reviews/2026-10-18-conversation-fidelity-acceptance.md: End a stopped Codex turn's background commands, or say they may still run (CF-8 live) #341 changed that file after Preserve historical CF8 candidate results and keyboard review kit #338's last commit (03:45 CDT). The queue will drop Preserve historical CF8 candidate results and keyboard review kit #338 with merge_conflict until its owner merges main into it.

Plans and trackers

[Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273: [P1][CF-1] Preserve native turn/item identity, message phases, and execution outcomes end to end #274 and [P1][CF-2] Harden Claude Stop, late-message draining, and cancelled-request cleanup on claude -p #275 are closed and ticked, and [P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276 to [P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281 are open and unticked, so the checkboxes match. The coordinator's checkpoint 15 (20:22Z) is the latest.
[P1][CF-6] Extend the existing event journal and paired-client replay to preserve typed conversation state #279 (CF-6): every review finding has landed. Its acceptance boxes wait on the live-provider, Store and iPhone gates.
[Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319 and [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 (CR-1) are unchanged. CR-2 (Add main's Code Review read service through gh (CR-2) #349) and CR-3 (Build the Code Review inbox and workspace host (CR-3) #346) are open.

Links

0 unresolved, across all 115 Markdown files on a full export of 7c4ef2e.

Worktrees

prune-queue-proof went through release and remove on squash-tree proof, with a recovery manifest. This session now owns no worktrees.
Others, left alone: cf7-verification (Say whether a check verified a turn's last change (CF-7 verification) #351), cf8-candidate (Preserve historical CF8 candidate results and keyboard review kit #338), cr2-read-service (Add main's Code Review read service through gh (CR-2) #349), cr3-inbox (Build the Code Review inbox and workspace host (CR-3) #346), jump-latest-pin (Hold Jump to latest when the composer grows or a Page Up is still moving #352). answer-focus is gone; its owner removed it.
The primary checkout was fast-forwarded to 7c4ef2e, so scripts/worktree.sh there runs the new proof.

Environment

Disk: 856 MiB free, down from 3.8 GiB at cycle 17. At this level test runs fail.
I deleted 391 MB of my own finished scratch: a CF-6 review build tree, old link-check exports, and Land main through a merge queue #345's CI artifacts.
Largest consumers, read only:

Path	Size
`~/Library/Developer/CoreSimulator/Devices`	20 GiB
`/private/tmp/claude-501` (other sessions' scratchpads)	15 GiB
`~/Library/pnpm`	7.9 GiB
`~/Library/Caches`	5.7 GiB
modex `.git`	4.6 GiB

The largest scratchpads are two OpenCoven/chat sessions (4.0 and 2.3 GiB) and two modex sessions (1.9 and 1.5 GiB).
~/.coven is 46 GiB. It is listed for information only and is never a cleanup candidate.

Questions for Val

Disk: which of the above may be cleared? The simulator devices are the largest. I'll only touch my own scratch unless you say otherwise.
Stray remote tag untagged-292dee8871c39450142a (points at 30ee279, v0.3.0 leftover): still there. Delete it?

Cycle 17's question 2, the keyboard failure, now has an owner in #352. Question 3 landed as #350.

Next action

Wait for Val's answers. The next cycle starts from 7c4ef2e. When #352 lands, check that Known rough edges no longer lists the keyboard failure.

Nothing was tagged, released, uploaded or flagged.

Coven-Harness: Claude Code / claude-opus-5-5

Checkpoint: cycle 19, 2026-10-20

Reviewed: main from 7c4ef2e to 9e4197e (#352). #351 landed during the cycle; cycle 20 covers it.

Landed by this session

PR	Merge	What
#353	`b654734`	"On main, unreleased" gains #352's Jump to latest fix.

GitHub shows the squash commit as verified.
The queue landed List #352's Jump to latest fix under On main, unreleased (steward cycle 19) #353 behind Say whether a check verified a turn's last change (CF-7 verification) #351. The worktree was removed on the new merge-tree proof from Prove a queue-landed branch merged when its tip merges cleanly into the landed tree #350, its first real use.

Drift

Fixed in List #352's Jump to latest fix under On main, unreleased (steward cycle 19) #353. Hold Jump to latest when the composer grows or a Page Up is still moving #352 changed what readers see: after a jump, the transcript no longer stops short of the bottom when the message box grows or a Page Up is still scrolling. Hold Jump to latest when the composer grows or a Page Up is still moving #352 moved the CI rough edge to the resolved list but added no unreleased entry, so the next release review would have missed it. The new entry sits beside CF-8's keyboard entry and links the resolved list's account of the causes.
Checked on Hold Jump to latest when the composer grows or a Page Up is still moving #352: Known rough edges no longer lists the keyboard failure, and the resolved list has it with both causes. This was the check cycle 18 set.
None left.

Open PRs

Build the Code Review inbox and workspace host (CR-3) #346, Preserve native question drafts across thread navigation (CF-3) #348, Add main's Code Review read service through gh (CR-2) #349 and Say whether a check verified a turn's last change (CF-7 verification) #351 merged cleanly onto 9e4197e. List #352's Jump to latest fix under On main, unreleased (steward cycle 19) #353 merged cleanly with Preserve native question drafts across thread navigation (CF-3) #348, Add main's Code Review read service through gh (CR-2) #349 and Say whether a check verified a turn's last change (CF-7 verification) #351, which also edit status.md.
Preserve historical CF8 candidate results and keyboard review kit #338 still conflicts in docs/reviews/2026-10-18-conversation-fidelity-acceptance.md, a change from End a stopped Codex turn's background commands, or say they may still run (CF-8 live) #341. Its owner needs to merge main.

Plans and trackers

No change. [Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273: [P1][CF-1] Preserve native turn/item identity, message phases, and execution outcomes end to end #274 and [P1][CF-2] Harden Claude Stop, late-message draining, and cancelled-request cleanup on claude -p #275 are closed, and [P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276 to [P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281 are open. [P1][CF-6] Extend the existing event journal and paired-client replay to preserve typed conversation state #279, [Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319 and [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 are unchanged.

Links

0 unresolved, across all 115 Markdown files on a full export of 9e4197e.

Housekeeping done on Val's request

Stray tag: deleted untagged-292dee8871c39450142a, remote and local. No release used it. v0.3.0 (signed tag 575e514 on 30ee279) is untouched.
Simulators: ran xcrun simctl erase all, so the iPhone simulators use 122 MB instead of 20 GiB. I erased rather than deleted them: apps/ios/ModexCompanion/Scripts/test-e2e.sh defaults to the device named iPhone 16 Pro (OS 26.5). Each device's first boot will be slow.
My scratchpad: 104 old logs deleted.
Disk: 29 GiB free, up from 856 MiB at cycle 18.

Worktrees

docs-steward-c19 went through release and remove, with a recovery manifest. This session owns no worktrees.
jump-latest-pin: Hold Jump to latest when the composer grows or a Page Up is still moving #352 merged, but the worktree is unreleased. It is clean and holds unknown ignored files. It belongs to session:jump-latest-pin, so I left it.
The others (cf7-verification, cf8-candidate, cr2-read-service, cr3-inbox) belong to sessions with open or just-landed PRs.

Next action

Cycle 20 runs now, from 9e4197e through b654734, covering #351 (CF-7 verification).

Nothing was released, uploaded or flagged. The only tag change was the deletion Val asked for.

Coven-Harness: Claude Code / claude-opus-5-5

Checkpoint: cycle 20, 2026-10-21

Reviewed: main from 9e4197e to 4e9d965 (#351, #353, #356).

Landed by this session

PR	Merge	What
#356	`4e9d965`	The CF-7 verification plan is marked done in #351.

GitHub shows the squash commit as verified.

Drift

Fixed in Mark the CF-7 verification plan done in #351 (steward cycle 20) #356. Say whether a check verified a turn's last change (CF-7 verification) #351 landed every task in docs/superpowers/plans/2026-10-20-cf7-verification.md, but the plan still read as not started.
- The five boxes are ticked from Say whether a check verified a turn's last change (CF-7 verification) #351's own evidence, including the full e2e suite in CI on the PR and in the queue (run 39508701952). The local e2e run stopped when the disk filled.
- A dated status line says that linking a check to a plan step stays open on [P2][CF-7] Connect a bounded task plan to project files, verification evidence, and a saved Note #280.
Checked on Say whether a check verified a turn's last change (CF-7 verification) #351:
- Its "On main, unreleased" entry matches the code and task-context.md's new Verification section.
- "Not done yet" now names the remaining gap: step linking, the context map, read-only reopen.
- The In flight conversation-fidelity block points to tracker [Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273, so nothing there went stale.
None left.

Plans and trackers

[P2][CF-7] Connect a bounded task plan to project files, verification evidence, and a saved Note #280 (CF-7): Say whether a check verified a turn's last change (CF-7 verification) #351 meets acceptance criterion 3 ("a relevant passing check at the recorded revision can support verification…"). The box stays unticked: ticking it is the tracker session's call. List the project files a turn touched, as the CLI reported them (CF-7 context map) #355 (context map) is the next CF-7 slice.
[Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273: [P1][CF-1] Preserve native turn/item identity, message phases, and execution outcomes end to end #274 and [P1][CF-2] Harden Claude Stop, late-message draining, and cancelled-request cleanup on claude -p #275 are closed, and [P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276 to [P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281 are open, so the checkboxes match.

Open PRs

Stacks: Correct Codex question evidence and acceptance claims #358 is stacked on Record a live Codex question: Modex reads it, Codex takes the answer (CF-3) #357 (CF-3), and Invalidate retained PR data after visibility loss #354 on Add main's Code Review read service through gh (CR-2) #349 (CR-2). The steward's pairwise check shows false conflicts for stacked PRs, so those pairs were checked against their bases instead.
Build the Code Review inbox and workspace host (CR-3) #346, Preserve native question drafts across thread navigation (CF-3) #348, Add main's Code Review read service through gh (CR-2) #349, Invalidate retained PR data after visibility loss #354, List the project files a turn touched, as the CLI reported them (CF-7 context map) #355 and Record a live Codex question: Modex reads it, Codex takes the answer (CF-3) #357 merge cleanly onto 4e9d965. Once you set the stacks aside, they also merge cleanly in sequence.
Preserve native question drafts across thread navigation (CF-3) #348: for a while it conflicted in status.md, because its CF-3 entry and Say whether a check verified a turn's last change (CF-7 verification) #351's CF-7 entry were both inserted at the top of "On main, unreleased". Its owner's 00:03Z push resolved that.
Preserve historical CF8 candidate results and keyboard review kit #338 still conflicts in docs/reviews/2026-10-18-conversation-fidelity-acceptance.md, because of End a stopped Codex turn's background commands, or say they may still run (CF-8 live) #341. Its owner needs to merge main.

Links

0 unresolved, across all 116 Markdown files on a full export of 4e9d965.

Worktrees

docs-steward-c20 went through release and remove on squash-tree proof, with a recovery manifest. This session owns no worktrees.
Since cycle 19, the owners of cf7-verification and jump-latest-pin removed them.
Open sessions' worktrees: cf3-live-question, cf7-context, cf8-candidate, cr2-read-service, cr3-inbox.

Environment

Disk: 29 GiB free.

Next action

None pending from Val. The next cycle starts from 4e9d965.

Nothing was tagged, released, uploaded or flagged.

Coven-Harness: Claude Code / claude-opus-5-5

Refreshed goal prompts — 2026-10-22 (`main` at `0e674e7`)

Since cycle 20 (4e9d965): #348, #354, #355, #357–#368 landed, so cycle 21 covers 4e9d965..0e674e7. Drift I noticed and haven't verified:

docs/status.md still reads "Last updated 2026-10-20".
Its In flight entry "Code Review inbox integration repair" describes work that landed in Preserve inbox page states and integrate authenticated CR3 navigation #362/Integrate Code Review inbox and workspace with main-owned PR reads #367, and its Code Review block may be stale after Add main-owned GitHub PR reads with production IPC coverage #366/Integrate Code Review inbox and workspace with main-owned PR reads #367.
[CR-3][P1] Build the PR inbox and resizable Code Review workspace #322 was closed automatically when Integrate Code Review inbox and workspace with main-owned PR reads #367 merged and was reopened by hand. Make sure no doc says CR-3 is done.
Preserve historical CF8 candidate results and keyboard review kit #338 was closed without merging. Confirm Reconcile historical CF8 evidence and keyboard review kit with main #359/Reconcile CF8 history with current main verification records #364/Preserve historical CF8 evidence and keyboard review kit #365 carry its content, and that no doc still points to it as open.

Worktrees whose PRs landed but are still checked out: cf3-evidence-repair, cf7-path-repair, cf8-main-sync, cf8-reconcile, cr1-verify, cr2-cache-denial, cr2-ipc-acceptance, cr2-review, cr3-inbox-repair. They belong to their owners, so report them and don't remove them.

Long-running goal prompt

Act as the docs-and-plans steward for TypeSafeAI/modex (tracker #290). Goal: every doc and plan tells the truth about current main; every plan is executing, explicitly blocked, or retired with a reason; Val can plan the next session from docs/status.md alone. Each cycle: fetch main and review every commit since the last checkpoint's SHA (cycle 21 starts at 4e9d965). For each landed PR, check that docs/status.md (On main, unreleased; In flight; Known rough edges), the feature docs, the plan files under docs/superpowers/plans and the tracker checkboxes (#273, #319) match the code and evidence. Fix drift in one docs-only PR per cycle from your own worktree (scripts/worktree.sh new docs-steward-cNN). Docs PRs land through the merge guard once CI is green; code goes back to its owner. Check for broken links across all Markdown files. Report merged-but-checked-out worktrees with their owners, and release or remove only your own, with proof. Check disk space. Ask Val on this issue before restructuring status.md, moving or deleting a doc, changing a spec decision or reordering Next. Never tag, release, upload or flip flags. End every cycle with a checkpoint comment: SHAs reviewed, PRs, drift fixed and left, plan states, questions for Val, and the next action.

Follow-up / resume prompt

Resume the steward from the newest checkpoint on #290. Fetch main; the next cycle covers last-checkpoint-SHA..origin/main. Re-check every drift item marked "left" in that checkpoint before starting new work. List open PRs and check they merge pairwise (compare stacked PRs with their base). Release and remove your previous cycle's worktree if it remains. Run the cycle and post its checkpoint.

Refreshed goal prompts — 2026-10-23 (`main` at `203a350`)

On main already (landed since 0e674e7, each with its own docs/status.md entry):

CF-3 ([P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276): the Plan-mode decision and question acceptance steps (CF-3: Plan-mode decision for Val, with a live capture; question acceptance steps #369); answers to an agent's questions from the iPhone through the Mac's broker (CF-3: answer an agent's questions from the iPhone, through the Mac's broker #373); Plan mode on Codex in Codex's native plan collaboration mode (Run Plan mode on Codex in Codex's native plan collaboration mode (CF-3) #385).
CF-6 ([P1][CF-6] Extend the existing event journal and paired-client replay to preserve typed conversation state #279): questions across replay and reconnect (Questions across replay and reconnect on the phone and paired desktops (CF-6) #389); Windows and Linux journal CI, an in-app live journal check and a mid-reply kill (Windows and Linux journal CI, an in-app live journal check and a mid-reply kill, on top of #389 (CF-6) #388). [P1][CF-6] Extend the existing event journal and paired-client replay to preserve typed conversation state #279 closed on 2026-10-23 and is ticked in [Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273.
CF-7 ([P2][CF-7] Connect a bounded task plan to project files, verification evidence, and a saved Note #280): a read-only task view for saved task notes (feat: reopen a saved task note as a read-only task view (#280) #375); step linking retired (docs: CF-7 step-linking decision note for Val (#280) #370); a decision on Claude plans after TodoWrite went away (docs: CF-7 Claude plans decision after TodoWrite left (#280) #387); live Codex captures, questions asked in a reply, and ending turns Codex never ends (feat: live Codex captures, questions asked in a reply, and ending turns Codex never ends (#280) #390); a "reported no plan" footer (feat: a turn says when its CLI reported no plan (#280) #391).

Still open:

No cycle 21 checkpoint. Cycle 20 stopped at 4e9d965, so the next cycle covers 4e9d965..203a350.
Drift from the 2026-10-22 refresh, still unchecked:
- docs/status.md still reads "Last updated 2026-10-20".
- Its In flight entry "Code Review inbox integration repair" describes work that landed in Preserve inbox page states and integrate authenticated CR3 navigation #362/Integrate Code Review inbox and workspace with main-owned PR reads #367.
Checked for this refresh: no doc on main names Preserve historical CF8 candidate results and keyboard review kit #338. status.md mentions CR-3 only as a development flag and doesn't claim it is done.
[Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319 (CR tracker): [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 (CR-1) and [CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321 (CR-2) closed on 2026-10-20 and 2026-10-21, but their boxes are unticked. The boxes say "accepted", so cite the acceptance evidence before ticking them, or ask the tracker owner. CR-3 to CR-8 ([CR-3][P1] Build the PR inbox and resizable Code Review workspace #322–[CR-8][P1] Verify Code Review fidelity, accessibility, recovery and end-to-end safety #327) are open.
[Tracker][LM] Integrate local models through native harnesses with verified compatibility #376 (LM tracker): nothing has landed. LM-1 ([P1][LM-1] Define local-model contracts and prove the native Codex integration seam #377) and LM-2 ([P1][LM-2] Add secure local connections and bounded model discovery #378) to LM-8 are open. LM-2's PR LM-2: secure local connections and bounded model discovery (#378) #386 is a draft with green CI. It adds a "Local models v1" In flight entry and docs/reviews/2026-10-22-lm2-connections-discovery.md, so status.md has no LM entry until it lands. Don't add a duplicate.
Open PRs: all six edit docs/status.md, so check they merge pairwise and leave their sections alone:
- fix: make Markdown overflow keyboard accessible and enforce CF-4 budget #371: CF-4 closeout;
- Add the CR-8 acceptance matrix, Electron journeys and cross-slice regressions (#327) #372: CR-8 acceptance matrix;
- LM-2: secure local connections and bounded model discovery (#378) #386: LM-2, draft;
- docs: record Val's deferral of the CF-6 iPhone gate and #279's close #392: CF-6 iPhone-gate deferral and [P1][CF-6] Extend the existing event journal and paired-client replay to preserve typed conversation state #279's close. It rewrites the CF-6 entry's "physical iPhone gate waits for Val".
- feat(code-review): add PR overview, activity and evidence cards #393: CR-4 overview, which also adds docs/superpowers/plans/2026-10-23-cr4-overview.md;
- feat: read Claude's task tools and Codex's proposed plan from live captures (#280) #394: CF-7 Claude task tools and Codex's proposed plan, which touches the In flight entry "Reading Claude's task tools".

These worktrees are still checked out after their PRs landed: cf3-evidence-repair (#358), cf3-phone-answers (#373), cf6-replay (#389), cf7-claude-plans (#387), cf7-codex-capture (#390), cf7-no-plan (#391), cf7-path-repair (#360), cf7-reopen (#375), cf7-step-linking (#370), cf8-main-sync (#364), cf8-reconcile (#359), codex-native-plan (#385), cr2-cache-denial (#354), cr2-ipc-acceptance (#361) and cr3-inbox-repair (#362). cf7-note-reopen belongs to #374, which closed without merging. Some worktrees have commits that are on no remote branch: lm1-contract (1), lm3-isolation (1) and cr5-remote-diffs (2). These have uncommitted changes and no PR: cf8-integrated-oct22, cr5-integration, lm6-runtimes and lm-foundation. They all belong to their owners, so report them and don't remove them.

Long-running goal prompt

Act as the docs-and-plans steward for TypeSafeAI/modex (tracker #290). Goal: every doc and plan tells the truth about current main; every plan is executing, explicitly blocked, or retired with a reason; Val can plan the next session from docs/status.md alone. Read AGENTS.md, ~/.claude/CLAUDE.md, this issue's original prompts and the cycle 20 checkpoint. Each cycle: fetch main and review every commit since the last checkpoint's SHA (cycle 21 covers 4e9d965..origin/main, which includes the CF-3, CF-6 and CF-7 merges #369–#391). For each landed PR, check its claims against code on main. Then make docs/status.md (On main, unreleased; In flight; Known rough edges; Next; the Last updated line), the feature docs (human-requests.md, task-context.md, event-journal.md, design/code-review.md), the plan files under docs/superpowers/plans and the tracker checkboxes (#273 CF, #319 CR, #376 LM) match. Tick a box only on cited evidence; #319's boxes say "accepted", so closed alone is not enough. Trim In flight entries whose work has landed, such as "Code Review inbox integration repair". When #386 (LM-2) lands, check its Local models v1 entry against #376 and its review. Before editing, list open PRs (gh pr list; gh pr diff <n> --name-only) and leave alone any section an open PR changes; today that is all six of #371, #372, #386, #392, #393 and #394. Fix drift in one small docs-only PR per cycle from your own worktree (scripts/worktree.sh new docs-steward-c21); never edit the primary checkout. Sign every commit (git commit -S) and confirm it with git log --show-signature. Docs-only PRs land through the merge guard once CI is green. Code PRs and anything outside docs wait for Val's go-ahead, and code fixes go back to their owners. Check for broken links across all Markdown files on a full export. Report merged-but-checked-out worktrees and unpushed commits with their owners, and release or remove only your own, with proof (docs/worktree-hygiene.md). Check free space on the data volume. Ask Val on this issue before restructuring status.md, moving or deleting a doc, changing a spec decision or reordering Next. Never tag, release, upload or flip flags. End every cycle with a checkpoint comment: SHAs reviewed, PRs, drift fixed and left, plan states with evidence, questions for Val, and the next action.

Follow-up / resume prompt

Resume the steward from the newest checkpoint on #290 (cycle 20 at 4e9d965 if no cycle 21 exists). Fetch main; the next cycle covers last-checkpoint-SHA..origin/main. Re-check every drift item marked "left" in that checkpoint and in the 2026-10-23 refresh before starting new work. List open PRs and check they merge pairwise (compare stacked PRs with their base); for each that merged since, check its status.md entry and tracker box. Release and remove your previous cycle's worktree if it remains. Run the cycle and post its checkpoint.

Checkpoint: cycle 21, 2026-10-30

Reviewed: main from 4e9d965 (cycle 20) to f51ef7e, 66 commits, including the v0.4.0 release (tag at 2eda138, published 2026-10-30), CF-3/CF-4/CF-6/CF-7 (#369–#394, #404, #427, #443–#454), CR-4 to CR-8 (#393, #395 via #404, #396, #410, #416, #422, #426, #430, #438, #440, #441), LM-1/LM-2/LM-3/LM-5 (#397, #400, #403, #405, #418, #423, #447, #448) and the OpenCode/MiMo backends (#432). The 2026-10-22 and 2026-10-23 refreshed prompts were never run; their drift items were re-checked here.

Landed by this session

PR	Merge	What
#460	`21f6a70`	docs-only: status.md reconciled, six plan ledgers given a state line, one resolved rough edge

Drift fixed (evidence in #460's diff)

status.md "Review repair delivery": the Fail refused ACP models and recover failed initialization #436, LM-5: signed policy delivery and truthful reroute termination #424/LM-3 and LM-5: signed delivery of isolated local execution and admission policy #423 and fix: retain Code Review ownership until process exit is confirmed #426 repairs shipped in v0.4.0 (all three precede 2eda138); folded into three bullets, keeping the unexplained hosted 2019 ms median on record.
status.md In flight, "protected delivery remains pending" while merged: Canvas readiness → test: wait for native Canvas readiness before Escape #446 2ce777c; LM-5 subagent admission → fix: verify disabled subagents before local admission #448 ba8f289; CF-7 failed-command follow-up → fix: preserve possible file writes from failed commands #453 f51ef7e; CR-8 review-flow fixture → Cover integrated review submission and restart reconciliation #441 eca2bd9 (v0.4.0).
Shipped in v0.4.0 and in its review, removed from In flight: Restore local credential hardening and isolate fake process signals #418 LM-2 credential repair, Run Code Review E2E against the selected packaged app #430 packaged harness, test: sequence native Markdown keyboard scroll actions #435 Markdown sequencing. Shortened: fix: keep Notes history and editor identities distinct #429 Notes history, the e25a554 candidate (Integrate Code Review and Markdown candidates with provenance fixes #404), the CF-4 closeout ([P1][CF-4] Render typed conversation turns with safe Markdown, code controls, and stable streaming UX #277 closed 2026-10-30 with Val's written deferral of C1–C5).
"v0.3.0 after publication" retired: both gates are superseded by v0.4.0 (installed Mac reports signed 0.4.0; Companion build 9 replaces build 8), per the v0.4.0 publication record.
Next → Windows/Linux: the KWallet5 pin is on main (ci: pin the qualified KWallet5 acceptance baseline #456 727e022), no longer "pending".
Known rough edges: the iPhone secure-answer focus item landed in test: wait for secure answer field keyboard focus #455 9f85d61; moved to reviews/2026-10-18-resolved-rough-edges.md.
Drift from the 2026-10-22/23 refreshes: "Last updated" is current (2026-10-30, set by docs: record v0.4.0 publication and refresh site fallback #445); the "Code Review inbox integration repair" entry already describes the landed Preserve inbox page states and integrate authenticated CR3 navigation #362/Integrate Code Review inbox and workspace with main-owned PR reads #367/CR-3: deliver acceptance follow-up with cancellation-safe sign-in retry #438 work with only Val's gates open, so it was left; LM-2: secure local connections and bounded model discovery (#378) #386 (LM-2) landed as LM-2: main-owned local connections and bounded model discovery (#378) #403 and its Local models entry matches reviews/2026-10-22-lm2-connections-discovery.md (no duplicate).

Plan and spec states

docs/superpowers/plans/2026-10-20-cf7-context-map.md: delivered (List the project files a turn touched, as the CLI reported them (CF-7 context map) #355 3e416ab, v0.4.0; follow-ups Name commands that may have changed unlisted files, and cut opening titles at a word break (#280) #450/fix: preserve possible file writes from failed commands #453).
2026-10-23-cr4-overview.md: delivered (feat(code-review): add PR overview, activity and evidence cards #393 6c40547, integrated by Integrate Code Review and Markdown candidates with provenance fixes #404, v0.4.0).
2026-10-23-cr5-integration.md: delivered (Integrate Code Review and Markdown candidates with provenance fixes #404 e3527f4, v0.4.0). 2026-10-23-cr5-remote-diffs.md: superseded by the integration plan.
2026-10-23-cr6-isolation.md: delivered (test(code-review): establish CR-6 native isolation limits #396 03ce71a; Phase 2 Add CR-6 agent review, PR questions and review instructions (#325) #422 b643617, repair fix: retain Code Review ownership until process exit is confirmed #426; v0.4.0).
2026-10-23-cr7-intents.md: delivered (Integrate Code Review and Markdown candidates with provenance fixes #404, Bind CR-7 intents to CR-5's shared line anchors #410 deb373a, Integrate and repair CR-7 review writes and guarded actions #416 b3aad0d; v0.4.0).
The three specs' status lines (approval rules, routines, sticky Auto) still match main; the 2026-10-18 plan state lines (OAuth, extensions, knowledge, pages, roadmap, hygiene, release candidates, sticky routing, images/plugins, Mac Store preview) are unchanged and still true.

Trackers

[Tracker] Native conversation fidelity: execution → transcript → recovery → Notes #273: ticked [P1][CF-4] Render typed conversation turns with safe Markdown, code controls, and stable streaming UX #277 (CF-4, closed completed 2026-10-30 with Val's deferral) and [P1][CF-5] Add provenance-aware citations, command evidence, and owned artifact actions to chat #278 (CF-5, closed completed 2026-10-20 after Show a Codex web search's query and the results Codex reported (CF-5) #344). Open: [P1][CF-3] Add typed native questions and a main-owned, race-safe human-request lifecycle #276, [P2][CF-7] Connect a bounded task plan to project files, verification evidence, and a saved Note #280, [P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281; the seven integration criteria stay unticked (no single receipt proves them yet; CF-8 [P1][CF-8] Prove the complete conversation-to-Notes workflow with E2E, native acceptance, and rollout gates #281 owns that).
[Tracker][LM] Integrate local models through native harnesses with verified compatibility #376: ticked [P1][LM-1] Define local-model contracts and prove the native Codex integration seam #377 on Val's 2026-10-24 narrow Gate A acceptance (reviews/2026-10-24-lm1-acceptance.md; the runtime probe is deferred, not failed). [P1][LM-2] Add secure local connections and bounded model discovery #378–[P1][LM-7] Prove local-model integration and publish the supported compatibility matrix #383 stay open.
[Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319: left untouched. [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 and [CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321 are closed completed, but their boxes say "accepted" and no comment records Val's acceptance of CR-1 or CR-2 as such. Question below.

Worktrees and disk

scripts/worktree.sh audit keeps all 31. Merged but still checked out (not-proven-merged only because the audit lacks fresh merge proof): cf4-full-suite (#454), cf8-main-sync (#364), cf8-reconcile (#359), cr2-cache-denial (#354), cr2-ipc-acceptance (#361), cr3-inbox-repair (#362). Dirty, no PR of mine: cf8-integrated-oct22 (also mid-git-operation), hosted-budget, lm-foundation, lm5-local-policy, release041-validation, strata-integration. codex-native-plan has a git operation in progress. Nothing was released or removed; none is mine except docs-steward-c21, which I will remove after this merge. Data volume: 6.0 TiB free of 7.3 TiB.

Drift left, and why

status.md is still a changelog in places. The "v0.4.0 source delivery and remaining acceptance" section (about 150 lines) repeats what the v0.4.0 review holds, and the CR-6/CR-7/CR-8, Mac App Store and Companion entries under In flight carry build-by-build history. Collapsing them is a restructure, so it waits for Val (question 1).
[Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319 boxes (question 2).
"On main, unreleased" was left exactly as Prepare v0.4.1: desktop version and release review #452 edits it. Once v0.4.1 publishes, that section needs the post-v0.4.0 code deliveries that are now only under In flight (test: wait for native Canvas readiness before Escape #446, fix: verify disabled subagents before local admission #448, fix: preserve possible file writes from failed commands #453, test: wait for secure answer field keyboard focus #455, ci: pin the qualified KWallet5 acceptance baseline #456, ci: retain desktop runner memory diagnostics #458) folded into the v0.4.1 row, and the v0.4.1 review leaves In flight.

Questions for Val

May the steward collapse status.md's "v0.4.0 source delivery and remaining acceptance" section to a short list of open gates linking the v0.4.0 review, and trim the In flight histories (CR-6/7/8, Mac App Store, Companion TestFlight, Approval rules) to current state plus links? Each is a restructure of a section another session wrote.
Do [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 (CR-1) and [CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321 (CR-2) count as "accepted" for [Tracker][CR] Reconstruct the screenshot Code Review workspace end to end #319's boxes on their closing evidence ([CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320: design page and manifest, R1–R4 originals confirmed gone; [CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321: Add main's Code Review read service through gh (CR-2) #349 with the authorized read-only smoke on Define the Code Review read contract and fixtures (CR-1) #334/Send a question or a plan to an older paired desktop as a notice it can read (CF-6) #330), or does acceptance need a note from you?
Prepare v0.4.1: desktop version and release review #452 (v0.4.1) merged as 6a97d23 while this cycle ran; its tag and Release workflow are yours. Before it merged, the optional "CLI and installers (linux)" job failed at "Exercise installed Debian and extracted AppImage" on a76cd3d. That is the release owner's call, not the steward's.

Next action

After v0.4.1 (#452) publishes: cycle 22 moves the v0.4.1 release into Shipped, folds #446/#448/#453/#455/#456/#458 into its row, and re-runs the link check. If Val answers question 1, the same cycle collapses the v0.4.0 section. Stop condition holds now: no unreconciled drift remains that is not waiting on Val's answer or on #452.

Steward cycle 22 checkpoint, 2026-11-01

SHAs reviewed: f51ef7e..32f1cb4 (#452, #459, #460, #461, #462, #465, #466, #464, #463).

PR: #468, docs-only, signed, auto-merge on; it enters the merge queue when CI is green.

Drift fixed in #468:

docs/status.md's In flight entry for deferred Markdown said Val's review was pending. Defer Markdown far from the view so long threads open fast #463 merged as 32f1cb4 on 2026-11-01.
The CR-3 entry did not name CR-3 acceptance re-run on main: populated journey captures and evidence (#322) #464, which merged as 3d72c6e.
On main, unreleased now lists Defer Markdown far from the view so long threads open fast #463 and CR-3 acceptance re-run on main: populated journey captures and evidence (#322) #464 after v0.4.1.

Drift left:

v0.4.1's Shipped row still names only Prepare v0.4.1: desktop version and release review #452; folding test: wait for native Canvas readiness before Escape #446, fix: verify disabled subagents before local admission #448, test: wait for secure answer field keyboard focus #455, ci: pin the qualified KWallet5 acceptance baseline #456 and ci: retain desktop runner memory diagnostics #458 out of In flight and Next waits on Q1, because it restructures status.md.
The deferred Markdown entry stays under In flight with a landed note; moving it is part of the same Q1 restructuring.

Questions for Val (unchanged from cycle 21):

May the steward collapse the "v0.4.0 source delivery and remaining acceptance" section and trim the In flight histories?
Are [CR-1][P1] Establish the Code Review reference contract and deterministic fixtures #320 and [CR-2][P1] Add a main-owned, revision-aware GitHub PR read service #321 accepted?

Trailer rule followed: this issue's body (Coven-Harness only), which takes precedence over the harness's Co-Authored-By reminder.

Next action: after #468 merges, remove the docs-steward-c22 worktree through scripts/worktree.sh remove.

docs: bring plan and tracker-doc status lines in step with main (#290)

[Tracker] Docs and plans steward #290

Description

Goal

How this issue is used

Long-running goal prompt — start

Follow-up long-running goal prompt — resume

Activity

BunsDev commented on Oct 18, 2026

Checkpoint: cycle 1, 2026-10-18

Plan states

Drift fixed (#292)

Drift left, and why

Plan step advanced

Questions for Val

Noticed, not acted on

Next cycle's first action

BunsDev commented on Oct 18, 2026

Follow-up: cycle 1 decisions, 2026-10-18

BunsDev commented on Oct 18, 2026

Checkpoint: cycle 2, 2026-10-18

What changed outside main

Val's decisions this cycle

Drift left, and why

Unverified claims from cycle 1, now settled

Plan step advanced

Questions for Val

Next cycle's first action

13 remaining items

BunsDev commented on Oct 18, 2026

Claim: post-v0.2.0 docs cycle, 2026-10-18

BunsDev commented on Oct 19, 2026

Checkpoint: cycle 13, 2026-10-18

Correction to cycle 12

Ownership

Landed

Drift fixed

Since cycle 12, besides my PRs

Pre-landing check

Plans

Plan step

Drift left

Question for Val

Next cycle's first action

BunsDev commented on Oct 19, 2026

Checkpoint: cycle 14, 2026-10-18

Since cycle 13

Confirmed on 396b17c

Plans

Process note

Questions for Val

Next action

BunsDev commented on Oct 19, 2026

Checkpoint: cycle 15, 2026-10-19

Landed by this session

Since cycle 14, besides mine

Drift fixed

Plans and trackers

Links

Worktrees

Environment

Questions for Val

Next action

BunsDev commented on Oct 19, 2026

Checkpoint: cycle 16, 2026-10-19

Landed by this session

Since cycle 15, besides mine

Drift

Plans and trackers

Links

Work by other sessions (claimed 23:26–23:29Z, left alone)

Worktrees

Environment

Questions for Val

Next action

BunsDev commented on Oct 20, 2026

Checkpoint: cycle 17, 2026-10-20

Landed by this session

Merge queue

Since cycle 16, besides mine

Drift

Confirmed on `396b17c`

Refreshed goal prompts — 2026-10-22 (`main` at `0e674e7`)

Refreshed goal prompts — 2026-10-23 (`main` at `203a350`)