Repository navigation
fix(session-flow): point the retro skill at its context files so they resolve from the skill directory - #16565
Conversation
…folder The retro skill named context/session.md and context/quick.md as bare backtick paths, and its "Spoke paths" note called them "the context/ files" beside <plugin-root>. Eval runs read that as <plugin-root>/context/session.md, failed, and globbed the plugin root, which is always denied. Every mention is now a relative Markdown link, the form this file already uses for reference/, and the note says the files sit in the skill's own context/ folder beside SKILL.md. Refs: #16368 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex Review Summary: ✅ CompletedThis 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-security-review reviewed Security review: no security issues foundI read both changed files in full:
Instruction surface. The change removes or weakens no rules. "Spoke paths" still tells the model to replace the Out of scope for this change:
PR context. The description matches the diff and contains no embedded instructions to this lane. The closing issue's body was left out by the wrapper because its author is not on the allowed list. Prior findings. This lane had not reviewed an earlier head ( critical-findings: 0 |
|
claude-review reviewed No findings. This is a docs-only change to two files: the retro skill's What I checked
Closing issue #16368, "Done when"
No embedded instructions in the PR context. No manifest changes. The change itself spends nothing; the author's eval cost was a one-off spend of prepaid credit. Verifier dropped 0 of 0 candidates. |
Closes #16368
Summary
/session-flow:retronamed its mode filescontext/session.mdandcontext/quick.mdas bare backtick paths, and its "Spoke paths" note called them "thecontext/files" next to<plugin-root>. Eval runs read that as<plugin-root>/context/session.md, got "File does not exist", then globbed the plugin root, which is always denied. Every mention is now a relative Markdown link, and "Spoke paths" says the files sit beside the skill's own SKILL.md.Fix
plugins/session-flow/skills/retro/SKILL.md: the sixcontext/mentions (mode table, Session settings, Multi-session awareness, Spoke paths, Next) are relative links, the form the file already uses forreference/..changes/session-flow/: patch changelog fragment.Verification
Tests, on the head after merging main:
plugins/skill-quality/scripts/check-skill.sh retro(the/skill-quality:checkgate):PASS, 0 errors, 5 warnings, none about the changed lines; its script testscripts/parse-transcript.test.shpassed.pytest plugins/session-flow/skills/retro/scripts/test_parse_transcript.py: 45 passed.context/mention inretro/SKILL.mdis a link (lines 66, 67, 107, 126, 193, 207).Eval record (pre-registered before any launch; one run of each batch, no reruns or rescoring):
re-mechanical-finding-check-proposal(the retro case with the most plugin-root Glob denials in an earlier eval: 7 of 12 Sonnet runs, both sides carrying the defect).claude plugin eval, Claude Code 2.1.325,--model sonnet, judgeclaude-opus-5-5, 6 runs a side. BEFORE = merge basead1c8576af6977687c604e784dabb424f06440a4; AFTER =be44a15353f16cf9604e0fd0ca2bbe15c03e23b7. Both copies carry the sameevals/tree; they differ only inskills/retro/SKILL.md.hands-off-to-codifycheck-not-prose-ruleno-early-codifyno-writeno-editno-bash-writesPass rules, fixed before launch: M1 both result files present with 6 launched runs each, neither void; M2 AFTER plugin-root runs = 0; M3 at least 5 AFTER runs reached the skill; M4 each grader passes in at least 5 of 6 launched AFTER runs; M5 no recorded stop. All five met: verdict MET. Total cost 3.3264 USD (judge cost included).
The BEFORE side denied in 1 of 6 runs, below the 4 of 6 the earlier eval saw for this case, so 1 against 0 is a small sample difference; the done-when bar (no AFTER-side plugin-root denial) is met.
Amendment to the launch path, recorded before any launch: the launch PATH filter strips mise's shim folder, and on the host node existed only as that shim. Retro session mode runs
node "<plugin-root>/skills/setup/scripts/setup-apply.mjs" --check, and the registration voids a result file oncommand not foundfor node, which would have failed M1 for a setup reason. The filter gained one line appending the node install folder at the end of PATH. No launch line, pass rule, bar, run count, exclusion, cap or eval flag changed.Main merged into the branch after the eval (
git merge, no conflict). In the measured paths (plugins/session-flow,plugins/evals) main changed, between the measured head and this head:retro/SKILL.mdGotchas: one sentence reordered so "(when thecorrectionsplugin is enabled)" sits beside/corrections:codify(fix(session-flow): place the retro's corrections condition and state the policy-class rule #16500).retro/reference/ecosystem-improvement-catalog.md: guardrail item 2 now also accepts a one-time policy-class finding (fix(session-flow): place the retro's corrections condition and state the policy-class rule #16500). This is text thecheck-not-prose-rulegrader's case can reach; the eval did not measure it.prerequisitesstubs' node notice wording,save_point.pyand its tests,continue-in-background/SKILL.md,reference/structure.md,reference/save-point.md, versions and changelogs. No eval case or grader changed.Related
🤖 Generated with Claude Code
https://claude.ai/code/session_01MWmASMXyUAaNx5GqGM3h6c