fix(ci): green main's Lint and unblock the 0.9.14 release gates

[Hmbown]

main's Lint has been red since 04:08Z today (CI runs 35054467418, 35054908721, 35050587333) while every PR rollup looked green. The reason is that four gates are declared

continue-on-error: ${{ github.event_name == 'pull_request' }}

so they are advisory on pull requests and fatal on push. A PR can go green, merge, and turn main red on the same change.

This fixes all four.

1. Blocking-call budget — shell.rs std_fs 2 → 3

#6208's Q6 (e522e4f93c) replaced if !path.exists() with std::fs::metadata(&path) in load_default_policy. The checker counts the textual std::fs:: form, so an equivalent stat that was previously uncounted became counted.

Net blocking work went down: that same stat now carries the cache key that removed a full file read, a TOML parse and a regex recompile from every shell execution. load_default_policy is a synchronous fn, so the site cannot move to spawn_blocking without restructuring its callers — this is exactly the "raise the budget with --update if the site can only run on synchronous code" case the checker names.

2. Runtime contract budget — Act/Operate full gain two tools

#5715 (eefdcb42f507) added session_get and session_search as bounded, read-only, workspace-scoped recall. Act/Operate full catalogs go 50 → 52 tools; Plan full is unchanged because the session tools are not offered there.

--update deliberately refuses identity changes (compare() raises before the update branch), so this is the explicit maintainer edit the _comment history calls for, written in that same style.

3. Feature release-note receipts — the actual release blocker

Seven feature commits since v0.9.13 referenced six issues with no changelog receipt. This is advisory in PR CI but blocking on every release path — release.yml, release-candidate.yml, auto-tag.yml and prepare-release.sh all run it non-advisory (check-versions.sh:33-38). 0.9.14 could not have been tagged without this.

Added entries for #5715, #6099, #6035 (Added) and #6033, #6085 (Changed), and attached #6151 to the two edit-gating entries whose commits carry that reference.

4. Stale [Unreleased] compare link

It still pointed at v0.9.12. check-versions.sh only validates the compare link for the current workspace version, so this survived a release.

Verification (macOS aarch64, source 55a9e1b778fa)

[runtime-contract-budget] PASS: all 55 metrics are exactly at budget.
blocking-call budget: 626 sites across 181 files, within budget
Feature release-note receipts OK: 16 linked issue reference(s) checked in v0.9.13..HEAD.
Version state OK: workspace=0.9.13, npm=0.9.13, npm-binary=0.9.13, lockfile in sync.
Ran 19 tests in 0.015s OK   (scripts/test_check_runtime_contract_budget.py)
Ran 3 tests in 0.000s OK    (scripts/test_measure_runtime_contract.py)

Not verified here

The runtime-contract byte/token ceilings are measured from macOS. The _comment history records that macOS measurements have under-counted before and had to be re-locked from a Linux Lint run (2026-09-13, after b4d48e9a4). If Lint reports numeric drift on these values, re-lock them from that run rather than re-measuring locally — the name-set identities and digests are host-independent and will not move.

No-Issue: CI hygiene — these four gates are red on main today and none of them has a tracking issue.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[claude]

Reviewing PR #6268

• Read repo instructions (CLAUDE.md/AGENTS.md) and PR context
• Inspect diff (git diff origin/main...HEAD)
• Review each changed file for correctness, provider/route safety, reuse, tests, security
• Post final review

View job run

[claude]

Claude finished @Hmbown's task in 3m 56s —— View job

---

Reviewing PR #6268

• Read repo instructions (CLAUDE.md/AGENTS.md) and PR context
• Inspect diff (git diff origin/main...HEAD)
• Review each changed file for correctness, provider/route safety, reuse, tests, security
• Post final review

Summary

This is a CI-hygiene/budget-file PR (no Rust source changes) that raises the four continue-on-error-on-PR-but-fatal-on-push gates to match current main. I checked each claim against the actual JSON/text diffs rather than trusting the PR body narrative.

Verified correct:

• check-blocking-calls-budget.json shell.rs std_fs: 2 → 3: confirmed against scripts/check-blocking-calls-budget.py's regex (std::fs::(?:...|File|...)). shell.rs now has std::fs::File (import, line 15), std::fs::metadata (line 126, from v0.9.14: Perf quick wins — cached sort keys, zero-alloc sniffing, shared HTTP client, precompiled matchers #6208's load_default_policy stat), and std::fs::metadata (line 3547) = 3 sites. The PR's attribution is accurate.
• session_get / session_search genuinely exist (crates/tui/src/tools/session.rs:80,165), so the Act/Operate full tool-count bump (50→52) and identity hash change are real, not fabricated.
• check-versions.sh (at scripts/release/check-versions.sh) does implement check Roadmap: surface "coherence" as a plain-language session health signal #6 (compare-link match) and Roadmap: simplify DeepSeek API key setup and remove auth/provider drift #12 (issue-linked feature commit receipts) as described.
• web/lib/changelog.generated.ts changes match what web/scripts/changelog-lib.mjs's clip()/itemsPerSection logic would produce from the new CHANGELOG.md entries — item counts (Added 6→9, Changed 2→4) line up with the added bullets.

One thing worth double-checking before merge: scripts/runtime-contract-budget.json — the diff changes more than the PR body accounts for. The PR says only "Act/Operate full gain two tools" and "Plan full is unchanged," but the numeric diff shows:

• act/operate/plan active bytes: 31438 → 31453 (+15 bytes / +4 tokens) — on a tool list that gained no new tool and whose identity_sha256 is unchanged.
• plan full bytes: 50801 → 50816 (+15 bytes / +3 tokens) — also with unchanged tool count/identity.

Since identity_sha256 didn't move for these, it's plausible this is legitimate drift from whatever else landed on main between the last lock-in and 55a9e1b (this PR's own stated source commit) — not something introduced by this PR's source changes. But the new _comment paragraph added to the JSON only explains the session_get/session_search addition; it doesn't attribute this separate +15-byte/+3-4-token shift anywhere. Given this file's own established convention (every prior entry in _comment explicitly source-attributes every byte delta, per the file's history) and AGENTS.md's "misconfiguration fails loud" / explicit-maintainer-decision rule for this exact budget, it'd be worth adding a sentence naming the source of that drift (or confirming it's inherited from an already-merged PR) so a future Lint re-lock doesn't have to re-derive it from scratch.

Minor nit: both CHANGELOG.md and crates/tui/CHANGELOG.md have a stray blank line inside the ### Changed list, between the new "launch card's MCP problems row" bullet and the pre-existing "Computer Use is the only..." bullet (e.g. CHANGELOG.md:102-104). Every other list boundary in that section is a tight list (no blank line); this one blank line makes it inconsistent (cosmetic only — CommonMark still renders it as one list, just slightly looser).

No provider/model-route safety concerns (no routing code touched), no dead code/duplication introduced, and no new tests needed for a pure budget/changelog PR.

[codewhale-agent]

Codewhale review

Complete review coverage: 1/1 passes, 5 file patches, sha256:861ba649dcef5926a530f605245485feec8fcd97919900305264e73e08ea5411.

Pass 1: Metadata-only CI PR: raises crates/tui/src/tools/shell.rs's std_fs blocking-call budget 2→3, adds session_get/session_search to the Act/Operate full runtime-contract catalogs (50→52) with a new name-set digest, records five new Unreleased changelog receipts plus #6151 on two existing edit-gating entries, and fixes the stale [Unreleased] compare link (v0.9.12→v0.9.13) in CHANGELOG.md and the generated web changelog. The edited JSON is internally consistent (name lists stay sorted, declared tools counts match list lengths, tokens_est is ceil(bytes/4), Act and Operate share the new digest because their name sets are identical) and the generated TypeScript matches the generator's observed behaviour (backtick stripping, per-item clipping, itemCount deltas). The only concrete problem found is an attribution gap in the new budget comment.

Findings

• [INFO] Budget comment claims Plan full is unchanged while its byte/token ceilings (and all active ceilings) were raised (scripts/runtime-contract-budget.json:89)
  The new _comment says "Plan full is unchanged: the session tools are not offered there" (and the PR description says only Act/Operate full changed), yet the same edit raises numeric ceilings on surfaces where nothing structural moved: the active ceilings of all three modes go 31438→31453 bytes / 7860→7864 tokens (lines 89, 189, 251) and Plan full goes 50801→50816 / 12701→12704 (line 270), while every active name set, its 11-tool count, its identity cf523fcd… and Plan full's identity ac8af1f4… are untouched. A +15-byte move with no name change can only come from a shared eager-tool schema or from measurement drift, and this file's own precedent (Linux run 34766373771, "Lock ceilings to those measured values with no margin") expects the attribution to be stated. Triggering path: any later maintainer running python3 scripts/check-runtime-contract-budget.py --update/re-locking after a Lint drift reads the comment, sees 'unchanged', and cannot tell whether the extra 15 bytes were source or host measurement. Nothing here breaks the gate on its own — a numeric ceiling above the measured value still passes (decreases pass) — so the impact is an unauditable receipt, not a functional failure.

Suggestions

• scripts/runtime-contract-budget.json:89 — Extend the v0.9.14 sentence in _comment to give the +15-byte (and +4-token) movement the same attribution as the rest of the file: say which shared eager-tool schema grew, or state plainly that the active/Plan-full ceilings were captured from this run and must be re-locked from the Linux Lint run before trusting them, and drop the unqualified 'Plan full is unchanged' wording so the comment and the numbers agree.

Assessment

Pass 1: No build, script or test was run for this review, and the numeric truth of the runtime-contract budget could not be independently reproduced: the new full name-set digest 1203d192…, the declared byte/token values, and the shell.rs std_fs count of 3 are all taken from the PR's own verification. What I could check structurally holds up: the added names sit in correct sorted position in both full catalogs, the declared tools counts equal the list lengths (52/52 act and operate, 30/30 plan, 11/11 active), tokens_est equals ceil(bytes/4) for every changed metric, Act and Operate legitimately share one digest, and the two changelogs plus the generated TS keep matching ordering/counts (Added 6→9, Changed 2→4, compareUrl v0.9.13). The generated web/lib/changelog.generated.ts is consistent with derive-changelog.mjs output so far as the diff shows (backticks stripped, *emphasis* retained as in pre-existing context, items clipped at ~470 chars on a word boundary), but it is a generated artifact and was not regenerated here, so if CI runs npm run prebuild plus a diff check, that step is the thing to confirm. Open questions I could not settle from the supplied material: (1) the +15-byte ceiling raise on active/Plan-full has no stated source attribution (see the info issue); (2) the issue-to-commit mapping claimed for #6151 and the five new receipts (and whether the seven feature commits are exactly those issues) cannot be verified without the commit range; (3) crates/tui/CHANGELOG.md is one of the two files not supplied in full, so I cannot tell whether it also carries a stale [Unreleased] compare link like the one fixed in the root file; (4) the blocking-call budget relies on shell.rs now containing exactly three counted std::fs:: sites, which I could not count.

---

Advisory review by Codewhale (codewhale review --pr 6268 --post, head 13853bd47f5e35cddfcbe8d44fdc89e0b7581166). Line-specific findings are also posted as inline review comments; mechanical fixes arrive as committable suggestions you can apply from the Files tab. CODEOWNERS approval still governs merge.

[codewhale-agent]

[INFO] Budget comment claims Plan full is unchanged while its byte/token ceilings (and all active ceilings) were raised

The new _comment says "Plan full is unchanged: the session tools are not offered there" (and the PR description says only Act/Operate full changed), yet the same edit raises numeric ceilings on surfaces where nothing structural moved: the active ceilings of all three modes go 31438→31453 bytes / 7860→7864 tokens (lines 89, 189, 251) and Plan full goes 50801→50816 / 12701→12704 (line 270), while every active name set, its 11-tool count, its identity cf523fcd… and Plan full's identity ac8af1f4… are untouched. A +15-byte move with no name change can only come from a shared eager-tool schema or from measurement drift, and this file's own precedent (Linux run 34766373771, "Lock ceilings to those measured values with no margin") expects the attribution to be stated. Triggering path: any later maintainer running python3 scripts/check-runtime-contract-budget.py --update/re-locking after a Lint drift reads the comment, sees 'unchanged', and cannot tell whether the extra 15 bytes were source or host measurement. Nothing here breaks the gate on its own — a numeric ceiling above the measured value still passes (decreases pass) — so the impact is an unauditable receipt, not a functional failure.

[codewhale-agent]

Extend the v0.9.14 sentence in _comment to give the +15-byte (and +4-token) movement the same attribution as the rest of the file: say which shared eager-tool schema grew, or state plainly that the active/Plan-full ceilings were captured from this run and must be re-locked from the Linux Lint run before trusting them, and drop the unqualified 'Plan full is unchanged' wording so the comment and the numbers agree.