* ci: audit closed-issue delivery claims — cited PRs must actually be merged
Follow-up to #3223: an issue closed as 'fixed in #N' while #N was closed
without merging (or does not exist) leaves a completion claim that does not
hold. The weekly maintenance workflow now runs a read-only audit over issues
closed in the last 30 days, extracts delivery-claim PR references from issue
bodies and close-window comments, verifies each cited PR's merge state, and
reports findings to the step summary.
Design constraints:
- Read-only: no comments, no closes, no labels; findings are advisory and
the run always stays green so the weekly audit never blocks maintenance.
- Unverifiable counts as unmerged: a 404 PR or an API failure fails the
verdict, because an unverifiable claim is exactly the failure mode.
- Pure decision functions are exported for tests; a first live run against
this repository surfaced one real finding (#3408 cites #3060, which does
not resolve), included here as the motivating sample.
8 unit/contract tests: extraction, windowing (close comments included),
the #3223 unmerged-citation failure mode, the all-clear summary, workflow
contract, and a no-write guard.
* ci: rework audit evidence contract — claims vs context, four PR states, visible completeness
Revision per the #3522 review (reviewer ran the CLI through 19 controlled
cases plus a bounded live check of #3408):
- Bind findings to asserted delivery claims. Closing wording ('Delivered
through #N'), claim verbs (fixed/shipped/landed/resolved/...), and bare
'PR #N' inside closing comments are claims; historical mentions, quoted
lines, fenced examples, negated claims, and tracking-only references are
reported as contextual references and never produce findings. On the live
#3408 the audit now recognizes the closing claim #3500 (merged) instead of
flagging the historical #3060 text.
- Separate observed states: merged / open-unmerged / closed-unmerged /
unknown (404, rate limit, transient). Unknown claims are advisory manual-
verification items, not failed claims; each row carries the claim snippet,
its source link, and the observation time. Checked claims are listed even
when passing, so the report is traceable rather than an opaque all-clear.
- Paginate issue and comment collection (--paginate --slurp), bound comment
requests (200) and PR checks (100), and surface skipped pages, rate limits,
and retrieval failures as an explicit incomplete audit instead of an empty
all-clear. A live-run maxBuffer overflow on the paginated issue list was
caught and fixed (64MB); the audit reported it as incomplete before the fix.
- Scope the job to read access: job-level contents/issues/pull-requests read.
13 contract/unit tests including a CLI fixture (stub gh) that places the
relevant issue on page two and forces a comment-fetch failure.
* ci: report incomplete audit at PR check cap
---------
Co-authored-by: chenhz01 <chenhz01@users.noreply.github.com>
Co-authored-by: haelyra <49814733+haelyra@users.noreply.github.com>
Prettier's config searcher also reads .prettierrc.json5, .prettierrc.ts/.mts/.cts and prettier.config.ts/.mts/.cts. resolve-formatter knew none of them, so a project configured only through prettier.config.ts was never formatted by post-edit-format, quality-gate or the Stop batch. config-protection's rc pattern stopped at .yaml/.toml, so an existing .prettierrc.ts or .prettierrc.json5 could be edited freely.
The Stop typecheck reported every tsc line that contained the edited path anywhere. A root index.ts has the relative path `index.ts`, which every `src/index.ts(...)` diagnostic contains, so a clean root file was reported with the errors of src/index.ts. The same happened to src/utils.ts with web/src/utils.ts, and to any file named in another file's message.
Compare the path that starts the diagnostic (`path(line,col): `), keeping the slash normalization for Windows.
A .gitignore rule outranks .git/info/exclude, so a source project that ships !FORK_REPORT.md made git add -A stage the report. Remove both reports from the index after staging, in the forker's first commit and in every fold-in amend. The history test now runs both documented blocks against a negating .gitignore, prints the Passed/Failed summary for run-all.js, and skips on Windows like the publish-guard test.
(cherry picked from commit f493d9ff66)
The "resolve output to absolute path" step only rebuilt
"$OUTPUT_DIR/$(basename ...)", so ./slides.pdf stayed relative. The
script then cd's into a temp dir, Playwright writes the PDF there, and
the temp dir is deleted. The script prints "PDF exported successfully!"
and then fails on du under set -e, with no PDF produced. This is the
script's own documented example.
Resolve the output directory with cd + pwd, as the input path already
does, and add a stubbed regression test (no network or browser).
Signed-off-by: Wen <jwcastillo@gmail.com>
(cherry picked from commit f91c14749e)
- SKILL.md invoked every script through ~/.claude/skills/ck, which does
not exist when ck is installed as part of the ECC plugin, so every
/ck:* command failed with MODULE_NOT_FOUND. Use ${CLAUDE_SKILL_DIR}
and tell non-Claude hosts to substitute the skill directory. Scripts
still run from the project directory, since they use it as the
project path.
- The documented SessionStart command quoted "~/...", which the shell
never expands. Use $HOME and explain the plugin-install case.
- /ck:init and /ck:save piped JSON through echo '<json>'. Any apostrophe
ends the quoted string, and init JSON carries README/package.json
text. Pass it through a quoted heredoc instead.
- session-start.mjs and shared.mjs resolve SKILL.md next to the code
(import.meta.url), and the hook expands ${CLAUDE_SKILL_DIR} in the
text it injects.
- encodeProjectPath only replaced '/', while Claude Code maps every
non-alphanumeric character to '-'. Memory entries for projects whose
path has a space or dot went to a directory Claude Code never reads.
Signed-off-by: Wen <jwcastillo@gmail.com>
(cherry picked from commit c920005c2b)
quality-gate.py never read stop_hook_active, so a condition the agent
cannot fix (disk below 15 GB, an unwritable growth-log) blocked every
Stop and kept the session generating turns. Exit 0 when Claude Code
reports that the Stop hook already forced a continuation.
The memory dir lookup only mapped '/', '\' and ':' to '-'. Claude Code
maps every non-alphanumeric character, so for any project path with a
space the memory dir was never found and enforcement silently became
warn-only.
Align SKILL.md with the code: disk thresholds (remind 50 / warn 30 /
block 15), the disk block applies to every session, the learning-lib
rule and its examples, stderr messages, and the hook timeout, which
Claude Code reads in seconds (5000 meant 83 minutes).
Signed-off-by: Wen <jwcastillo@gmail.com>
(cherry picked from commit a1385c4397)
The compiled OpenCode tools import `tool` from `@opencode-ai/plugin/tool`
at module top level, and the hooks plugin imports those tools statically.
The package was only a devDependency, so production installs (OpenCode's
npm plugin cache, `npm install --omit=dev`) could not resolve it and the
whole plugin failed to load.
Move it to `dependencies` and add a build-opencode regression test that
fails when the built bundle imports a package not declared as a runtime
dependency.
Fixes#3543
(cherry picked from commit 106024070e)
The platform-configs module lists `.pi`, but PLATFORM_SOURCE_PATH_OWNERS
had no entry for it, so isForeignPlatformPath() returned false and every
install-apply target (e.g. `--target opencode`) copied the Pi harness
adapter (.pi/README.md, .pi/extensions/*) into its config root and
recorded it as ECC-managed.
Pi consumes ECC through `pi install`, not install-apply, so map `.pi` to
the `pi` owner. No install target has that id, so the directory is now
filtered out everywhere, like `.codex`, `.cursor`, and `.zed`.
Add regression tests for the opencode plan and for every registered
target.
Fixes#3558
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: haelyra <49814733+haelyra@users.noreply.github.com>