Doc-alignment roster-tail campaign — durable save-point (2026-08-04 clean-stop)

[kyle-sexton]

Durable save-point for the doc-alignment roster-tail campaign (session of 2026-08-04, clean-stop sweep). A cold agent can resume from this issue alone; the machine-local .work/doc-alignment/ROSTER.md (memory-tier, never committed by policy) carries the full per-row evidence if the disk survives.

Where the campaign stands

• Roster: 242 doc rows total; 187 closed (DONE/aligned), ~34 untouched, ~15 partial, escalation markers all either resolved or carried by the open PRs below.
• All 13 Dependabot alerts remediated: chore(deps-dev): bump postcss from 8.5.18 to 8.5.25 in /plugins/knowledge/skills/course-digest/extraction #1902, chore(deps): bump undici from 7.28.0 to 7.29.0 in /plugins/ai-briefing/skills/generate/output/build #1903, chore(deps): bump fast-uri from 3.1.4 to 3.1.5 in /plugins/miro #1904, chore(deps-dev): bump the npm-minor-patch group with 2 updates #1921, chore(deps): bump fast-uri to 3.1.5 and brace-expansion to 5.0.9 #1926, chore(deps): bump ip-address from 10.2.0 to 10.4.0 in /plugins/miro #1901 merged; chore(deps): bump hono from 4.12.29 to 4.13.0 in /plugins/miro #1905 (hono 4.13.0) verified MERGE-SAFE, awaiting final merge — verifier evidence in this issue's session PRs.
• Merged today from this campaign wave: docs(knowledge): align model-matching table with dateless model-ID scheme #1928 (knowledge 0.10.22), docs(prompts): add model-deprecation trigger to loop-lane re-resolve list #1929 (loop-lane deprecation trigger), docs(claude-memory): run the C2 deletion test per line and codify C1's symptom-first diagnostic #1932 (claude-memory 0.5.4 / criteria 1.5.0).

Open PRs from this campaign (each self-describing per the repo PR body contract; independent Fable verifier verdicts already relayed on 1930/1934/1935/1937 — all PASS/SHIP)

• docs(plugin-quality): cover plugin-shipped harness config in the config lens #1930 plugin-quality 0.5.0 config lens — verifier SHIP 7/7, merge pending threads/cooldown
• docs(playbooks): re-resolve fable-5 model adaptation on Opus fallback #1931 playbooks fable-5 re-resolve on Opus fallback (row 240 chain)
• docs(claude-memory): re-align auto-memory reference facts with live memory doc #1933 claude-memory row-127 memory-doc re-alignment — HOLD (see PR comment): version collision with merged docs(claude-memory): run the C2 deletion test per line and codify C1's symptom-first diagnostic #1932; needs rebase + re-claim 0.5.5/criteria 1.5.1 preserving row-60 content, then fresh verifier on rebased diff
• docs(autonomy): document the lane-stop gate's deliberate server-managed settings exclusion #1934 autonomy 0.12.1 server-managed veto exclusion — verifier PASS 4/4, merge pending
• docs(topics): supersede channel-adapters Gap 5 — channels ship the native feed trigger #1935 channel-adapters Gap-5 supersession — verifier VERIFIED 4/4, merge pending
• docs(knowledge): add the tag-exempt class to the applicability filter #1936 knowledge 0.10.23 tag-exempt class (row 230) — verifier pending
• docs(visualization): note connector-backed live data in the artifact tier #1937 visualization 0.1.2 connector-backed artifacts (row 57) — verifier PASS 4/4, merge pending

In-flight uncommitted work (machine-local; producers killed by the 2026-08-04 session limit, resets 5:30pm ET; all re-derivable from live docs + this issue)

• D:/repos/.worktrees/row229-pricing — 3 dirty files (playbooks bump + orchestration.md pricing-literal fix), producer died writing CHANGELOG
• D:/repos/.worktrees/claude-ops-changelog-fix — 7 dirty files (claude-ops: series-agnostic changelog grep, raw-.md fetch channel, ultracode lane-effort acceptance in lanes config + launcher + test, cost-caveat line), producer died awaiting test suites
• .work/fable-5-system-card/ (memory-tier) — row-240 digest COMPLETE: 11 digests, 957/957 quotes verified, config-attribution discipline applied; producer died on final INDEX correction. PR docs(playbooks): re-resolve fable-5 model adaptation on Opus fallback #1931 carries the shipped half; the digest set is machine-local only.
• Row-230 (PR docs(knowledge): add the tag-exempt class to the applicability filter #1936 open): retag-sweep producer died mid-classification of 561 api-only rows (659 tagged rows confirmed); memory-tier sweep record incomplete.

Remaining roster rows (resume with one-agent-per-doc + independent Fable verifier, non-negotiable)

• Lane C (claude-config): 234, 154, 98, 99, 112, 113, 135, 134, 132
• Lane P (playbooks): 240 (PR docs(playbooks): re-resolve fable-5 model adaptation on Opus fallback #1931 open), 221, 235, 108, 159, 79, 80, 220, 198, 199, 200, 218, 114, 102, 54, 52, 145 (escalated: loop-lane RC requirements + boris advanced.md RC availability)
• Lane Q: 142, 139, 156, 206, 213, 238 (141 = docs(plugin-quality): cover plugin-shipped harness config in the config lens #1930)
• Lane H: 229 (in flight); 225/227 closed
• Lane O: 62/77/86 fix PR (in flight, worktree above); 130/197 closed
• Lane M: 127 (docs(claude-memory): re-align auto-memory reference facts with live memory doc #1933), then row-74 fix (claude project purge pointer into claude-memory:stateless, next patch above 0.5.4)
• Batches still open: B11 (83, 89, 96, 97, 105, 107, 111, 118, 125, 126), B13 (157, 162, 165, 169, 189), B15 (209, 210, 212, 215, 216, 217, 219, 223, 224, 226, 231, 232)

Owner-gated parked ledger (decisions only the owner can make — unchanged, batched)

1. I17–I22 gate-sweep mechanization — needs filing authorization
2. Frozen-campaign set: DOC-38 (re-parked with refreshed 4-row content for ~/.claude/docs/upstream-docs-fetch-channels-and-drift.md), CG-13 upstream doc-defect filings, CG-14 register destination, DOC-36 F02/F12 snippets
3. Interactive chezmoi apply (deployed ~/.claude/CLAUDE.md lags source)
4. Dotfiles CLAUDE.md findings: fast-mode third cache key; model-routing-lane wording (overlap fix, fable-cost reason, context-first step); advisor-substitute standing-state wording; classifier-premise wording ("both Fable 5 and Opus 5 carry classifiers, calibrated differently"); optional steering-is-cache-safe note (corroborated)
5. Upstream filings: schemastore effortLevel stale; harness-vs-platform advisor availability contradiction (Claude Platform on AWS)
6. Ops: DISK_HYGIENE_GUARD_WATCHDOG_SECONDS raise (destructive_guard 10s timeouts recur under agent parallelism); ~/.claude/keybindings.json dotfiles-tracking candidate
7. Awareness only: /team-onboarding, /powerup uncovered (enablement-only); boris distillation currency caveats (keybindings "every key" claim, /terminal-setup list)

Fix-batch candidates (no lane owns them; one batch near end of run)

1. autonomy routines.md glossary goal row "budget cap trips" vs no upstream budget cap
2. discovery research-deep "subagent cannot reach Workflow tool" — missing fork nuance
3. session-flow orchestrate "(main-session-only)" — same fork nuance

[kyle-sexton]

Clean-stop update — 2026-08-04, pausing until the 5-hour limit resets (~18:20 ET), then resuming on Opus 5

Two model-usage-limit waves (≈16:28 and ≈21:14 UTC) killed roughly 30 agents mid-run. Everything they produced is now committed and pushed; nothing of substance remains only on local disk. Session pauses here and resumes with Opus 5 after the reset.

Roster position

189 rows DONE/aligned, 33 untouched, 11 partial, 10 carrying an escalate marker (each escalation is either already routed to a lane or has an open PR below). Roster file stays memory-tier at .work/doc-alignment/ROSTER.md.

Merge queue — verified, blocked only on permission

Merging is blocked session-wide: the auto-mode permission classifier denies gh pr merge and the resolveReviewThread GraphQL mutation. Nothing about these PRs is unhealthy — they are verified, green, and waiting.

Ready to merge the moment merge permission exists (verifier verdicts already in):

PR	What	Verifier
#1930	plugin-quality 0.5.0 config lens	SHIP 7/7
#1934	autonomy 0.12.1 server-managed veto exclusion	PASS 4/4
#1935	channel-adapters Gap-5 supersession (docs-only)	VERIFIED 4/4
#1937	visualization 0.1.2 connector-backed artifacts	PASS 4/4
#1905	dependabot hono 4.13.0 (last open alert)	MERGE-SAFE 6/6
#1936	knowledge 0.10.23 tag-exempt class	PASS 3/3
#1942	playbooks 0.6.18 pricing-snapshot note (row 229)	PASS 7/7

.work/doc-alignment/MERGE-RUNBOOK-2026-08-04.md (machine-local) holds the verbatim execution sequence: the one thread-resolution mutation for #1930 (thread PRRT_kwDOTCGFQM6WZO-G, already dispositioned with evidence), merge order, per-PR hand-assert literals, roster edits, and guarded worktree removals. #1931 must merge before #1942 (playbooks version serialization).

PRs still needing a verifier pass

• docs(playbooks): re-resolve fable-5 model adaptation on Opus fallback #1931 playbooks row-240 Fable 5 system card — digest complete (11 digests, 957/957 quotes verified); its verifier was killed twice.
• docs(claude-memory): re-align auto-memory reference facts with live memory doc #1933 claude-memory row 127 — held: rebase onto main, re-claim 0.5.5 / criteria 1.5.1 preserving docs(claude-memory): run the C2 deletion test per line and codify C1's symptom-first diagnostic #1932's landed row-60 content, then verify. See the PR comment.
• docs(adr): record that plugins-reference now documents body-text placeholder substitution #1944 row-142 ADR-0007 amendment (docs-only) — real drift fixed: the ADR claimed upstream forbids ${CLAUDE_PLUGIN_ROOT} substitution in prose body text; the live substitution table refutes it.
• fix(claude-ops): series-agnostic changelog grep, raw-md channel, ultracode lane effort #1945 claude-ops 0.27.1 — series-agnostic changelog grep, raw-.md fetch channel, ultracode lane effort, cost-caveat line (rows 62/77/86).
• feat(mcp-tools): audit Claude Code _meta annotations and server-instruction truncation #1946 mcp-tools 0.2.3 — the three anthropic/* _meta annotations plus server-instruction truncation (row 125).
• docs(claude-memory): point stateless purge at the official project-purge command #1947 claude-memory 0.5.6 — claude project purge pointer in stateless (row 74); rebase after docs(claude-memory): re-align auto-memory reference facts with live memory doc #1933.
• fix(work-items): count main checkout local settings in worktree preflight coverage #1948 work-items — worktree preflight approval scoping (row 189); incomplete: version bump, CHANGELOG entry, and a test run are all still owed.
• docs(claude-config): corroborate I6 and I15 from the context-engineering blog #1949 claude-config 0.21.8 / criteria 1.16.0 — I6/I15 corroboration from the context-engineering blog (row 234).

Verified this session but not yet written into the roster

Row 225 (resources hub): verifier PASS 3/3 — hub byte-identical to the 2026-07-31 slice (MD5 be524ae2bee88758203958f215254bdb), all three sweep hits correct as-is (the Opus 5 system-card redirect's 15,994,568-byte length matches the repo's recorded figure), and /docs/for-claude still serves /llms.txt byte-for-byte so its drift watch has not fired. DOC-38 stays owner-gated and re-parked.

Remaining roster work

• Lane C (claude-config): 154, 98, 99, 112, 113, 135, 134, 132
• Lane P (playbooks): 220, 198, 199, 200, 218, 114, 102, 54, 52, and row 145 — reserved as one PR carrying the loop-lane Remote-Control requirements fix, the boris advanced.md:21 availability fix, plus two approved riders (row-221 dating refresh, row-80 link text)
• Lane Q: 139, 156, 206, 213, 238
• Lane M: 127 (via docs(claude-memory): re-align auto-memory reference facts with live memory doc #1933), then 74 (via docs(claude-memory): point stateless purge at the official project-purge command #1947)
• Lane K: row 230's memory-tier retag sweep (561 api-only rows to classify; 659 tagged rows confirmed)
• Fix batch: autonomy routines glossary budget-cap wording, discovery research-deep and session-flow orchestrate fork-nuance wording

Owner decisions waiting

1. Merge permission — the whole queue above is blocked on it. Adding Bash(gh pr merge:*) and Bash(gh api graphql:*) permission rules, or merging by hand, unblocks seven verified PRs immediately.
2. DISK_HYGIENE_GUARD_WATCHDOG_SECONDS — the disk-hygiene destructive_guard hook's 10-second watchdog timed out and denied benign Bash calls repeatedly under agent parallelism; nearly every agent reported it.
3. The parked ledger is unchanged: I17–I22 mechanization authorization, DOC-38 dotfiles landing (content refreshed and ready), CG-13 upstream filings, CG-14 register destination, DOC-36 snippets, interactive chezmoi apply, and the dotfiles CLAUDE.md wording items (fast-mode cache key, model-routing lane overlap, advisor standing-state, classifier premise).

Machine-local state that dies with this disk

• .work/ in full, including the roster, the merge runbook, the invalidation recheck, and the row-240 Fable 5 system-card digest set (11 digests, 957 verified quotes) — memory-tier by policy, never committed.
• Five pre-existing stashes in the main checkout, all predating this session and unrelated to it.
• Worktrees under D:/repos/.worktrees/ are all clean now; none is pruned, since each still backs an open PR.

[kyle-sexton]

Campaign save-point — resumed session, 2026-08-05T00:45Z

Resumed from .work/handoffs/20260805T000616Z-handoff-roster-tail-merge-queue-and-lanes.md after a session clear. This comment records what a cold agent would need that the handoff does not already carry.

Drift check: clean

origin/main is still f668526f76. All 15 campaign PRs still open, none merged out-of-band. Merge-runbook Step 0 freshness guard re-run and passes: head SHAs for #1930/#1934/#1935/#1937/#1905/#1936 all match the runbook table exactly, and the thread recount is exact — #1930 carries its one known unresolved thread PRRT_kwDOTCGFQM6WZO-G, all others zero. The runbook is execution-ready the moment authorization lands.

Merge block re-tested and still live

Both denials reproduced directly this session, from the main session:

• gh pr merge 1934 --squash --delete-branch → denied by the auto-mode permission classifier.
• gh api graphql -f query='mutation{resolveReviewThread(...)}' → denied.

New detail worth recording: gh api graphql READS are permitted — the freshness guard's thread-count queries ran fine. Only mutations are blocked. So the block is narrower than "graphql is denied", and a permission rule scoped to mutations would be enough if such scoping is possible.

Fable 5 is still exhausted — the handoff's reset estimate is wrong

The handoff predicted a reset around 18:20 ET. That time has passed and Fable still refuses. Every Fable spawn dies instantly on You've reached your Fable 5 limit.

Consequence discovered the hard way: a Fable-PINNED teammate session cannot be revived by instruction. Lane P2 died on quota, was messaged with a restart brief, and died again on the same error the instant it woke — messaging it only wakes it into another quota death. Telling such a session to "spawn subagents with opus" does not help, because the session's own model is the one being refused. Fable-pinned lanes are unrecoverable until the quota resets, full stop.

Lane P2's work is fully accounted for — nothing lost

Probed its worktrees directly rather than trusting the failure notification. P2 held the fix batch, not roster rows, and never started:

• D:/repos/.worktrees/fixbatch-autonomy on docs/fixbatch-goal-budget-cap
• D:/repos/.worktrees/fixbatch-discovery on docs/fixbatch-nonfork-workflow
• D:/repos/.worktrees/fixbatch-sessionflow on docs/fixbatch-orchestrate-nonfork

All three: zero commits ahead of origin/main, git status --porcelain --ignored empty. Clean re-dispatch, no recovery owed.

Roster integrity finding — rows 79, 159, 235 were NOT closed

Raised by lane P and confirmed by direct read of ROSTER.md. Row 79 (Code Review) is partial ?, row 159 (Create custom subagents) is untouched, row 235 (Fable field guide) is partial ?. Rows 221, 108, 80 are closed with evidence in the same file, so the roster is otherwise current — these three specifically were reported closed but never written back.

A provenance probe (PRs in all states, the campaign merged set, all 100 worktrees, unmerged remote branches) returned:

• Row 159 — DONE-BUT-UNRECORDED. Its disposition is recorded in the predecessor handoff's "Decisions already settled": review:fanout's "cannot dependably" and boris's configurable-depth doctrine are complementary readings, not a contradiction. But no verifier record exists. Ruling: a prior session's conclusion is not an independent sign-off, and the close rests on a judgment call rather than a mechanical fact — a proportional fresh-context verifier was commissioned before the close is written.
• Rows 79 and 235 — GENUINELY OPEN, never worked. Row 79 owes a live re-check of the four availability facts docs(playbooks): ship DOC bundle 3 — verification surfaces, channel authority, consistency rationale #1893 already shipped as constraints (research preview; Team/Enterprise only; unavailable under ZDR; per-repo Owner enablement) — verification, not authoring. Row 235 owes a custody record, applicability tags, and an attestation pass.
• Row 235 is boundary-parked. That work is docpage-digest profile content, which the standing rules reserve to the knowledge lane (rows 2/8/13 precedent). Routed to lane K rather than absorbed by a doc-alignment lane.

No stranded on-disk work exists for any of the three.

ROSTER.md summary counts are known-stale — do not trust them

The head-count line and the section-6 "Revised totals" both still read 1 aligned / 27 partial / 214 untouched despite dozens of rows now carrying DONE. This is by design: LANE-BRIEF-STANDING-RULES.md rule 27 says "Edit only your row's line", so closing agents deliberately leave the totals alone. Any agent needing real counts must tally the rows itself.

Two design decisions converged on independently by two lanes

Both lane C and lane Q split their lanes into read-only diagnosis first, ship second, for different reasons:

• Lane C: a PR-per-row pipeline built before diagnosis pressures each producer to manufacture a change to justify its row. Many rows in this campaign legitimately close no commit owed.
• Lane Q: the version-serialization rule says take patch+1 then rebase after it lands — and with merge blocked, that second half cannot execute, so versions must be allocated centrally from the assessors' returns instead of by five colliding producers.

Two lanes reaching the same shape from different premises is strong evidence it is correct under current conditions. Adopted. One adjustment issued: because read-only producers never write to disk, their returns are the only copy — lanes persist each verdict and proposed diff into their .work/ ledger as it arrives, so a mid-lane death costs one row rather than the whole fan-out.

Verifier tier is degraded, and it is disclosed per row

Standing rules specify a Fable fresh-context verifier. Fable is unavailable, so verifiers run on opus. Lanes record the substitution in every row's evidence cell so the tier drop is visible rather than silent. The part the owner declared NON-NEGOTIABLE — producer never self-certifies, rationale withheld, binary criteria — is intact.

Review threads: dispositioned-but-unresolvable

Since resolveReviewThread is denied, the standing procedure this session is: fix the substance, reply in-thread with the commit SHA and evidence, and record the thread as dispositioned-but-unresolvable with its thread ID in the lane ledger. That ID list is a deliverable — whoever gets merge authorization resolves them in one pass instead of rediscovering them.

New throughput blocker: the subagent concurrency cap

CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS is 20 and the pool is saturated session-wide. Lane P has 7 rows queued behind it; the orchestrator's own dispatches (fix batch, #1948 completion, the version-collision discriminator, the eight PR verifiers) are queued behind slots rather than behind any real dependency. Freed slots are reclaimed within seconds. The lever is owner-scoped and has been surfaced to the owner; no agent touches it.

Related load symptom, now corroborated in four independent lanes: the disk-hygiene destructive_guard PreToolUse watchdog (10s) expires under this parallelism and denies benign Bash calls, a bare echo has exceeded 30 seconds, and a plain grep over ROSTER.md exceeded a 180-second tool timeout. DISK_HYGIENE_GUARD_WATCHDOG_SECONDS (raisable to 50) remains parked with the owner and the evidence for it is now strong.

Owner decisions outstanding

1. Merge authorization — add Bash(gh pr merge:*) and Bash(gh api graphql:*) permission rules, or merge the six verified PRs by hand (docs(plugin-quality): cover plugin-shipped harness config in the config lens #1930 needs its one thread resolved first, then docs(autonomy): document the lane-stop gate's deliberate server-managed settings exclusion #1934, docs(topics): supersede channel-adapters Gap 5 — channels ship the native feed trigger #1935, docs(visualization): note connector-backed live data in the artifact tier #1937, chore(deps): bump hono from 4.12.29 to 4.13.0 in /plugins/miro #1905, docs(knowledge): add the tag-exempt class to the applicability filter #1936, all squash), or decline and the roster tail stops rather than producing more unmergeable PRs. Note the breadth: those rules grant merge authority over any PR the session touches and arbitrary GraphQL mutations.
2. Concurrency cap — whether to raise CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS.
3. The pre-existing parked-decisions ledger is unchanged and still owner-gated.

[kyle-sexton]

Save-point — 2026-08-05 (resumed session, post-merge wave)

MERGES LANDED (first in three sessions). origin/main f668526 -> c308e57. Merged: #1930 (plugin-quality 0.5.0), #1934 (autonomy 0.12.1), #1935 (Gap-5 supersession), #1937 (visualization 0.1.2), #1905 (hono 4.13.0, dependabot alert 20 fixed), #1944 (ADR-0007), #1936 (knowledge 0.10.23). 11/11 post-merge hand-asserts PASS; ROSTER rows 141/152/63/64 closed, row 57 fix appended; 6 merged-PR worktrees removed. Runbook Step 0 re-verified before execution; both dispositioned threads resolved.

Remediation state: #1947 round-2 done+pushed (67d4b94); #1946 fully pushed (4fc0a05); both under distinct-lens gate verification now. #1933 round-2 and #1945 round-2 uncommitted survivor work being finished in place; #1949 remediation being finished (worktree commit 8c6311f + dirty). #1948 CI now fully green; gate verify queued. Fix batch: autonomy 0.12.2 + session-flow 0.17.24 bumps pushed; discovery relabel+0.9.3 in flight. Dometrain shipped as PR #1955 (0.1.2). #1864 merge-update to 0.12.3 in flight.

Roster bookkeeping: unit-04 retag merged into decisions.tsv (44 rows content-verified, holding file deleted); closes 113/99 written; corrections 109/114/222/235/239 landed; new rows 243-246 filed (lane-C escalations + official-corroboration mischaracterization + row-227 spin-off). #1950/#1951 threads dispositioned in-reply.

Version allocation table: autonomy 0.12.1 MERGED / 0.12.2 fixbatch / 0.12.3 #1864; discovery 0.9.3; session-flow 0.17.24; dometrain 0.1.2 (#1955); claude-config 0.21.8 (#1949) / 0.21.9 (#1951); claude-memory 0.5.5 + criteria 1.5.1 (#1933); plugin-quality 0.5.0, visualization 0.1.2, knowledge 0.10.23 MERGED.

Next: gate-verify everything pushed, merge SHIP verdicts, lane successors (C/P/Q/K), invalidation shards A/B re-running, OFFICIAL-DOCS date pass at campaign end.