Dream Cycle 2026-09-15: cross-host-conformance — reconcile ADR-084 with the 09-11/12 Codex-parity hooks

[stuinfla]

Rotation

SLOT=0 (20260915 % 5 == 0) → DEEP=cross-host-conformance, SCAN=codex-parity,stranger-project-behaviour. Session commit 35fbe038afe657610d0c4924b335312aa351041c. No bonus deep-dive (20260915 % 25 = 15, % 75 = 40).

Ledger Check

docs/dream-cycle/LEDGER.md on main last updated 2026-08-31 (the "ledger row ships in the candidate PR" convention — 14 nights since, main just doesn't have their rows because none merged). Re-checked recent fates via GitHub MCP, not assumed: #142/#143, #147/#148, #149/#150 (08-19/20), #178 (08-26), #215 (08-31) — all MERGED. Nothing has merged since #215 (2026-08-31) — 15 days, zero. 12 open draft dream/* PRs as of tonight (oldest #269, 2026-09-08).

Issue Disposition

Zero open dream-cycle issues on this DEEP surface (checked: 4 open issues total repo-wide, all on brain-currency/enforcement-integrity/memory-durability — none on cross-host-conformance). PR #278 (2026-09-10) already re-audited this exact surface and found no parity gap.

Deep Dive

Since #278, real human-authored change landed directly on main (2026-09-11/12: 41bbd5fb, 76632b15, 9c45d408, 7b8e6e73, 5168f03a, ef2b8e12) extending decision-gate's write route and grounding-stamp to Codex, plus a new Stop-time "answered without searching" gate. The commits are unusually rigorous — "MEASURED, not assumed" — and self-disclose one remaining sub-gap (decision-gate's bash/exec_command route, still Claude-only) that this container's missing codex CLI could not extend tonight; re-deriving the author's own already-stated conclusion would not be new work.

What tonight's research found instead: docs/adr/0084-the-three-user-invariants.md governs: exactly those changed files and still stated the Codex-parity gap as fully open, three days later. node scripts/doc-currency.mjs --check flags this presumed-stale/BLOCK on unmodified main — a real, pre-existing, reproducible finding, not manufactured for this row.

Hypothesis (frozen before verification)

Given the 2026-09-11/12 hook changes, when the governing documentation for those files is checked against main's own doc-currency.mjs evaluator, then a reproducible gap will be found, subject to: any gap found must be reproduced against current main source, not inferred from the diff alone.

Candidate

docs/adr/0084-the-three-user-invariants.md only — no production code, one file, ~13 lines. Reconciled Invariant 2's "Known gap" paragraph with current source (partial closure via ef2b8e12, verified; the one still-open sub-gap stated accurately, not omitted) and added a Currency log row.

Evaluation Receipt

• node scripts/doc-currency.mjs --check --changed main: baseline (main@35fbe03) → adr/0084 presumed-stale · 1 BLOCK · 2 warn. Candidate (59e5315e) → current · 2 warn (BLOCK clears; remaining 2 warns pre-existing/unrelated).
• npx vitest run across 7 targeted files (codex-claude-hook-parity, grounding-turn-gate, ruvnet-gate1-pattern, hook-contracts-doctor, hook-registry-lint, hook-conformance-both-hosts, codex-lifecycle-hooks): 108 passed / 21 skipped, 0 failed.
• Full test:integration (49 files): 9 failed files / 23 failed tests, all pre-existing/environmental (native sqlite3 under this container's root user, global Ruflo is required — not installed here, cross-encoder model network download) — none touching hooks/codex/conformance code, confirmed by direct inspection.
• npm run claims:verify: 4 PASS / 3 SKIP (brain-not-installed class, consistent with every prior night).
• npm run eval:gate: EVALUATED=blocked (no brain at /root/.cache/ruvnet-brain/kb, stores 0 dark 0 — not a credentials block, OPENROUTER_API_KEY present).
• npm run wired:check: exit 0.

Darwin

Not run — no numeric benchmark axis; a doc-currency BLOCK/clear evaluator, not a metric to search.

Reward-Hack Check (self-critique — single session, not a separate agent; disclosed, not overclaimed)

No benchmark/gold data touched, no threshold changed, no cost hidden. The still-open bash-route sub-gap is stated explicitly rather than omitted, even though omitting it would let the candidate read as full closure. Verdict CLEAR.

Security Review

Out of scope — one markdown file under docs/adr/, no code/credentials/MCP/permission surface touched.

ADR

None created — this amends existing ADR-084 via its own Currency log convention, not a new decision.

Gist

LOCAL — no gh CLI or gist-creation MCP tool available this session (same limitation every night since 2026-08-19). Full report: docs/dream-cycle/2026-09-15-cross-host-conformance-report.md.

Witness

SESSION_COMMIT = 35fbe038afe657610d0c4924b335312aa351041c
REPORT_HASH    = dd77efc289b90f54d3de7eecf2064f2cd8103b1bed7177db4fadf2a5a42ee282
WITNESS        = b205a45a83eac4e2939a156e1aca23090ea51bafd7c46f56a5d0462c330a133f

5-step verifier in the committed report's own Witness section.

The standing finding this row does not fix

Zero dream-cycle PRs have merged since #215 (2026-08-31) — 15 days. 12 open draft PRs are stacked up as of tonight, spanning every DEEP surface, most self-certified ACCEPT with clean evaluation receipts. The mechanism is working exactly as designed (autoMerge: false held every night); the review step is now the bottleneck, not the research. Per the compiled routine's own STEP 1.1 learning signal (zero of the last 14 candidate PRs merged → bias to a tiny, easily-reviewable candidate), tonight's candidate was deliberately kept to one file. Worth the owner's direct attention independent of this PR's content.

Merge Policy

Human review required. This session never self-merges and never autonomously promotes candidate state. autoMerge: false held throughout.

🤖 Generated with Claude Code

https://claude.ai/code/session_0191Fjn2uKtXNgFPbyGippSV

---

Generated by Claude Code

[vercel]

Deployment failed for project explainer with the following error:

Redirect at index 0 has invalid `source` pattern "^/(?!api/|metrics)(.*)$".

Learn More: https://vercel.link/invalid-route-source-pattern

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
explainer	Ready	Preview	Oct 2, 2026 6:45pm UTC
ruvnet-brain	Ready	Preview	Oct 2, 2026 6:45pm UTC

[stuinfla]

Not this PR's failure — pre-existing on main, reproduced elsewhere.

The vercel[bot] deployment-failed comment above (explainer project: Redirect at index 0 has invalid \source` pattern "^/(?!api/|metrics)(.)$") is not caused by this PR. This PR touches only docs/adr/0084-.md, docs/dream-cycle/LEDGER.md, and the new report file — it never touches explainer/vercel.json. The identical error is already posted on [PR #290](https://github.com/stuinfla/ruvnet-brain/pull/290#issuecomment-5664707418) (a dependabot PR from 2026-09-14, also unrelated to explainer/`), confirming this is base-branch-wide, not introduced here.

Root cause (checked against Vercel's own docs, "Correcting Route Source Patterns in vercel.json"): explainer/vercel.json's redirect source is ^/(?!api/|metrics)(.*)$ — a negative lookahead followed by a separate capture group. Vercel's path-to-regexp-based redirect matcher requires the lookahead to be wrapped inside the same group as the rest of the pattern it's supposed to gate, not placed before it. The existing attempt to fix this class of issue (a9cbc939, 2026-09-11, "fix(vercel): correct regex for metrics route exclusion") didn't land the wrapped form and the deployment is still failing 4 days later.

Proposed patch (not pushed here — out of scope for a cross-host-conformance documentation-currency fix; this repo's own review discipline this cycle is "keep the candidate to what the hypothesis needs"):

-      "source": "^/(?!api/|metrics)(.*)$",
+      "source": "^/((?!api/|metrics).*)$",

This is the sibling ruvnet-brain Vercel project's own deployment that stays Ready regardless (see the same bot comment) — only the explainer preview fails, and it has been failing on every PR since at least 2026-09-14. Flagging here rather than silently re-checking, since it's real and unowned, not a flake.

---

Generated by Claude Code

---

Generated by Claude Code

[stuinfla]

Fixed: canonical-qa's release-source-identity check failed (manifest is stale; run npm run convergence:write) on 22f934f — this one was this PR's: the new report file and the ADR-084/LEDGER.md edits changed sourceIdentity without regenerating data/convergence-manifest.json. Confirmed via a worktree check that main@35fbe03 passes (ok:true) and this branch's tip did not, before the fix. Pushed 56ba183 regenerating the manifest with npm run convergence:write, matching this repo's own established convention. Re-verified locally: convergence-manifest.mjs now ok:true, doc-currency.mjs's ADR-084 finding still current, sync-version.mjs --check still agrees on 4.3.25.

The Vercel – explainer failure from the earlier comment remains unrelated/pre-existing (reproduces on #290 too) — nothing further to do there from this PR.

---

Generated by Claude Code

---

Generated by Claude Code

[stuinfla]

Merge conflict resolved. main advanced 35fbe03..c81f2b6 (13 commits, unrelated corpus-seed/ADR-0086 oracle work) and this PR went mergeable_state: dirty. Merged main in (986506f); the only real conflict was data/convergence-manifest.json — a generated file both sides regenerated independently — resolved by re-running npm run convergence:write, not by hand, matching this repo's convention. Nothing else conflicted; no file this candidate actually touches (docs/adr/0084-*.md, docs/dream-cycle/*) was in contention.

Re-verified post-merge, against the merged-in code (not assumed to still hold): doc-currency.mjs (itself updated by the merge) still reports adr/0084 → current; sync-version.mjs --check still agrees on 4.3.25; the 7 targeted hook/conformance test files still 108 passed/21 skipped, 0 failed.

---

Generated by Claude Code

---

Generated by Claude Code

[stuinfla]

Second merge-conflict resolution. main advanced again (c81f2b6..965ee55, unrelated). Same pattern as before: only data/convergence-manifest.json conflicted (both sides regenerate it independently), resolved via npm run convergence:write, not by hand. Re-verified: doc-currency.mjs's ADR-084 finding still current, sync-version.mjs still 4.3.25, the 7 targeted test files still 108 passed/21 skipped, 0 failed. Pushed d8f0d3d.

Note for visibility: main is moving fast right now (active corpus-seed/ADR-0086 work landing directly), so this draft PR may need another rebase before it's reviewed — not a problem, just flagging the pattern rather than re-explaining it each time it recurs.

---

Generated by Claude Code

---

Generated by Claude Code

[stuinfla]

Rebase note: new content drift on this PR's own target file

Mechanical rebase (af725fc6 → 8c6bb18d, main through 6e0f9623, 9 commits). Only conflict was the generated data/convergence-manifest.json, resolved via regeneration as usual.

Flagging for reviewer rather than resolving unilaterally: one of the 9 merged commits (2229745, "feat(capture): record every turn's outcome to AgentDB and distill it, on Claude and Codex") touched plugin/hooks/codex-hooks.json and plugin/scripts/continuity-hook-policy.mjs — both governed by docs/adr/0084-the-three-user-invariants.md, the exact doc this PR's candidate reconciles.

node scripts/doc-currency.mjs --check --changed main now flags adr/0084 presumed-stale again, for a reason distinct from the 09-11/12 gap this PR already fixed. doc-currency isn't wired into any .github/workflows/*.yml, so this doesn't block CI or merge — but it's new drift on this PR's own subject matter, introduced by main after this candidate's evaluation was frozen, and judging whether it actually invalidates anything in Invariant 2's text is an editorial call I'm leaving to you rather than expanding this candidate's scope to chase a moving target.

Re-verified after merge: sync-version.mjs --check (4.3.37, all surfaces agree), and the 7 targeted vitest files (109 passed / 21 skipped, 0 failed — note hook-conformance-both-hosts.test.mjs moved tests/unit/ → tests/integration/ in the merged commits).

---

Generated by Claude Code