[Dream Cycle 2026-09-09] developer-experience: verdictColor() leaks ANSI color through --no-color + cli,tui scan

[ruvnet]

Rotation

DEEP=developer-experience · SCAN=cli,tui · SLOT=4 (20260909 % 5 == 4). No bonus deep dive (20260909 % 25 == 9, % 75 == 34).

Ledger Check

docs/dream-cycle/LEDGER.md on main (39 rows) inspected. GitHub-authenticated (get_me → ruvnet) — re-checked the last 7 rows' PR fates directly via pull_request_read. Found: the GitHub MCP list_pull_requests tool reports merged: false for every PR in this repo, including ones with merged_at set and independently confirmed merged: true via pull_request_read (get) on #21, #29, #33, #46, #55, #91 — the same defect class PR #89 already fixed inside this repo's own learningSignals(). Recomputed ledger signals --merged <confirmed list>: zeroMergeStreak=false (a large batch-merge landed 2026-09-07, ~40 PRs). Not actionable as a code change here (the defect is in the calling GitHub MCP server, not this repo) — recorded for cross-night memory; see report's Next Steps §3.

Deep Dive

packages/cli/src/tui.ts's renderDashboard(ledgerMd, { noColor: true }) still leaks a raw ANSI SGR color escape in the verdict column. PR #21 (2026-08-19, merged) explicitly flagged this exact defect as a known, unfixed follow-up ("verdictColor() bypasses the no-color proxy, leaking a raw ANSI escape into --no-color output"). Reproduced live tonight against main @ 3edd426 before writing any fix:

$ node -e "renderDashboard(md, { noColor: true })"
contains ANSI escape with noColor:true -> true
"est                 \x1b[32mACCEPT         "

Root cause: verdictColor(v: string): string closed over the module-level, always-colored C constant directly, instead of the noColor-aware c (either C itself, or a Proxy mapping every key to '') that renderDashboard already resolves and threads through every other cell in the same row.

Hypothesis

Given renderDashboard(ledgerMd, { noColor: true }), when verdictColor() is changed to accept and use the resolved no-color-aware color map (c) instead of the raw C constant, then noColor: true output should contain zero ANSI escapes for any verdict value, while noColor: false output stays byte-identical to today. Confirmed.

Evaluation Receipt

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

	Baseline (3edd426)	Candidate
vitest	616	618 (+2, 0 removed/modified)
governance	81	81
Total	697	699
Lint	clean	clean
Build	clean (7 pkgs)	clean (7 pkgs)

Live: pre-fix leaks \x1b[32m in noColor:true output; post-fix does not; colored (noColor:false) output unchanged (\x1b[32m/\x1b[31m still present). Self-hosted dream.config.json compile output byte-identical before/after (out of scope, verified anyway).

Darwin Results

Not run — DARWIN=not-applicable. Single deterministic color-dispatch parameterization, no evolvable population (same judgment as PRs #11/#21/#29/#33/#46).

Evidence

OBSERVATION (grade A, live, pre-fix code read + reproduction) → OBSERVATION (grade A, PR #21's own body names this exact defect, "not fixed") → MEASUREMENT (grade A, live noColor leak before/after) → MEASUREMENT (grade A, 697→699 tests) → DECISION (parameter-thread c rather than add a second no-color check inside verdictColor).

Reward-Hack Check

Independent critic (fresh subagent, no shared authoring context): CLEAR. Confirmed the bug is real by reading pre-fix code and reproducing the leak first-hand (not on faith); confirmed the fix fully closes it (grepped for any remaining direct C.xxx reference in renderDashboard or its helpers — zero); no weakened/skipped tests, no gold-data edits, no threshold changes; verified by hand both new tests are bug-exercising (the noColor test fails pre-fix, passes post-fix; the colored-path test passes on both, correctly proving no regression); no regression to colored output (c === C when noColor is false); scope minimal (4-line function + 1-line call site + 2 additive tests).

Security Review

No security-sensitive surface: no prompt-injection vector (no LLM calls in this path), no MCP/tool-authority change, no credential exposure, no filesystem/network scope change, no evaluator/gate/safety-constant touched. Pure ANSI-escape string-formatting logic.

Scan Findings

• cli: the GitHub MCP list_pull_requests tool's merged field is unreliable in this environment (always false); see Ledger Check above. Not a dream-machine code defect.
• tui: no other direct C.xxx reference bypassing the c no-color proxy remains in renderDashboard or its helpers (confirmed by grep, both by this session and independently by the critic).

Competitors

Sakana AI Scientist (plain-text CLI output, structurally can't leak — grade C, general repo inspection); OpenHands/rich-based TUIs (same bypass-a-plain-mode-wrapper footgun is a recognized recurring class across rich/ANSI-wrapper tools generally — grade C); DSPy/GEPA (keeps machine-readable logs entirely separate from any colorized renderer, sidestepping the bug class — grade B, documented format); SWE-agent (routes color through a single per-line wrapper, reducing but per public issue history not eliminating the same risk — grade C). Full rows with grades in the committed report.

Gist

GIST=LOCAL — no gh CLI or gist-creation MCP tool available this session (consistent with every prior night). Full report committed at docs/dream-cycle/2026-09-09-developer-experience-report.md.

Witness

report_sha256 : 69a97031631f19d559ebc5c0d7601d36cb02efa7708fda9776eb1616e7372ebe
session_commit: 3edd426f6c9c4b1e80235f7447dc863e749345cc
witness       : 16c867ef6fc58d714ac698888e99b25e24e6f39a21e45bf4fcc7e03098b23ee0

Recommendation

evaluated: accepted — sufficient evidence to recommend human review. PR to follow, draft, human-review-only per this repo's merge policy.

---

Generated by Claude Code