minhtran3124/agent-harness/context-propagation-audit

Context-Propagation Audit

Run only when the changed diff trips the workflow-engine hard gate in harness-manifest.json: skills, skill dispatch/reviewer/scorer prompts, agents/, or rules/. It is a delivery oracle, separate from runtime correctness and intent review.

Audit matrix

For each changed instruction or policy, enumerate every relying consumer and record:

Use graph tools to find consumers, then corroborate every load-bearing claim against source with grep/read. A no-consumer result is unknown unless it names the search surface. Main-session evidence never proves a fresh child context.

Fail conditions and repair

Fail the audit when a load-bearing instruction is assumed/unconfirmed, child delivery is inferred from main-session behavior, or an inline policy subset is unanchored. Prefer an explicit Read; an inline copy is allowed only when complete and protected by a drift test. Do not flag a complete pasted authority, explicit Read, or registry-linted copy.

Write the matrix and PASS/FAIL to SUMMARY.md under ### Context-Propagation Audit. List each failed row and repair; route an unfixable Rule-4 issue to ESCALATIONS.md. Standalone audits return the same matrix inline. Preserve the known escape probes in evals/skills/review-chain/fixtures/context-rule-unread/ and evals/skills/review-chain/fixtures/stale-inline-policy/.

The audit is invoked before correctness review by SDD only when the trigger fires. Workflow-engine changes are high-risk at intake; do not lower the lane to bypass this proof.