compiler-parity: autoMerge accepts a non-boolean, silently flipping the compiled merge policy (issue #183)

[ruvnet]

Closes #183.

Hypothesis

Given a dream.config.json whose autoMerge value is present but not a boolean (e.g. the string "no", "false", or any other truthy/falsy non-boolean), when validateConfig() is extended to reject it, then dream-machine compile should report a structured ValidationResult error instead of silently compiling a prompt whose STEP 20-25 merge-policy section reflects JS truthiness of the raw value rather than the config author's intent — subject to: zero behavior change for any well-formed boolean value, including this repo's own self-hosted dream.config.json (autoMerge: false), and zero regression in the existing test suite or golden snapshots. Confirmed, not amended after evaluation began.

Candidate

packages/compile/src/config.ts (validateConfig()): one new check, mirroring the existing buildStep.degradeOnWasmFailure guard —

if (config.autoMerge !== undefined && typeof config.autoMerge !== 'boolean') {
  errors.push('autoMerge must be a boolean');
}

packages/compile/src/index.test.ts: +7 tests — non-boolean rejection ('no', 'false', 0, 1, 'true'), boolean acceptance, optional-omission, plus three compile()-level tests proving the real-world consequence (pre-fix a truthy non-boolean compiled the guarded-auto-merge-ENABLED policy text).

docs/dream-cycle/2026-10-15-compiler-parity-report.md: full evidence report (GIST=LOCAL, committed per STEP 17-18's best-effort policy — no gh CLI auth or gist MCP tool this session).

+11/-0 net lines across the two source files, well under the 300-line target, one conceptual change.

Evaluation Receipt

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

	vitest	governance
Parent (baseline, b2abcc1)	809 passed	195/195 passed
Candidate	820 passed (+11, all new)	195/195 passed (unchanged)

npm run lint/npm run typecheck clean. Self-hosted dream.config.json compiles byte-identical before/after, independently re-confirmed by an adversarial critic subagent from a separate run.

Baseline

Parent commit b2abcc14a632d7e0948962065a7c6c8920cf306c (main), measured live via git stash/git stash pop on this session's own checkout — not inferred from logs. Baseline's own fresh shallow clone initially failed 3/195 governance tests on the already-filed, already-open issue #149/PR #150 (BASELINE_COMMIT git-show oracle on a shallow clone); resolved session-locally via git fetch --depth=1 origin <that exact pinned SHA> before the baseline run, so it is not conflated with this candidate's regression count (same handling as the 2026-10-10 report).

Darwin Lineage

DARWIN=not-applicable — single validation-shape fix, no evolvable population or numeric objective, consistent with unbroken precedent (#29, #79, #105, #111, #127, #142/#143, #164/#165).

Evidence

OBSERVATION (zero type check, bare-truthy downstream read) → OBSERVATION (live pre-fix repro: autoMerge: "no" compiles "guarded auto-merge ENABLED") → HYPOTHESIS (frozen above) → MEASUREMENT (809/195 → 820/195, live before/after) → INFERENCE (independent critic's own grep found the only other autoMerge consumer, scripts/automerge-guard.mjs, already uses strict === true equality, never exposed) → DECISION (ship; ACCEPT).

Reward-Hack Check

Independent adversarial critic (separate subagent context, no authorship stake, read the live diff and repo state directly): CLEAR on first pass. Verified autoMerge: null is also rejected (closing an obvious bypass: null !== undefined and typeof null !== 'boolean'). Verified the only other real consumer of this field, scripts/automerge-guard.mjs's evaluateGuard/inspectPullRequest, already uses strict boolean equality (config.autoMerge !== true) and so was never exposed to this defect. Independently rebuilt and reran the full suite (820/820 vitest, 195/195 governance, 0 regressions) and re-compiled the self-hosted config twice (byte-identical both times). Confirmed via its own search_issues/search_pull_requests calls this is not a duplicate. No fixture/threshold/gold-answer touched at any point; no test weakened.

Security Review

Pure in-memory type validation — no new I/O, credential, network, or tool/MCP authority surface. Closes a (low-severity, config-authoring-time) integrity gap in this self-hosting system's own authoritative control-plane compiler. Does not grant, and could never grant, auto-merge or self-promotion authority — only prevents the textual description of that policy from silently inverting in a compiled prompt. The one other real-world consumer of this field (CI-facing scripts/automerge-guard.mjs, which gates actual auto-merge eligibility) was confirmed never exposed, since it already uses strict equality.

Regression Analysis

0 regressions: vitest 809→820 (+11, all new), governance 195/195 unchanged. No existing test weakened, no threshold changed, no gold data touched. Independently re-verified by the critic from a from-scratch rebuild.

ADR

None — parameter/validation-shape fix, consistent with every prior compiler-parity fix of this shape (#29, #79, #105, #111, #127, #142/#143, #152, #164/#165 — none has an ADR).

Gist

GIST=LOCAL — no gh CLI (invalid token this session) or gist-creation MCP tool available. Full report committed at docs/dream-cycle/2026-10-15-compiler-parity-report.md.

Issue

#183

Witness

report_sha256 : 3169cdad4ee28c9b2fd78f98e1ced4ffd5ff918ca72b3729dc4724aef26cef9a
session_commit: b2abcc14a632d7e0948962065a7c6c8920cf306c
witness       : 316c77d1902ce2bcbe5eab6f13165e41c2385df39279ada2d25fd705e7cec886

Reproduce: node packages/cli/dist/bin.js witness verify-report docs/dream-cycle/2026-10-15-compiler-parity-report.md --commit b2abcc14a632d7e0948962065a7c6c8920cf306c

Merge Policy

Human review required. This session never self-merges and never autonomously promotes candidate state. No automerge-safe label applied.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01DQn3topashe98CEmrcK2zm

---

Generated by Claude Code