health-check

/health-check — 36-Point Primitive Validation

Run all 36 checks against the current harness. Report pass/fail for each check with details on failures.

How to Run

1. Determine the harness directory (current working directory by default)
2. Run each check below
3. Report results in the format shown at the bottom

Checks

Identity (5 checks)

• I1: CLAUDE.md exists — Check for CLAUDE.md in the harness root. It must contain a purpose statement.
• I2: Persona defined — harness-config.json has identity.persona with at least name and tone.
• I3: Authority hierarchy present — .claude/rules/identity.md exists and contains the authority hierarchy (Base training > Operator > User > Tool results).
• I4: Domain constraints defined — harness-config.json has identity.hardcoded_constraints with at least one entry.
• I5: Conflict resolution order present — .claude/rules/identity.md contains "Safe > Honest > Compliant > Helpful".

Context Contract (5 checks)

• C1: SessionStart hook configured — hooks/hooks.json (or settings) registers a SessionStart handler.
• C2: Tier 1 files exist — Every item in harness-config.json → context.tier1 has a corresponding file in .claude/memory/topics/.
• C3: Tier 2 capabilities registered — Tier 2 items from config have corresponding skills, MCPs, or documented capabilities.
• C4: Absence declarations configured — The SessionStart hook script includes absence declaration logic.
• C5: Context budget rule present — .claude/rules/context-budget.md exists with tier budgets defined.

Honesty Protocol (5 checks)

• H1: Honesty rule with 7 properties — .claude/rules/honesty.md exists and references all 7 honesty properties.
• H2: PreToolUse hook configured — hooks/hooks.json registers a PreToolUse handler.
• H3: Confidence threshold calibrated — harness-config.json has honesty.confidence_threshold set and it matches domain_risk level.
• H4: Abstention triggers defined — harness-config.json has honesty.abstention_triggers with at least one entry.
• H5: Anti-epistemic cowardice rule — .claude/rules/honesty.md contains anti-epistemic cowardice directive.

Memory Protocol (7 checks)

• M1: Memory directory exists — .claude/memory/ directory exists with subdirectories.
• M2: Session summary initialized — .claude/memory/session-summary.md exists.
• M3: Topic categories created — Each item in memory.topic_categories has a directory in .claude/memory/topics/.
• M4: PreCompact hook configured — hooks/hooks.json registers a PreCompact handler.
• M5: Hot tier under size cap — .claude/memory/session-summary.md is under 200 lines.
• M6: WIP ledger rule present — .claude/rules/wip-ledger.md exists. If absent, the harness predates the WIP ledger primitive and should be backported via the 2026-04-13-wip-ledger patch.
• M7: WIP ledger not stale — If .claude/memory/wip.json exists, its status must not be in_progress or interrupted with a last_updated older than 24 hours. A stale active ledger is evidence of an unrecovered interruption — the user should be told and asked whether to resume, archive, or abandon. Ledgers with status of completed or abandoned must have been archived to .claude/memory/wip/archive/; if present at the active path, this is a fail.

Observability (5 checks)

• O1: PostToolUse hook configured — hooks/hooks.json registers a PostToolUse handler.
• O2: Loop detection configured — PreToolUse hook script includes loop detection logic.
• O3: Observability rule present — .claude/rules/observability.md exists.
• O4: Log directories exist — .claude/memory/logs/ and .claude/memory/metrics/ directories exist.
• O5: SessionEnd hook configured — hooks/hooks.json registers a SessionEnd handler.

Learning Protocol (5 checks)

• L1: Stop hook configured — hooks/hooks.json registers a Stop handler.
• L2: Consolidate skill available — /consolidate command exists at .claude/commands/consolidate.md in the harness.
• L3: Learning rule with signal classification — .claude/rules/learning.md exists with signal classification guidance.
• L4: Learning constraints applied — harness-config.json has learning.never_learn with at least one entry.
• L5: Drift check references identity — .claude/commands/consolidate.md includes drift checking against CLAUDE.md.

Self-Containment (4 checks)

The harness must be fully self-contained. It must run and evolve itself without the meta-harness plugin being installed. This category enforces constitution principle 6 (Self-Containment After Instantiation, added 2026-04-13).

• SC1: No runtime plugin references — Grep for CLAUDE_PLUGIN_ROOT across hooks/, .claude/rules/, .claude/commands/, and schemas/ in the harness. Zero matches expected. Any match is a FAIL — the harness has a silent runtime dependency on the plugin. Fix: copy the referenced file into the harness and rewrite the reference to use $CLAUDE_PROJECT_DIR or a harness-relative path.
• SC2: Hook scripts are harness-local — Parse hooks/hooks.json. Every command field must reference a path under $CLAUDE_PROJECT_DIR (or a harness-relative path), not ${CLAUDE_PLUGIN_ROOT}. Every referenced script must exist inside the harness and be executable. FAIL if any command resolves outside the harness, if any referenced script is missing, or if any referenced script is not executable.
• SC3: Schemas copied into harness — $CLAUDE_PROJECT_DIR/schemas/harness-config.schema.json and $CLAUDE_PROJECT_DIR/schemas/mutability-tags.json both exist. These are required for /evolve-harness and harness-config validation to work without the plugin. FAIL if either is absent — recommend applying the 2026-04-13-plugin-decoupling patch.
• SC4: Operational commands copied into harness — $CLAUDE_PROJECT_DIR/.claude/commands/health-check.md, consolidate.md, evolve-harness.md, and context-status.md all exist. FAIL if any is absent. These commands are how the harness manages itself independently of the plugin.

Output Format

=== META-HARNESS HEALTH CHECK ===

✓ IDENTITY (5/5)
  ✓ I1: CLAUDE.md exists with purpose statement
  ✓ I2: Persona defined
  ✓ I3: Authority hierarchy present
  ✓ I4: Domain constraints defined
  ✓ I5: Conflict resolution order present

✓ CONTEXT CONTRACT (5/5)
  ✓ C1: SessionStart hook configured
  ...

RESULT: 36/36 PASS — All primitives intact and harness is self-contained

For failures, show:

✗ H3: Confidence threshold calibrated
  FAIL: confidence_threshold is 0.5 but domain_risk is "high" (expected >= 0.9)
  FIX: Update honesty.confidence_threshold in harness-config.json

Error Handling

• If harness-config.json is not found, skip harness-config dependent checks (I2, I4, C2, C3, H3, H4, M3, L4) and report them as SKIPPED with reason
• If harness-config.json is malformed JSON, report parse error and skip all config-dependent checks
• If hook scripts exist but are not executable, report this as a FAIL with fix instruction: chmod +x <script-path>
• If the harness directory has no .claude/ subdirectory, report all rule-dependent checks as FAIL