[ruvnet]

Nightly Dream Cycle, 2026-10-07. DEEP=evaluation-adapters, SCAN=flywheel,darwin. Full report: docs/dream-cycle/2026-10-07-evaluation-adapters-report.md. Issue: #155.

Hypothesis (frozen before implementation)

Given dream.config.json's own evaluatorEntrypoints.darwin value is an &&-chain (rm -rf .metaharness && npx @metaharness/darwin evolve . --sandbox mock), when verify-entrypoints gains splitAndChain() — recognizing a command chained only by && with no empty segment, and running each resulting segment via execFile in order, short-circuiting on the first nonzero exit — then verify-entrypoints run against this repo's real config should classify darwin as live instead of a fixed blocked: compound command, subject to: any other control operator (;, |, ||, &), or any empty-segment malformed chain, must still be refused outright exactly as before; a plain non-compound command is byte-identical; npm test stays green with 0 regressions.

Candidate

+39/-0 in packages/cli/src/entrypoint.ts (splitAndChain), +77/-20 in packages/cli/src/index.ts (wiring + verdictExitCode dedup), plus test updates. One conceptual change. looksLikeCompoundCommand's existing refusal is the fallback whenever splitAndChain returns null (any non-&& operator present, or any empty segment — including a quoted-empty-token segment, caught by an independent critic and fixed same session).

This does not duplicate open PR #137 (bin.ts's execFile spawn-failure diagnostics) or #148 (classifyEntrypointResult's stdout fallback) — both read in full first, neither touches looksLikeCompoundCommand/verify-entrypoints' dispatch loop.

Evaluation Receipt

Real evaluator: npm test (vitest run && npm run test:governance).

	Baseline (main@23a5770)	Candidate
vitest	769	778 (+9 new, 0 regressions)
governance	150	150 (unchanged)
Total	919	928
typecheck / lint	clean	clean

Live, against the real dream.config.json, run twice:

$ node packages/cli/dist/bin.js verify-entrypoints dream.config.json
bench: live (exit 0) — produced output
darwin: live (exit 0) — &&-chain, all 2 segments ran: produced output

Identical the second run (the chain's own rm -rf .metaharness && prefix keeps it idempotent). git status --porcelain clean after both.

Baseline

Parent commit 23a577082aad39750051dc7ef0ec9d0d72900adf (main). This session's own fresh checkout was shallow; npm test showed 3 unrelated governance failures (scripts/benchmark-ticket-codec.test.mjs's pinned git show <sha> oracle, object not reachable at depth 50) — confirmed environmental (git rev-parse --is-shallow-repository → true), fixed with git fetch --unshallow, then 919/919 green. This exact class is already owned by open draft PR #150 (security-adversarial, same base commit) — not re-proposed here.

Darwin Lineage

DARWIN=not-applicable — a classifier/dispatch-layer change with a fixed contract, no evolvable population (same precedent as #65/#91/#116/#132/#137/#148).

Evidence

OBSERVATION (live repro: darwin permanently blocked: compound command on the real config) → OBSERVATION (read #137/#148/#150 in full, confirmed no overlap) → HYPOTHESIS (frozen above) → MEASUREMENT (769→778 tests, 0 regressions) → MEASUREMENT (live pre/post real-CLI repro, run twice) → INFERENCE (independent critic subagent, separate context, instructed to try to break it) → DECISION (critic CLEAR + 1 real non-exploitable gap found and fixed same session) → DECISION (verdict ACCEPT).

Reward-Hack Check

Independent adversarial-critic subagent (separate context, no shared authoring state): CLEAR, one real but non-security, low-severity gap found and fixed same session — splitAndChain's original empty-segment guard checked segment length only, so a && "" && b slipped through as a [''] segment (a falsy executable name). Fixed (!s[0] instead of s.length === 0) with a dedicated regression test. execFile is never shell-invoked regardless, so this was never shell-exploitable. The two pre-existing tests this PR replaces had their behavior assertions updated (refuse → correctly execute each segment), not their safety invariant weakened: "never dispatch an executable with wrong/garbage arguments" still holds, now exercised by exact per-segment file/args assertions, including a dedicated short-circuit-on-failure test.

Security Review

No new credential/network/MCP-authority surface. execFile is still never given a shell — only the literal && token is interpreted, and only in this process. Each && segment gets its own correctly-tokenized argv, never mixed with another segment's tokens. Any operator without exact &&-equivalent semantics (;, |, ||, &) is still refused outright, unchanged. Trust boundary unchanged: dream.config.json is repo-committed, PR-reviewed configuration, never runtime-attacker-controlled.

Regression Analysis

0 pre-existing tests removed or weakened. Two pre-existing tests had their assertions updated to match the new, intentional capability (see Reward-Hack Check); their safety property is preserved and more precisely exercised. 9 new tests added across entrypoint.test.ts/index.test.ts. npm run lint/npm run typecheck clean. npm run build clean, no wasm/NAPI degradation.

ADR

None — a dispatch-layer capability extension to an existing classifier, not a new architectural decision (same judgment as every prior single-conceptual-change night on this surface).

Gist

GIST=LOCAL — gh auth status reports an invalid GH_TOKEN this session (same as #148/#150); no gist-creation MCP tool available. Full report committed at docs/dream-cycle/2026-10-07-evaluation-adapters-report.md.

Issue

#155

Witness

report_sha256 : bcbd53b851ce2e213ddf71215d4b62e9199a7b2d7b12fdf023fa090c277eb752
session_commit: 23a577082aad39750051dc7ef0ec9d0d72900adf
witness       : 3e529171e74805a9a2d22e03a1513a7e01880772355cf7ad939d2a902afb24e5

Reproduce: node packages/cli/dist/bin.js witness verify-report docs/dream-cycle/2026-10-07-evaluation-adapters-report.md --commit 23a577082aad39750051dc7ef0ec9d0d72900adf (✓ VALID, self-contained).

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_01LC7cMu6kUD4CYe9vTuSVKQ

---

Generated by Claude Code

[ruvnet]

Dream Cycle — CONSOLIDATE refresh, 2026-10-27

MODE=CONSOLIDATE (ADR-0011 STEP 1.5). Review queue: 20 open dream/* candidate PRs (well over the 6-PR limit; pipelineMergeRate ≈ 0.591, 39 merged / 66 candidate). This PR was the oldest eligible draft (own-receipt ACCEPT, surface ≠ routine-self-review, no prior dream-consolidate marker for today's SHA) — selected over #145/#148/#150/#152/#154, which already carry a marker for this exact SHA.

Actions taken:

• Merged main@3fc3222d2e5566a203331d3686c4d2eaf06713a8 into this branch (merge commit b27eedf, never rebased/force-pushed). One textual conflict (docs/dream-cycle/LEDGER.md, append-only — resolved by date-ordered concatenation, no content changed).
• Re-ran the real evaluator on the merged head: npm test 1081 (baseline main@3fc3222, measured in an isolated worktree) → 1090 (+9, 0 regressions); lint/typecheck clean.
• Live re-confirmed verify-entrypoints against the real, now-drifted (version-pinned + -y) dream.config.json darwin entry — still classifies live, idempotent on a second run.
• Fresh independent critic (separate subagent, no shared authoring context): CLEAR — traced splitAndChain by hand/execution against the drifted config, confirmed execFile-only dispatch (no shell reachability), re-confirmed the reward-hack check and the merge-drift analysis independently. One disclosed, non-blocking, pre-existing gap (not introduced by this refresh): no single test exercises the exact current pinned/-y darwin string through splitAndChain end-to-end.
• Pushed, then CI's dependency-audit job went red on a newly disclosed high-severity advisory (GHSA-r825-w8mm-5pxw + related Vite/launch-editor CVEs) in the locked vite 8.3.0-8.3.2 — confirmed unrelated to this PR's diff (reproduces identically on the unmodified merged lockfile; dependency-security, the production-reachable gate, stayed green throughout). Fixed with the same minimal, dev-toolchain-only repair as this PR's own prior commit (source-map-js) and compiler-parity: buildStep.cmd unvalidated, compiles literal "undefined" into an executed bash block (issue #142) #143's precedent: bumped the exact-pinned vite devDependency 8.3.0 → 8.3.5. npm test unchanged (1090/1090) post-bump.
• Refreshed the report's Evaluation Receipt and Witness section (docs/dream-cycle/2026-10-07-evaluation-adapters-report.md) accordingly; new witness bound to tonight's session commit, verified self-contained VALID.

Result: verdict ACCEPT unchanged. All 9 required checks green on the final head (6d40de4): test (22), test (24), dependency-security, dependency-audit, no-optional-deps, software-evidence, CodeQL/analyze, Human merge policy guard/guard. Critic clear. Marked ready for review per the CONSOLIDATE rule (ACCEPT holds ∧ all required checks green ∧ critic clear).

Human review and merge decision are still required — this session never merges or self-promotes. autoMerge: false; no automerge-safe label applied.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ReiDH5vqu3wMKGXHSCcjvY

---

Generated by Claude Code

[ruvnet]

Check-ins paused — no change across 3 checks (still green, mergeable, waiting on human review). The next GitHub event on this PR, or a message to this session, will bring it back.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ReiDH5vqu3wMKGXHSCcjvY

---

Generated by Claude Code