skills/pbakaus_impeccable
1
0
Fork 0
mirror of https://github.com/pbakaus/impeccable.git synced 2026-09-20 01:56:37 +03:00
Code Issues Packages Projects Releases Wiki Activity

Fix: Setup survives a refused launcher (#750)

Browse Source 
Preserve Setup context and reference loading after launcher refusal, disclose the failure before editing, and limit Claude skill-directory substitution to SKILL.md. Add plugin-path and denied-launcher behavior regressions. Addresses part of #744 without closing its remaining scope.

AI assistance: Cursor on the original contribution; Codex on maintainer-directed follow-up fixes and validation.
This commit is contained in:
Abdul Wahab
2026-09-06 13:07:29 -07:00
committed by GitHub
parent 720628f4ea
commit 3e653377de
7 changed files with 195 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/skill-behavior/README.md
+22
View File
@@ -75,6 +75,28 @@ The trace is the source of truth, not the model's free-form reply.
| 16 | existing surface, with and without PRODUCT.md; asks where to start | loads `routing.md`, delivers advice, and does not edit project files, start an interview, archive a critique, or run menu scans |
| 17 | existing surface; asks whether critique is required before polish | loads `routing.md` and both command references, then delivers advice without executing the playbooks |
| 18 | existing surface; explicitly requests polish followed by a next-command recommendation | loads `polish.md` rather than substituting workflow advice for the requested work |
| 19 | tiny spacing edit with PRODUCT.md + DESIGN.md; Bash denied, plus a real-loader success control | actually reads playbook and craft floor before editing; denial also requires direct context-file reads and a user-visible warning before the edit |
## Setup launcher-failure branch (2026-09-06, PR #750)
Scenario 19 injects a host permission error before any shell command executes,
including retries and compound commands. File tools remain available; the
staged skill is read-only. Assertions require successful reads, an actual UI
edit, unchanged context files, and disclosure before the first write in the
assistant message sequence. Failed read attempts and shell commands merely
mentioning a reference do not count as loading it. The success control allows
only the real context-loader command through Bash and requires exit 0.
The suite measures continuation after a tool refusal, not Claude's skill
activation or plugin substitution. Those are separate loader/path checks.
It also does not establish behavior for every missing-binary or runtime error.
Focused baseline (2026-09-06): both scenario 19 cases passed on
`claude-sonnet-5`, one run each. The original wording also continued after
denial in the comparison run, but disclosed the failure only in its final
summary. This does not reproduce the reporter's complete reference-loading
failure or establish a multi-provider pass. An earlier scenario 6 result used
attempt-based reference assertions and is not counted as a success control.
## Workflow-advice baseline (2026-09-05, PR #737)