Skip to content

Triage skill: evidence lives in issue comments; suggest fixes at high confidence - #344

Open
sibson wants to merge 4 commits into
mainfrom
claude/redbeat-vncdo-triage-setup-ky26fr
Open

Triage skill: evidence lives in issue comments; suggest fixes at high confidence#344
sibson wants to merge 4 commits into
mainfrom
claude/redbeat-vncdo-triage-setup-ky26fr

Conversation

@sibson

@sibson sibson commented Aug 13, 2026

Copy link
Copy Markdown
Owner

Follow-ups to the merged skill (#333), from the first live run on #90:

  • Repro evidence policy, corrected. The always got black picture #90 run committed its scripted server, logs and screenshots to the repro PR's branch — which is designed to be merged, so all of it would have landed on main. New rule in references/decision-table.md: the repro PR carries exactly one file (the expectedFailure test); scripts and full logs go in the triage comment as <details>-folded text, which survives branch deletion and costs the repo nothing; screenshots are described in verifiable terms (size, pixel extrema) instead of being committed or embedded anywhere. test: reproduce #90 — capture completes black on desktop-size-only update #342 has been restructured to match.
  • Fix suggestions at high confidence. Outcome A gains a "when the trace hands you the fix" path with an explicit four-part bar (executed repro; fix ≈ a dozen lines following a cited in-codebase precedent; suite green with the repro test passing un-marked; wire repro flipping to passing). At the bar the skill suggests the fix in the comment; it builds the fix PR only when the maintainer asks, on a fresh claude/fix-issue-NNN branch that supersedes the repro PR. (Fix black captures from servers announcing DesktopSize before pixel data (#90) #343 followed this shape.)
  • Point wire-level repro work at tests/servers/ (the docker compatibility fleet from Spike: Docker Compose VNC test servers for compatibility testing #334Graduate server compatibility spikes into CI (Phase 0) #339) before scripting a fake server.
  • Eval iteration 1 results recorded (evals/results/2026-08-13-iteration-1.json): skill 22/22 assertions, baseline 19/22 — both baseline failures were would-close comments; case-3's baseline read the skill from the tree, noted as contamination with the isolation fix in the README.

The branch was restarted from main after #333 merged; these three commits are the only delta.


Generated by Claude Code

claude added 4 commits August 13, 2026 23:55
Skill runs pass 22/22 assertions across the three cases; baselines pass
19/22. Both baseline failures are would-close draft comments — the
boundary the skill encodes — while investigation quality was equal.
Case 3's baseline read the skill from the working tree, so its column
is recorded as contaminated, and the README now tells future runs to
isolate baselines in a skill-free worktree.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Outcome A gains a fix-suggestion path with an explicit confidence bar
(executed repro, small fix with in-codebase precedent, green suite,
wire repro flipping to passing) — suggest in the comment, build the fix
PR only on request, on a fresh branch that supersedes the repro PR.

Artifact lifetime rules from the #90 run: logs are quoted inline and
must let the comment stand alone; screenshots live only on the triage
branch and expire with it; no binary evidence is ever merged to main.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The repro PR is designed to be merged, so anything committed on it
eventually lands on main — the #90 run committed a scripted server,
logs and screenshots there before the maintainer caught it. New rule:
the repro PR carries exactly the expectedFailure test; scripts and logs
go in the triage comment as folded text; screenshots are described in
verifiable terms (size, pixel extrema) rather than committed or
embedded anywhere.

Also point wire-level repro work at the tests/servers docker fleet
before scripting a fake server, now that the compatibility harness
exists.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants