[ruvnet]

Nightly Dream Cycle, 2026-10-02. DEEP=evaluation-adapters, SCAN=flywheel,darwin. Full report: docs/dream-cycle/2026-10-02-evaluation-adapters-report.md. Issue: #147.

Hypothesis (frozen before implementation)

Given classifyEntrypointResult() receives an ExecResult from a real evaluator-entrypoint failure where the underlying tool (Node's node --test, the governance half of this repo's own bench entrypoint) writes its failure diagnostics to stdout and leaves stderr empty, then falling back to r.stdout (clearly labeled) when r.stderr is empty on a nonzero exit should surface that diagnostic instead of the current "exited N with no stderr", subject to: a failure with real stderr content must produce a byte-identical reason to today (stdout is never consulted when stderr is non-empty); a failure with both streams empty still produces a clear, distinct generic message; the existing stale-state carve-out (which only inspects stderr) is unchanged and runs first; npm test stays green with 0 regressions. Not modified after evaluation began.

Candidate

packages/cli/src/entrypoint.ts (+19/−1) and packages/cli/src/entrypoint.test.ts (+34 lines, 4 new tests). One conceptual change: a stdout fallback in classifyEntrypointResult's blocked branch, gated strictly behind stderr being empty.

// after:
if (!stderr) {
  const stdout = r.stdout.trim();
  if (stdout) {
    return { verdict: 'blocked', code: r.code, reason: `(stdout) ${stdout}` };
  }
}
return {
  verdict: 'blocked',
  code: r.code,
  reason: stderr || `exited ${r.code} with no output on either stream`,
};

Distinct from open PR #137 (bin.ts's execFile catch-block for a spawn failure, e.g. ENOENT — its fallback only fires when err.code is non-numeric, inert for a normal nonzero exit). Read #137's full diff first to confirm no overlap.

Evaluation Receipt

Real evaluator: npm test (this repo's own bench entrypoint).

	Baseline (main@23a5770)	Candidate
vitest	769	773 (+4 new, 0 regressions)
governance	150	150 (unchanged)
Total	919	923
npm run typecheck	clean	clean
npm run lint	clean	clean

Live pre/post-fix proof through the real built CLI:

$ node packages/cli/dist/bin.js verify-entrypoint stream-probe --cmd "node --test <failing-file>"
# BASELINE (pre-fix):
stream-probe: blocked (exit 1) — exited 1 with no stderr
# CANDIDATE (post-fix):
stream-probe: blocked (exit 1) — (stdout) TAP version 13
not ok 1 - intentionally failing to observe which stream node:test writes to
  ...
  error: 'STREAM_PROBE_MARKER\n\n1 !== 2\n'

Stream-separation proof backing the hypothesis: a throwaway failing node --test file produced 1040 stdout bytes / 0 stderr bytes; npm run test:governance on this repo's own real governance suite produced 27057 stdout bytes / 0 stderr bytes.

Darwin entrypoint: not exercised live tonight — this session's own execution-permission policy denied npx @metaharness/darwin ... outright before completion. Recorded as DARWIN_LIVE_CHECK=blocked (sandbox policy), not worked around. LLM_EVAL was not blocked (OPENROUTER_API_KEY present, unused by this candidate).

Baseline

Parent commit 23a577082aad39750051dc7ef0ec9d0d72900adf (main). Initial npm test on this session's shallow checkout showed 3 unrelated failures in scripts/benchmark-ticket-codec.test.mjs (pinned-commit git show on a path absent from that commit); confirmed environmental (shallow clone — both ci.yml and dream-nightly.yml already set fetch-depth: 0 for this exact "pinned historical codec differential oracle"), fixed with git fetch --unshallow, then 919/919 green.

Darwin Lineage

DARWIN=not-applicable — a single-branch fallback in one classifier function has no evolvable population to fitness-search, same precedent as every prior single-conceptual-change night on this surface (#65, #91, #116, #132, #137).

Evidence

OBSERVATION (classifyEntrypointResult's blocked branch reads only r.stderr) → OBSERVATION (live repro: failing node --test → 1040/0 stdout/stderr bytes) → OBSERVATION (same pattern on real npm run test:governance: 27057/0) → MEASUREMENT (pre/post real-CLI repro) → MEASUREMENT (919→923 tests, 0 regressions) → INFERENCE (independent critic, separate context) → DECISION (critic CLEAR, one disclosed note) → DECISION (verdict ACCEPT).

Reward-Hack Check

Independent adversarial-critic subagent (separate context): CLEAR. Verified via git diff HEAD that the change is purely additive — one new conditional block in entrypoint.ts, four new tests in entrypoint.test.ts, zero existing tests modified/weakened/removed. The fix checks r.stderr.trim() truthiness generically, not special-cased to any fixture string; one new test actually strengthens the contract (stdout must NOT leak into reason when stderr is present). Confirmed the only two call sites (index.ts) just log check.reason, no hidden string-matching coupling elsewhere in the repo. No benchmark/gold data touched; no hidden cost; no undocumented cache; no threshold changed.

Security Review

classifyEntrypointResult never executes anything — it only classifies an already-completed ExecResult; no new network/filesystem/credential surface. Disclosed, non-blocking: this modestly widens what can appear in a blocked reason — previously only stderr content, now stdout too when stderr is empty, with no truncation. Mirrors the trust boundary already applied to stderr (same consumer, same lack of a size cap) — a widening of an existing boundary, not a new exposure class. No evidence either configured entrypoint (bench, darwin) emits secrets to stdout on failure.

Regression Analysis

0 pre-existing tests modified or removed. All 919 baseline tests pass unchanged; 4 new vitest tests added. npm run lint/npm run typecheck clean. npm run build clean, no wasm/NAPI degradation.

ADR

None — a diagnostic-completeness fix to an existing classifier's blocked branch, not a new architectural decision or repo-wide invariant.

Gist

GIST=LOCAL — no gh CLI gist capability (gh auth status reports an invalid GH_TOKEN) or MCP gist-creation tool available this session. Full report committed at docs/dream-cycle/2026-10-02-evaluation-adapters-report.md.

Issue

#147

Witness

report_sha256 : 643ad95981d507ff636d8a14a0525dc59dcff80158f8d4a722414783cdd07272
session_commit: 23a577082aad39750051dc7ef0ec9d0d72900adf
witness       : 29c998e2cef84d47a01469e871f3322d3dc5d7c0622128548b3f276b13c0a57b

Reproduce: node packages/cli/dist/bin.js witness verify-report docs/dream-cycle/2026-10-02-evaluation-adapters-report.md --commit 23a577082aad39750051dc7ef0ec9d0d72900adf

Merge Policy

Human review required. This session never self-merges and never autonomously promotes candidate state. dream.config.json's autoMerge: false; no automerge-safe label applied.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01TAHRTU6Nrq1aPvuS2eBoK1

---

Generated by Claude Code

[ruvnet]

Consolidation refresh — 2026-10-23 (MODE=CONSOLIDATE)

20 open dream/* candidate PRs tonight, well over implementationTrack.reviewBacklogLimit=6. This was the oldest eligible draft (verdict ACCEPT in its own Evaluation Receipt, no prior dream-consolidate marker for today's SHA).

Done: merged main@3fc3222d2e5566a203331d3686c4d2eaf06713a8 into this branch (merge commit 5d29d638, no rebase, no force-push; one conflict in LEDGER.md, resolved by keeping both sides' rows). Re-ran the real evaluator on the new head: npm test 886→890 vitest (+4, this candidate's own tests, 0 regressions), governance 195 unchanged; typecheck/lint/check:edge-contracts all clean. Re-ran an independent critic + reward-hack pass on the diff (byte-identical to the one originally reviewed — no competing changes landed on main for entrypoint.ts/entrypoint.test.ts since baseline): CLEAR. Refreshed the Evaluation Receipt and Witness in docs/dream-cycle/2026-10-02-evaluation-adapters-report.md (new witness f0fe18ae…, verified self-contained via dream-machine witness verify-report). Verdict ACCEPT still holds.

Not done — staying draft. The dependency-audit check (npm audit --audit-level=high) fails on this head. This is not this PR's failure: package-lock.json on this branch is byte-identical to main's (git diff origin/main...HEAD -- package-lock.json is empty), and running npm audit --audit-level=high locally against that exact unmodified lockfile reproduces the identical two findings (@vitest/mocker moderate, vite high — both transitive npm-registry advisories, not a lockfile change on this branch). It's the same class of pre-existing, already-tracked dev-toolchain audit drift as issue #37 (open, unresolved, no fix PR exists yet to port — remediation there is a semver-breaking vitest/vite major bump, out of scope for this single-classifier-function diagnostic fix). Not re-running the check: this isn't a flake (reproduced deterministically, twice, against an unmodified lockfile), so a re-run would just fail the same way. Per policy this PR never self-merges and never gets automerge-safe; leaving it draft rather than ready-for-review since not every required check on the head is green, even though the gate the Dream Cycle itself controls (evaluation, critic, reward-hack, witness) all clear.

---

Generated by Claude Code

[ruvnet]

Consolidation refresh — 2026-11-08 (MODE=CONSOLIDATE)

18 open dream/* candidate PRs tonight (#98,#101,#137,#141,#145,#148,#152,#154,#156,#161,#180,#182,#184,#189,#192,#199,#202,#204), well over implementationTrack.reviewBacklogLimit=6. #145/#156 are already ready-for-review (non-draft) and #199/#202/#204 are routine-self-review PRs — both excluded from eligibility, same as #98 (compound multi-verdict portfolio receipt, not a single PR-level ACCEPT) and #101 (incompatible legacy ledger-schema conflict, per last night's #145 writeup). This PR (#148) was again the oldest eligible draft: verdict ACCEPT in its own Evaluation Receipt, not self-review, no dream-consolidate marker for today's default-branch SHA (the prior marker, from 2026-10-23, named 3fc3222…, which main has since moved past).

Done: merged main@71d607bea6608180b8c41980dc5386526d3f1f16 into this branch (merge commit 3bd4c4fd9cf3817974e97f544aafea60c7aaccaa, no rebase, no force-push; one conflict in LEDGER.md, resolved by taking the union of both sides' rows, de-duplicated by date, chronological order — dream-machine ledger verify --since-row 38 reports ✓ ledger OK — 60 rows). packages/cli/src/entrypoint.ts/entrypoint.test.ts (this candidate's only files) received zero competing changes on main since the last refresh — the diff is still byte-identical to the one originally reviewed on 2026-10-02.

Re-ran the real evaluator on the new head: npm test 908→912 vitest (+4, this candidate's own tests, 0 regressions), governance 197 unchanged; typecheck/lint/check:edge-contracts/check:development-policy all clean. Ran a fresh independent critic (separate subagent, no prior context, instructed to try to kill the candidate) + reward-hack pass on the diff: CLEAR (full checklist and two non-blocking disclosures in the refreshed report, docs/dream-cycle/2026-10-02-evaluation-adapters-report.md). Darwin ran live tonight, bounded mock sandbox: Winner: g2_v5 · Delta over baseline: +0.110, exit 0 — generic control-plane check, DARWIN=not-applicable to this single-conceptual-change candidate (unchanged precedent). Refreshed the Evaluation Receipt and Witness (new witness 8c431495…, self-verified via dream-machine witness verify-report). Verdict ACCEPT still holds.

The dependency-audit blocker from the two prior consolidation rounds (2026-10-23/10-24/10-28) is resolved. npm audit --audit-level=high — the exact command ci.yml's dependency-audit job runs — now exits 0 on this head (10 moderate-only dev-toolchain findings remain, 0 high/critical; cleared by #205's vite 8.3.6/vitest 5.1.0 bump, already on main, confirmed via issue #197). npm audit --omit=dev and the production-scoped audit-gate both report 0 findings at any severity.

CI on the pushed head (3bd4c4f). All 9 checks green: guard, test (22), test (24), no-optional-deps, dependency-security, dependency-audit, software-evidence, CodeQL/analyze. analyze (CodeQL) failed on the first run with a transient HTTP 504 from GitHub's own CodeQL-bundle CDN during codeql-action/init (before any code analysis ran — confirmed via job logs, three retries, all 504, nothing candidate-related); re-ran just that job (rerun_failed_jobs) and it passed clean on retry. This is the first night this exact PR has had every required check green.

Decision. evaluation_complete ∧ effect_positive ∧ significance_sufficient ∧ no_material_regression ∧ tests_green ∧ reward_hack_clear ∧ critic_clear ∧ witness_valid ∧ receipt_reproducible all hold, and every required check on the exact pushed head is green → marking ready for review. This is not approval — human review required, this session never self-merges and never applies automerge-safe. Ledger row appended: docs/dream-cycle/LEDGER.md (2026-11-08).

No merge, no close, no force-push, no self-promotion.

---

Generated by Claude Code