feat(#473): /launch-check --workflow — opt-in parallel + adversarially-verified audit#474
Merged
atlas-apex merged 1 commit intoJun 1, 2026
Conversation
…y-verified audit Adds a third, opt-in mode to /launch-check. The 10 dimensions are independent and read-only — the textbook fan-out shape — so --workflow authors and runs a Claude Code Workflow that: - evaluates all applicable dimensions concurrently (one agent each, reusing each dimension's existing PASS/WARN/FAIL criteria; non-applicable dims pre-resolved) - adversarially verifies every FAIL/WARN finding as it returns (pipeline, no barrier) — a refuted finding downgrades to PASS (the false-positive cut) - synthesizes the SAME verdict table + four-state vocabulary and persists via the SAME _lib-audit-history.sh (audit_run_persist / audit_render_trend), so trend history and the render-trend.sh chart are continuous across serial↔workflow runs Opt-in by design: a workflow run spawns ~10-20 agents, so the serial path stays the default; no flag → no Workflow invocation. Degrades to serial if the Workflow tool is unavailable. No new skill/hook/role — counts unchanged. Documents the canonical workflow script (phases, dimension + verdict schemas, synthesis + persist) in-skill so each run authors a consistent workflow. AgDR-0055 records the decision and the pattern for other multi-dimension audits. Closes #473 Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
atlas-apex
commented
Jun 1, 2026
atlas-apex
left a comment
Collaborator
Author
There was a problem hiding this comment.
Code Review: PR #474
Commit: a8e4428872daca197e258c75c85d714bb2f8f8c2
Summary
Adds an opt-in --workflow mode to /launch-check. Docs-only: the skill (SKILL.md) gains a frontmatter argument-hint, a third mode in the modes list, and a new "Workflow mode" section documenting a canonical Claude Code Workflow script (parallel fan-out of the 10 dimensions + per-finding adversarial verify, pipelined). Adds AgDR-0055 capturing the decision, plus a one-line CLAUDE.md skill-blurb update. No application/domain code, no new skill/hook/role/agent.
Checklist Results
- Architecture & Design: Pass — opt-in flag, serial default preserved; reuses existing persistence/trend rather than forking it
- Code Quality: Pass (docs) — the illustrative workflow script is internally consistent
- Testing: Pass — no executable code; markdownlint/site-counts/lychee/shellcheck are the relevant gates
- Security: Pass — read-only audit mode; no secrets, no new attack surface
- Performance: Pass — opt-in precisely because it costs ~10–20 agents vs one pass; documented and gated behind the flag
- PR Description & Glossary: Pass — Glossary present (4 terms), ticket linked (Closes #473)
- Summary Bullet Narrative: Pass — every bullet is what + why; no advisory finding
- Technical Decisions (AgDR): Pass — AgDR-0055 linked via
<!-- agdr: ... -->, covers the decision (opt-in vs default, Workflow vs /fan-out, graceful degrade) - Adopter Handbooks: N/A — docs-only skill change; architecture/general handbooks have no trigger here, no
.ts/.tsxin diff so the TypeScript handbook doesn't load, no domain handbooks present
Issues Found
None blocking.
Verification notes
- Serial path + Step 6 unchanged: confirmed against the diff. Steps 1–2 and Step 6 (
_lib-audit-history.sh/audit_run_persist/audit_render_trend/render-trend.sh) are untouched; the only edit near the default path is adding the word "serially" to the modes-list bullet — descriptive, not behavioural. The no-flag default is byte-for-byte preserved. No regression. - Cross-references resolve: the new section cites "Verdict logic" (SKILL.md L73), "Output format" (L40), and the four-state vocabulary go / go-with-warnings / conditional-go / no-go (L364) — all exist and match. The persistence claim (
scores{}from each row'sscore,findings[]from WARN/FAIL rows,audit_run_persist) lines up with the documented Step 6 superset payload. - Workflow script coherence: judged as skill documentation an agent would author from, not runtime-compiled.
export const meta(name/description/phases),agent(prompt, opts)with label/phase/schema/agentType,pipeline(items, mapFn, reduceFn), JSON-schemaschemaoption, and the pipeline-with-adversarial-verify shape (PASS short-circuits with nothing to refute; WARN/FAIL spawn a refute agent;upheld === falsedowngrades to PASS with a note) are all internally consistent.DIM_SCHEMA/VERDICT_SCHEMAare well-formed.return { results: [...preResolved, ...evaluated.filter(Boolean)] }matches the "After the workflow returns" contract. An agent could author a correct workflow from this. - Counts: 3 files (SKILL.md modified, CLAUDE.md 1 line, AgDR added). No new skill/hook/role/agent → site-counts unaffected, as the PR claims.
Suggestions
- nit (non-blocking): the "After the workflow returns" step says "apply the same verdict logic" — the serial verdict-logic table keys on PASS/WARN/FAIL counts, which the workflow rows carry, so this is fine. Worth a one-line note in a future pass that refuted FAILs downgrade to PASS before the count, so the verdict reflects the post-verify state (the script already does this via the downgrade; just not spelled out in the synthesis step). Purely a clarity nit.
Verdict
APPROVED
🤖 Reviewed by Rex (Code Reviewer Agent)
📌 Reviewed commit: a8e4428872daca197e258c75c85d714bb2f8f8c2
7 tasks
me2resh
added a commit
that referenced
this pull request
Jun 5, 2026
…y-verified audit (#474) Adds a third, opt-in mode to /launch-check. The 10 dimensions are independent and read-only — the textbook fan-out shape — so --workflow authors and runs a Claude Code Workflow that: - evaluates all applicable dimensions concurrently (one agent each, reusing each dimension's existing PASS/WARN/FAIL criteria; non-applicable dims pre-resolved) - adversarially verifies every FAIL/WARN finding as it returns (pipeline, no barrier) — a refuted finding downgrades to PASS (the false-positive cut) - synthesizes the SAME verdict table + four-state vocabulary and persists via the SAME _lib-audit-history.sh (audit_run_persist / audit_render_trend), so trend history and the render-trend.sh chart are continuous across serial↔workflow runs Opt-in by design: a workflow run spawns ~10-20 agents, so the serial path stays the default; no flag → no Workflow invocation. Degrades to serial if the Workflow tool is unavailable. No new skill/hook/role — counts unchanged. Documents the canonical workflow script (phases, dimension + verdict schemas, synthesis + persist) in-skill so each run authors a consistent workflow. AgDR-0055 records the decision and the pattern for other multi-dimension audits. Closes #473 Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
--workflowmode to/launch-check([Feature] /launch-check --workflow — opt-in parallel + adversarially-verified audit #473). The 10 readiness dimensions are independent, read-only, and gate-free — the textbook fan-out shape — so--workflowauthors and runs a Claude CodeWorkflowinstead of the serial single-context pass._lib-audit-history.sh(audit_run_persist/audit_render_trend), same superset JSON schema, samerender-trend.shchart. Trend history is continuous across serial↔workflow runs; no schema change.Workflowprimitive requires explicit opt-in — the--workflowflag is that opt-in. No flag → serial path, noWorkflowinvocation. The default behaviour is byte-for-byte unchanged.Workflowisn't available in the installation, the skill says so in one line and falls back to serial.--workflowis an optimisation, not a hard dependency.test_site_counts.sh) are unchanged.Testing
markdownlint-cli2(CI ruleset) — 0 errors on the changed files (SKILL.md, AgDR-0055,CLAUDE.md). The only flags from a newer local markdownlint are MD060, a rule absent from CI's pinned version and already firing on pre-existing tables repo-wide.bash .claude/hooks/tests/test_site_counts.sh— green (counts unchanged: this PR adds no skill/hook/role).--workflowonly swaps the evaluation of the 10 dimensions for a fan-out, then rejoins Step 6.DIM_SCHEMA/VERDICT_SCHEMA, pipeline-with-adversarial-verify, after-the-workflow synthesis + persist) so each run authors a consistent workflow.Glossary
/launch-check --workflowpath that authors + runs a Claude CodeWorkflowto evaluate the 10 dimensions in parallel and adversarially verify findings.scores{}map, sorender-trend.shkeeps plotting unchanged.Closes #473