Add /swarm:review local mixture-of-agents pipeline (P2) - #24
Merged
Conversation
- agents.sh: add grok-composer-2.5-fast as a second grok voice via a defensive-parse path (strict-JSON prompt, no --json-schema/--effort); prefer the findings-bearing object over the first decodable one - workflows/swarm-review.js: registered pipeline — scope+gate → 4-voice fan-out (Claude lenses + codex + grok-build + composer) → (file,mechanism) merge with family-aware consensus → verify solos → output-gated synthesis - skills/review/SKILL.md: /swarm:review — builds diff + fenced prompt, probes live backends, runs the workflow, renders the ranked report - minimal security: fence diff as data, output gate (secret scrub over all findings), error != empty transport marker; P1 adapter floor stays - bump swarm 0.1.0 -> 0.2.0; update README + knowledge Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017rpoJ2DokY1SMfv1dSfZwe
Follow-up from a full 4-voice /swarm:review of the composer diff (cross-family consensus flagged the first two): - validate each composer finding against finding.schema.json — composer is not CLI-schema-enforced like codex/grok-build, so a malformed item now ERRORs (error != empty) instead of reaching merge/verify - prefer a non-empty findings object over an empty/preamble object that would otherwise mask the real answer - guard the assembled prompt+schema argv word against MAX_ARG_STRLEN (the prompt cap didn't account for the appended schema on this path) - note on stderr that --effort is ignored for grok-composer-2.5-fast Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017rpoJ2DokY1SMfv1dSfZwe
A full 4-voice /swarm:review of PR #24 surfaced real issues in the new code; fix the confirmed/agreed ones (deferred: trusted-source path interpolation, self-pwn pathspec, haiku-transport fundamental limit). Security / output gate: - SKILL.md: per-run nonce in the diff fence — a fixed marker could be forged by diff content to close the fence and inject reviewer instructions - swarm-review.js: output gate scrubs ALL string fields (was missing verifier evidence + cluster mechanism); refuted findings now scrubbed too - agents.sh: composer valid_item mirrors finding.schema.json exactly (exact key set + maxLength), not a type-only subset Correctness: - SKILL.md: resolve the real default branch (origin/HEAD) instead of hardcoded main/master; warn instead of silently narrowing to `git diff HEAD` - SKILL.md: oversize prompt drops external voices (Claude-only) to match its own warning, instead of passing voices the adapter only rejects - swarm-review.js: Claude lens failures marked error != empty; zero-member clusters dropped; sort comparator made antisymmetric - agents.sh: cap composer stdout before the O(n^2) brace-scan (a multi-MB blob could spin past the timeout) - external prompt requests [lens] tags so external findings map to real lenses Presentation + docs: - SKILL.md: encode the locked P4 report layout (model-named Agents column, correct severity icons, Verifier vs main-session Verdict, balance block); note the P5 --fix/--loop ask-on-ambiguous-fix rule - swarm-review.js: per-backend model+count rollup for the balance Agents line - sync CLAUDE.md + top-level README to swarm 0.2.x / P2 / /swarm:review Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017rpoJ2DokY1SMfv1dSfZwe
Round 1 of a self-review loop over the P2 branch (the prior 16 findings all resolved, none recurred). Fix the agreed findings; option B for the schema duplication — add the missing caps + a drift-warning comment to each mirror. - SKILL.md: `|| true` on the DEFBR pipeline — a regression from the previous round where pipefail aborted the default-branch fallback loop (critical) - swarm-review.js: merge coverage guard — recover pool findings the merge agent leaves unclustered as solo clusters instead of silently dropping them - schema caps (B): add maxLength + maxItems:100 to the workflow FINDING_ITEM and the composer validator so all three schema copies match; drift-warning comment in finding.schema.json / swarm-review.js / agents.sh - output gate: redact the whole PEM block (BEGIN..END), not just the header, in BOTH the JS gate and the adapter scrub; scrub backendErrors + gate text - balance.agents: a single failed lens no longer marks the whole backend down - SKILL.md: stop advertising --fix/--loop as runnable (P5, not built) - docs: composer is a grok model, not a third CLI — fix plugin.json, marketplace.json, READMEs, CLAUDE.md Deferred (documented residuals): verifier diff-offset line numbers, trusted path interpolation, composer schema-append asymmetry, defensive arg defaults. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017rpoJ2DokY1SMfv1dSfZwe
A control run without Claude lenses (codex + grok-build + composer only) surfaced real gaps the with-Claude run missed — notably a real secret-scrub drift. - output gate: add the explicit aws_secret_access_key rule the JS scrub was missing (the generic \bsecret\b rule can't match inside the underscored key name) — JS gate and adapter scrub now redact identically; also redact a PEM key truncated by a field cap (END marker optional), in both scrubs - merge: first-wins disjoint member_indices so a finding can't be emitted twice or inflate consensus; CLUSTER_SCHEMA now carries FINDING_ITEM's length caps - pool skips ok:false voices; sort tolerates unknown severity (no NaN); the input-guard early return returns the full result shape - SKILL.md: default scope now includes untracked files (git diff omits them) - workflow: add args.claude:false for an external-only control run (no Claude finder lenses; merge/verify still in-session), documented in the skill Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017rpoJ2DokY1SMfv1dSfZwe
Third control round (external-only: codex + grok-build + composer) caught follow-on gaps in earlier fixes plus a few pre-existing ones. - output gate: scrub gate.run + skip[].lens too (not just change_kind/why); PEM redaction now handles encrypted keys (Proc-Type/DEK-Info metadata) AND field-cap-truncated keys via alternation, in both JS and adapter scrubs - gate/merge agents get a .catch — a failure degrades gracefully (gate → all lenses, merge → coverage guard recovers every finding) instead of rejecting - honor an explicit empty gate run:[] as "no lenses" (was falling back to all) - SKILL.md: untracked-file append is now flag-gated (INCLUDE_UNTRACKED) so a ref/--staged review doesn't pull in out-of-scope untracked files; align the oversize cutoff with the adapter's 122880-byte cap - EXTERNAL_SCHEMA gains maxItems:100 to match the other schemas - knowledge: record the claude:false control run + the per-lens external-prompt idea as opt-in P3 (the ~5x call cost rules it out as a default) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017rpoJ2DokY1SMfv1dSfZwe
gering
added a commit
that referenced
this pull request
Jul 16, 2026
Second dogfooding round at the deepest profile (/swarm:review --max --fix: per-lens split live — gate pruned style, 10 xhigh finders + codex gpt-5.6-sol@xhigh + grok-4.5 + composer; 24 findings, 10 consensus). Fixes the agreed findings: - Never auto-accept an all-untagged consensus cluster: no tagged lens backs it, so it is verified like a solo. Verify/auto-accept now derive from ONE needsVerify predicate (structural exactly-once partition); kind vote in clearer every-form (#1, consensus critical; #21, #22) - Design verifier sees the finding's recommendation — the proposal the applicability rubric actually tests — and carries an escape hatch: a genuine defect mis-filed under a design lens is not refuted away (#12, #3); "solo" dropped from verifier prompts (#4) - Untagged findings from multi-lens cluster finders fall back to 'unspecified' (safe defect bucket), not lenses[0] (#7); merge-agent free-text lens validated, majority-member fallback (#9) - Improvement invitation scoped to design finder units — defect-lens finders stay defect-only (#8); merge prompt clusters by issue, not only defect (#13); schema descriptions generalized for design findings (#14) - LENS_BRIEF startup assertion (#6); gate prompt interpolates LENS_CLUSTERS.design (#19); new test_lens_sync.py guards all lens mirrors: SKILL HDR prompt, LENS_BRIEF, pr-post DESIGN_LENSES (#5) - Workflow assigns stable finding num (defects first, shared sequence); presenter/pr-post render it verbatim (#20) - pr-post.py: design lens is the backup kind signal when the handoff drops kind (explicit defect still wins); single-pass partition (#17, #24) - Balance: REFUTED is its own segment (refuted ⊄ solo since design consensus can be refuted) (#2); LOCKED design-table column precedence in --loop rounds clarified (#11); "verifies solos" doc sweep across manifests/README/knowledge (#4); cluster failure-isolation trade-off documented (#10); knowledge index line trimmed (#15) Declined: #16 (JS test harness for sandbox code; lens-sync test covers the drift class), #18 (gate-fail under --max runs all lenses — the documented never-silently-narrower degrade), #23 (micro-opt vs readability). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126cGxsBYxqgEFH4AcpriNk
gering
added a commit
that referenced
this pull request
Jul 17, 2026
Second dogfooding round at the deepest profile (/swarm:review --max --fix: per-lens split live — gate pruned style, 10 xhigh finders + codex gpt-5.6-sol@xhigh + grok-4.5 + composer; 24 findings, 10 consensus). Fixes the agreed findings: - Never auto-accept an all-untagged consensus cluster: no tagged lens backs it, so it is verified like a solo. Verify/auto-accept now derive from ONE needsVerify predicate (structural exactly-once partition); kind vote in clearer every-form (#1, consensus critical; #21, #22) - Design verifier sees the finding's recommendation — the proposal the applicability rubric actually tests — and carries an escape hatch: a genuine defect mis-filed under a design lens is not refuted away (#12, #3); "solo" dropped from verifier prompts (#4) - Untagged findings from multi-lens cluster finders fall back to 'unspecified' (safe defect bucket), not lenses[0] (#7); merge-agent free-text lens validated, majority-member fallback (#9) - Improvement invitation scoped to design finder units — defect-lens finders stay defect-only (#8); merge prompt clusters by issue, not only defect (#13); schema descriptions generalized for design findings (#14) - LENS_BRIEF startup assertion (#6); gate prompt interpolates LENS_CLUSTERS.design (#19); new test_lens_sync.py guards all lens mirrors: SKILL HDR prompt, LENS_BRIEF, pr-post DESIGN_LENSES (#5) - Workflow assigns stable finding num (defects first, shared sequence); presenter/pr-post render it verbatim (#20) - pr-post.py: design lens is the backup kind signal when the handoff drops kind (explicit defect still wins); single-pass partition (#17, #24) - Balance: REFUTED is its own segment (refuted ⊄ solo since design consensus can be refuted) (#2); LOCKED design-table column precedence in --loop rounds clarified (#11); "verifies solos" doc sweep across manifests/README/knowledge (#4); cluster failure-isolation trade-off documented (#10); knowledge index line trimmed (#15) Declined: #16 (JS test harness for sandbox code; lens-sync test covers the drift class), #18 (gate-fail under --max runs all lenses — the documented never-silently-narrower degrade), #23 (micro-opt vs readability). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126cGxsBYxqgEFH4AcpriNk
gering
added a commit
that referenced
this pull request
Jul 17, 2026
Second dogfooding round at the deepest profile (/swarm:review --max --fix: per-lens split live — gate pruned style, 10 xhigh finders + codex gpt-5.6-sol@xhigh + grok-4.5 + composer; 24 findings, 10 consensus). Fixes the agreed findings: - Never auto-accept an all-untagged consensus cluster: no tagged lens backs it, so it is verified like a solo. Verify/auto-accept now derive from ONE needsVerify predicate (structural exactly-once partition); kind vote in clearer every-form (#1, consensus critical; #21, #22) - Design verifier sees the finding's recommendation — the proposal the applicability rubric actually tests — and carries an escape hatch: a genuine defect mis-filed under a design lens is not refuted away (#12, #3); "solo" dropped from verifier prompts (#4) - Untagged findings from multi-lens cluster finders fall back to 'unspecified' (safe defect bucket), not lenses[0] (#7); merge-agent free-text lens validated, majority-member fallback (#9) - Improvement invitation scoped to design finder units — defect-lens finders stay defect-only (#8); merge prompt clusters by issue, not only defect (#13); schema descriptions generalized for design findings (#14) - LENS_BRIEF startup assertion (#6); gate prompt interpolates LENS_CLUSTERS.design (#19); new test_lens_sync.py guards all lens mirrors: SKILL HDR prompt, LENS_BRIEF, pr-post DESIGN_LENSES (#5) - Workflow assigns stable finding num (defects first, shared sequence); presenter/pr-post render it verbatim (#20) - pr-post.py: design lens is the backup kind signal when the handoff drops kind (explicit defect still wins); single-pass partition (#17, #24) - Balance: REFUTED is its own segment (refuted ⊄ solo since design consensus can be refuted) (#2); LOCKED design-table column precedence in --loop rounds clarified (#11); "verifies solos" doc sweep across manifests/README/knowledge (#4); cluster failure-isolation trade-off documented (#10); knowledge index line trimmed (#15) Declined: #16 (JS test harness for sandbox code; lens-sync test covers the drift class), #18 (gate-fail under --max runs all lenses — the documented never-silently-narrower degrade), #23 (micro-opt vs readability). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126cGxsBYxqgEFH4AcpriNk
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
1 participant
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
/swarm:review(Swarm P2): a local mixture-of-agents code review that fans one diff across four voices — Claude lenses +codex+grok-build+composer— and merges them into one ranked report.composer(grok-composer-2.5-fast) joins as a fast fourth voice via a defensive-parse adapter path. Consensus counts model families, so the two grok voices agreeing is one vote — aCONSENSUStag needs ≥2 of claude/openai/grok.Changes
scripts/agents.sh): composer voice via--model grok-composer-2.5-fast— strict-JSON prompt (no--json-schema/--effort), defensive parser that prefers the non-empty findings object, per-itemfinding.schema.jsonvalidation, argv-size guard, and an--effort-ignored note.workflows/swarm-review.js): scope+gate → 4-voice fan-out →(file, mechanism)merge with family-aware consensus → adversarial 3-state verify of solos → output-gated ranked synthesis.skills/review/SKILL.md):/swarm:review [ref | --staged | pathspec]— builds the diff + fenced prompt, probes live backends, runs the workflow, renders the report.0.1.0 → 0.2.0(plugin.json + marketplace.json), README, two knowledge entries.Readiness
/swarm:review, pipeline + security sections)check-structure.pygreen (0 errors)Test plan
/swarm:reviewon a real diff with all backends live; confirm 4 voices, a ranked report, and a balance footerCONSENSUStag appears only when ≥2 of claude/openai/grok agree (composer + grok-build alone stays solo)python3 scripts/check-structure.pystays green🤖 Generated with Claude Code
https://claude.ai/code/session_017rpoJ2DokY1SMfv1dSfZwe