Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions .changes/verification/fix-6916-usage-pointer-a85e2c17.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Fixed

- `/verification:confirm` points its worker usage line at the section that defines the Agent result, and its model override at the subagent model section.
6 changes: 3 additions & 3 deletions plugins/verification/skills/confirm/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -176,8 +176,8 @@ Read the criterion context file for the dispatched mode, then run the flow below
- **Merged report.** The missing-entry-point check above reads the merged per-entry-point report of every worker: an entry point no worker reported holds the verdict exactly as it would in a single drive.
- **Measured usage per worker.** For each worker the report lists its entry points, its result and its usage as measured: the tokens and wall time the Agent tool's result reports for that worker, never an estimate, and never a price or cost figure.
- **Cloud host.** When this run is on a cloud execution target (`CLAUDE_CODE_REMOTE` is `true` in the session, or the stage's `execution_target` in `docs/conventions/execution-target.yaml` resolved to a cloud value), the report adds the line `live-UI step unverified on this host`.
- **Pointer**: the `CLAUDE_CODE_REMOTE` row of <https://code.claude.com/docs/en/env-vars>; for the subagent result whose usage each worker line reports, <https://code.claude.com/docs/en/sub-agents>.
- **As of**: 2026-10-04
- **Pointer**: the `CLAUDE_CODE_REMOTE` row of <https://code.claude.com/docs/en/env-vars>; for the subagent result whose usage each worker line reports, <https://code.claude.com/docs/en/agent-sdk/typescript#agent-2>.
- **As of**: 2026-10-31
- **Recheck trigger**: the env-vars row changes, or a Claude Code release changes the Agent tool result or its usage format.

**Three rules for the live drive.**
Expand All @@ -190,7 +190,7 @@ Read the criterion context file for the dispatched mode, then run the flow below
5. **Evidence collection**. Stage-1 results, E2E results, test names + assertions proving the claimed behavior. For UI changes: the UI evidence artifacts per [context/outcome.md](context/outcome.md) (pre/action/post snapshot, console, network, behavior assertion, and the render-check results: axe, geometry, pixel baseline, vision review as leads only. "screenshot looks fine" is NOT an assertion). When the plan states a measurable goal: the `/verification:measure` comparison table. Rank the test evidence by independence and label each test the producing session wrote as self-authored, per [context/outcome.md](context/outcome.md) "Evidence ranking".
6. **Report + verdict**. Emit the outcome report (the proof-level line, the worker line, the proof boxes under `strict`, intent-match table, mechanical results with every skip named and its reason, E2E + UI-evidence tables when triggered, evidence table, measurements when applicable, and the lessons the run learned, each with its scope and route in the Lessons table) and a `CONFIRMED` / `NEEDS WORK` / `NOT VERIFIED` verdict. Report template and verdict criteria in [context/outcome.md](context/outcome.md). The report also states the effort level this run used: `${CLAUDE_EFFORT}`. The skill pins `effort: high`, a verdict lane's level, and no higher; an environment variable or an effort cap can still put the run at another level, which is why the report states it (pointer: the `effort` field in <https://code.claude.com/docs/en/skills#frontmatter-reference> and the `CLAUDE_EFFORT` row in <https://code.claude.com/docs/en/skills#available-string-substitutions>; as of 2026-10-02; recheck trigger: either row is renamed or removed, or the field stops overriding the session level). We pin the model-config row the pointer here names, and the fresh-context verifier below is raised to the model that produced the work whenever that model is known and stronger than this session's (see Independence of the verdict; pointer: the `high` row of <https://code.claude.com/docs/en/model-config#choose-an-effort-level> and the advisor capability rule in <https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool#model-compatibility>; as of 2026-10-10; recheck trigger: next model release).

**Independence of the verdict.** This skill usually runs in the context that produced the changes, and that context carries the assumptions that produced any defect, converging on approval rather than detection. Stage 1's mechanical pass/fail is objective and needs no escalation, but for the Stage 2 outcome verdict on anything beyond a mechanical, behavior-preserving change, render `CONFIRMED` / `NEEDS WORK` from an agent that did NOT produce the artifact: dispatch a fresh-context verifier with the acceptance criteria and the diff, withholding your rationale so it audits the artifact and not your story. The verifier checks the provenance of every test it cites (`/testing:test-value`): a test added or changed in the diff under review is a producer claim until its expected value names an independent source. This is stricter than the `implementation:phase-verifier` agent, which reports an unsourced expected value as a finding without changing its verdict: here a criterion proved only by such a test is not COMPLETE ([context/outcome.md](context/outcome.md) "Evidence ranking"). When the work was produced on a model stronger than this session's (a security-surface phase routed to the frontier alias, say) and you know that model, pass the dispatch a per-invocation `model` at or above it; this override routes upward only and keeps the checked-work rule in the marketplace's `docs/plugin-philosophy.md` "Model tiers". When the producing model is not known, the report says the verifier's model was not matched to it. Where the outcome is high-stakes and correlated blind spots are the risk, prefer a cross-vendor advisor **when one is installed and set up**, e.g. the OpenAI Codex plugin, when its documented surface can take this artifact, invoked per its own docs, with the fresh-context same-vendor verifier above as the stated fallback, never a route to a command that may not resolve (per `docs/plugin-philosophy.md` "Fresh-eyes checkpoints" in the marketplace repository).
**Independence of the verdict.** This skill usually runs in the context that produced the changes, and that context carries the assumptions that produced any defect, converging on approval rather than detection. Stage 1's mechanical pass/fail is objective and needs no escalation, but for the Stage 2 outcome verdict on anything beyond a mechanical, behavior-preserving change, render `CONFIRMED` / `NEEDS WORK` from an agent that did NOT produce the artifact: dispatch a fresh-context verifier with the acceptance criteria and the diff, withholding your rationale so it audits the artifact and not your story. The verifier checks the provenance of every test it cites (`/testing:test-value`): a test added or changed in the diff under review is a producer claim until its expected value names an independent source. This is stricter than the `implementation:phase-verifier` agent, which reports an unsourced expected value as a finding without changing its verdict: here a criterion proved only by such a test is not COMPLETE ([context/outcome.md](context/outcome.md) "Evidence ranking"). When the work was produced on a model stronger than this session's (a security-surface phase routed to the frontier alias, say) and you know that model, pass the dispatch a per-invocation `model` at or above it (for the aliases that call accepts, see <https://code.claude.com/docs/en/sub-agents#choose-a-model>, as of 2026-10-31; recheck when that list or the resolution order changes); this override routes upward only and keeps the checked-work rule in the marketplace's `docs/plugin-philosophy.md` "Model tiers". When the producing model is not known, the report says the verifier's model was not matched to it. Where the outcome is high-stakes and correlated blind spots are the risk, prefer a cross-vendor advisor **when one is installed and set up**, e.g. the OpenAI Codex plugin, when its documented surface can take this artifact, invoked per its own docs, with the fresh-context same-vendor verifier above as the stated fallback, never a route to a command that may not resolve (per `docs/plugin-philosophy.md` "Fresh-eyes checkpoints" in the marketplace repository).

**Record the verdict for the head it judged.** Once the verdict is final (including a
`NOT VERIFIED` from the Stage 1 gate), store it for the commit it judged, so the pull request
Expand Down
Loading