docs(upstream): answer "What a task costs on Opus 5.5" with routing, observability and pointer updates - #5827
Conversation
Add docs/upstream/opus-5-5-task-cost.md: every post item mapped to the official docs section it rests on, the post's figures labelled vendor-reported, 12 post-vs-docs conflicts with recheck triggers, and the stated deviations. Keep CLAUDE_CODE_SUBAGENT_MODEL_FORCE declined in claude-code.md and note that the plain variable does not move built-in Plan or Explore. Commit the approved plan and design for #5763. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add a compare action that reads two sessions from the local OTEL store: tokens by type (cache writes as their own type) split by model and effort, and api_request event cost reconciled against the claude_code.cost.usage metric, flagging an event undercount. A store with non-delta metrics exits 2 instead of reconciling. cc-otel.sql is unchanged; effort and temporality are read from raw attributes. Refs #5763 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…lementer Add a Model column (sonnet | opus | frontier) to the plan template's per-phase routing table and a scoped-implementer agent (sonnet, medium effort). implement-dispatch spawns it with an explicit model: sonnet for sonnet rows, falls back to implementer when a non-Anthropic provider would resolve the sonnet alias to an older model, and keeps implementer for unrouted, opus, unknown and frontier rows. Both agents carry the worker contract inline; a sync test fails when the copies diverge. The work skill's fix worker and the loop-lane tier roles say the same. Refs #5763 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…gaps Row 7 now says only accepted claims must be HIGH confidence; a MEDIUM or LOW claim listed in the Gaps section is not accepted, so an artifact of honestly listed vendor-only claims passes. The verifier, discipline 11 and the phases context say the same. Four wording trims keep the effort-ceiling sentence inside the re-attach slice the contract test asserts. Refs #5763 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Link the Claude Code prompt-caching page and the costs doc's "Why usage climbs in a long session" from the prompt-caching chapter and draft-goal-condition; name /doctor prompt-audit for Claude Code configuration beside /claude-api prompt-audit in audit-instructions, audit-pass and the bundled-claude-api reference. Add a path-scoped cost-claims rule: cost claims point at the costs and pricing docs, and only docs/upstream records list vendor figures. Instruction-placement discovery skips eval fixture trees, so the AGENTS.md rules index and the glob gate no longer pick up fixture rules. Refs #5763 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…onnet 5.5 effort guidance Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-philosophy and the loop-lane profile Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…sts as correlates Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ded-thinking section Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…he task-cost record Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…records Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Claude finished @kyle-sexton's task in 2m 42s —— View job Review completeScoped to the diff against Findings1. Stale agent-count enumeration in docs/plugin-philosophy.md#L1379 states:
This PR adds Fix: update the sentence to "Five pin What checked out
No other correctness issues found in the reviewed diff. The author-claimed verification results in the PR description (test pass counts, anchor checks, etc.) are reported as claimed, not independently re-run here. |
|
Claude finished @kyle-sexton's task in 1m 41s —— View job Security review complete
No security issues found. This PR is almost entirely documentation and agent/skill prompt content (routing tables, pointer updates, changelogs, plan templates) plus two new shell/SQL additions. I reviewed the security-relevant surface specifically:
|
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6b66e32eea
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…enter's effort pin A session with token points but no claude_code.cost.usage point counted its token points as metric points, so it reconciled as "events exceed metric" instead of exiting 2. Count cost points only, with a regression case. plugin-philosophy's medium-effort list now names scoped-implementer (five). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 1m 43s —— View job Security review complete
No security issues found. Reviewed the diff against What I looked at:
|
|
Claude finished @kyle-sexton's task in 2m 10s —— View job Review completeRan
FindingsNone. No inline comments posted. Both issues raised by the prior reviews on commit
I independently re-checked the routing logic this PR introduces, since it's the highest-risk surface:
I did not re-run the test suites or |
…y turn cost (#5834) No related issue: follow-ups found while landing #5827. The two findings that need a design decision are filed as #5832 and #5833. ## Summary - `harness-config:audit-instructions` said that `/doctor prompt-audit`'s write posture was undocumented. The upstream memory page now documents it. - `fleet:reach`'s relay probe table stated a per-turn dollar figure. It now links the official cost-tracking docs instead. ## Fix - `plugins/harness-config/skills/audit-instructions/SKILL.md`, `doctor` Boundary section: - The mutation gate states our decision: applying the proposed edits is the person's call, and the skill never chains into `/doctor`. - The availability line notes that the subcommand needs the bundled `claude-api` skill to be on. SKILL.md stays at 499 lines. - `reference/native-doctor.md`: the "undocumented" row is replaced by a links-only record. It points to <https://code.claude.com/docs/en/memory#audit-your-instruction-files>, carries the as-of date 2026-10-02, and gives a recheck trigger. - `plugins/fleet/skills/reach/reference/relay.md`: the Cost row keeps what the cost went to (SessionStart hooks and context loading) and links <https://code.claude.com/docs/en/costs#track-your-costs>. - Version bumps with CHANGELOG entries: harness-config to 1.3.2 (#5827 holds 1.3.1) and fleet to 0.2.2. ## Verification - `scripts/check-changelog-parity.sh --check`, `--check-bump origin/main`, `--check-order` and `--check-preserved origin/main` all pass. - `scripts/check-changed-skills.sh origin/main`: 2 skills, 0 errors. The one warning is pre-existing, in the untouched `reference/criteria.md`. - markdownlint-cli2 reports 0 issues on the 5 touched markdown files. `typos` is clean. - `run-plugin-tests.sh` passes on the 6 audit-instructions suites plus `check-doc-citations` and `check-structure`. ## Related - #5827: the session that found these. - #5832: the mod-authoring rule glob matches no file. - #5833: the discovery research gate passes an artifact with zero accepted claims. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Renumber this branch's plugin versions above main's: harness-config 1.3.3, harness-ops 2.3.0, instruction-placement 0.19.1, knowledge 0.15.2, planning 0.61.0, work-items 0.44.2. This branch's CHANGELOG entry stays first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 2m 54s —— View job Code review complete — no new findings
Skip-gate hit: this head already has a successful review from this lane that still applies.
That |
|
Claude finished @kyle-sexton's task in 1m 47s —— View job Security review complete
No security issues found. HEAD Re-verified the security-relevant surface still holds at this HEAD:
No changes touch GitHub Actions workflows, secrets handling, auth, or untrusted external input. |
Main took harness-ops 2.3.0, so this branch's entry moves to 2.4.0 and stays first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 1m 45s —— View job Security review complete
No security issues found. Skip-gate hit: this head already has a successful security review that still applies.
All 11 files are either entirely outside this PR's own change set ( Re-confirmed the security-relevant surface from the prior passes is untouched at this HEAD (
No changes touch GitHub Actions workflows, secrets handling, auth, or untrusted external input. |
|
Claude finished @kyle-sexton's task in 2m 19s —— View job Code review complete — skip gate hit, no new findings
Skip gate: this head already has a successful review from this lane that still applies.
Re-confirmed the two fixes from the prior review passes (at
No inline comments posted — there is no new or changed content on this head to anchor findings |
|
Claude finished @kyle-sexton's task in 49s —— View job Code review complete — skip gate hit, no new findings
Skip gate: this head already has a successful review from this lane that still applies.
I diffed Re-confirmed the two previously-fixed items are still intact at this HEAD:
No inline comments posted — there is no new or changed content on this head to anchor findings to. Note: this sandbox's working tree has |
|
Claude finished @kyle-sexton's task in 51s —— View job Security review complete
No security issues found. Skip-gate hit: this head already has a successful security review that still applies.
Every one of these files belongs to Re-confirmed the security-relevant surface from the prior passes is still intact at this HEAD:
Note: this sandbox's working tree shows No changes touch GitHub Actions workflows, secrets handling, auth, or untrusted external input. |
…5895) No related issue: adopts the claude.dev eval-design and hillclimbing post across six plugins; follow-ups are #5878 and #5879. ## Summary Applies what this repository takes from the eval-design and hillclimbing post, with every eval score taken only from a run the new validity gate calls VALID. One commit per plugin: - **evals 0.4.0**: noise report, run-validity gate, judge calibration, offline grader tests, routing by repository kind, a 31-case suite with labelled samples for every `llm` grader, and six settings. - **skill-quality 0.26.0**: hard-case reasons in the evals schema and trigger-noise reporting. - **discovery 0.26.0**: a first-party content claim with one publisher passes the research gate as `HIGH (single source)`, with the flag kept visible. - **harness-config 1.3.4**: the `audit-instructions` records for the bundled `claude-api` and `doctor` skills are links-only. - **context-budget 0.7.4**: skill pruning routes to `/skill-doctor`, with its boundary recorded. - **playbooks 0.17.2**: skill-authoring says how to carry an eval failure back into a skill, and allows an old-patterns names table. ## Fix - `plugin-eval` now runs three scripts before a score counts. `run-validity.py` prints VALID or INVALID from the result and the kept traces. `noise-report.py` prints intervals and a paired with-versus-without delta. `calibrate-judge.py` scores each `llm` judge against labelled samples and fails it below 90% agreement. - The suite grew from 3 to 30 cases. Each hard case says why it is hard, and four controls must not invoke an evals skill. - Records that point at Anthropic pages state our decision, then a pointer, an as-of date and a recheck trigger. They no longer restate the upstream text. - `.claude/rules/cost-claims.md` (added by #5827) gains a second exception. A skill that prices its own runs before spending may state run costs it measured on its own suite, each dated, with the version, the setup and a recheck trigger. `plugin-eval`'s sizing anchors are that case. **Reviewer: please confirm this fits the rule's intent.** Its author's session has closed. - The evals options follow the plugin option naming convention: noun-phrase boolean titles, and `options` pickers for the three fixed-set strings. ## Verification - Before the rebase onto main (job 11, 72507f118): **VALID**. With the plugin 0.99 (95% CI 0.97 to 1.00), without 0.31 (0.16 to 0.47), delta **+0.68 (+0.52 to +0.83)**. The three remaining misses are judge errors on calibrated graders (follow-up #5879). - Every `llm` grader is at or above 90% agreement with its labels. - The earlier tip (276c5fd), full suite, 3 runs, sonnet judge: **VALID**. With the plugin 1.00 (95% CI 0.99 to 1.00), without 0.29 (0.14 to 0.44), delta **+0.71 (+0.56 to +0.85)**; 30 of 31 cases at 1.00. The one miss is the recurring judge error in #5879. An earlier run the same day was INVALID: every no-plugin run in one case hit a sandbox-probe denial, and one with-plugin run did not fire the skill. Neither recurred. - Compressing `plugin-eval/SKILL.md` for cost (two `/claude-api hillclimb` iterations) was **not adopted**. The second, run beside that same-day control, was VALID and cut agent cost per request 11.1%. It dropped two explanatory clauses that graders depend on, though, and two cases regressed (with-arm 0.97 against 0.99). A version with both clauses restored is follow-up #5956. - This tip (50c2769), after the respelling and review fixes below: the 13 `llm` graders whose samples were respelled recalibrate at 100% agreement. Full suite, 3 runs, sonnet judge: **VALID**. With the plugin 0.98 (95% CI 0.94 to 1.00), without 0.28 (0.13 to 0.44), delta **+0.69 (+0.54 to +0.84)**; 30 of 31 cases at 1.00. The one miss is `measurable-criterion` (0.33): in two of three runs the reply set a target without stating its basis. These commits changed only that case's calibration samples, which no run reads, so the miss is model variance, not a regression. Later merges from main change nothing under `plugins/evals`. - Wording changes after job 11: option titles and descriptions, three links-only rewordings, one caveat clause, two line merges in `plugin-eval/SKILL.md`. None changes a rubric, a sample or a routing rule. - A security review of this diff found two issues in scripts it adds, both fixed with tests in 6ca9f6a. `calibrate-judge.py build` could write outside `--out` through a crafted `case.yaml` grader name, and `render-review.py` Markdown cells left link and image syntax live. Neither script runs inside an eval, so the eval result above still applies. - After the ready flip, CI's typos (en-us) and machine-specific-paths checks failed on this PR's eval files. 8c792eb respells them to American English, renames the `expected-behaviour` grader to `expected-behavior`, and swaps `/home/dev/` example paths for `/srv/`. No label, expected value or rubric criterion changed. A fresh-context verifier confirmed that. - Review fixes. 8055109: `calibrate-judge.py` no longer counts a run it could not show reproduced the labeled sample toward agreement (Codex P1, security review), and the `md_cell` docstring now states the bare-URL contract. 1479903: `measure-invocation.sh compare` builds its interval from paired per-probe deltas (Codex P2). Both have tests. - Local checks on the merged head (after merging main): changelog parity (check, order, bump), `validate-plugin-contracts.mjs`, `check-changed-skills.sh origin/main`, `render-index.sh check`, `overlap.py generate --check`, `check-spoke-plugin-root.sh --check`, `validate-cases.py` (0 FAIL), the affected shell suites, and every Python test under `plugins/evals`. The only failures are outside this diff: htmlhint is not installed locally, and two hook timing tests (guardrails, pr-body-linkage-gate) went over their 8000 ms budget under load. ## Related - #5878: the eval sandbox reads only a skill's hub `SKILL.md`. - #5879: calibration gaps for the two judges that missed in job 11, and the must-fail samples the agent will not reproduce. - #5956: the v2 `plugin-eval/SKILL.md` compression with the two clauses restored. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…review, session-flow and telemetry (#5920) No related issue: first of two PRs carrying this repo's share of the decisions from digesting the "Spending your effort" post (https://claude.dev/blog/spending-your-effort); the second PR follows once this one merges. ## Summary Make effort an explicit, recorded choice across lanes, review fan-out, session-flow advice and telemetry, and add a check that flags when the upstream effort guidance changes. - **harness-ops 3.0.0 (breaking):** the lane launcher refuses a lane whose config names no `effort`, always passes `--effort`, and warns once per run when `CLAUDE_CODE_EFFORT_LEVEL` may override lane levels and agent pins. Migration: add `lanes[].effort` to every lane in `lanes.json`. The session event log records the effort level (`n/a` on events that never carry one, `unset` when missing) and an allowlist of documented non-content hook input fields, read only from the payload's top level. - **harness-config:** new `check-effort-pins.sh` drift check (audit `effort-pins` scope, also run by the changelog skill) hashes model-config's effort sections against a committed baseline and flags pins in agents, skills, lane configs and Workflow scripts without editing them. unhobble records the session effort. - **knowledge:** `check-html-rows.py`, a standing gate for quotes taken from a page's HTML, with a synthetic negative-control suite. - **review:** fanout's generic slice and extract agents run at an explicit effort; named agents keep their pins; the report shows each leaf as `label@level` (the level requested). - **session-flow:** workflow gives per-stage effort advice by reading model-config's live table and never sets effort; continue-in-background picks the resumed task's level from that table and passes `--effort`. - **songwriting:** object-writer pinned at `medium`, provisional until an eval compares levels. - **work-items, source-control:** the work-loop and babysit-loop state blocks record the effort a lane actually ran at. ## Fix One commit per area, a links-only conformance pass (upstream specifics are pointed at, never restated, per `.claude/rules/skill-bodies-state-current-rules.md`), and the version bumps and CHANGELOG entries as the last commit. ## Verification - Each phase was checked by a fresh-context phase verifier against the plan's acceptance criteria (all PASS after fixes). - Suites: `check-effort-pins.test.sh` (67 checks), `lane-launcher.test.sh` (232 cases), `session-event-log.test.sh` (139), `check-html-rows.test.sh` (18), plus the reader suites for the event log; spawn ratchets unchanged. - `scripts/validate-plugins.sh`, `scripts/check-changelog-parity.sh --check` and `--check-bump origin/main`, and `scripts/check-changed-skills.sh origin/main` pass. - `scripts/run-plugin-tests.sh`: every suite for the changed plugins passes. Two failures outside this branch's files: `plugins/code-tidying/scripts/evals-fixtures.test.sh` (fails the same way on main, missing tooling) and a wall-clock budget in `plugins/guardrails/hooks/block-root-delete-target.test.sh` that overran under a machine load average of ~30; this branch changes nothing under `plugins/guardrails`. ## Related - #5767 and #5827 (merged peer work this builds on) - #5682 (replay-sweep scope approved alongside this work) 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
… OTEL columns, with the upstream record (#5992) No related issue: second of two PRs carrying this repo's share of the decisions from digesting the "Spending your effort" post (https://claude.dev/blog/spending-your-effort); the first was [#5920](#5920). ## Summary Finish the effort work that had to wait for #5920 and its two sibling PRs ([#5767](#5767), [#5827](#5827)) to merge, and record every decision the post drove. - **planning, session-flow:** the interview's session-config recommendation names an implement level and a verify level, each from model-config's effort table, neither below `medium`, and no level when the page cannot be read. Orchestrate's imperative 7 keeps code-changing, verifying and likely-edge-case work off the lower effort. - **harness-ops:** an opt-in `session_event_log_content` (default off) adds the payload's top-level content strings to event-log rows, cut at 64 KB with `<key>_truncated` or a row-level `content_truncated` marker. Rows over 4000 bytes now append under an exclusive-create lock: bash writes long lines in 4 KB pieces, so concurrent long appends interleaved (15 to 18 of 33 parallel 64 KB appends corrupt before, 0 after). Short rows keep the single unlocked write. The OTEL store exposes every attribute the monitoring page documents as a typed column, effort included; cold macros union a typed stub with older Parquet by name. A new `CC_OTEL_COLD_KEEP_CONTENT` (default keep, `=0` scrubs) covers content columns beyond user prompts, which `CC_OTEL_COLD_KEEP_USER_PROMPTS` still governs; `session-compare` reports a missing level as `unset`. - **harness-config:** the audit checklist runs the effort-pin drift check. - **implementation:** the upward-only binding now covers Workflow `effort` and `model` for the implementer and phase verifier. Pins stay as #5885 set them (implementer `medium`, phase-verifier `high`). - **knowledge:** map-corpus's effort gotcha points at the Workflow probe record in `docs/plugin-philosophy.md`. - **playbooks:** prompt-caching's effort note points at Claude Code's own cache page. - **docs:** `plugin-philosophy.md`'s low-effort sweep rule also excludes likely-edge-case work; `docs/upstream/claude-dev-spending-your-effort-blog.md` records each decision and where it landed. ## Fix One commit per area, then version bumps (harness-config 1.6.0, harness-ops 3.3.0, implementation 0.21.0, knowledge 0.18.0, planning 0.63.0, playbooks 0.17.5, session-flow 0.47.0, each with a CHANGELOG entry), with the upstream record last. No breaking change. Deliberate departures from the plan: the implementer keeps `effort: medium` as [#5885](#5885) set it (the plan predates it); the event-log lock and `content_truncated` marker were added after a fresh verifier found interleaved long rows and a silently dropped cut field. ## Verification - Two fresh-context phase verifiers checked every phase against the plan's acceptance criteria on the committed diff: 23/27 then 22/22, with the one failure fixed and two criteria superseded (#5885) or stale (three newer workflow-only agents leave effort unpinned on purpose). - `scripts/run-plugin-tests.sh --jobs 4`: all suites pass except `plugins/code-tidying/scripts/evals-fixtures.test.sh`, which fails the same way on main (missing tooling); this PR does not touch code-tidying. - Suites touched here: `session-event-log.test.sh` 168/168, `cc-otel.test.sh` 22/22, `prune-otel-store.test.sh` 355 passed, `session-compare.test.sh` 38/38, `claude-observability.test.sh` 73/73, `check-effort-pins.test.sh` 69/69, `interview-defenses.test.sh` 172/172, `agent-contract-sync.test.sh` 5/5; spawn ratchets unchanged. - `measure-hook-log-budget.sh --samples 5`: 0 corrupt lines at 4, 16 and 64 KB on two runs (15 to 18 of 33 at 64 KB before the lock). - `scripts/validate-plugins.sh` and `scripts/check-changelog-parity.sh --check-bump origin/main` pass. ## Related - [#5920](#5920): first PR of this pair. - [#5767](#5767), [#5827](#5827): sibling PRs this one waited on. - [#5885](#5885): set the implementation agents' pins this PR keeps. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
1 participant
Closes #5763
Summary
This PR applies the claude.dev post What a task costs on Opus 5.5 to the repo. The changes come from a verified digest of the post and a full planning interview. The repo acts on the official docs each item rests on, not on the post itself.
docs/upstream/opus-5-5-task-cost.mdmaps every item in the post to its docs section and records the decision made for it.Fix
docs/upstream/opus-5-5-task-cost.md: maps every item in the post to a docs section. The post's figures appear only in the record and are labelled vendor-reported..claude/rules/cost-claims.md: guidance links the costs and pricing docs and states no prices or per-task figures./doctor prompt-auditand defer to it where the two overlap. The 12 overlapping criteria rows stay, because the two checks complement each other (Q42).prompt-caching.mdanddraft-goal-conditionlink costs#why-usage-climbs-in-a-long-session instead of repeating it.Modelcolumn.sonnetruns on the newimplementation:scoped-implementer(sonnet, effort medium), spawned with an explicitmodel: sonnet. If the provider has no Sonnet alias, it falls back toimplementation:implementer.implementation:implementer(opus, high).docs/plugin-philosophy.md, the loop-lane README and the loop-lane prompts state the split.harness-ops:observability comparecompares two sessions' tokens by type, split by model and effort, and reconciles them against the cost metric (TDD,session-compare.test.sh).Two decisions differ from the docs or from another session, and both are logged in the deviation record:
scoped-implementerruns at medium effort, not high. This follows the Sonnet 5.5 effort guidance for well-specified agentic coding. The "Spending Your Effort" session's user adopted this design (its Q59).Verification
Results on
b1218703b, with origin/main merged:check-changelog-parity.sh:--check,--check-bump origin/main,--check-orderand--check-preserved origin/maineach pass.validate-plugin-contracts.test.sh: 138 passed, 0 failed. It includes the gate that every shipped agent setsmodel:.agent-contract-sync.test.sh: 5 passed, 0 failed. It checks thatimplementerandscoped-implementercarry the same contract, skills and tools.session-compare.test.sh: 36 passed, 0 failed.plugins/discovery/scripts/contract.test.sh: passes, including the 20000-byte re-attach slice.check-skill: passes for implementation, harness-config, harness-ops and planning. audit-instructionsSKILL.mdis 499 lines and audit-pass is 497, under the 500 cap.affected-tests.sh(after the first main merge): every shell suite passes, and the 8 delegated Python suites pass (989 passed, 1 skipped).check-script-contract.test.shcould not run locally because htmlhint is not installed.implement-dispatcheval id 13 covers the three routes, including the explicitmodel: sonnet. The route has not run live yet, because the installed plugin predatesscoped-implementer.Known gaps that this branch doesn't cause:
.claude/rules/mod-authoring.md, whoseplugins/*/types/**glob matches no files.The plan and the deviation log are on #5763. The removed
docs/topics/slice is still in history at1dc4ccceb.Related
effort: high(Q9) and the implementer pin-note edits (Q28, Q29, Q48).🤖 Generated with Claude Code