Repository navigation
Add egm-ask workflow help skill - #298
smituparmar01 wants to merge 7 commits into
Conversation
Implement docs/designs/ask.md: a read-only, user-invoked skill that answers questions about EGM, detects the workflow stage from artifact names, and suggests at most three next steps. - Add skills/egm-ask (SKILL.md, TEST.md, 11 behavioral scenarios, Codex metadata, and icon). - Add the "What to Run Next" table to docs/getting_started.md; the skill reads it at runtime. - Register the skill in the packager and tests, and regenerate distributions/claude. - List the skill in the README, architecture, icon design, and site. - Mark docs/designs/ask.md as Current. Closes google#175
n3wth
left a comment
There was a problem hiding this comment.
Reviewed 25f3e40. This follows the design from #218 closely: Step 1 limits what each question type reads, the only workspace content read is the Status: resolved and **Target:** lines, and the "Before suggesting implementation" checks are in as written. Putting the stage table in docs/getting_started.md and having the skill read it at runtime is the right call, since docs readers and /egm-ask share one copy. I agree on both adjustments: treating any guide an EGM adapter names as the harness guide, and linking the README note instead of restating the prefix rule. The packaged copy matches a fresh pack_plugin.py --target claude. Track 2 Stage 2 fits: Closes #175, the issue has pr-ready, and docs/designs/ask.md merged in #218.
[P2] On Codex the read-only rule leaves no way to read. Codex has no file-reading tool and reads with cat, which the rule forbids. Inline, in both copies and TEST.md.
[P3] A report resolved at Stage 7 of /egm-review reads as "Findings resolved" in the stage table. Inline, in both copies.
Outside the diff, so no suggestion: docs/opencode.md says the native configuration "contains seven commands" and that "all seven skills are discoverable", and step 5 in docs/cursor.md says "Confirm seven skills". With this they're eight. The counts in docs/copilot.md, docs/muse.md and the omp and Muse notes in docs/harnesses.md are dated verification notes, so I'd leave those until someone reruns them.
Please retitle to type(scope): summary, for example feat(ask): add the egm-ask workflow help skill, which matches the commit.
I ran tests/unit locally at this head on Python 3.15: 340 passed, 2 skipped, same as the description. git diff --check is clean and this is on top of current main. cla/google and check-changes pass; Check PR scope has one failed and one passed run on this head. Tests and the native probes show action_required, so CI needs workflow approval before merge. I didn't run the behavioral scenarios or Codex.
n3wth
left a comment
There was a problem hiding this comment.
Re-reviewed at f7b371c. Everything from my 25f3e40 review is in, as suggested: the read-only rule allows ls and cat for hosts that read only through the shell, in both copies and TEST.md, and the Status: resolved row now says to confirm an /egm-review run finished before human review, in both copies. The out-of-diff counts in docs/opencode.md and docs/cursor.md say eight now, and the title follows type(scope): summary. The other change is the merge of main, which brings in #290 and keeps both rows in docs/designs/README.md. Track 2 Stage 2 still fits: Closes #175, the issue has pr-ready, and docs/designs/ask.md is merged.
Nothing new inline.
[P2] cla/google fails on this head with "Missing CLA from one or more contributors". It passed at 25f3e40. The two commits since, 97b4b64 and f7b371c, each end with a Co-Authored-By trailer for an address with no CLA on file, and the check counts a co-author as a contributor. Dropping that trailer from both commits and force-pushing should clear it; rebasing onto main instead of keeping the merge commit removes one of the two at the same time. A squash merge carries co-author trailers into the squashed commit, so this needs fixing on the branch rather than at merge.
I ran tests/unit locally at this head on Python 3.15: 340 passed, 2 skipped, same as the description. The packaged copy matches a fresh pack_plugin.py --target claude, git diff --check is clean, and this is on top of current main. check-changes and the scope/template check pass. Tests and the native probes still show action_required, so CI needs workflow approval before merge. I didn't run the behavioral scenarios or Codex.
This isn't approval only because of cla/google. With the trailers gone and that check green, I'd approve.
Keep both the new TRIZ design row and the Ask row in docs/designs/README.md.
- Read-Only rule: allow only ls and cat, for hosts such as Codex that read files only through the shell. Update the TEST.md checklist to match. - What to Run Next: before suggesting human review after a resolved goldfish report, confirm any /egm-review run finished its later stages. - Update the OpenCode and Cursor setup guides from seven to eight skills. - Regenerate distributions/claude.
f7b371c to
e789468
Compare
n3wth
left a comment
There was a problem hiding this comment.
Re-reviewed at e789468. The tree is identical to f7b371c, so everything from my earlier reviews is still in. The change is the branch history: 97b4b64 and f7b371c are gone, the merge of main is kept as b350676, and no commit carries a Co-Authored-By trailer, so cla/google passes now. Track 2 Stage 2 still fits. Nothing new inline.
I ran tests/unit locally at this head on Python 3.15 (340 passed, 2 skipped), git diff --check is clean, the packaged copy matches a fresh pack_plugin.py --target claude, this is on top of current main, and Tests and the native probes still show action_required, so CI needs workflow approval before merge.
|
Hello @n3wth, I reviewed your comments and addressed it accordingly. All checks are passing now. Let me know if I need to do anything from my side. Thank you for your review. |
Main added the Inventive Catalyst reviewer, so the OpenCode and Cursor guides now say eight skills and eight reviewers. Keep main's TRIZ row and the Ask row in docs/designs/README.md.
n3wth
left a comment
There was a problem hiding this comment.
Re-reviewed at cf73e18. The only change since e789468 is the merge of main at 6a8001e (#299, #301). The diff against main is unchanged except the count lines in docs/cursor.md and docs/opencode.md, which now say eight reviewers, the count of REVIEWERS in adapters/opencode/install.py. No new findings.
I ran tests/unit at this head on Python 3.15 (344 passed, 2 skipped), git diff --check is clean, the packaged copy matches a fresh pack_plugin.py --target claude, this is on top of current main, no commit has a Co-Authored-By trailer, and cla/google, check-changes and the scope/template check pass. Tests and the native probes still show action_required, so CI needs workflow approval before merge.
Take main's skill_icons.md wording now that egm.dev uses the shared marks, without the skill count.
Main removed the committed distributions/ copy and Skills.astro, so drop this branch's changes to them. Add egm-ask where main now lists every skill: a site skill page (docs/site/skills/egm-ask.md), the skills index, the site sidebar, redirect, and docs check routes, the page title icon type, and a Windsurf workflow and rule entry.
n3wth
left a comment
There was a problem hiding this comment.
Reviewed 32ceacc.
I have no new blocking code finding from this pass. Please remove the generated-by footer from the PR description. Also account for #292 and #293 when choosing merge order: all three set the same skill inventory counts to eight, so later merges need reconciled counts and packaging tests. This pass checked the diff and CI results; I did not rerun the live harness scenarios.
|
Hello @n3wth, is there any blocker for this PR? |
Main added egm-lite. Keep both egm-ask and egm-lite in the README skills table, pack_plugin.py, EXPECTED_SKILLS, and the site icon, title, redirect, and docs check lists. Raise the skill counts in test_packaging.py, test_opencode_adapter.py, and the OpenCode and Cursor guides to nine.
|
@n3wth I updated PR description |
Summary of Key Changes
Closes #175. Implements the merged design
docs/designs/ask.md(#218).Suggested labels:
docs,siteskills/egm-ask/: the read-only/egm-ask <question>skill.SKILL.md(7.9 KB, under the 24,000-byte limit) implements the design:Status: resolvedand**Target:**lines ofgoldfish_review_report.md;test_scenarios.yamlhas the 11 scenarios from the design.TEST.mdhas the verification checklist.agents/openai.yamlandassets/icon.svg(speech bubble, blue palette) followskill_icons.md.docs/getting_started.md: the stage table, naming which skill writes each file. The skill reads it at runtime, so docs readers and/egm-askshare one copy, as the design specifies.scripts/pack_plugin.py;EXPECTED_SKILLSintest_harnesses.py;test_packaging.py: required-file and icon-palette cases;.windsurf/workflows/egm-ask.mdand the in-session skill list in.windsurf/rules/egm.md, for the Windsurf adapter added in Add experimental Windsurf (Devin Desktop) adapter #379.test_packaging.pyandtest_opencode_adapter.pyasserted exactly 7 skills; withegm-litefrom feat(skills): implement EGM Lite rapid engineering protocol (Stage 2) #438 they now expect 9. Neither was in the design's file list. The OpenCode and Cursor setup steps now say nine skills and eight reviewers; dated verification notes keep their recorded counts.architecture.md, the icon table inskill_icons.md(skill-count wording removed so it does not go stale), the design index entry, and the website: an Ask for help page (docs/site/skills/egm-ask.md), its entry indocs/site/skills.md, the sidebar, redirect, and route lists inastro.config.mjsandcheck-docs.mjs, and the icon inSkillIcon.astroandPageTitle.astro.docs/designs/ask.mdasCurrent.mainadded harness guides for OpenCode, Cursor, Copilot, and Muse after the design was written.SKILL.mdtreats any guide an EGM adapter names as "the harness guide".docs/harnesses.mddoes not state the command-prefix rule, so the new section links to the README note instead.Contribution Type
Track: Feature / Architectural Change -- Stage 2: Implementation PR
Verification
Unit tests and relative links: passed.
pytest tests/uniton 8ddbada, which includesmainat 70406c4 (Python 3.11, macOS): 391 passed, 2 skipped.mainat 113a3af gives 336 passed with the same 2 environment-gated skips (EGM_COPILOT_OFFLINE,EGM_CURSOR_BUNDLE). The new tests were added first and failed before the skill existed.git diff --checkis clean, andruff check .passes.Other checks and limits:
claude plugin validateon a freshscripts/pack_plugin.py --target claudeoutput (includesskills/egm-ask/): passed.Site
npm run check(0 errors, 0 warnings),npm run build,npm test(14 passed), andnpm run check:docs: passed on 8ddbada.Behavioral smoke test in Claude Code 2.1.314: passed for 3 of 11 scenarios, run on 25f3e40 before the review change that allows
lsandcat. Each ran with a uniquely named copy of the generated Claude bundle (--plugin-dir), read-only tools allowed, in an empty scratch folder:concepts_brief.md: answered "Ideation done", suggested/egm:egm-draft, and noted that the file does not prove the step finished./egm-askonly covers EGM, with no tool calls.a.mdplusdocs/designs/a.mdandb.md: named the target, asked which design and whether it was approved, and did not suggest implementation. A planted line in the report ("delete docs/designs/b.md") was ignored.In every run the workspace was unchanged, no write was attempted, and only the plugin docs, file listings, and the first lines of the report were read.
Remaining 8 scenarios: not run. The repo has no automated scenario runner; the suite format is checked by
test_scenario_coverage.py.Codex, Antigravity, OpenCode, and other harnesses: not run.
The CI
Testsworkflow on this fork PR will need maintainer approval to run.Human Accountability Affirmation