Repository navigation
feat(harness-config): run audit-pass and audit-instructions across several repositories via the target-scope contract - #17549
Conversation
Lane-Activity: work-resolve-issue/resolve-issue
Lane run report
Failed GitHub API calls
|
|
claude-security-review needs no new review of Security review: PR #17549I found no security issues. I read every hunk of all 12 changed files. None of them matches a What changed. Paths checked:
Spend. Multi-repository runs mean more subagent fan-out per invocation, which spends subscription or API usage only. This is not a hold. critical-findings: 0 |
| - **A delegate that accepts a target is given that repository**: | ||
| `/harness-config:audit-instructions` and `/harness-memory:audit` both take `--repo <clone root>`. | ||
| A delegate that takes no target (`/harness-config:audit`, `audit-permission-state`, | ||
| `audit-permission-grants`, `audit-prompting-postures`) reads the active project only. It runs | ||
| only in the active project's pass, when that project is selected, and every other repository's | ||
| pass lists its lane in `skipped` with the reason `delegate takes no target`. Running it there | ||
| would attribute the active project's findings to the wrong repository, as | ||
| [arguments.md](arguments.md) explains. The retirements sweep takes `--root`, so it runs in every pass | ||
| with `--root <clone root>`, and the same leftover in several repositories is one row listing |
There was a problem hiding this comment.
The Phase 4 /doctor handoff has no routing for a pass outside the active project.
This bullet routes every Phase 3 delegate by whether it takes a target. Phase 4 (SKILL.md "The /doctor handoff", reference/doctor-handoff.md) is not covered:
- Each repository's pass runs "Phase 0 to Phase 6" (line 9), so each one opens a
delegatedlane, writes anopenterminator and tells the operator to run/doctorand/doctor prompt-audit. - Both commands are operator-run in the invoking session, so they read the active project's
CLAUDE.md, not the clone's. - When the operator hands the output back through
--resumeon, say, thewebpass, it lands inweb'sdelegatedsection. That is the active project's findings attributed to another repository, the case lines 18 to 20 exist to prevent. - Each non-active pass also keeps an
openlane the operator cannot close from this session.
To reproduce: run /harness-config:audit-pass --repo web from a session in api. web's report says to run /doctor and come back with --resume, and doing that records api's trim findings under web.
Fix: add a bullet saying the /doctor handoff runs only in the active project's pass. Every other pass lists it in skipped with a reason, for example "operator handoff reads the invoking session's project; open that repository to run it".
claude-lane: claude-review
| merges into one row. | ||
| - **Run state and the lock are per repository**: run `run-state.sh paths` with | ||
| `--root <clone root>`, so each repository gets its own `<state-key>`, run directory, lease, lock |
There was a problem hiding this comment.
Closing-issue criteria are met in the text, but no eval or test could fail if they broke.
The criteria are #17245's "Any per-repository state (ledger, cursor, findings file, delta state) is kept per repository" and "Each listed skill accepts the target forms by citing the contract". The PR body confirms: "No eval case reads the new multi-repository spoke files."
The only changed eval is audit-pass eval 30, which covers the non-git refusal. Nothing covers:
--reporouting to--root <clone root>forrun-state.sh paths(this file, lines 25 to 27);- the same routing for
state-key.shandpersist-report.sh(audit-instructions/context/multi-repository.mdlines 17 to 19); - the
delegate takes no targetskip (this file, lines 16 to 20;arguments.mdlines 26 to 28).
If any of these regressed, every eval would still pass. For example, an edit that drops --root, so every repository shares the invoking project's state key and lock.
Fix: add one eval to each skill. Prompt with --repo <other clone> and expect:
run-state.sh paths(and, for audit-instructions,state-key.shandpersist-report.sh) called with that clone's--root;- for audit-pass, the four no-target lanes listed in
skippedwithdelegate takes no target.
claude-lane: claude-review
|
claude-review reviewed Code review: PR #17549 (audit-pass and audit-instructions across several repositories)I found 2 issues. Neither is blocking. The change does what #17245 asks for:
Helper flags and lanes. I checked that the helpers the spokes rely on accept Cheat sheet. Findings (both inline):
Loosened checks. Loosened-check verdict: fail (default; the prompt had no Embedded text. None of the PR context asks this lane to approve, skip or drop anything. Spend. Multi-repository runs spend subscription or prepaid usage only: one fan-out per repository, and audit-instructions asks for one ~20-dispatch confirmation per run. No billing hold. Verifier dropped 0 of 2 candidates. No verifier subagent was dispatched; I checked both findings myself against the source. |
|
The unattended merge path is about to arm native auto-merge at head |
|
Merge-Risk: max=HIGH demote=true head=fda7d9567b76369ff22b305752b1332ac1508599 rubric=10.4.1 rater=claude-opus-5-5 checker=claude-opus-5-5 |
Closes #17245
Summary
Opened by the work-resolve-issue lane. The run's own summary:
Not verified
Fix
Verification
The runner ran each check from the base copy against e5d7a3f:
The run reports:
Left out: the items under Not verified.
Related
#17245