dream(swarm): #3061 bound MessageBus retry attempts, fix unreachable message.failed (evaluated, ACCEPT)

[ruvnet]

1. Hypothesis

Given a Ruflo MessageBus with a subscriber whose callback synchronously throws on every invocation, when addToQueue()'s re-queue path preserves the failed entry's attempts count (instead of resetting it to 0 on every re-queue), then the number of delivery attempts for that message should be bounded by config.retryAttempts and a message.failed event should fire exactly once, relative to the current unbounded-retry baseline, subject to: (1) delivery behavior for non-throwing subscribers is completely unchanged; (2) existing test suite remains green; (3) zero LLM/API cost.

Frozen before evaluation began; not modified after seeing results.

2. Candidate

v3/@claude-flow/swarm/src/message-bus.ts — MessageBus.handleDeliveryError() (:491-509) increments a queue entry's attempts counter and, if still under config.retryAttempts, re-queues via addToQueue(). But addToQueue() (:343-363) always constructed a fresh MessageQueueEntry with attempts: 0, discarding the count. A permanently-throwing subscriber retried forever instead of being bounded, and the message.failed event was structurally unreachable.

Fix: addToQueue() gains an optional existingAttempts: number = 0 parameter; the retry-path call now passes entry.attempts through instead of the implicit 0. The two fresh-message call sites in enqueue() are untouched (default preserved). Net diff: ~10 lines + a disclosure comment for a related pre-existing limitation found during adversarial review (see §9). Plus a new test file, v3/@claude-flow/swarm/__tests__/message-bus.test.ts (2 tests, ~110 lines).

3. Evaluation Receipt

evaluated: accepted. Real evaluator: vitest run, deterministic, $0, zero LLM calls. Full receipt in docs/dream-cycle/dream-gist-2026-08-19.md.

4. Baseline Comparison

Metric	Baseline (pre-fix, git stash)	Candidate (post-fix)
Callback invocations (500ms window)	96–98 (unbounded, still climbing)	3
message.retry events	96 (≈1:1 with invocations — every attempt retried)	2
message.failed events	0	1
Stability after +300ms	n/a (still growing)	Unchanged (3/2/1)
Healthy-subscriber control	1 delivery, 0 retry/failed (unaffected)	1 delivery, 0 retry/failed (unaffected)

Baseline independently reproduced twice: authoring session (98 invocations) and, separately, the adversarial critic via its own git stash (96 invocations, plus the additional observation that callbackInvocations === retryEvents exactly — the failure branch was provably unreachable).

5. Darwin Lineage

Skipped — scope mismatch. This is a deterministic correctness fix (thread an existing integer through one re-queue call), not a continuous/categorical parameter; @metaharness/darwin's real interface (darwin <config.json> --execute) evolves scoped tunable parameters against a benchmark corpus. Same class of skip as 4 of the last 5 dream-cycle nights.

6. Flywheel Evidence

No signed @metaharness/flywheel bundle — bespoke deterministic test, not an LLM-task corpus the replay/verify tooling targets, same as every algorithmic candidate the last 5 nights. Evidence retained as the committed test file + git-stash-verified before/after counts + the issue/gist.

7. Reward Hack Check

Manual checklist, independently re-verified by an adversarial critic (fresh session, no authoring context): no existing test weakened (new file only, 218 pre-existing tests untouched); no gold-answer leakage (test asserts public event counts / callback invocations only); no cherry-picking (both failure-path and healthy-path reported); no seed manipulation (fully deterministic); $0 cost. Critic specifically tried to construct a no-op/unrelated hack that would satisfy the exact-count assertions and concluded the actual diff is the minimal, principled fix — not a special-cased one.

8. Security Review

Not security-sensitive — confirmed by the critic. Purely in-process integer accounting (an optional parameter threaded through a private method, always called with a value already tracked in-process). No network/filesystem/credential/MCP-authority surface touched.

9. Regression Analysis

Full package suite: 220/220 passing (218 pre-existing + 2 new, 0 regressions), independently re-run by the critic. tsc --noEmit shows pre-existing environmental errors (missing @types/node/tsconfig lib config) identical with or without this diff (confirmed via git stash) — not attributable to this candidate, same class as multiple prior nights' documented environmental gap.

Disclosed, not fixed here: the critic's review surfaced a separate, pre-existing bug (present since commit 1d365d4, 2026-08-11, not introduced or worsened by this diff) — for a broadcast message, the retry re-queue uses the literal string 'broadcast' as the queue key (since message.to stays 'broadcast' on the message object) instead of the real per-subscriber agentId enqueue() fanned out to, so a failing broadcast subscriber's retries silently vanish into a queue nobody drains. Now flagged with an in-code comment at the fix site; not in scope for tonight's single-conceptual-change diff.

10. ADR

None created — single-file correctness fix to an existing retry-accounting mechanism, not an architectural decision, matching repo convention for scoped bug fixes.

11. Research Gist

docs/dream-cycle/dream-gist-2026-08-19.md (committed on this branch; no gist-creation tool available in this session, consistent with every dream-cycle night since 2026-08-14).

12. Issue

Closes #3061 (full research: 5-role parallel fan-out, ledger recovery for a 5-night append gap, ruview/ruvector scan findings, competitor comparison).

13. Witness

Field	Value
Session commit	44b67bccdd23a99482aa093fbc52a0516bca7257
Gist SHA-256	5e61580eb87fe98f7a489e55eb1fa6dc6493dc78ebd7759b8d8aedb77b15fca9
Witness stamp	62c4fdf7750c8bda9cb3101a732fbe1ea39ca56cbd484a972ce5d8cb126ef052

Verifier procedure: fetch docs/dream-cycle/dream-gist-2026-08-19.md from this branch, SHA-256 it, concatenate with the session commit above, SHA-256 again — result must equal the witness stamp.

14. Merge Policy

Human review required. Do not self-merge. Do not autonomously promote Flywheel state. Verdict: ACCEPT — deterministic, independently-reproduced-twice evidence for a real correctness bug fix; adversarial critic verdict CONFIRMED-WITH-CAVEATS (caveats addressed via disclosure, not hidden); zero regressions; zero new attack surface.

---

Generated by Claude Code