Bullswarm codebase audit — 2026-09-09

Workspace: /Users/cowcow02/Repo/bullswarm-worktrees/audit-0.27, branch audit/codebase-2026-09-09, HEAD b8487b1. package.json says 0.26.0; CHANGELOG.md:3 says 0.27.0.

Words used in this report, defined once:

• Tier A — dead: nothing outside its own tests refers to it. Safe to delete in a patch release.
• Tier B — superseded but still wired: an older way of doing something is still connected to the live code path. Deleting it needs a decision, not just a patch.
• Tier C — refactor: the code works, but one job is implemented more than once.
• Reachable — the file can be arrived at by following import statements from a real entry point (bin/bullswarm.js, mcp/server.mjs, scripts/*.mjs, the paths in package.json, and the node commands in .github/workflows/*.yml).
• Refuted — a lens called something dead, and an experiment proved it is used.
• Pool — one configured provider account Bullswarm can send work to.
• Rung — the (pool, effort tier) → (model, reasoning level) choice, added in 0.26.0.

---

Table of contents

1. Verdict
2. Tier A — dead code to delete in 0.27.1
3. Tier B — superseded but still wired
4. Tier C — refactor opportunities, best value first
5. Command surface
6. README and documentation
7. Refuted candidates — looked dead, is used
8. Audit-run incidents — what the audit learned about its own workers
9. How this was measured
10. Appendix A — full command-surface table (all 64 paths)
11. Appendix B — analyzer scripts, verbatim

---

1. Verdict

Very little of this codebase is dead: 64 source files are reachable and 2 are not, and both of those two are executed by worker processes through a shell command rather than an import, so they are alive too. The genuinely dead material is 187 lines across 11 items — five test-only helper functions, twelve one-line alias exports nobody imports, one unused constant, one dead string in a frozen list, one dead arithmetic operand, one if-statement whose two branches are identical, and a 36-line fixture pack — and the refutation pass deleted all of it at once for a green suite: 659 tests before, 651 after, 0 failures. That is the whole 0.27.1 patch scope, and eight verdicts from the three code lenses were reversed or downgraded before reaching it, including two files that every worker prompt executes.

Three things need a decision rather than a patch: the 90 %-of-quota burst gate exists in two generations and the older one wins, so two pools at 95 % of their 5-hour window make the CLI return no eligible pool and exit 1 while the newer router hands back a pick; state.json is a last-writer-wins file with no lock, and a concurrent strategy set-provider beta off --yes during a 9-second run was silently undone; and run --dry-run, documented as writing nothing, added 186 lines to state.json and performed a live datapack download.

The command surface is structurally strong and weak on validation — all 64 help paths render, exit 0 and carry all seven required sections, but every command silently ignores unknown flags (pools --bogus-flag and workflow runs list --limit nope both exit 0) and three --json flags (health, workflow capabilities, strategy inventory) produce byte-identical output with and without them. Documentation is mostly accurate: of 17 files under docs/, 2 are stale-misleading, 3 are current and 12 are historical-keep of which 10 need a banner. Its two worst defects are that package.json says 0.26.0 while the changelog, README and AGENTS.md all describe 0.27.0 as shipped, and that README promises a bundled data/openrouter-benchmarks.json that does not exist.

The audit’s own workers failed in two fixable ways. A worker’s stdout outgrew Node’s maximum string length and threw RangeError: Invalid string length inside src/lib/watch.js, killing the detached kernel and every worker under it — twice — and nothing in events.jsonl, result.json or state.json records it, because the trace exists only in ~/.bullswarm/goals/<runId>/stderr.log, so workflow watch showed nothing but a run that stopped moving. Separately, one worker on the command-code pool spent 4 minutes 43 seconds to return a 253-byte statement of intent instead of findings, which the content gate correctly rejected — and every number in this verdict reappears below with the command that produced it.

---

2. Tier A — dead code to delete in 0.27.1

The refutation pass deleted each item in an rsync copy of the tree at /tmp/bullswarm-audit/copy/ and ran the suite. Baseline in that copy: # tests 659 # pass 659 # fail 0.

The table

File or export	Lines	Superseded by	Only referenced by	Evidence
recordAgentAction — src/lib/agent-events.js:179-218	40	0.10.0 agent-event/TUI generation	tests/agent-events.test.js:7,80,100	sed -n '179,218p' src/lib/agent-events.js \| wc -l → 40. Case R9: delete → 655 tests, 654 pass, 1 fail, and the one failure is its own test file
aggregateUsage — src/lib/usage.js:135-183	49	0.8 usage accounting; watch.js uses estimateInvocationUsage	tests/usage.test.js:4,65,77,87	sed -n '135,183p' src/lib/usage.js \| wc -l → 49. Case R11: → 654 / 653 / 1, own test file only
fixtures/openrouter/{benchmarks,models}.json	36	never superseded — never wired	nothing	wc -l fixtures/openrouter/*.json → 23 + 13 = 36. grep -rIn 'fixtures/openrouter' . → no output. Case R5: → 659 / 659 / 0
parseEvidenceOutput — src/workflow/evidence-output.js:141-157	17	durable candidate JSON + bin/check-v2-evidence.js; runtime imports readEvidenceCandidate	tests/workflow-evidence-output.test.js:11,34,36,37	sed -n '141,157p' src/workflow/evidence-output.js \| wc -l → 17. Case R6: → 652 / 651 / 1, own test file only
classifyAgentProgress — src/lib/agent-events.js:220-233	14	same generation as recordAgentAction	tests/agent-events.test.js:7,109,110	sed -n '220,233p' src/lib/agent-events.js \| wc -l → 14. Case R10: → 655 / 654 / 1
12 alias exports, 6 files (list below)	14	each is a one-line re-export of a live symbol	nothing — src=1 tests=0 docs=0 each	Case R14: all 12 at once → 659 / 659 / 0, no test touched
fiveHourTier — src/lib/route.js:99-103 plus its comment 94-98	11	0.25.5 — the tier is recomputed inline on the forecast at route.js:317	tests/route.test.js:4,129-134	grep -rn '\bfiveHourTier\b' --include='*.js' . \| grep -v node_modules \| grep -v '^\./tests/' → one line, the definition. Case R1b: fn + its test → 658 / 658 / 0
REASONING_DEFAULT_TIERS — src/lib/strategy.js:85	1	0.26.0 — the wizard’s tier step now writes nothing on Enter	nothing at all	sed -n '85p' src/lib/strategy.js → export const REASONING_DEFAULT_TIERS = Object.freeze({ high: 'xhigh', medium: 'high', low: 'medium' });. git grep -n REASONING_DEFAULT_TIERS 4b18d2b -- src/ → that line only. Case R2: → 659 / 659 / 0
Inert --json ternary — src/strategy-cli.js:684	1, in place	—	—	sed -n '684p' src/strategy-cli.js → console.log(opts.json ? JSON.stringify(value, null, 2) : JSON.stringify(value, null, 2)); — both branches identical. Case R17: collapse → 659 / 659 / 0
'workflow-v1' inside ASSIGNMENT_SOURCES — src/lib/assignments.js:31	1, in place	0.27.0 — the V1 engine is gone (e8c841a)	prose at skill/references/operations.md:217	sed -n '31p' src/lib/assignments.js → export const ASSIGNMENT_SOURCES = Object.freeze(['run', 'workflow-v1', 'workflow-v2']);. Every live writer passes 'workflow-v2' or 'run'. Case R3: → 659 / 659 / 0
pool.weeklyUsedPct operand — src/lib/spend.js:381	1, in place	—	only that read	sed -n '381p' src/lib/spend.js → return num(pool?.weeklyUsedPct) ?? rate.windowUsedPct ?? null;. grep -rn weeklyUsedPct --include='*.js' --include='*.json' . → one line, the read itself; config.js writes usedPct and fiveHourUsedPct, never weeklyUsedPct. Case R4: → 659 / 659 / 0

The 12 alias exports, each verified present by sed:

src/workflow/v2-scheduler.js:167  export const scheduleActions = scheduleV2Actions;
src/workflow/v2-scheduler.js:168  export const getReadySet = scheduleV2Actions;
src/workflow/v2-scheduler.js:169  export const selectReadyActions = scheduleV2Actions;
src/workflow/v2-state.js:8        export const GOAL_SCHEMA_VERSION = V2_GOAL_SCHEMA_VERSION;
src/workflow/v2-state.js:9        export const STATE_SCHEMA_VERSION = V2_STATE_SCHEMA_VERSION;
src/workflow/v2-state.js:617      export const V2_PLANNER_MODES = Object.freeze([...PLANNER_MODES]);
src/workflow/v2-state.js:649      export const assertV2ResumeCompatible = assertV2Resume;
src/workflow/ownership.js:173     export const verifyOwnership = checkOwnership;
src/workflow/v2-runtime.js:1308   export const runAutonomousV2 = runV2AutonomousWorkflow;
src/workflow/evidence-output.js:138  export const validateV2EvidenceOutput = validateEvidenceOutput;
src/workflow/evidence-output.js:139  export const validateEvidenceEnvelope = validateEvidenceOutput;
src/setup.js:134-136              export function integrationBlock() { return awarenessBlock(); }

Eleven one-liners plus a three-line pass-through = 14 lines. Sibling aliases that do have importers were kept: createV2State (37 test references), serializeV2ResultEnvelope (12), parseV2PlannerResponse (3), changedManifestPaths (2).

Proposed 0.27.1 patch scope

Delete these exports and functions. recordAgentAction, classifyAgentProgress, aggregateUsage, parseEvidenceOutput, fiveHourTier and its doc comment, REASONING_DEFAULT_TIERS, and the 12 alias exports listed above.

Edit these three lines in place. Remove 'workflow-v1' from ASSIGNMENT_SOURCES (assignments.js:31); remove the num(pool?.weeklyUsedPct) ?? operand (spend.js:381); collapse the identical ternary (strategy-cli.js:684).

Delete this directory. fixtures/openrouter/ — but read the risk note below first.

Tests to delete or rewrite — 8 tests, 4 files.

Test file	Change
tests/agent-events.test.js	drop the import and 3 tests (:72, :89, :107); keep everything covering createAgentEventDecoder
tests/usage.test.js	drop the import and 3 tests (:62, :74, :85); keep estimateInvocationUsage coverage
tests/route.test.js	drop the import and the 5h tier: test (:127-136), but move its FIVE_HOUR_NEAR_LIMIT_PCT === 75 assertion into the next test — that constant is still live and README:249 documents the number
tests/workflow-evidence-output.test.js	drop the import and 1 test (:33)

Total lines removed: 187. This is the sum of the per-file wc -l deltas the refuter measured when it deleted the whole set at once (case R20): route.js 571→560 (11), evidence-output.js 187→167 (20), agent-events.js 233→177 (56), usage.js 183→133 (50), v2-scheduler.js 169→166 (3), v2-state.js 649→645 (4), setup.js 582→578 (4), ownership.js 173→172 (1), v2-runtime.js 1308→1307 (1), strategy.js 776→775 (1), fixtures/openrouter/ 36→0 (36) — which sums to exactly

1. Test files shrink too: route.test.js 458→448, workflow-evidence-output.test.js 115→107, agent-events.test.js 122→78, usage.test.js 98→61 — 99 test lines.

One arithmetic discrepancy, stated plainly: the itemised column in the table above sums to 185, not 187. The gap is 2 lines and has two causes — three rows are in-place edits that remove zero net lines, and the per-file deltas include blank separator lines that an itemised count of declarations does not. The 187 is the empirical wc -l figure and is the number to trust; the exact line-by-line reconciliation of the last 2 lines is unverified because I did not re-run case R20 myself.

Expected test-count change: 659 → 651. Measured in this workspace today:

$ env -u CLAUDE_CONFIG_DIR -u FORCE_COLOR -u NO_COLOR npm test
# tests 659
# suites 0
# pass 659
# fail 0
# cancelled 0
# skipped 0
# todo 0
# duration_ms 12048.127875
exit 0

So the answer to “659 or 661” is 659 in this workspace. The 661 figure is real but belongs to a different commit: 16dddb0 on branch feat/one-workflow-engine says “Suite 661/661” in its message, and that commit is not an ancestor of this HEAD (git merge-base --is-ancestor 16dddb0 HEAD → not an ancestor). A patch built on feat/one-workflow-engine should therefore expect 661 → 653.

Risks.

1. fixtures/openrouter/ is a working input, not rot. It is the only key-free way to exercise the datapack builder, and it works today: node scripts/refresh-openrouter-benchmarks.mjs --benchmarks-file fixtures/openrouter/benchmarks.json --models-file fixtures/openrouter/models.json --output /tmp/… → exit 0, wrote 1 models and 2 benchmark records. Deleting it is safe for the suite and removes the only offline exercise of that script. Preferred: add one test that runs the script in fixture mode, and keep the directory. That test would also have caught the missing data/openrouter-benchmarks.json (see Tier B).
2. pool.weeklyUsedPct is a choice, not a cleanup. Deleting the operand makes the fall-through to rate.windowUsedPct honest. Writing the field in buildPools would instead fix projectedWeeklyPct returning null for pools that do have a pacing reading — a null that currently reads as “unmeasured” when the measurement exists. Pick one deliberately.
3. fiveHourTier deletion must not take the threshold with it. FIVE_HOUR_NEAR_LIMIT_PCT is live and its value (75) is documented at README.md:249. Deleting the function alone without fixing the test import fails loudly: case R1 gave not ok 17 - tests/route.test.js, 631 tests, 1 failure, because the whole test file failed to import.
4. Deleting 'workflow-v1' changes behaviour by design. The ledger will start rejecting a workflow-v1 record instead of accepting one no live writer can produce. If any installed home holds pre-0.27.0 assignment records with that source, reading them becomes an error. Not verified against a real installed home — unverified.
5. The alias exports may be someone’s API. All 12 have zero importers in this repo. A static audit cannot see an external import { runAutonomousV2 } from 'bullswarm/…'. Since src/ is not a declared entry point in package.json, the risk is low, but it is a judgement call, not a measurement.

Acceptance checks the patch must pass.

1. env -u CLAUDE_CONFIG_DIR -u FORCE_COLOR -u NO_COLOR npm test → # pass 651 # fail 0 (or 653 on feat/one-workflow-engine).
2. tests/route.test.js still asserts FIVE_HOUR_NEAR_LIMIT_PCT === 75.
3. grep -rn 'fiveHourTier\|REASONING_DEFAULT_TIERS\|aggregateUsage\|recordAgentAction\|classifyAgentProgress\|parseEvidenceOutput' --include='*.js' src bin mcp scripts tests → no output.
4. node bin/bullswarm.js run --lane build --effort medium --dry-run --json --no-caller --prompt hi under a fixture home still routes and still reports nearFiveHourLimit and forecastGated.
5. node /tmp/bullswarm-audit/graph.mjs (run from the repo root) still reports 64 reachable and 2 unreachable — the two bin/check-v2-*.js files, which this patch does not touch.
6. node bin/bullswarm.js strategy inventory --json still emits a parseable JSON document after the ternary collapse.
7. skill/references/operations.md:217 no longer names workflow-v1 as an assignment source.

---

3. Tier B — superseded but still wired

Each entry: what the current flow uses instead, who still references the old thing, the exact code path that changes, and a recommendation.

B1 — Two generations of the 90 % burst gate, and the older one wins

• Current flow uses: route.js:321 gates on the forecast — gated: forecast.forecasted && forecast.forecast != null && forecast.forecast >= BURST_BLOCK_PCT — and route.js:373-388 keeps a least-loaded fallback so work is never stranded.
• Still referenced by: the older reading-based pre-filter, at src/cli.js:241-242 (const gated = pools.filter((p) => p.burstGate); / const ungatedPools = gated.length ? pools.filter((p) => !p.burstGate) : pools;) and src/workflow/v2-dispatch.js:51 (if (pool.enabled === false || pool.burstGate === true || isQuarantined(pool, now)) continue;).
• What breaks: the pre-filter removes gated pools before pickPool ever sees them, so pickPool’s own “if every capable pool is gated, still name one” branch is unreachable from the CLI. Measured with two pools at 95 % of their 5-hour window: the CLI returned {"ok": false, "why": "no eligible pool"} and exit 1, while pickPool called directly with the same pools returned pick: alpha, why: every capable pool is forecast-gated at/above 90% of its 5h window; least loaded wins. Opposite outcomes from the two generations.
• Exact code path that would change: make route.js gate on forecast.forecast ?? raw (drop the forecasted guard), then delete cli.js:241-242 and the burstGate test in v2-dispatch.js:51, keeping pool.burstGate only as a display flag for cli.js:92.
• Recommendation: fold into route.js. This is the highest-value item in the section. Not deletable as-is — the reading gate is currently the only thing covering a pool with no spend model, because route.js:321 requires forecast.forecasted, which is false without a rate or a producer projection. Doing it makes the 0.25.5 promise at CHANGELOG.md:210-212 true on both entry points.

B2 — strategy recommend alias, with zero test coverage

• Current flow uses: strategy refresh. sed -n '794p' src/strategy-cli.js → if (sub === 'refresh' || sub === 'recommend') { — one handler, identical behaviour.
• Still referenced by: src/help.js:391 calls it an alias, and help.js:568-578 documents it as a separate command with its own usage block and its own HELP_PATHS entry.
• Exact code path: drop || sub === 'recommend', delete strategyRecommendText (help.js:568-578) and its HELP_PATHS entry, re-run tests/help.test.js.
• New finding: the alias has no test at all. Case R19 deleted the branch and the suite still passed 659/659. A documented command path with zero coverage can be broken silently.
• Recommendation: keep the alias, add one dispatch test — or delete alias and help block together. What must not continue is documenting a command that nothing tests.

B3 — strategy show is presented as read-only but refreshes on a cold cache

• Current flow uses: strategy inventory and strategy rungs as the structured read surfaces.
• Still referenced by: strategy-cli.js:806-810. On a cold cache show calls refreshStrategy(useOpenRouter: true), which spawns model discovery (strategy-cli.js:207), downloads the datapack with force: true (:211, bypassing the 24-hour TTL) and writes state.json (:217-220). src/help.js:601 already admits this.
• Exact code path: make show print no cached report; run \strategy refresh` instead of refreshing. That is a behaviour change, so help.js, tests/strategy-cli.test.js and the strategy tui bootstrap (strategy-cli.js:592-593`) move together.
• Recommendation: keep the command, split the refresh out. Not re-probed in the refutation pass because confirming it needs the live network fetch the routing lens already exercised — the cold-cache network behaviour is therefore carried from one lens run, not independently reproduced.

B4 — Three inert --json flags

• Current flow uses: nothing — the flag changes no output.
• Still referenced by: help.js:422,429 document --json on strategy inventory/routes as an option whose default is a different format. There is no human renderer for either verb.
• Measured today, fixture home /tmp/bullswarm-audit/report-home.8wkEbT:

$ node bin/bullswarm.js health          # exit 1
$ node bin/bullswarm.js health --json   # exit 1
$ cmp -s h1.txt h2.txt && echo identical
HEALTH BYTE-IDENTICAL (7 lines)

$ node bin/bullswarm.js workflow capabilities         # exit 0
$ node bin/bullswarm.js workflow capabilities --json  # exit 0
CAPABILITIES BYTE-IDENTICAL (239 lines)

$ node bin/bullswarm.js strategy inventory  vs  --json   # capturedAt normalised
INVENTORY IDENTICAL (2404 lines)

cmdHealth(opts) at src/cli.js:474 never reads opts.json — zero matches for opts.json in lines 474-500. The static flag analyzer agreed: counts.json lists documentedButUnread: 1 for both cli health and workflow capabilities.

• Exact code path: either delete the flag from the handlers and from help.js:422-431, making JSON the only documented format, or write the human renderer the help promises.
• Recommendation: delete the flags and fix the help for strategy inventory/routes; keep --json on health and workflow capabilities as accepted no-ops but say in help that output is always JSON, because agents already pass the flag. The one-line dead ternary at strategy-cli.js:684 is Tier A and goes now.

B5 — assignments[tier] versus modelTiers — the largest two-generation overlap

• Current flow uses: both. assignments[tier] = {pool, model} is a hard pool pin that becomes pickPool’s preferredPool (cli.js:260); modelTiers[pool][model] = [tiers] is the allow-list that becomes allowedModels. 0.26.0 says rungs are a projection of strategy.modelTiers and strategy.reasoning (CHANGELOG.md:148-150) and never mentions assignments.
• Still referenced by: 8-9 non-test read sites — cli.js:243,260, v2-dispatch.js:64,125,222, strategy.js:378,704, config.js:79 — plus five hand-written delete state.strategy.assignments[tier] invalidations at strategy-cli.js:502,720,783,899 and strategy-dashboard.js:335.
• Exact code path: set-rung would first have to gain a pool-pin (or preferredPool be derived from modelTiers), then delete assign/clear-assignment, the eight read sites, the five invalidation lines, and the workflow capabilities field at cli.js:1203.
• Recommendation: keep both for now. They are not interchangeable: only assignments can pin a pool, only modelTiers can express an allow-list, and applyStrategyRecommendations writes both. This is a project, not a patch. Fix C7 (one clearTierAssignment helper) as the cheap step.

B6 — TIER_CONTEXTS, plus a third inline copy

• Current flow uses: DEFAULT_EFFORT_BY_LANE and KIND_DEFAULTS (src/workflow/action-validator.js:10-25) as the canonical lane/effort tables.
• Still referenced by: src/lib/strategy.js:632-648, 3 non-test references, plus an inline third copy at strategy-cli.js:366: const context = report.suggestions?.[tier]?.requirements ?? { lane: ({ high: 'analyze', medium: 'build', low: 'chore' })[tier], capabilities: [] };
• The disagreement: TIER_CONTEXTS says high→analyze, medium→build, low→chore; DEFAULT_EFFORT_BY_LANE says analyze→medium, build→medium, chore→low; KIND_DEFAULTS puts high on analyze and on build.
• Exact code path: delete the inline fallback at strategy-cli.js:366 and derive TIER_CONTEXTS[tier].lane from the 0.26 tables.
• Recommendation: keep TIER_CONTEXTS, delete the inline copy. TIER_CONTEXTS also carries per-tier capability requirements that nothing else does, so it cannot simply go.

B7 — preferredConcurrency promises a cap nothing enforces (moved down from Tier A)

• Current flow uses: nothing enforces it. git grep -n preferredConcurrency e8c841a^ -- src/ → src/workflow/runtime.js:202: const preferred = Number(connector.preferredConcurrency); — the V1 authored-graph executor that e8c841a deleted in 0.27.0. 0.25.5’s load spread only prefers a quieter pool.
• Still referenced by: src/setup.js:270 copies the field forward on upgrade, and tests/setup.test.js:162 asserts that. grep -rn preferredConcurrency --include='*.js' src/ → only setup.js:270.
• Why it moved out of Tier A: case R13 deleted the key and its $comment- and the suite went 659 → 658 pass, 1 fail: not ok 176 - connector metadata upgrades additive provider concurrency preferences.
• Exact code path: either re-implement the cap in route.js/v2-dispatch.js, or delete the key, its $comment-, the setup.js:270 entry and the setup assertion together.
• Recommendation: decide, do not drift. The connector’s own $comment- documents a real OpenCode failure mode (sibling workers killing each other). Right now connectors/opencode2.json advertises a constraint the engine does not apply. Which way to go depends on real OpenCode behaviour, which a read-only repo audit cannot settle — unverified.

B8 — connectors/_schema.json is dead to code but is the only connector-format reference

• Current flow uses: nothing reads it. setup.js:89 skips _-prefixed files, and case R12 deleted it with 659 / 659 / 0.
• Still referenced by: package.json files[] ships connectors/, and docs/studies/portal-token-diet.md:78 cites it.
• Recommendation: keep the file, fix its content — see B9. It is documentation that happens to live in a .json.

B9 — json-field is advertised but never implemented

• Advertised: sed -n '17p' connectors/_schema.json → "strategy": "stdout|stdout-tail|json-field|file|event-stream",
• Implemented: src/lib/watch.js:255-283 has exactly four case labels — event-stream, stdout, stdout-tail, file. No json-field. A connector declaring it would fall through silently.
• Related: shipped connectors declare only event-stream (5 of them) and stdout (echo), so stdout-tail and file are implemented-but-unused compatibility paths.
• Recommendation: delete the json-field option from the schema line. Implementing it has no demand — no connector asks for it.

B10 — Deleting a verification fixture silently removes a test case

• Current flow uses: tests/verify.test.js:12-21 readdirSyncs fixtures/failures/ and fixtures/real/.
• The hole: case R16 deleted fixtures/failures/01-rate-limit-head.json and the suite reported 658 tests, 658 pass, 0 fail — coverage shrank by one case with no failure. The only guard is a floor at verify.test.js:23 (fixtures.length >= 17), and there are 23 fixtures (find fixtures/failures fixtures/real -type f | wc -l → 23).
• Recommendation: raise the floor to the current count, or enumerate the fixtures explicitly. Cheap, and it closes a way to lose coverage without noticing.

B11 — The claude alias in READERS is dead in this repo only

• Still referenced by: src/meters/registry.js:19 and readerFor at :58. Case R18 removed both with 659 / 659 / 0. git log -- connectors/claude.json → no commits, ever; no packaged connector is named claude (the packaged names are claude-code, codex, command-code, echo, grok, opencode2).
• Recommendation: hold. Only an already-installed, hand-written home carrying a claude.json could still need it. Whether such a home exists is unverified — a repo audit cannot see installed homes. 2 lines; not worth the breakage risk without that check.

Tier B items re-checked and confirmed still wired — keep as they are

execution-policy.js (23 lines, 6 non-test importers) · workspace-report.js (33 lines, imported at v2-runtime.js:35) · v2-presentation.js (101 lines, 4 non-test refs — Tier C candidate, not deletion) · steering.js (74 lines, 5 non-test refs) · preflight scout (7 non-test, 8 test refs) · gap planning (automaticGapRounds: false for programs, but gap boundaries live at v2-outcome.js:176-219) · workflow capabilities.engines.authoredGraphs (asserted at tests/workflow-goal.test.js:194-198; deleting it would turn explicit retirement into an ambiguous missing field) · --strict-orchestrator and the planner compatibility flags (8 non-test refs) · top-level runs alias (cli.js:694-695) · workflow watch --classic and workflow events (20 non-test, 10 test refs) · integrate retire-legacy (11 non-test refs; it moves, does not delete) · bumpVersion (used internally at release.js:36 — only the export keyword is surplus) · verify.js content gate (judgeContent, 6 non-test and 7 test refs) · invocation usage estimation (estimateInvocationUsage, 3 non-test refs) · health versus workflow runs result · setup/strategy overlap · strategy set-subscription (consumed by config.js:75-78) · claude-accounts.js and opencode-relay.js (both are the current generation of their mechanism — see section 7).

Behaviour findings that are not dead code but should ship with the patch

These came out of the lenses’ live probes. They are not deletions, and two of them are more serious than anything in Tier A.

#	Finding	Where	How it was seen
D5	state.json is last-writer-wins with no lock, and a long run holds a stale copy for its whole duration. A concurrent operator write is silently discarded.	src/lib/state.js:45-51 (bare writeFileSync, while atomicWriteFileSync already exists at src/workflow/fsjson.js:18-23); src/cli.js:206,218,417; src/workflow/v2-dispatch.js:110-121	With a 9-second run in flight, strategy set-provider beta off --yes set beta.enabled = false (confirmed mid-run); after the run’s own saveState, beta.enabled was true again, and only 1 decision-log entry survived. The same mechanism can erase a quota quarantine.
D3	run --dry-run can rewrite state.strategy and hit the network.	maybeRefreshStrategy at src/cli.js:217, 51 lines before the dryRun check at :268	On a home with autoApplyRecommendations: true and a stale lastRefreshedAt, a run --dry-run --json added 186 lines to state.json, and the persisted report’s openRouter.cache read "refreshed" — a live HTTPS fetch. Contradicts CHANGELOG.md:238-240. The decision-log half of that promise is true; the strategy half is not.
D1	Read-only commands mutate state.json on first use, and the fixture migration force-disables a pool the operator explicitly enabled.	src/cli.js:672; src/setup.js:288-311,366-372	One bullswarm pools turned echo.enabled from true to false and added testFixturesMigrated: true. The comment at setup.js:285-287 claims explicit choices are respected; it cannot tell a post-migration choice from a legacy default.
D6	Disabled pools are still polled.	src/lib/config.js:161-164 polls, :89 discards	With beta.enabled = false, an injected getReadings was still asked for ["alpha","beta"]. On a real home that means a disabled pool’s credentials are read and its usage endpoint called on every pools, every run, and every 15-second V2 refresh.
D7	A misleading routing reason when the tier allow-list is empty.	src/lib/route.js:356-367	strategy routes reported medium → "no eligible pool with capabilities: code-reading, file-editing" while all three pools declared both. The true cause, from strategy rungs --json, was "modelSource": "tier-selection-empty".
—	The content gate accepts error-shaped structured output.	hasStructuredAnswer() at src/lib/verify.js:139-153	judgeContent('{"error":"provider unavailable"}') → {"verdict":"pass","why":"content passed all gates"}, and a controlled watchOnce returning that body produced {"ok":true,"why":"verified"}. Related: scanForFailure() only inspects the first 400 characters of outputs ≥600 characters (verify.js:92-96), so a 401-character prefix followed by Error: provider unavailable passed.
—	MCP exposes no workflow tool.	mcp/server.mjs:12-39	Handshake returned version 0.26.0 and exactly three tools: bullswarm_run, bullswarm_health, bullswarm_pools. No way to launch or read a workflow over MCP. Either add one or say the limit is intentional.

---

4. Tier C — refactor opportunities, best value first

Ordered by value over risk: cheap and contained first.

#	Step	Size	Risk	Why it is worth it
C1	Delete the inline third lane table at strategy-cli.js:366 and derive TIER_CONTEXTS[tier].lane from the 0.26 tables.	1 line deleted, ~5 changed	Low	Three tables describe one relation and two disagree (B6). One-line fix, removes a whole class of drift between the strategy preview and the V2 validator.
C2	One clearTierAssignment(strategy, tier) helper to replace five hand-written delete state.strategy.assignments[tier] lines (strategy-cli.js:502,720,783,899, strategy-dashboard.js:335).	~5 → 1	Low	Every writer of modelTiers must currently remember to clear the overlapping assignments entry. Forgetting once leaves the two stores disagreeing about routing.
C3	Export one strict finiteOrNull and delete the five copies. meters/framework.js:148 numberOrNull · route.js:114 num · spend.js:64 num · config.js:130 finiteOrNull · openrouter-models.js:15 finite · epoch-benchmarks.js:62 finite.	~30 → ~6	Low	Not just noise: four of the six reject '' and booleans (the “a missing measurement never becomes a zero” rule) and two do not, so Number('') === 0 slips through in the two datapack loaders. This is a latent correctness gap.
C4	Derive GENERIC_QUOTA_SIGNATURES from DEFAULT_QUOTA_SIGNATURES instead of restating 6 of 14 phrases by hand (quota.js:28-44, :54-61).	~8	Low	A phrase added to one list and not the other changes quota detection asymmetrically.
C5	Extract one eligibleModelsFor(pool, tier, …) and build both projections from it — buildStrategy currently scores the same candidates twice (strategy.js:663-687 and :688-730 inside a 104-line function).	68 of 104	Low-medium	Two nested pools × tiers loops that both call recommendationModels, both filter on autoRecommend !== false, both apply the exclusion set, and differ only in whether the winner is per-pool or global. Blast radius is contained to tests/strategy.test.js.
C6	Extract one renderTable(header, rows) and one formatRoute(route); strategy-cli.js:48-66 render() and strategy-dashboard.js:113-125 routeLines() print the same three route facts with independently written padding, and the pad-to-column-width algorithm exists twice (strategy-cli.js:335-343, strategy-dashboard.js:108-111).	32 duplicated, across 911 + 551 lines	Medium	Widest formatting overlap in the codebase; the dashboard then keeps only its cursor and inverse-video concerns.
C7	Extract the shared “strip vendor prefix, lower-case, collapse separators” step from the three model-id normalisers (epoch-benchmarks.js:143-151, openrouter-models.js:208-221, usage.js:49-61), and replace openRouterModelKey’s hard-coded 7-vendor regex table with a lookup over the datapack’s own model keys.	34	Medium	A new vendor currently has to be added by hand in two of the three. The datapack already contains every id the table could match.
C8	Centralise lifecycle status classification. src/workflow/status.js defines terminal/delivered workflow statuses; v2-state.js:17-20 separately defines planner, action, attempt and lifecycle sets; dashboard and watch each add local terminal sets.	4 sites	Medium	Preserve legacy completed_with_concerns and the V2 action statuses as distinct compatibility values — the point is one classifier, not one vocabulary.
C9	Split src/workflow/short-id.js into ID/resolution, legacy history, and runner-liveness modules.	276	Medium	It currently combines short-ID generation, run resolution, legacy detection, run listing, process liveness, heartbeat liveness, and ongoing-run classification.
C10	Split src/help.js into per-surface modules, or generate parts from dispatcher metadata.	1,349	Medium	Must preserve HELP_PATHS, the aliases, and the parser-drift tests — those are what make the 64-path guarantee testable.
C11	Extract shared liveness/event primitives from the watch layers: src/lib/watch.js 449, src/workflow/watch-cli.js 831, src/lib/run-heartbeat.js 59, src/lib/agent-events.js 233.	1,572	Medium-high	They serve genuinely different contracts (single-delegate transport, workflow event watch, heartbeat formatting, connector event decoding) but share liveness and status vocabulary. Do not merge the user-facing modes. Note the incident in section 8 lived in exactly this layer.
C12	Make paceSnapshot the single producer of “is this pool near or over its 5h limit”, have config.js:141-151 copy its fields instead of recomputing what framework.js:98-99 already put on the reading, then fold the pre-filter into route.js per B1. Five places currently decide this.	layers total 1,509 (route 571, spend 452, framework 209, config 166, forecast 111)	High	Highest-value correctness fix in the codebase — it resolves B1 — and the highest risk, because it sits directly on dispatch. Requires the routing and forecast suites green before and after.
C13	Split src/workflow/v2-runtime.js: extract planner-boundary handling, action execution, and terminal publication behind tested interfaces.	1,308	High	Currently combines leases, resume reconciliation, scout dispatch, caller/dispatched planning, steering, action execution, evidence transport, isolated workspaces, heartbeats, cancellation, result publication and finalization. Do not attempt without the SIGTERM/SIGKILL, resume, cancellation, evidence and concurrent-action tests.
C14	Split src/workflow/dashboard.js into data projection, timeline formatting, static detail output, and interactive terminal control.	1,920	High	Largest file in the repo. Still uses V1-era “orchestrator” vocabulary while rendering caller-planner runs. Preserve legacy read-only rendering and the current keyboard behaviour.
C15	Separate src/workflow/cli.js goal/planning commands from observation and legacy-compatibility commands.	1,482	High	Parses flags, resolves planner modes, builds goals, launches detached kernels, manages resume/cancel/steer, prints capabilities, runs watch/events, and handles legacy refusal — all in one file.
C16	Generate connector schema/reference documentation instead of maintaining 24 top-level $comment* keys across 7 JSON files plus a 96-line _schema.json.	7 files	Medium	Keeps connector quirks declarative and gives B8/B9 a home that cannot drift from the code.

Do C1-C5 in the patch. Treat C12 as the next real project, because it is the one that fixes a wrong answer rather than tidying a right one.

---

5. Command surface

64 command paths, from node -e "import('./src/help.js').then(m => console.log(JSON.stringify(m.HELP_PATHS)))". All 64 rendered with exit 0, and all 64 contained all seven required rich-help sections (usage, purpose, args/commands, options, safety, example, next). From help-probes.json: help paths: 64 exit0: 64 complete: 64, and the list of paths with any defect is empty.

No public dispatcher verb lacks a help path, and no declared public help path lacks a dispatcher. Five internal dispatcher-only paths stay undocumented on purpose: --version, strategy help, integrate help, workflow plan help, and the detached workflow goal --request/--run-id/--quiet.

Rows with findings

Full 64-row table in Appendix A. Legend: 2U = exit 2 with usage guidance · 2V = exit 2 with validation guidance · 1V = validation message but exit 1, which is wrong · I = bogus flags silently ignored · J* = JSON always emitted or the flag has no effect.

Path	Missing / drift	--json	Bad input	Unknown flag	Fix
run	help omits the accepted --no-caller	yes	missing lane exits 1 with FAIL unknown lane undefined; malformed values correctly exit 2	I	High priority: require the lane and exit 2
workflow goal	none	yes	missing program exits 2 with next commands	I — a bogus flag is accepted; the launch itself was not reproduced by the appendix probe (its case exits 2 for a missing program), so “launched a run” is unverified	High priority: reject unknown flags
workflow plan validate	none	yes	2U on missing/invalid program	I — valid command plus bogus flag still exit 0	High priority: strict flags
workflow runs / runs list	none	yes	2U on time filters, but --limit nope exits 0	I	Reject invalid numeric limits
workflow capabilities	usage omits the advertised --json	J* always JSON	1 on live-pool failure	I	Add [--json] to usage, or delete the flag
health	none	J* always JSON	semantic unhealthy result exits 1	I	Document that output is always JSON
version	package reports 0.26.0, help and release context say 0.27.0	no	—	I	Reconcile version metadata
workflow (root)	one root example omits the required --program, --scout or --orchestrator	no	non-TTY prints help, exits 2	I	Fix the invalid example
integrate + all 4 subcommands	parent help omits install/remove --yes context	yes	1V on bad agent and on missing --yes	I	Change approval/bad-agent errors to exit 2
strategy set-provider, set-model, reset-tier, set-reasoning, set-rung, reset-reasoning, configure, apply, assign, clear-assignment, exclude-model, include-model, set-subscription, auto, auto status, auto off (16 paths)	no advertised JSON, but output is always JSON	not offered; always JSON	2U / 2V	I	Document the output format, or add --json
strategy inventory / strategy routes	none	yes, but inert — see B4	1V/2V on discovery errors	I	Delete the flag and fix help.js:422-431
strategy rungs	none	yes	--pool with no value becomes unknown pool "true", exit 2	I	Improve the missing-value message
strategy tui	none	no	exits 1: requires a TTY	I	Message is clear; exit 2 would be more correct for a usage error
workflow tui	non-TTY without an ID requires --json or an ID; help is broad	yes	bare non-TTY exits 1	I	Clarify the non-TTY refusal in help
setup, pools, assignments, doctor	none	yes	nothing meaningful / readiness failure exits 1	I	Strict flags

The unknown-flag problem, reproduced today

Fixture home /tmp/bullswarm-audit/report-home.8wkEbT, BULLSWARM_DISABLE_CLAUDE_PROFILES=1 BULLSWARM_DISABLE_OPENCODE_RELAY=1, no provider dispatched:

$ node bin/bullswarm.js run --lane analyze --bogus-flag --dry-run probe
exit=0
OK (keep-on-caller) caller pool won the lane; keep work in-session

$ node bin/bullswarm.js pools --bogus-flag
exit=0
claude-code    cost=4 lanes=analyze/build/chore unmetered surplus=0 inflight=0 ready

$ node bin/bullswarm.js strategy rungs --bogus-flag
exit=0
no rungs yet: enable a provider pool and configure an effort tier …

$ node bin/bullswarm.js workflow runs list --bogus-flag
exit=0
no ongoing runs (try --all to see historical)

$ node bin/bullswarm.js assignments --bogus-flag
exit=0
no in-flight assignments

And the two malformed-input cases:

$ node bin/bullswarm.js run --add-dir . --dry-run probe
exit=1
FAIL unknown lane undefined            # should be exit 2 with a required-lane fix

$ node bin/bullswarm.js workflow runs list --limit nope
exit=0                                  # invalid numeric input silently ignored

behavior-probes.json records the same 26 probes from the command lens, including two that show the validation is right where it was written: run --lane analyze --heartbeat nope → exit 2, --heartbeat must be a number..., and run --lane analyze --effort extreme → exit 2, --effort must be high, medium, or low. The gap is that only selected fields are checked.

Non-TTY behaviour is sane everywhere it was tested: piped run --dry-run produced the normal human summary without prompting; bare non-TTY workflow printed help and exited 2; strategy tui exited 1 with strategy tui requires an interactive terminal; workflow tui without an ID exited 1 with workflow dashboard requires a TTY, or pass a run ID for a static text tree.

Overlap decisions

Overlap	Decision
workflow runs vs top-level runs	Keep workflow runs canonical; retain runs as a short alias. The skill should teach the canonical form.
strategy refresh vs recommend	Keep refresh. Retain recommend — but add the missing test (B2).
strategy show vs inventory	Keep both. show is the last strategy report; inventory is the agent-facing full machine document.
strategy routes vs rungs	Keep both. routes answers “what will route now”; rungs answers “which model/reasoning/evidence/record backs each pool-tier choice”. rungs is also the only one that gives the true cause when a tier allow-list is empty (D7).
health vs workflow runs result	Keep both. health is a global verifier/quarantine signal; runs result is the stable per-run delivery envelope.
doctor vs integrate status vs pools	Keep all three: installation readiness, agent integration, routing/quota/load.
workflow capabilities vs pools vs strategy inventory	Keep all three: engine constraints, live quota state, models and assignments.

Skill cross-check

28 unique command forms extracted from skill/SKILL.md and skill/references/operations.md were run with --help. All 28 returned exit 0 (skill-command-help.json: entries: 28 exit0: 28), including equals-form flags such as --cwd=<abs-dir> and --program=plan.json. The documented agent path is covered with no detours: choose run or workflow goal → workflow plan contract → workflow plan validate → workflow goal --program → workflow watch --next → repeat with --after/--since → workflow runs result --json.

Two problems with the skill text:

1. One contradiction. operations.md:60-63 says --classic is a no-op for legacy runs; the CLI and implementation refuse legacy watching with exit 2 before polling. Keep the fail-closed CLI behaviour and fix the skill.
2. One omission that matters for agent safety. The skill does not warn that unknown flags are currently ignored. An agent that typos a flag gets exit 0 and no signal.

Also confirmed from skill-help-checks.tsv: workflow capabilities --json --help prints Usage: bullswarm workflow capabilities — the synopsis really does omit [--json].

Ease-of-use verdict

The hierarchy and the help renderer are genuinely strong. All 64 paths are discoverable, structurally complete, and machine-readable, and the agent flow reads cleanly end to end. Ease of use is moderate for a careful operator and unsafe for a typo-prone agent, for one reason: bogus flags pass through silently and some malformed values return exit 0 or exit 1 where exit 2 is the contract. An agent that misspells a flag on workflow goal does not get an error — it gets a launched run.

The five fixes that matter most

1. Reject unknown flags centrally. Report unknown flag --name, print the relevant help path, exit 2. This single change fixes the I column on all 64 rows.
2. Validate required and typed inputs at the boundary — especially run --lane, workflow runs --limit, flags given without a value, and integration agent names.
3. Align help with actual behaviour: add run --no-caller, add --json to workflow capabilities usage, document the always-JSON strategy mutations, and replace the invalid bare workflow goal example.
4. Correct the --classic statement in skill/references/operations.md.
5. Reconcile 0.26.0 against 0.27.0 across package.json, bullswarm version, the changelog, README and AGENTS.md.

---

6. README and documentation

README.md findings

Every line number below was re-read in this workspace.

| Line | Finding | Evidence | Fix | |—|—|—|—| | README.md:116 | Claims the runs alias includes cleanup operations. It does not. | sed -n '116p' README.md → \| `runs` \| Short alias for `workflow runs`, including list, show, result, delete, and cleanup operations. \|. src/workflow/runs-cli.js:41-49 dispatches only list, show, result, delete. node bin/bullswarm.js workflow runs cleanup → exit 2 | Remove “and cleanup” | | README.md:91 | Says bare bullswarm launches an interactive setup wizard on first run. True only on a TTY. | sed -n '91p' README.md → bullswarm # first run: interactive setup wizard. src/cli.js:678 → if (opts.yes \|\| !process.stdin.isTTY) return cmdSetup({ ...opts, yes: true }); — non-TTY callers self-initialize silently | Qualify as “on a TTY” | | README.md:183-208 | Claims two replaceable benchmark assets with a bundled fallback. The OpenRouter bundled file does not exist. | ls data/ → README.md, epoch-benchmarks.json only. src/lib/openrouter-models.js:13 builds the path and :159 uses it as the bundledFile default. Offline probe: OpenRouter cache=miss models=0; Epoch cache=bundled records=547 | Commit a seed pack, or drop the bundledFile tier so offline behaviour is honest | | README.md:342-346 and README.md:505-510 | The soft planning-target explanation is duplicated almost verbatim. | Both passages open `--max-agents`, `--max-actions`, and `--max-expansion-rounds` are soft V2 planning targets | Keep one canonical explanation; link to it from the other section | | README.md:572 | Comment says --name filters “by workflow”; it is an exact goal/workflow-name filter. | sed -n '572p' README.md → bullswarm workflow runs --name audit-code # filter by workflow; src/workflow/runs-cli.js:6-8 | Say “filter by exact goal/name” | | README.md:152-313 and 612-780 | A 162-line rungs section and a 169-line dashboard section duplicate detailed command help. | Line ranges measured in the file | Keep concise operational guidance; link to command help | | README.md:585, 651 | Legacy-run behaviour and the 0.27.0 executor-removal message match the implementation but conflict with the version the CLI reports. | sed -n '585p' README.md → removed in 0.27.0; files remain under <dir> and exits 2; node bin/bullswarm.js version → 0.26.0 | Resolve with the version reconciliation | | README.md:249, 264 | **Correct.** 75 % near-limit and 90 % forecast gate match the code. | sed -n ‘249p;264p’ README.md → FIVE_HOUR_NEAR_LIMIT_PCT (75) and BURST_BLOCK_PCT (90) ; both live in src/meters/framework.js:47,49 | none | | README.md:511-517, 599-603 | **Correct.** Shared-workspace program execution and the separate verified qualification match src/workflow/cli.js:482-489, 1159-1179. | as cited | none | | README.md:282-298 | **Correct.** In-flight load and the 500-line meter history match src/meters/registry.js:153-166. | as cited | none | | README.md:14-15, 31-35, 98-101, 319-340, 430-463, 568-580 | **Correct.** The documented run, workflow goal, workflow plan, runs, watch, tui, events and action show` surfaces all exist. | checked against the help tree and fixture execution | none |

Other required files

File:line	Finding	Fix
AGENTS.md:34-36	Describes 0.27.0 behaviour as installed. sed -n '34,36p' AGENTS.md → “Their executor was removed in 0.27.0”, while package.json:3 and bullswarm version both report 0.26.0.	Align the version metadata before release, or stop documenting unreleased behaviour as installed
AGENTS.md:41-71	Correct. Test, doctor, workflow, run-history, recursion, dependency and release guidance match the current CLI. The full suite is offline.	none
GOAL.md:26-32	Still describes a four-verb prototype as the intended outcome. sed -n '26,32p' GOAL.md → “bullswarm CLI with four verbs” listing setup, run, health, pools. The CLI now has 64 help paths and exactly two work-entry points.	Mark as a historical prototype charter, or rewrite
GOAL.md:112-119	“No measured routing table yet” is stale against the current strategy, meter, rungs and invocation-telemetry implementation.	Move to historical project rationale
data/README.md:7-12	Says both datapacks have a bundled fallback. sed -n '7,12p' data/README.md → “then the copy bundled in this directory”. The OpenRouter bundled file is missing.	Restore the asset, or correct the statement
data/README.md:42-48	Correct. The Epoch refresh command and source paths are consistent with .github/workflows/refresh-benchmarks.yml.	none
skill/SKILL.md:8-15, 17-23, 44-115	Correct. Caller-planner, shared-workspace, evidence-envelope, routing and observation flow are consistent with the CLI. No stale behaviour claim found.	none
skill/references/operations.md:60-63	--classic is described as a no-op for legacy runs; it is rejected with exit 2 before polling.	Say it applies only to V2 and that legacy runs are refused
skill/references/operations.md:217	Names workflow-v1 as an assignment source — the only remaining mention of the string the Tier A patch deletes.	Update with the patch
skill/references/operations.md:3-31, 67-84, 161-185, 274-278, 311-331	Correct. Workflow, resume, isolation, evidence, recovery and strategy guidance match the implementation.	none

Cross-document contradictions

#	Contradiction	The two sides
1	Release version	AGENTS.md:35 and README legacy text say executor removal happened in 0.27.0, and CHANGELOG.md:3 declares 0.27.0 — but grep -n '"version"' package.json → 3: "version": "0.26.0", and node bin/bullswarm.js version → 0.26.0. The repository is split between its release notes and its installed metadata.
2	Number of entry points	README.md:11-26 and skill/SKILL.md:17-23 both say exactly two ways to start work. GOAL.md:26-32 still presents the four-verb prototype.
3	Planner mode	README.md:316-340 says caller-authored --program is the default and --orchestrator is explicit. docs/claude-dynamic-workflow-mechanics.md:42-46 still says workflow goal “runs an LLM at every checkpoint” without limiting that to the older engine.
4	Workspace policy	README.md:511-517 and src/help.js:770 say shared workspace by default. docs/experiments/2026-09-06-caller-planner-evaluation.md:118-120 says isolated worktrees were the default. Historical, but needs a banner.
5	Action vocabulary	README.md:378-402 defines the V2 kind values. docs/workflow-design.md:70-116, 190-197 still presents run, fanout, verify, decide as the active schema.
6	bullswarm delegate	docs/design/2026-09-06-caller-first-cli.md:123-134 documents delegate as current — sed -n '123,127p' → “## bullswarm delegate” / “delegate keeps its single-agent branch unchanged”. But CHANGELOG.md:81-92 states it was removed, and node bin/bullswarm.js delegate → unknown verb "delegate". Run "bullswarm --help" for the list of commands.

docs/ classification

File	Size / date	Class	Note
docs/claude-dynamic-workflow-mechanics.md	26,589 B; dated 2026-08-29, updated 2026-09-08	Stale-misleading	:42-46 describes every workflow goal as an LLM checkpoint loop, despite caller-planner program execution being current. Rewrite that paragraph to name it as pre-caller-planner behaviour.
docs/design/2026-09-06-caller-first-cli.md	8,606 B; dated 2026-09-06	Stale-misleading	:123-134 presents bullswarm delegate as live after its removal. Remove the section or mark it superseded.
docs/workflow-simplification.md	12,875 B; live acceptance dated 2026-09-08	Current	Claims match current V2 source and tests.
docs/studies/portal-token-diet.md	17,793 B; no self-dated sentence, committed 2026-09-09	Current	Add a study date for reproducibility. Also the only citer of connectors/_schema.json (see B8).
docs/dynamic-workflow-handoff.md	24,899 B; historical 2026-08-21	Historical-keep	Already has an explicit historical banner. Retain as is.
docs/workflow-design.md	10,521 B; created 2026-08-21	Historical-keep	Lines 3-7 already identify it as historical. Retain.
docs/dynamic-workflow-qa.md	6,601 B; verified 2026-08-27	Historical-keep	Add a one-line banner: :61 still shows the old --detach zero-graph launch.
docs/dynamic-workflow-v2-execution-plan.md	44,393 B; written 2026-08-31, baseline 0.21.0	Historical-keep	Add a banner saying it is a dated plan, not the current capability matrix.
docs/integration-audit-2026-08-31.md	25,704 B; dated 2026-08-31, version 0.20.0	Historical-keep	Add a historical banner.
docs/planner-prompt-audit-2026-08-29.md	12,802 B; dated 2026-08-29	Historical-keep	Add a historical banner.
docs/workflow-agent-usability-audit-2026-08-27.md	13,961 B; dated 2026-08-27	Historical-keep	Add a banner; its 0.7.2/0.4.0 version-skew findings are not current product status.
docs/experiments/2026-08-28-trending-ai-autonomy.md	11,965 B; dated 2026-08-28	Historical-keep	Add a historical banner.
docs/experiments/2026-08-29-dogfood-bullswarm-builds-bullswarm.md	49,163 B; dated 2026-08-29, documents 0.13-0.21 runs	Historical-keep	Add a historical banner.
docs/experiments/2026-08-29-ultracode-vs-bullswarm.md	41,870 B; dated 2026-08-29, versions 0.10.9-0.13.1	Historical-keep	Add a historical banner.
docs/experiments/2026-08-31-v2-component-probes.md	3,421 B; dated 2026-08-31, candidate revision 0030e12	Historical-keep	Add a historical banner.
docs/experiments/2026-09-06-caller-planner-evaluation.md	19,291 B; dated 2026-09-06, versions 0.23.2→0.24.0	Historical-keep	Add a historical banner; it is side 2 of contradiction 4.
docs/audits/2026-09-09-codebase-audit.md	this file	Current	—

17 files: 2 stale-misleading, 3 current, 12 historical-keep. Ten of the twelve need a banner — docs/dynamic-workflow-handoff.md and docs/workflow-design.md already have one.

The ten documentation fixes that matter most

1. Align package.json, bullswarm version, CHANGELOG.md, README and AGENTS.md around the actual release version. This is the single most-reported defect: all three code lenses, the command lens and the docs lens each found it independently.
2. Remove the nonexistent runs cleanup claim (README.md:116).
3. Qualify the bare-bullswarm setup behaviour by TTY versus non-TTY (README.md:91).
4. Restore data/openrouter-benchmarks.json or correct the bundled-datapack claim in both README.md:183-208 and data/README.md:7-12.
5. Move GOAL.md to historical project rationale, or rewrite its prototype outcome and non-goals.
6. Correct or supersede the live delegate section in docs/design/2026-09-06-caller-first-cli.md.
7. Clarify the stale checkpoint-loop paragraph in docs/claude-dynamic-workflow-mechanics.md:42-46.
8. Add explicit historical banners to the nine dated audits, experiments, plans and acceptance documents listed above.
9. Consolidate the duplicated README planning-target material and shorten the oversized rungs and dashboard sections.
10. Correct the --classic legacy-run statement in skill/references/operations.md:60-63, and add the warning that unknown flags are currently ignored.

---

7. Refuted candidates — looked dead, is used

Nine candidates the lenses called dead survived an experiment. Where a lens and the refuter disagree, the refuter’s empirical result wins, and it is said so explicitly in each row below.

Candidate	Called dead by	What actually keeps it alive	Result of deleting it
bin/check-v2-evidence.js (26 lines)	the run lens and the reachability report — both wrong; the workflow lens got it right	Not imported by anything. evidence-output.js:174 builds the path with new URL('../../bin/check-v2-evidence.js', import.meta.url), and v2-runtime.js:365 embeds the resulting shell command in every worker prompt. A worker process runs it.	Case R7 → not ok 391 - CLI returns deterministic statuses for valid, invalid, malformed, and usage input
bin/check-v2-plan.js (36 lines)	the run lens and the reachability report — both wrong	Same mechanism: v2-planner.js:508 plus v2-runtime.js:831, embedded in every planner prompt	Case R8 → not ok 577 - planner checker enforces the exact scout unit handoff used by runtime
connectors/command-code.json	the reachability report’s “unreferenced assets” list (which flagged this caveat itself)	Dynamically scanned by src/setup.js:85-112 — nothing names the file as a literal	Case R15 → 648 tests, 644 pass, 4 failures across 3 test files
The 23 fixtures/failures/* and fixtures/real/* files	the reachability report’s “unreferenced assets” list	tests/verify.test.js:12-21 readdirSyncs both directories	Case R16 → 658 tests, 658 pass, 0 fail — the suite went green while losing a test case. This is the most dangerous kind of false positive, and it is now B10.
workflow flags concurrency, max-actions, max-agents, max-expansion-rounds, no-scout, retry-attempts	the reachability report’s flags.json “documented-but-unread” list — a false positive	All six are read, through a rename table in goalSettings() that uses computed access opts[flag] at src/workflow/cli.js:118-131. --no-scout is read as opts.noScout at cli.js:481. A static flag analyzer cannot follow computed property access.	not deleted — refuted by reading the code
connectors/opencode2.json:31-32 preferredConcurrency	the routing lens as Tier A3	src/setup.js:270 copies the field forward on upgrade, and tests/setup.test.js:162 asserts it	Case R13 → not ok 176 - connector metadata upgrades additive provider concurrency preferences. Downgraded to B7. The routing lens’s underlying observation still stands: no routing code reads it.
connectors/_schema.json (96 lines)	the reachability report as Tier A	Genuinely dead to code (setup.js:89 skips _-prefixed files) and case R12 gave 659/659/0 — but package.json files[] ships connectors/, docs/studies/portal-token-diet.md:78 cites it, and it is the only connector-format reference	Downgraded to B8: fix it, do not delete it.
src/lib/claude-accounts.js (202 lines)	the audit brief’s premise (“versus the claude-code:<name> pool mechanism”)	This file is that mechanism. poolNameForSlug (:48-50) mints claude-code:<slug>; expandClaudeAccountConnectors (:160-201) clones the connector per login; config.js:34 calls it on every loadConnectors; registry.js:12,26-40 routes the meter per seat	not deleted — there is no separate mechanism to compare against
src/lib/opencode-relay.js (131 lines)	the audit brief	Live via config.js:35 and registry.js:13,42-55	not deleted. One thing to watch: retargetOpenCodeModel (:31-46) hard-pins --model <providerId>/gpt-5.6-luna into connector.spawn.cmd at load time, and a 0.26 rung appends its own --model later, so two model flags coexist on one command line. Which one the OpenCode CLI honours is unverified — it needs a real provider.
data/openrouter-benchmarks.json	the routing lens as Tier A6	Nothing to delete — the file is already absent. openrouter-models.js:13 builds the path, :159 uses it as the bundledFile default. It is a missing asset, not dead code.	reclassified: a documentation and packaging defect (section 6, fix 4)
Every src/meters/* module (914 lines across 5 files)	the audit brief (“meter frameworks for providers no longer configured”)	Every meter maps to a packaged connector: claude.js→claude-code, codex.js→codex, command-code.js→command-code, grok.js→grok, relay.js→opencode2*, via READERS at registry.js:18-24 and readerFor at :57-65	No orphans. Only two leftovers, 1-2 lines each: the claude alias key (now B11) and monthlyWindowMs (framework.js:155), documented as “Copilot/cmd period-end semantics” while no Copilot connector exists — the function itself is live for command-code.
reconcileInterruptedRun(s)	the audit brief, as a candidate to judge	Already gone. e8c841a removed the V1 stale-owner reconciliation with the V1 executor. The only match left is a historical test comment at tests/workflow-dashboard.test.js:1039.	no action remains
~34 “never-imported” exports in src/lib/route.js, spend.js, quota.js, state.js, strategy.js, forecast.js, reasoning.js, verify.js, watch.js and the meters	the reachability report’s never-imported list	They all have real in-file callers — only the export keyword is surplus, not the code. Examples: route.js:53 elapsedPct used at :84; route.js:67 costOf at :426; route.js:189 inflightLoad at :305; route.js:239 isExhausted at :300,414; state.js:18 DEFAULT_STATE at :32,36,38,41; strategy.js:21 isModelExcluded at :196,217,224,232,698	not deleted — “never imported” is not the same as “never called”
src/lib/release.js:17 bumpVersion	the reachability report and the workflow and run lenses, as a tests-only export	Used internally at release.js:36	keep the function; only the export keyword is surplus
src/help.js:1280 HELP_PATHS	flagged tests-only by the analyzer — the reachability report offered this as its own deliberate false-positive example	It drives the help test enumeration and is exported for that contract on purpose	keep

Eight of these were verdicts the refutation pass reversed or downgraded — the first seven rows of the table above plus data/openrouter-benchmarks.json. The remaining rows in the table refute a premise in the audit brief rather than a lens verdict, or record a deliberate false-positive example.

The pattern worth naming. Every single false positive came from one of three blind spots: a path built at runtime with new URL(...) and executed as a shell command; a directory scanned with readdirSync; or a flag read through computed property access opts[flag]. A static analyzer cannot see any of the three. That is why the refutation pass existed.

---

8. Audit-run incidents — what the audit learned about its own workers

Two things went wrong while producing this report. Both are findings about worker reliability per pool, and both are recorded here because they are more actionable than anything in Tier A.

Incident A — a flooding worker killed the kernel twice, and watch could not tell you

What happened. A command-code worker’s stdout grew past Node’s maximum string length. The stream handler in src/lib/watch.js appends each chunk to one accumulating string, so the append threw, inside an event handler, in the detached kernel process — which killed the kernel and every worker under it. It happened twice during the reachability action.

The trace, from ~/.bullswarm/goals/wf-mttj9k9w-baca81/stderr.log, which contains this block exactly twice:

file:///Users/cowcow02/Repo/bullswarm/src/lib/watch.js:207
      stdout += d;
                ^

RangeError: Invalid string length
    at Socket.<anonymous> (file:///Users/cowcow02/Repo/bullswarm/src/lib/watch.js:207:17)
    at Socket.emit (node:events:518:28)
    at addChunk (node:internal/streams/readable:561:12)
    at readableAddChunkPushByteMode (node:internal/streams/readable:512:3)
    at Readable.push (node:internal/streams/readable:392:5)
    at Pipe.onStreamRead (node:internal/stream_base_commons:189:23)

Node.js v22.17.0

The unbounded append is still present in this workspace:

$ sed -n '205,207p' src/lib/watch.js
    child.stdout.on('data', (d) => {
      stdout += d;

What the operator saw instead. Two interrupted attempts with a generic reason:

reachability  attempt 1  interrupted  03:22:56.433Z → 03:51:36.300Z
              why: "runner stopped before the attempt reached a durable terminal state"
reachability  attempt 2  interrupted  03:51:36.447Z → 04:22:51.149Z
              why: "runner stopped before the attempt reached a durable terminal state"

The observability gap, measured. The string RangeError appears 0 times in events.jsonl, 0 times in result.json and 0 times in state.json for that run. And the event stream shows the shape of the loss directly:

$ grep -o '"type":"[a-z.-]*"' events.jsonl | sort | uniq -c | sort -rn
   9 "type":"action.finished"
   8 "type":"attempt.started"
   8 "type":"action.started"
   6 "type":"attempt.finished"
   2 "type":"workflow.resumed"
   1 "type":"workflow.started"
   1 "type":"workflow.finished"
   1 "type":"planner.finished"

Eight attempts started, six finished. The two killed attempts never emitted attempt.finished, so workflow watch — which reads events.jsonl — had nothing to show but a run that stopped moving. The only record of why was a file watch never reads.

The fix, on the PR branch. Commit 16dddb0 on feat/one-workflow-engine:

$ git show --stat 16dddb0
commit 16dddb00689acd67413f1902b9b7b25b6b029092
Date:   Wed Sep 9 12:22:30 2026 +0800

    fix(watch): bound worker stream capture so a flooding worker cannot kill the kernel
    …
 CHANGELOG.md        |  15 ++++++-
 src/lib/watch.js    | 113 ++++++++++++++++++++++++++++++++++++++++++----------
 tests/watch.test.js |  47 +++++++++++++++++++++-
 3 files changed, 153 insertions(+), 22 deletions(-)

It introduces a bounded head-plus-tail capture:

+export const MAX_CAPTURED_STREAM_BYTES = 32 * 1024 * 1024;
+export class BoundedCapture {
+  constructor(limit = MAX_CAPTURED_STREAM_BYTES) {
…
+      captureTruncated: { stdout: stdoutCapture.dropped, stderr: stderrCapture.dropped },

32 MiB per stream, first and last half kept, dropped characters counted on the observation, and a throw while reading a worker now fails the attempt, not the kernel. Fatal-signature matching already used only the last 4,000 characters, so quota and auth detection are unchanged.

The timeline supports the fix: 16dddb0 is dated 2026-09-09 12:22:30 +0800 = 04:22:30Z; the kernel’s last resume launched at 04:22:47.671Z (launcher.json), 17 seconds later; and attempt 3 of reachability then succeeded (04:22:51.305Z → 04:35:37.441Z, why: "verified").

This fix is not in the audited tree. git merge-base --is-ancestor 16dddb0 HEAD → not an ancestor, and grep -c 'BoundedCapture' src/lib/watch.js → 0. Anything running from this branch, or from the main checkout at 890b4b9, still has the unbounded append.

Two things to carry forward.

1. Land 16dddb0 before 0.27.1. A worker that talks too much should not be able to kill the supervisor and its siblings. It is a denial-of-service by verbosity, from the inside.
2. Surface stderr.log on dead-kernel detection. When the kernel dies without writing a terminal attempt state, watch and runs show should say so and print the tail of ~/.bullswarm/goals/<runId>/stderr.log. Today the reason exists on disk and no command shows it, which turns a five-second diagnosis into an hour of guessing. Attach it to the existing interrupted reason — that reason is where an operator already looks.

Incident B — a worker announced its plan instead of doing the work, and burned 4m43s

What happened. The lens-routing action, attempt 1, was rejected by the content gate:

lens-routing  attempt 1  failed
  pool: command-code   model: gpt-5.6-luna
  reasoning: { requested: "max", applied: "max", source: "strategy-pool", clamped: false }
  startedAt:  2026-09-09T04:35:39.767Z
  finishedAt: 2026-09-09T04:40:22.520Z
  why: "announcement without substance"
  usage.tokens: { standardRead: 1626, output: 64, totalKnown: 1690 }
  tokenSource: "estimated:utf8-bytes/4"
  cost.estimatedUsd: 0.000402

Elapsed = 04:40:22.520 − 04:35:39.767 = 282.75 s = 4 m 42.8 s. (The audit brief rounds this to 4m45s; the timestamps give 4m43s, and the timestamps are the source.)

The entire output was 253 bytes — out-lens-routing-attempt-1.md, quoted in full:

I’ll seed the temporary home with connectors/echo.json, a configured state, and no provider credentials. Then I’ll run pools, strategy inventory --json, strategy rungs, and run --dry-run, capturing exit codes, outputs, and file-tree diffs.

That is a statement of intent. It contains no finding. 64 output tokens at estimated:utf8-bytes/4 ≈ 256 bytes, consistent with the 253 measured bytes.

The gate worked. semantic: announcement without substance is exactly the right verdict, and it is worth noting that this is the failure mode the fixtures/failures/13-announcement-477b.json fixture exists to cover — one of the 23 fixtures that case R16 showed can be deleted without turning the suite red (B10).

The attempt table for the whole run — every attempt of wf-mttj9k9w-baca81, from node bin/bullswarm.js workflow runs show e8ffk2 --json:

Action	Attempt	Pool	Model	Reasoning	Started → finished	Status	Why	Output tokens
reachability	1	command-code	gpt-5.6-luna	max	03:22:56.433 → 03:51:36.300	interrupted	runner stopped before the attempt reached a durable terminal state	—
lens-commands	1	command-code	gpt-5.6-luna	max	03:22:56.453 → 03:34:34.020	succeeded	verified	3,977
lens-docs	1	command-code	gpt-5.6-luna	max	03:22:56.469 → 03:31:00.807	succeeded	verified	2,937
reachability	2	command-code	gpt-5.6-luna	max	03:51:36.447 → 04:22:51.149	interrupted	runner stopped before the attempt reached a durable terminal state	—
reachability	3	command-code	gpt-5.6-luna	max	04:22:51.305 → 04:35:37.441	succeeded	verified	6,617
lens-routing	1	command-code	gpt-5.6-luna	max	04:35:39.767 → 04:40:22.520	failed	announcement without substance	64
lens-workflow	1	command-code	gpt-5.6-luna	max	04:35:39.787 → 04:44:50.830	succeeded	verified	3,316
lens-run	1	command-code	gpt-5.6-luna	max	04:35:39.804 → 04:50:21.554	succeeded	verified	3,381

Every attempt of this run landed on command-code / gpt-5.6-luna / reasoning max, with source: "strategy-pool". Eight for eight. There was no variation to compare against.

What this says about the medium rung. Five of eight attempts succeeded on this rung and produced substantial work — the lens-commands, lens-docs, lens-run, lens-workflow and reachability outputs in this report are all its output, between 2,937 and 6,617 tokens each. So the rung is capable. But one attempt in eight returned an announcement, and the announcement cost the same 4-minute wall clock as real work while producing 64 tokens — 1.6 % of the median successful output. Three observations follow, and the limits of each:

1. The routing evidence is not a comparison. Because all eight attempts used one pool, model and reasoning level, this run says nothing about whether another rung would have done better. Any claim that command-code gpt-5.6-luna is worse or better than an alternative for lens work is unverified — there is no control group in this data.
2. reasoning: max did not prevent the failure. The failed attempt requested and applied max, unclamped, from strategy-pool. Turning reasoning up is not a defence against a worker that decides to describe the task instead of doing it.
3. The gate is the load-bearing part, and it earned its keep. verify.js’s content gate caught this in one attempt for $0.0004, and the retry on the same rung produced the 42.6 KB routing lens this report’s sections 3, 4 and 5 draw on. Keep the gate (it is on the Tier B keep list), and keep it covered — which is exactly why B10 matters, since its fixtures can currently be deleted without a red test.

One process observation. The refute action was the highest-value action in this workflow and it is the reason nine candidates in section 7 are not in the patch scope. An audit that had stopped at the three lenses would have proposed deleting two files that every worker prompt executes.

---

9. How this was measured

The four analyzers

All four live under /tmp/bullswarm-audit/ and are plain zero-dependency Node. Their source is copied verbatim into Appendix B so this report stands alone. Each writes its JSON next to itself (outDir is the script’s own directory), never into the repository.

Script	Lines	What it does	Output
graph.mjs	224	Parses every .js/.mjs under src/, bin/, mcp/, scripts/ for static import … from, export … from and dynamic import('…') of relative paths. Entry roots: bin/bullswarm.js, mcp/server.mjs, every scripts/*.mjs, every path in package.json bin/scripts, and every node … command in .github/workflows/*.yml. Then, per exported symbol, every non-test file that imports it by name.	reachability.json, exports.json
identifiers.mjs	82	For every top-level function/const/class in src/**, counts bare-identifier occurrences (word boundary) across src, bin, mcp, scripts, connectors, skill, tests, .github, README.md, AGENTS.md.	identifiers.json
assets.mjs	38	For every file under fixtures/, data/, connectors/, checks whether its basename or relative path appears as a literal anywhere in src/tests/scripts/.github/package.json.	assets.json
flags.mjs	122	Per command, collects flag names the parser accepts versus flag names src/help.js documents, and reports documented-but-unread and read-but-undocumented sets.	flags.json
probe.mjs	58	Runs 26 CLI cases under a fixture BULLSWARM_HOME and records exit code, stdout, stderr, whether stdout parses as JSON, and line count.	behavior-probes.json

Re-run today, in this workspace

$ cd /Users/cowcow02/Repo/bullswarm-worktrees/audit-0.27
$ node /tmp/bullswarm-audit/graph.mjs
{ "reachable": 64, "unreachable": 2, "neverImported": 103, "testsOnly": 121, "entries": 4 }

$ node /tmp/bullswarm-audit/identifiers.mjs
{ "definitions": 1067, "singleReference": 13, "testsOnly": 16 }

$ node /tmp/bullswarm-audit/assets.mjs
{ "assets": 35, "referenced": 9, "unreferenced": 26 }

Every count matches counts.json exactly: {"reachable":64,"unreachable":2,"neverImported":103,"testsOnly":121,"singleReferenceIdentifiers":13,"testsOnlyIdentifiers":16,"unreferencedAssets":26}. The two unreachable files are bin/check-v2-evidence.js and bin/check-v2-plan.js — both refuted (section 7).

Known limitations of the heuristics

These are not caveats for form’s sake. Each one produced a wrong verdict that the refutation pass had to reverse.

1. graph.mjs cannot see a path built at runtime. new URL('../../bin/check-v2-evidence.js', import.meta.url) followed by a shell invocation is invisible to it. This is why both bin/check-v2-*.js files were reported unreachable when a worker process executes them on every run.
2. assets.mjs matches literals only, so any readdirSync directory looks unreferenced. 26 of 35 assets came back unreferenced and at least 26 of those are false positives — connectors/command-code.json (scanned by setup.js:85-112) and the 23 verification fixtures (scanned by tests/verify.test.js:12-21). The script’s own report flagged this caveat; R15 and R16 confirmed it.
3. flags.mjs cannot follow computed property access. opts[flag] through a rename table made all six workflow planning flags look documented-but-unread. All six are read.
4. “Never imported” is not “never called.” 103 exports have no importer; ~34 of those in the routing and meter layers have real in-file callers, so only the export keyword is surplus.
5. “Tests-only” is sometimes the point. HELP_PATHS exists to be enumerated by the help test. The analyzer offered this as its own deliberate false-positive example.
6. A static audit cannot see an installed home or an external consumer. Three decisions rest on that and are labelled unverified in this report: the claude reader alias (B11), the real OpenCode behaviour behind preferredConcurrency (B7), and whether any external package imports the 12 alias exports.
7. One published count is wrong. The reachability report states src/**/*.js: 16,881 lines. That figure omits the six top-level src/*.js files, because a shell src/**/*.js glob without globstar matches only one directory level:

$ wc -l src/*.js | tail -1
    4385 total
$ find src -name '*.js' | xargs wc -l | tail -1
   21266 total

16,881 + 4,385 = 21,266. The real src/ total is 21,266 lines, not 16,881. The per-file numbers in line-counts.txt are correct; only the total is short. All Tier A and Tier C line counts in this report were taken from per-file wc -l, so none of them inherit this error. Similarly, the reachability report sized strategy-dashboard.js at 278 lines and src/help.js against a 1,724-line figure from the task brief; the measured values here are 551 and 1,349.

The empirical removal results

This is the part that separates a guess from a finding. The refutation pass copied the tree to /tmp/bullswarm-audit/copy/ with rsync -a --exclude .git, restored it from /tmp/bullswarm-audit/pristine/ between every case (diff -rq clean), and ran the suite each time. Baseline in the copy: # tests 659 # pass 659 # fail 0 # duration_ms 11669, exit 0.

Case	Deleted	tests	pass	fail	Reading
R1	fiveHourTier alone	631	630	1	own test file fails to import
R1b	fiveHourTier + its test	658	658	0	confirmed A
R2	REASONING_DEFAULT_TIERS	659	659	0	confirmed A, no test to change
R3	'workflow-v1'	659	659	0	confirmed A
R4	weeklyUsedPct operand	659	659	0	confirmed A
R5	fixtures/openrouter/	659	659	0	confirmed A
R6	parseEvidenceOutput	652	651	1	own test file only → confirmed A
R7	bin/check-v2-evidence.js	659	658	1	refuted
R8	bin/check-v2-plan.js	659	658	1	refuted
R9	recordAgentAction	655	654	1	own test file only → confirmed A
R10	classifyAgentProgress	655	654	1	own test file only → confirmed A
R11	aggregateUsage	654	653	1	own test file only → confirmed A
R12	connectors/_schema.json	659	659	0	dead to code; kept for other reasons (B8)
R13	preferredConcurrency + $comment-	659	658	1	downgraded to B7
R14	12 alias exports (14 lines)	659	659	0	confirmed A
R15	connectors/command-code.json	648	644	4	refuted
R16	one fixtures/failures file	658	658	0	refuted — coverage shrank silently (B10)
R17	inert --json ternary	659	659	0	dead branch removable
R18	claude alias in READERS	659	659	0	dead in-repo; hold (B11)
R19	\|\| sub === 'recommend'	659	659	0	alias has zero test coverage (B2)
R20	the whole confirmed Tier A set at once	651	651	0	187 source lines + 99 test lines, 8 tests

Where the refuter overruled a lens

Stated explicitly, because the refuter’s empirical result wins in each case:

1. bin/check-v2-evidence.js and bin/check-v2-plan.js are not Tier A. The run lens and the reachability report both said unreachable. The workflow lens said they are dynamically invoked and was right. R7 and R8 settled it with failing tests.
2. The flags.json “documented-but-unread” list for workflow is a false positive. The reachability report’s own data was misleading; reading cli.js:118-131 settled it.
3. connectors/command-code.json and the 23 fixtures are not deletable. R15 and R16.
4. connectors/_schema.json moved from Tier A to Tier B, and preferredConcurrency from Tier A3 to Tier B — R12 and R13.
5. strategy-dashboard.js is 551 lines, not 278, and src/help.js is 1,349, not 1,724.
6. data/openrouter-benchmarks.json is not deletable code — it is an absent asset.
7. The 659 baseline is this workspace’s. The 661 figure in 16dddb0’s message belongs to feat/one-workflow-engine, which this HEAD does not contain.

Three inputs to this report were not independently reproduced and are flagged where they appear: B3’s cold-cache network refresh (needs the live fetch the routing lens performed), B7’s real OpenCode concurrency behaviour, and B11’s installed-home question.

Everything that was read to produce this report

Confirmed with ls -la before writing; all read in full.

/Users/cowcow02/.bullswarm/workflows/wf-mttmenoy-b96de2/out-refute-attempt-1.md   10.1K
/tmp/bullswarm-audit/lenses/refute.md                                             21.5K
/tmp/bullswarm-audit/lenses/lens-routing.md                                       42.6K
/tmp/bullswarm-audit/lenses/out-lens-workflow-attempt-1.md                        13.0K
/tmp/bullswarm-audit/lenses/out-lens-run-attempt-1.md                             13.2K
/tmp/bullswarm-audit/lenses/out-lens-commands-attempt-1.md                         15.5K
/tmp/bullswarm-audit/lenses/out-lens-docs-attempt-1.md                             11.5K
/tmp/bullswarm-audit/lenses/out-reachability-attempt-3.md                          25.8K
/tmp/bullswarm-audit/{graph,identifiers,assets,flags,probe}.mjs
/tmp/bullswarm-audit/counts.json  help-probes.json  behavior-probes.json
/tmp/bullswarm-audit/skill-command-help.json  skill-help-checks.tsv  line-counts.txt
/tmp/bullswarm-audit/{reachability,exports,identifiers,assets,flags}.json
/Users/cowcow02/.bullswarm/workflows/wf-mttj9k9w-baca81/state.json                 91.0K
/Users/cowcow02/.bullswarm/workflows/wf-mttj9k9w-baca81/events.jsonl               34.6K
/Users/cowcow02/.bullswarm/workflows/wf-mttj9k9w-baca81/out-lens-routing-attempt-1.md  253B
/Users/cowcow02/.bullswarm/goals/wf-mttj9k9w-baca81/stderr.log                      1.1K
/Users/cowcow02/.bullswarm/goals/wf-mttj9k9w-baca81/launcher.json                    327B

No provider was dispatched to produce this report. .relay-api-key* and .openrouter-api-key were never read. Every CLI command run for this section used a temporary BULLSWARM_HOME under /tmp/bullswarm-audit/ with BULLSWARM_DISABLE_CLAUDE_PROFILES=1 and BULLSWARM_DISABLE_OPENCODE_RELAY=1, and only --dry-run, --help and read verbs were used.

---

Appendix A — full command-surface table (all 64 paths)

Copied verbatim from the command-surface audit (/tmp/bullswarm-audit/lenses/out-lens-commands-attempt-1.md). Legend: 2U = exit 2 with usage/fix guidance · 2V = exit 2 with validation guidance · 1V = validation message but incorrect exit 1 · I = bogus flags silently ignored · J* = JSON is always emitted or the flag has no effect.

Path	Dispatcher	Help complete?	Missing fields / drift	--json	Non-TTY	Bad input	Unknown flag	Verdict / fix
root	yes	yes	none	no	help/status path	2U for unknown verb	rejects as unknown verb	good
setup	yes	yes	none	yes	auto-setup, no prompt	2V for invalid setup options	I	good; strict flags needed
integrate	yes	yes	parent help omits install/remove --yes context	yes	human or JSON	2U subcommand; 1V bad agent	I	good surface; fix exit code
integrate status	yes	yes	none	yes	normal	1V unknown agent	I	fix bad-agent exit to 2
integrate install	yes	yes	none	yes	normal	1V missing --yes	I	fix approval error to 2
integrate remove	yes	yes	none	yes	normal	1V missing --yes	I	fix approval error to 2
integrate retire-legacy	yes	yes	none	yes	normal	1V missing --yes	I	fix approval error to 2
run	yes	yes	omits accepted --no-caller	yes	human or JSON	missing lane exits 1: FAIL unknown lane undefined; malformed values exit 2	I	high priority: require lane and exit 2
health	yes	yes	none	J* always JSON	JSON	semantic unhealthy result exits 1	I	behavior documented
pools	yes	yes	none	yes	human or JSON	none meaningful	I	good; strict flags needed
assignments	yes	yes	none	yes	human or JSON	none meaningful	I	good; strict flags needed
doctor	yes	yes	none	yes	human or JSON	readiness failure exits 1	I	good; strict flags needed
version	yes	yes	package reports 0.26.0, help/task context says 0.27.0	no	normal	2U only for unrelated dispatcher errors	I	reconcile version metadata
release	yes	yes	none	no	normal	2U missing bump kind	I	good maintainer surface
strategy	yes	yes	none	yes	defaults to non-TTY show	2U bad subcommand	I	good
strategy tui	yes	yes	none	no	exits 1: requires TTY	1V/2U depending dashboard failure	I	message is clear; code could use 2 for usage
strategy inventory	yes	yes	none	yes, always JSON	JSON	1V/2V underlying discovery errors	I	useful machine surface
strategy routes	yes	yes	none	yes, always JSON	JSON	1V/2V underlying discovery errors	I	useful narrow surface
strategy set-provider	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U approval/args	I	document output or add --json
strategy set-model	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U approval/args	I	document output or add --json
strategy reset-tier	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U approval/args	I	document output or add --json
strategy set-reasoning	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy rungs	yes	yes	none	yes	human or JSON	--pool without value becomes unknown pool "true" and exits 2	I	improve missing-value message
strategy set-rung	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy reset-reasoning	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy configure	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy refresh	yes	yes	none	yes	human or JSON	2U/2V	I	keep as canonical refresh
strategy recommend	yes	yes	none	yes	human or JSON	2U/2V	I	retain as compatibility alias
strategy apply	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U approval	I	document output or add --json
strategy show	yes	yes	none	yes	human or JSON	1V discovery failure	I	distinct cached-report view
strategy assign	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy clear-assignment	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy exclude-model	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy include-model	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy set-subscription	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy auto	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy auto status	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U/2V	I	document output or add --json
strategy auto off	yes	yes	no advertised JSON; output is always JSON	not offered; always JSON	JSON	2U approval	I	document output or add --json
workflow	yes	yes	one root example omits required --program, --scout, or --orchestrator	no	non-TTY prints help and exits 2	2U bad subcommand	generally I	fix invalid example
workflow goal	yes	yes	none	yes	detached launch; not TTY-dependent	missing program exits 2 with next commands	I; bogus flag accepted (launch unverified)	high priority: reject unknown flags
workflow plan	yes	yes	none	no	normal	2U bad/missing subcommand	I	good parent help
workflow plan contract	yes	yes	none	J* always JSON	JSON	2U/1V cwd	I	good
workflow plan validate	yes	yes	none	yes	human or JSON	2U missing/invalid program; 1 cwd	I; valid bogus flag still exit 0	high priority: strict flags
workflow plan show	yes	yes	none	yes	human or JSON	2U missing ID; 1 missing/non-paused run	I	good
workflow plan submit	yes	yes	none	yes	human or JSON	2U/2V; 1 relaunch/runtime error	I	good
workflow capabilities	yes	yes	usage omits advertised --json	J* always JSON	JSON	1 live-pool failure	I	add [--json] to usage
workflow tui	yes	yes	non-TTY without ID requires --json or an ID; help is broad	yes	bare non-TTY exits 1; ID gives static tree	1 dashboard/runtime error	I	clarify non-TTY refusal
workflow watch	yes	yes	none	no; use --jsonl	normal follow loop	2U/2V typed values	I	good agent loop
workflow events	yes	yes	none	J* always JSON	JSON	2U missing/bad cursor	I	good
workflow steer	yes	yes	none	yes	human or JSON	2U missing message/ID; 1 runtime	I	good
workflow cancel	yes	yes	none	yes	human or JSON	2U missing ID; 1 missing run	I	good
workflow resume	yes	yes	none	yes	human or JSON	2U invalid combinations; 1 missing run	I	good
workflow action	yes	yes	none	no	normal	2U bad subcommand	I	good parent help
workflow action show	yes	yes	none	J* always JSON	JSON	2U missing args; 1 missing run/action	I	good
workflow runs	yes	yes	none	yes	human or JSON	2U time filters; --limit nope exits 0	I	reject invalid numeric limits
workflow runs list	yes	yes	none	yes	human or JSON	2U time filters; --limit nope exits 0	I	reject invalid numeric limits
workflow runs show	yes	yes	none	yes	human or JSON	2U missing ID; 1 missing run	I	good
workflow runs result	yes	yes	none	yes	human or JSON	2U missing ID; 1 missing/not-ready run	I	canonical stable result
workflow runs delete	yes	yes	none	yes	human or JSON	2U missing --yes/ID; 1 missing run	I	good destructive guard
runs	yes	yes	alias-specific text only	yes	same as workflow runs	same as canonical	I	retain compatibility alias
runs list	yes	yes	none	yes	same as canonical	same as canonical	I	retain compatibility alias
runs show	yes	yes	none	yes	same as canonical	same as canonical	I	retain compatibility alias
runs result	yes	yes	none	yes	same as canonical	same as canonical	I	retain compatibility alias
runs delete	yes	yes	none	yes	same as canonical	same as canonical	I	retain compatibility alias

---

Appendix B — analyzer scripts, verbatim

Five zero-dependency Node scripts, copied byte-for-byte from /tmp/bullswarm-audit/ so this report can be reproduced without that directory. Each writes its JSON output next to itself, never into the repository. Run them from the repository root:

$ cd /Users/cowcow02/Repo/bullswarm-worktrees/audit-0.27
$ node /tmp/bullswarm-audit/graph.mjs
$ node /tmp/bullswarm-audit/identifiers.mjs
$ node /tmp/bullswarm-audit/assets.mjs
$ node /tmp/bullswarm-audit/flags.mjs
$ BULLSWARM_HOME=<fixture-home> node /tmp/bullswarm-audit/probe.mjs

Each accepts the repository root as argv[2] and falls back to process.cwd().

B.1 — graph.mjs (224 lines) — import graph, reachability, per-symbol importers

import { existsSync, readdirSync, readFileSync, writeFileSync } from 'node:fs';
import { dirname, extname, join, relative, resolve } from 'node:path';
import { fileURLToPath } from 'node:url';

const repo = resolve(process.argv[2] ?? process.cwd());
const outDir = dirname(fileURLToPath(import.meta.url));
const sourceRoots = ['src', 'bin', 'mcp', 'scripts'];
const isTest = (p) => p === 'tests' || p.startsWith('tests/') || /\.test\.[cm]?js$/.test(p);
const rel = (p) => relative(repo, p).replaceAll('\\', '/');

function walk(dir) {
  if (!existsSync(dir)) return [];
  const result = [];
  for (const entry of readdirSync(dir, { withFileTypes: true }).sort((a, b) => a.name.localeCompare(b.name))) {
    if (entry.name === 'node_modules' || entry.name === '.git') continue;
    const path = join(dir, entry.name);
    if (entry.isDirectory()) result.push(...walk(path));
    else result.push(path);
  }
  return result;
}

const files = sourceRoots.flatMap((root) => walk(join(repo, root)))
  .filter((p) => ['.js', '.mjs'].includes(extname(p)))
  .map(rel)
  .sort();
const consumerFiles = [...new Set([...files, ...walk(join(repo, 'tests'))])]
  .filter((p) => ['.js', '.mjs'].includes(extname(p)))
  .map(rel)
  .sort();
const fileSet = new Set(files);

function resolveRelative(from, specifier) {
  if (!specifier.startsWith('.')) return null;
  const base = resolve(repo, dirname(from), specifier);
  const candidates = [base, `${base}.js`, `${base}.mjs`, join(base, 'index.js'), join(base, 'index.mjs')];
  return candidates.map(rel).find((candidate) => fileSet.has(candidate)) ?? null;
}

function masked(source) {
  let out = '';
  let mode = 'code';
  let quote = '';
  for (let i = 0; i < source.length; i += 1) {
    const c = source[i];
    const n = source[i + 1];
    if (mode === 'line') {
      out += c === '\n' ? '\n' : ' ';
      if (c === '\n') mode = 'code';
      continue;
    }
    if (mode === 'block') {
      out += c === '\n' ? '\n' : ' ';
      if (c === '*' && n === '/') { out += ' '; i += 1; mode = 'code'; }
      continue;
    }
    if (mode === 'string') {
      out += c === '\n' ? '\n' : ' ';
      if (c === '\\') { if (source[i + 1] === '\n') out += '\n'; i += 1; out += ' '; continue; }
      if (c === quote) mode = 'code';
      continue;
    }
    if (c === '/' && n === '/') { out += '  '; i += 1; mode = 'line'; continue; }
    if (c === '/' && n === '*') { out += '  '; i += 1; mode = 'block'; continue; }
    if (c === '\'' || c === '"' || c === '`') { out += ' '; quote = c; mode = 'string'; continue; }
    out += c;
  }
  return out;
}

function importSpecs(source) {
  const specs = [];
  const staticRe = /\bimport\s+(?!\()(?:(?:[\s\S]*?)\sfrom\s*)?["']([^"']+)["']/g;
  const exportRe = /\bexport\s+(?:(?:[\s\S]*?)\sfrom\s*)["']([^"']+)["']/g;
  const dynamicRe = /\bimport\s*\(\s*["']([^"']+)["']\s*\)/g;
  for (const re of [staticRe, exportRe, dynamicRe]) {
    for (const match of source.matchAll(re)) specs.push({ specifier: match[1], kind: re === dynamicRe ? 'dynamic' : re === exportRe ? 're-export' : 'static' });
  }
  return specs;
}

function exportedSymbols(source) {
  const code = masked(source);
  const symbols = new Map();
  const add = (name, line = 1) => { if (name && !symbols.has(name)) symbols.set(name, { name, line }); };
  for (const match of code.matchAll(/\bexport\s+(?:async\s+)?function\s+([A-Za-z_$][\w$]*)/g)) add(match[1], source.slice(0, match.index).split('\n').length);
  for (const match of code.matchAll(/\bexport\s+(?:const|let|class)\s+([A-Za-z_$][\w$]*)/g)) add(match[1], source.slice(0, match.index).split('\n').length);
  for (const match of code.matchAll(/\bexport\s*\{([^}]*)\}/g)) {
    const line = source.slice(0, match.index).split('\n').length;
    for (const item of match[1].split(',')) {
      const cleaned = item.trim().replace(/\/\/.*$/, '');
      if (!cleaned) continue;
      const parts = cleaned.split(/\s+as\s+/);
      add((parts[1] ?? parts[0]).trim(), line);
    }
  }
  return [...symbols.values()].sort((a, b) => a.name.localeCompare(b.name));
}

function importBindings(source) {
  const bindings = [];
  const staticRe = /\bimport\s+([\s\S]*?)\s+from\s*["']([^"']+)["']/g;
  const exportRe = /\bexport\s*\{([^}]*)\}\s+from\s*["']([^"']+)["']/g;
  for (const match of source.matchAll(staticRe)) {
    const clause = match[1].trim();
    const values = [];
    const named = clause.match(/\{([^}]*)\}/);
    if (named) {
      for (const item of named[1].split(',')) {
        const parts = item.trim().split(/\s+as\s+/);
        if (parts[0]) values.push({ imported: parts[0].trim(), local: (parts[1] ?? parts[0]).trim(), kind: 'named' });
      }
    }
    const namespace = clause.match(/\*\s+as\s+([A-Za-z_$][\w$]*)/);
    if (namespace) values.push({ imported: '*', local: namespace[1], kind: 'namespace' });
    const defaultPart = clause.replace(named?.[0] ?? '', '').replace(namespace?.[0] ?? '', '').replace(',', '').trim();
    if (defaultPart && /^[A-Za-z_$][\w$]*$/.test(defaultPart)) values.push({ imported: 'default', local: defaultPart, kind: 'default' });
    bindings.push({ specifier: match[2], bindings: values });
  }
  const dynamicNamedRe = /\b(?:const|let|var)\s+\{([^}]*)\}\s*=\s*(?:await\s+)?import\(\s*[\"']([^\"']+)[\"']\s*\)/g;
  for (const match of source.matchAll(dynamicNamedRe)) {
    const values = [];
    for (const item of match[1].split(',')) {
      const parts = item.trim().split(/\s*:\s*/);
      if (parts[0]) values.push({ imported: parts[0].trim(), local: (parts[1] ?? parts[0]).trim(), kind: 'dynamic-named' });
    }
    bindings.push({ specifier: match[2], bindings: values });
  }
  const dynamicNamespaceRe = /\b(?:const|let|var)\s+([A-Za-z_$][\w$]*)\s*=\s*(?:await\s+)?import\(\s*[\"']([^\"']+)[\"']\s*\)/g;
  for (const match of source.matchAll(dynamicNamespaceRe)) {
    bindings.push({ specifier: match[2], bindings: [{ imported: '*', local: match[1], kind: 'dynamic-namespace' }] });
  }
  for (const match of source.matchAll(exportRe)) {
    const values = [];
    for (const item of match[1].split(',')) {
      const parts = item.trim().split(/\s+as\s+/);
      if (parts[0]) values.push({ imported: parts[0].trim(), local: (parts[1] ?? parts[0]).trim(), kind: 're-export' });
    }
    bindings.push({ specifier: match[2], bindings: values });
  }
  return bindings;
}

const graph = Object.fromEntries(files.map((file) => [file, { imports: [], importers: [] }]));
const exportMap = Object.fromEntries(files.map((file) => [file, exportedSymbols(readFileSync(join(repo, file), 'utf8'))]));
const allBindings = [];
for (const file of consumerFiles) {
  const source = readFileSync(join(repo, file), 'utf8');
  for (const item of importSpecs(source)) {
    const target = resolveRelative(file, item.specifier);
    if (!target) continue;
    if (graph[file]) graph[file].imports.push({ file: target, specifier: item.specifier, kind: item.kind });
    graph[target].importers.push(file);
  }
  for (const item of importBindings(source)) {
    const target = resolveRelative(file, item.specifier);
    if (target) allBindings.push({ importer: file, target, bindings: item.bindings });
  }
}
for (const file of files) {
  graph[file].imports = [...new Map(graph[file].imports.map((x) => [`${x.file}:${x.kind}`, x])).values()].sort((a, b) => a.file.localeCompare(b.file));
  graph[file].importers = [...new Set(graph[file].importers)].sort();
}

const entryCandidates = new Set(['bin/bullswarm.js', 'mcp/server.mjs', ...files.filter((file) => file.startsWith('scripts/') && file.endsWith('.mjs'))]);
const pkg = JSON.parse(readFileSync(join(repo, 'package.json'), 'utf8'));
for (const value of Object.values(pkg.bin ?? {})) if (typeof value === 'string') entryCandidates.add(value.replace(/^\.\//, ''));
for (const value of Object.values(pkg.scripts ?? {})) {
  for (const match of value.matchAll(/(?:^|\s)node(?:\s+[^\s]+)*\s+([^\s]+)/g)) entryCandidates.add(match[1].replace(/^\.\//, ''));
}
for (const yml of walk(join(repo, '.github')).filter((p) => /\.ya?ml$/.test(p))) {
  const text = readFileSync(yml, 'utf8');
  for (const match of text.matchAll(/\bnode\s+(?:--[^\s]+\s+)*([^\s'"`\\|;&]+)/g)) entryCandidates.add(match[1].replace(/^\.\//, ''));
}
const entries = [...entryCandidates].filter((file) => fileSet.has(file)).sort();
const reachable = new Set();
const queue = [...entries];
while (queue.length) {
  const file = queue.shift();
  if (reachable.has(file)) continue;
  reachable.add(file);
  for (const next of graph[file].imports.map((x) => x.file)) if (!reachable.has(next)) queue.push(next);
}

const exports = {};
for (const file of files) {
  const imported = new Map();
  for (const item of allBindings.filter((x) => x.target === file)) {
    for (const binding of item.bindings) {
      for (const symbol of exportMap[file]) {
        if (binding.imported === symbol.name || binding.imported === '*') {
          const list = imported.get(symbol.name) ?? [];
          list.push({ file: item.importer, importedAs: binding.local, kind: binding.kind, namespace: binding.imported === '*' });
          imported.set(symbol.name, list);
        }
      }
    }
  }
  const all = exportMap[file].map((symbol) => ({ ...symbol, importers: [...(imported.get(symbol.name) ?? [])].sort((a, b) => a.file.localeCompare(b.file)) }));
  exports[file] = all;
}
const neverImported = [];
const testsOnly = [];
for (const [file, symbols] of Object.entries(exports)) {
  for (const symbol of symbols) {
    const nonTest = symbol.importers.filter((x) => !isTest(x.file));
    const test = symbol.importers.filter((x) => isTest(x.file));
    const item = { file, symbol: symbol.name, line: symbol.line, importers: symbol.importers, evidenceCommand: `rg -n "\\b${symbol.name}\\b" src bin mcp scripts tests` };
    if (!nonTest.length && !test.length) neverImported.push(item);
    else if (!nonTest.length && test.length) testsOnly.push(item);
  }
}
const output = {
  generatedAt: new Date().toISOString(),
  repo,
  heuristic: 'Static relative imports/re-exports/dynamic imports only; package and node imports are ignored. Reachability graph nodes are src/**, bin/**, mcp/**, and scripts/**; tests/** are scanned only as export consumers. Namespace imports count as importing every named export found in the target. Test files are tests/** or *.test.js/*.test.mjs.',
  entries,
  files: graph,
  reachable: [...reachable].sort(),
  unreachable: files.filter((file) => !reachable.has(file)),
};
writeFileSync(join(outDir, 'reachability.json'), `${JSON.stringify(output, null, 2)}\n`);
writeFileSync(join(outDir, 'exports.json'), `${JSON.stringify({ generatedAt: output.generatedAt, repo, heuristic: output.heuristic, exports, neverImported, testsOnly }, null, 2)}\n`);
console.log(JSON.stringify({ reachable: output.reachable.length, unreachable: output.unreachable.length, neverImported: neverImported.length, testsOnly: testsOnly.length, entries: entries.length }, null, 2));

B.2 — identifiers.mjs (82 lines) — single-reference and tests-only identifiers

import { existsSync, readdirSync, readFileSync, writeFileSync } from 'node:fs';
import { extname, join, relative, resolve } from 'node:path';
import { fileURLToPath } from 'node:url';

const repo = resolve(process.argv[2] ?? process.cwd());
const outDir = resolve(fileURLToPath(new URL('.', import.meta.url)));
const roots = ['src', 'bin', 'mcp', 'scripts', 'connectors', 'skill', 'tests', '.github'];
const isTest = (p) => p === 'tests' || p.startsWith('tests/') || /\.test\.[cm]?js$/.test(p);
const rel = (p) => relative(repo, p).replaceAll('\\', '/');
function walk(dir) {
  if (!existsSync(dir)) return [];
  const result = [];
  for (const entry of readdirSync(dir, { withFileTypes: true }).sort((a, b) => a.name.localeCompare(b.name))) {
    if (entry.name === 'node_modules' || entry.name === '.git') continue;
    const path = join(dir, entry.name);
    if (entry.isDirectory()) result.push(...walk(path)); else result.push(path);
  }
  return result;
}
const files = [...new Set(roots.flatMap((root) => walk(join(repo, root))).concat([
  join(repo, 'README.md'), join(repo, 'AGENTS.md'),
]))].filter((p) => existsSync(p)).map(rel).sort();
function mask(source) {
  let out = ''; let mode = 'code'; let quote = '';
  for (let i = 0; i < source.length; i += 1) {
    const c = source[i]; const n = source[i + 1];
    if (mode === 'line') { out += c === '\n' ? '\n' : ' '; if (c === '\n') mode = 'code'; continue; }
    if (mode === 'block') { out += c === '\n' ? '\n' : ' '; if (c === '*' && n === '/') { out += ' '; i += 1; mode = 'code'; } continue; }
    if (mode === 'string') { out += c === '\n' ? '\n' : ' '; if (c === '\\') { i += 1; out += ' '; continue; } if (c === quote) mode = 'code'; continue; }
    if (c === '/' && n === '/') { out += '  '; i += 1; mode = 'line'; continue; }
    if (c === '/' && n === '*') { out += '  '; i += 1; mode = 'block'; continue; }
    if (c === '\'' || c === '"' || c === '`') { out += ' '; quote = c; mode = 'string'; continue; }
    out += c;
  }
  return out;
}
function definitions(file) {
  if (!file.startsWith('src/') || !['.js', '.mjs'].includes(extname(file))) return [];
  const source = readFileSync(join(repo, file), 'utf8');
  const code = mask(source);
  const result = [];
  let depth = 0;
  for (const [index, line] of code.split('\n').entries()) {
    if (depth === 0) {
      const functionMatch = line.match(/^(?:export\s+)?(?:async\s+)?function\s+([A-Za-z_$][\w$]*)/);
      const classMatch = line.match(/^(?:export\s+)?class\s+([A-Za-z_$][\w$]*)/);
      const valueMatch = line.match(/^(?:export\s+)?(?:const|let)\s+([A-Za-z_$][\w$]*)\s*=/);
      if (functionMatch) result.push({ file, line: index + 1, name: functionMatch[1], kind: 'function' });
      if (classMatch) result.push({ file, line: index + 1, name: classMatch[1], kind: 'class' });
      if (valueMatch) result.push({ file, line: index + 1, name: valueMatch[1], kind: 'const/let' });
    }
    depth += (line.match(/\{/g) ?? []).length - (line.match(/\}/g) ?? []).length;
    if (depth < 0) depth = 0;
  }
  return result;
}
const texts = Object.fromEntries(files.map((file) => [file, readFileSync(join(repo, file), 'utf8')]));
const defs = files.flatMap(definitions);
const singleReference = []; const testsOnly = []; const all = [];
for (const definition of defs) {
  const escaped = definition.name.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
  const pattern = new RegExp(`(?<![\\w$])${escaped}(?![\\w$])`, 'g');
  const references = [];
  for (const [file, text] of Object.entries(texts)) {
    const matches = [...text.matchAll(pattern)];
    for (const match of matches) references.push({ file, line: text.slice(0, match.index).split('\n').length, context: text.split('\n')[text.slice(0, match.index).split('\n').length - 1].trim() });
  }
  const nonDefining = references.filter((reference) => !(reference.file === definition.file && reference.line === definition.line));
  const item = { ...definition, count: references.length, references, evidenceCommand: `rg -n "\\b${definition.name}\\b" src bin mcp scripts connectors skill tests .github README.md AGENTS.md` };
  all.push(item);
  if (references.length === 1) singleReference.push(item);
  if (nonDefining.length > 0 && nonDefining.every((reference) => isTest(reference.file))) testsOnly.push(item);
}
const output = {
  generatedAt: new Date().toISOString(), repo,
  heuristic: 'Top-level declarations are line-oriented declarations at brace depth zero in src/**/*.js or src/**/*.mjs. Bare word-boundary counts include comments, strings, JSON, YAML, and same-name unrelated symbols; only references outside the defining declaration line are used for tests-only classification.',
  singleReference: singleReference.sort((a, b) => `${a.file}:${a.line}:${a.name}`.localeCompare(`${b.file}:${b.line}:${b.name}`)),
  testsOnly: testsOnly.sort((a, b) => `${a.file}:${a.line}:${a.name}`.localeCompare(`${b.file}:${b.line}:${b.name}`)),
  definitions: all.sort((a, b) => `${a.file}:${a.line}:${a.name}`.localeCompare(`${b.file}:${b.line}:${b.name}`)),
};
writeFileSync(join(outDir, 'identifiers.json'), `${JSON.stringify(output, null, 2)}\n`);
console.log(JSON.stringify({ definitions: defs.length, singleReference: singleReference.length, testsOnly: testsOnly.length }, null, 2));

B.3 — assets.mjs (38 lines) — unreferenced fixtures, data and connectors

import { existsSync, readdirSync, readFileSync, writeFileSync } from 'node:fs';
import { basename, join, relative, resolve } from 'node:path';
import { fileURLToPath } from 'node:url';

const repo = resolve(process.argv[2] ?? process.cwd());
const outDir = resolve(fileURLToPath(new URL('.', import.meta.url)));
const rel = (p) => relative(repo, p).replaceAll('\\', '/');
function walk(dir) {
  if (!existsSync(dir)) return [];
  const result = [];
  for (const entry of readdirSync(dir, { withFileTypes: true }).sort((a, b) => a.name.localeCompare(b.name))) {
    if (entry.name === 'node_modules' || entry.name === '.git') continue;
    const path = join(dir, entry.name);
    if (entry.isDirectory()) result.push(...walk(path)); else result.push(path);
  }
  return result;
}
const assets = ['fixtures', 'data', 'connectors'].flatMap((root) => walk(join(repo, root))).map(rel).sort();
const searchFiles = ['src', 'tests', 'scripts', '.github'].flatMap((root) => walk(join(repo, root))).concat([join(repo, 'package.json')]).filter(existsSync).map(rel);
const texts = Object.fromEntries(searchFiles.map((file) => [file, readFileSync(join(repo, file), 'utf8')]));
const findings = assets.map((asset) => {
  const tokenSet = [...new Set([basename(asset), asset])];
  const matches = [];
  for (const [file, text] of Object.entries(texts)) {
    const tokens = tokenSet.filter((token) => text.includes(token));
    if (tokens.length) matches.push({ file, tokens });
  }
  return { file: asset, basename: basename(asset), relativePath: asset, matches, referenced: matches.length > 0, evidenceCommand: `rg -n -F -e ${JSON.stringify(basename(asset))} -e ${JSON.stringify(asset)} src tests scripts .github package.json` };
});
const output = {
  generatedAt: new Date().toISOString(), repo,
  heuristic: 'A literal substring match of either the asset basename or repository-relative path in src/, tests/, scripts/, .github/, or package.json counts as referenced. Runtime directory scans, generated paths, JSON keys, and semantic fixture loading are not inferred.',
  searchFiles,
  assets: findings,
  unreferenced: findings.filter((item) => !item.referenced),
};
writeFileSync(join(outDir, 'assets.json'), `${JSON.stringify(output, null, 2)}\n`);
console.log(JSON.stringify({ assets: assets.length, referenced: output.assets.length - output.unreferenced.length, unreferenced: output.unreferenced.length }, null, 2));

B.4 — flags.mjs (122 lines) — documented-but-unread and read-but-undocumented flags

import { existsSync, readFileSync, writeFileSync } from 'node:fs';
import { join, relative, resolve } from 'node:path';
import { pathToFileURL, fileURLToPath } from 'node:url';

const repo = resolve(process.argv[2] ?? process.cwd());
const outDir = resolve(fileURLToPath(new URL('.', import.meta.url)));
const fileText = (file) => readFileSync(join(repo, file), 'utf8');
const sourceFiles = ['src/cli.js', 'src/strategy-cli.js', 'src/workflow/cli.js', 'src/workflow/runs-cli.js', 'src/setup.js', 'src/integrate.js'];

function functionBody(source, name) {
  const start = source.search(new RegExp(`(?:export\\s+)?(?:async\\s+)?function\\s+${name}\\s*\\(`));
  if (start < 0) return '';
  let parens = 0;
  let signatureEnd = -1;
  for (let i = source.indexOf('(', start); i >= 0 && i < source.length; i += 1) {
    if (source[i] === '(') parens += 1;
    if (source[i] === ')') { parens -= 1; if (parens === 0) { signatureEnd = i; break; } }
  }
  const open = source.indexOf('{', signatureEnd < 0 ? start : signatureEnd);
  if (open < 0) return '';
  let depth = 0; let quote = ''; let mode = 'code';
  for (let i = open; i < source.length; i += 1) {
    const c = source[i]; const n = source[i + 1];
    if (mode === 'line') { if (c === '\n') mode = 'code'; continue; }
    if (mode === 'block') { if (c === '*' && n === '/') { mode = 'code'; i += 1; } continue; }
    if (mode === 'string') { if (c === '\\') { i += 1; continue; } if (c === quote) mode = 'code'; continue; }
    if (c === '/' && n === '/') { mode = 'line'; i += 1; continue; }
    if (c === '/' && n === '*') { mode = 'block'; i += 1; continue; }
    if (c === '\'' || c === '"' || c === '`') { quote = c; mode = 'string'; continue; }
    if (c === '{') depth += 1;
    if (c === '}') { depth -= 1; if (depth === 0) return source.slice(start, i + 1); }
  }
  return source.slice(start);
}

function flagsRead(source) {
  const result = new Set();
  for (const match of source.matchAll(/\bopts(?:\[['"]([A-Za-z0-9_-]+)['"]\]|\.([A-Za-z0-9_-]+))/g)) result.add(match[1] ?? match[2]);
  for (const match of source.matchAll(/['"]([A-Za-z0-9_-]+)['"]\s*:/g)) {
    if (['mappings', 'settings', 'next', 'status', 'message'].includes(match[1])) continue;
  }
  return [...result].sort();
}

function flagsAccepted(source) {
  const result = new Set();
  for (const match of source.matchAll(/(?:BOOLEAN_FLAGS|valueFlags|new Set\(\s*\[)([\s\S]*?)(?:\]|\)\s*;)/g)) {
    for (const item of match[1].matchAll(/['"]([A-Za-z0-9_-]+)['"]/g)) result.add(item[1]);
  }
  for (const match of source.matchAll(/(?:===|has\()\s*['"]--?([A-Za-z0-9_-]+)['"]/g)) result.add(match[1]);
  return [...result].sort();
}

const help = await import(pathToFileURL(join(repo, 'src/help.js')).href);
function documented(paths) {
  const result = new Set();
  const texts = [];
  for (const path of paths) {
    const text = help.helpText(path);
    texts.push({ path: path.join(' '), text });
    for (const line of text.split('\n')) {
      const match = line.match(/^\s{2}(--[a-z0-9-]+)/);
      if (match) result.add(match[1].slice(2));
    }
  }
  return { flags: [...result].sort(), texts };
}
function unionBodies(source, names) {
  return names.map((name) => functionBody(source, name)).join('\n');
}
function entry({ command, source, paths, bodies = null, acceptedSource = null, note = '' }) {
  const sourceText = fileText(source);
  const readSource = bodies ? unionBodies(sourceText, bodies) : sourceText;
  const read = flagsRead(readSource);
  const accepted = flagsAccepted(acceptedSource ?? sourceText);
  const docs = documented(paths);
  return {
    command, source, paths: paths.map((path) => path.join(' ')),
    parserAccepted: accepted,
    directlyRead: read,
    documented: docs.flags,
    documentedButUnread: docs.flags.filter((flag) => !read.includes(flag)),
    readButUndocumented: read.filter((flag) => !docs.flags.includes(flag)),
    heuristic: note || 'Flags are inferred from opts.foo/opts["foo"] in the selected handler bodies; parserAccepted also includes explicit parser literals. Help flags come from option lines in helpText() for the mapped command paths.',
  };
}
const rootPaths = [[], ['setup'], ['run'], ['health'], ['pools'], ['assignments'], ['doctor'], ['release']];
const workflowPaths = [
  ['workflow'], ['workflow', 'goal'], ['workflow', 'plan'], ['workflow', 'plan', 'contract'], ['workflow', 'plan', 'validate'],
  ['workflow', 'plan', 'show'], ['workflow', 'plan', 'submit'], ['workflow', 'cancel'], ['workflow', 'resume'],
  ['workflow', 'capabilities'], ['workflow', 'tui'], ['workflow', 'watch'], ['workflow', 'events'], ['workflow', 'steer'],
  ['workflow', 'action'], ['workflow', 'action', 'show'],
];
const entries = [
  entry({ command: 'cli top-level', source: 'src/cli.js', paths: rootPaths, note: 'The top-level parser accepts arbitrary --name tokens. This entry unions the root command help paths and all cli.js handler reads; delegated strategy/workflow/integrate flags are analyzed in their own entries.' }),
  entry({ command: 'cli setup', source: 'src/cli.js', paths: [['setup']], bodies: ['cmdSetup'] }),
  entry({ command: 'cli run', source: 'src/cli.js', paths: [['run']], bodies: ['cmdRun', 'emit'] }),
  entry({ command: 'cli pools', source: 'src/cli.js', paths: [['pools']], bodies: ['cmdPools'] }),
  entry({ command: 'cli assignments', source: 'src/cli.js', paths: [['assignments']], bodies: ['cmdAssignments'] }),
  entry({ command: 'cli health', source: 'src/cli.js', paths: [['health']], bodies: ['cmdHealth'] }),
  entry({ command: 'cli doctor', source: 'src/cli.js', paths: [['doctor']], bodies: ['cmdDoctor'] }),
  entry({ command: 'cli release', source: 'src/cli.js', paths: [['release']], bodies: ['cmdRelease'] }),
  entry({ command: 'strategy', source: 'src/strategy-cli.js', paths: [
    ['strategy'], ['strategy', 'tui'], ['strategy', 'inventory'], ['strategy', 'routes'], ['strategy', 'set-provider'], ['strategy', 'set-model'], ['strategy', 'set-rung'], ['strategy', 'reset-tier'], ['strategy', 'set-reasoning'], ['strategy', 'reset-reasoning'], ['strategy', 'rungs'], ['strategy', 'configure'], ['strategy', 'refresh'], ['strategy', 'recommend'], ['strategy', 'apply'], ['strategy', 'show'], ['strategy', 'assign'], ['strategy', 'clear-assignment'], ['strategy', 'exclude-model'], ['strategy', 'include-model'], ['strategy', 'set-subscription'], ['strategy', 'auto'], ['strategy', 'auto', 'status'], ['strategy', 'auto', 'off'],
  ], bodies: ['cmdStrategy'] }),
  entry({ command: 'workflow', source: 'src/workflow/cli.js', paths: workflowPaths, bodies: ['cmdWorkflow', 'wfGoal', 'goalSettings', 'buildNewGoalDocument', 'resolvePlanning', 'planContract', 'planValidate', 'planShow', 'planSubmit', 'wfCancel', 'wfResume', 'wfCapabilities', 'wfEvents', 'wfWatch', 'wfSteer', 'wfAction'], note: 'The workflow module has shared parsing and helper functions; this entry unions direct opts reads across dispatch and command helpers.' }),
  entry({ command: 'workflow capabilities', source: 'src/workflow/cli.js', paths: [['workflow', 'capabilities']], bodies: ['wfCapabilities'] }),
  entry({ command: 'workflow events', source: 'src/workflow/cli.js', paths: [['workflow', 'events']], bodies: ['wfEvents'] }),
  entry({ command: 'workflow watch', source: 'src/workflow/cli.js', paths: [['workflow', 'watch']], bodies: ['wfWatch'] }),
  entry({ command: 'workflow steer', source: 'src/workflow/cli.js', paths: [['workflow', 'steer']], bodies: ['wfSteer'] }),
  entry({ command: 'workflow action show', source: 'src/workflow/cli.js', paths: [['workflow', 'action', 'show']], bodies: ['wfAction'] }),
  entry({ command: 'workflow runs', source: 'src/workflow/runs-cli.js', paths: [['workflow', 'runs'], ['workflow', 'runs', 'list'], ['workflow', 'runs', 'show'], ['workflow', 'runs', 'result'], ['workflow', 'runs', 'delete']], note: 'Runs flags are parsed in parseRunsFlags and consumed by several handlers; this entry uses the complete module read set.' }),
  entry({ command: 'setup module', source: 'src/setup.js', paths: [['setup']], bodies: ['runWizard'], note: 'setup.js is not the CLI parser; cli.js owns setup flag dispatch. This reports opts directly read by setup.js only.' }),
  entry({ command: 'integrate', source: 'src/integrate.js', paths: [['integrate'], ['integrate', 'status'], ['integrate', 'install'], ['integrate', 'remove'], ['integrate', 'retire-legacy']], bodies: ['cmdIntegrate'] }),
];
const output = {
  generatedAt: new Date().toISOString(), repo,
  heuristic: 'Best-effort static comparison. Help flags are parsed from generated option lines in src/help.js. Handler-scoped entries use opts property reads in named functions; module-scoped entries union all reads because dispatch is split across helpers. A generic parser may accept unknown flags syntactically even when no handler reads them. --help/-h contextual handling is not treated as a command option unless it appears in the command help option block.',
  entries,
};
writeFileSync(join(outDir, 'flags.json'), `${JSON.stringify(output, null, 2)}\n`);
console.log(JSON.stringify(Object.fromEntries(entries.map((item) => [item.command, { documentedButUnread: item.documentedButUnread.length, readButUndocumented: item.readButUndocumented.length }])), null, 2));

B.5 — probe.mjs (58 lines) — the 26 CLI behaviour probes

import { spawnSync } from 'node:child_process';
import { writeFileSync } from 'node:fs';
const node = process.execPath;
const repo = process.cwd();
const home = process.env.BULLSWARM_HOME;
const baseEnv = { ...process.env, BULLSWARM_HOME: home, HOME: home, PATH: `${process.env.PATH}:${home}/bin` };
const cases = [
  ['run-missing-lane', ['run', '--add-dir', '.', '--dry-run', 'probe']],
  ['run-bogus-flag', ['run', '--lane', 'analyze', '--bogus-flag', '--dry-run', 'probe']],
  ['run-json-dry', ['run', '--lane', 'analyze', '--add-dir', '.', '--dry-run', '--json', 'probe']],
  ['run-piped', ['run', '--lane', 'analyze', '--add-dir', '.', '--dry-run', 'probe'], true],
  ['pools-bogus-flag', ['pools', '--bogus-flag']],
  ['pools-json', ['pools', '--json']],
  ['strategy-rungs-bogus-flag', ['strategy', 'rungs', '--bogus-flag']],
  ['strategy-rungs-json', ['strategy', 'rungs', '--json']],
  ['strategy-set-rung-bogus-flag', ['strategy', 'set-rung', 'echo', 'low', '--model', 'echo-local', '--bogus-flag']],
  ['strategy-set-rung-missing-model', ['strategy', 'set-rung', 'echo', 'low']],
  ['workflow-goal-bogus-flag', ['workflow', 'goal', 'probe', '--program', 'missing.json', '--bogus-flag']],
  ['workflow-goal-no-program', ['workflow', 'goal', 'probe']],
  ['workflow-plan-validate-bogus-flag', ['workflow', 'plan', 'validate', 'probe', '--program', 'missing.json', '--bogus-flag']],
  ['workflow-plan-validate-missing-program', ['workflow', 'plan', 'validate', 'probe']],
  ['workflow-runs-list-bogus-flag', ['workflow', 'runs', 'list', '--bogus-flag']],
  ['workflow-runs-list-json', ['workflow', 'runs', 'list', '--json']],
  ['health-bogus-flag', ['health', '--bogus-flag']],
  ['health-json', ['health', '--json']],
  ['doctor-bogus-flag', ['doctor', '--bogus-flag']],
  ['doctor-json', ['doctor', '--json']],
  ['assignments-bogus-flag', ['assignments', '--bogus-flag']],
  ['assignments-json', ['assignments', '--json']],
  ['workflow-capabilities-json', ['workflow', 'capabilities', '--json']],
  ['workflow-events-missing-run', ['workflow', 'events']],
  ['workflow-watch-missing-run', ['workflow', 'watch']],
  ['workflow-runs-result-missing-run', ['workflow', 'runs', 'result']],
];
const results = [];
for (const [name, args, pipe] of cases) {
  const r = spawnSync(node, ['bin/bullswarm.js', ...args], {
    cwd: repo,
    env: baseEnv,
    input: pipe ? '' : undefined,
    encoding: 'utf8',
    timeout: 30000,
  });
  const stdout = r.stdout ?? '';
  let json = false;
  try { JSON.parse(stdout); json = true; } catch {}
  results.push({
    name, args, piped: Boolean(pipe), exit: r.status, signal: r.signal,
    stdout, stderr: r.stderr ?? '', jsonDocument: json,
    stdoutLines: stdout.trimEnd() ? stdout.trimEnd().split('\n').length : 0,
  });
}
writeFileSync('/tmp/bullswarm-audit/behavior-probes.json', JSON.stringify(results, null, 2));
for (const r of results) {
  const out = r.stdout.trim().replaceAll('\n', ' \\n ').slice(0, 220);
  const err = r.stderr.trim().replaceAll('\n', ' \\n ').slice(0, 220);
  console.log(`${r.name}\texit=${r.exit}\tjson=${r.jsonDocument}\tstdout=${JSON.stringify(out)}\tstderr=${JSON.stringify(err)}`);
}

---

Audit produced 2026-09-09 against audit/codebase-2026-09-09 at b8487b1. Read-only: the only file written was this one.