v2.86.0 — the harness stops giving confident answers where it does not know

The harness stops giving confident answers where it does not know

One theme across the release: every place the system could fail silently and
render its failure as a clean, confident answer now says what it actually knows.
Informed by the open problems in arXiv:2605.18747 "Code as Agent Harness"
(oracle adequacy, semantic verification, self-evolving harnesses, shared state).

Board always-on (ADR-007)

• great-cto board ensure — idempotent health gate: starts the board only when
  it is not already answering; never kills a healthy instance; restarts a live
  pid whose port is hung (the case a liveness supervisor cannot see).
• great-cto board install-daemon / uninstall-daemon — per-user OS service
  (launchd / systemd --user / schtasks) with start-at-login + restart-on-crash.
  Explicit opt-in, never a side effect of init. --dry-run prints the unit.

Board reads space-path / tasks.md-backed projects

• The tasks.md fallback now parses the Markdown-table dialect the pipeline
  writes when beads cannot open a path (e.g. one containing a space) — with
  header-mapped columns, escaped-pipe handling, and owner-debris filtering.
• Gate approve/reject falls back to a tasks.md status write-back when bd is
  unavailable, and reports which path it used (via: beads|tasks.md).
• Metrics and session-log parsers tolerate real-world heading/format variance;
  the session-log fallback scopes verdicts to the project instead of showing
  every project the same global list.
• The project switcher hides $HOME (the global config dir is not a project) and
  dead registrations with no data source.

Silent failures now speak (the release theme)

• readSafe/parseSafe keep "absent", "unreadable" and "here it is" apart —
  the shared root of five separate board bugs where emptiness rendered as truth.
• GET /api/tasks and /api/logs carry an X-Board-Degraded header when a
  read failed; the UI mounts a persistent banner (visible on every tab), and
  the landing view no longer claims "All clear" / "Nothing urgent" over data it
  could not read. Round-trip verified: break the file → banner; fix it → gone.
• Session-log reads distinguish ENOENT (normal) from a real read failure.

Quality oracle tells the truth (bench)

• parseTestCounts reads vitest and jest summaries, not only node:test — an
  unparsed suite was previously scored as "1 test, and it failed".
• "ran" now means "we obtained real counts"; assess() returns
  overall: null when the suite was not measured, and evaluateGate(null)
  blocks with a reason. A product whose suite exited 143 can no longer carry a
  published grade.
• scripts/bench-env.mjs provisions each product's declared database (env-file
  discovery → compose fallback), relocates onto a port we own when the declared
  one is somebody else's cluster, and migrates the schema with the product's own
  script. Measured on one product: 62/132 without env → 131/132 with it.

Pipeline state survives interruption

• scripts/pipeline-state.mjs reconstructs completed / failed / next /
  remaining stages from the verdict logs; /resume runs it instead of the
  operator hand-writing "these stages are done". Exit 3 when a mandatory stage
  (QA / security) has no terminal verdict.

Learning loop disarmed until it earns trust

• /crystallize approve is gated: a stamped eval regression can never be
  activated (no override); no evidence blocks unless --no-eval "<reason>",
  and the reason is logged. Noisy deltas activate as "not shown to help".
• The GP frontmatter writer now emits the fix:/verification: keys its
  consumers grep — all four live patterns were being injected with an empty
  "→ apply:" line. Consumers fall back to detection steps when no remedy exists.
• continuous-learner delegates cross-project promotion to lessons-merge.mjs
  instead of a prose reimplementation that had already drifted.

Decisions log is per-project (ADR-008)

• Gate decisions land in <project>/.great_cto/decisions.md; with no scope the
  write is refused rather than redirected to the global file. One project's gate
  vocabulary can no longer reach another project's agent context. The global
  file keeps only promoted ## pattern: entries.

Misc

• LLM router default moves to the newest cheap-lane Kimi (k2-0905: 2× context
  for +5%/+9% per token); pricier tiers stay explicit opt-in.
• bench-run.sh can bill against an API key from ~/.great_cto/secrets.env
  (out of repo) instead of the subscription's shared session limit.
• pre-push private terms load from an out-of-repo file
  (~/.great_cto/private-terms), so the leak detector no longer ships the very
  names it guards against.

Tests: 99 board · 199 top-level · 368 lib · 281 CLI — all green.

---