v0.23.0

[0.23.0] - 2026-06-10

Schema Compatibility

• Two new schemas added (backward compatible, additive only):
  golden-trace-case (.harness/schemas/golden-trace-case.schema.json) and
  structural-test-result (.harness/schemas/structural-test-result.schema.json).
  Both are registered in the stable schema policy (now 17 contracts). No
  existing schema changed.

Added

• migrate-settings command (alias migrate-config): repairs older
  installs — moves legacy harness.config.json, restores the
  Edit/Write/MultiEdit baseline permissions, merges unmerged hooks and
  compiled permission hints into .claude/settings.json. Doctor now fails
  with the remediation command when it detects this drift.
• Golden-trace regression gate (check:golden-traces): 7 recorded cases
  replay deterministic harness surfaces hermetically and exit 2 naming the
  case and assertion when behavior regresses; wired into the generated CI
  workflow.
• Structural-test --json on all six language adapters (TypeScript,
  Python, Go, Rust, Swift, Kotlin) emitting one structural-test-result
  object for dashboards and PR annotations.
• Loop detection (PostToolUse): repeated edits to one file inside a
  window inject a doom-loop warning and a loop_detected telemetry row.
• Symbol index (symbol-index.mjs): where-is-X navigation to file:line
  across 8 languages with incremental PostToolUse updates.
• Context budget (context-budget.mjs + contextBudget.totalChars):
  explicit per-source allocation for SessionStart context with line-boundary
  truncation and a usage report.
• Observational memory (observation-compress.mjs): PreCompact and
  SessionEnd compress transcripts into dense observations; SessionStart
  re-injects the three most recent.
• MCP skill server (mcp-skill-server.mjs): dependency-free MCP stdio
  server exposing list_skills/load_skill to any MCP client.
• Terrain-correct value-prop benches: eval tasks 05–07 (weak-model,
  greenfield, adversarial) plus docs/benchmarks/value-prop-protocol.md
  replacing the 1-shot A/B design that returned five consecutive nulls.
• Doctor checks for eslint-plugin-boundaries presence/version and for
  settings drift (baseline permissions, unmerged hooks).

Fixed

• Codex advisor runner retries transient failures (timeout, non-zero
  exit) with jittered backoff and records each attempt in the runtime-proof
  JSONL; a hung Codex CLI no longer blocks the task until manual recovery.
• Mistargeted advisor decisions are rejected: any decision whose
  taskId differs from the active task fails instead of being persisted
  into the active task's decision path.
• Telemetry rotation is concurrency-safe: mkdir-as-lock with PID-unique
  temp files and stale-lock recovery; concurrent rotations no longer clobber
  each other's snapshot.
• Task contracts with typo'd layer names now get a denial naming the
  valid layers from the domains config instead of a confusing scope message.
• Advisor decision writes are proof-checked at write time: a decision
  referencing a missing or invented runtime proof is denied by the
  PreToolUse guard with contextual guidance instead of failing later at the
  Stop gate.
• SubagentStop structural checks are cached on a tree fingerprint, so
  unchanged trees skip the workspace-wide rerun (failed runs are never
  cached).

Changed

• src/core/upgrade.mjs split into phase modules under src/core/upgrade/
  behind an unchanged re-export facade.

---

Full history: CHANGELOG.md
Install: npx agent-harness-kit@0.23.0 init