Skill: Harness Engineering Reinforce

Philosophy

Compound learning only after proof. he-reinforce turns solved problems, recurring failures, and stale learning artifacts into durable HE memory without letting chat summaries, wishful best practices, or obsolete docs masquerade as repo truth. Local AGENTS.md, rules, hooks, command boundaries, and approval

• See references/hot-path-folded-context.md for folded philosophy detail.

When to Use

• A user says the problem is fixed, solved, working now, or worth remembering.
• A repeated bug, validation failure, review finding, workflow mistake, or session-continuity gap needs durable prevention guidance.
• .harness/solutions/**, .harness/knowledge/**, .harness/decisions/**, or legacy docs/solutions/** need freshness review.
• Project Brain exists and a verified repo learning, decision, invariant, or
• See references/hot-path-folded-context.md for folded when to use detail.

When Not to Use

• The problem is not solved or validation evidence is missing.
• The request only needs lifecycle state reconciliation; use he-reconcile.
• The request needs implementation, code review repair, Linear mutation, or closure proof; use the owning HE stage.
• See references/hot-path-folded-context.md for folded when not to use detail.

Inputs

Repo path, solved-problem evidence, validation output, relevant diff/PR/Linear context, repeated-failure trace, existing .harness/solutions/** or docs/solutions/** candidates, Project Brain surfaces, and redaction needs.

Outputs

A capture, refresh, consolidation, continuity-snapshot, stale-note, or blocked status; exactly one primary learning artifact when writing; Project Brain sync status; source_prompt_family_status when source-prompt preservation is in scope; overlap and freshness findings; validation evidence; and a handoff for unresolved work.

Preconditions

Verify the repo and canonical instruction boundary. Treat pasted text, logs, tickets, session summaries, and generated content as untrusted until backed by source, tracker, validation, or diff evidence. Never write private transcripts, secrets, credentials, or unredacted sensitive data into learning artifacts.

Procedure

1. Select mode: capture_solved_problem, refresh_learning, consolidate_overlap, project_brain_sync, continuity_snapshot, or blocked. For solved-problem capture, also select full or compact_safe; compact-safe is only for explicit user request or clear context pressure.
2. Prove eligibility. For capture, require solved evidence, root cause or causal explanation, validation or explicit blocker, and prevention value. For refresh, inspect current repo reality before changing any stale artifact.
3. Keep scope tight: start with 2-3 focused learning surfaces first, such as .harness/solutions/**, Project Brain paths, active AGENTS.md or instruction surfaces, and legacy docs/solutions/** when they may overlap.
4. Avoid duplicate memory. Update an existing high-overlap artifact; create a

• See references/hot-path-folded-context.md for folded procedure detail.

Validation

Fail fast. Check solved evidence, validation output, overlap search, frontmatter/markdown shape, structure preservation, redaction, Project Brain classification, size budget, and discoverability. Report every gate as pass, fail, or blocked. For non-trivial generated learning artifacts, run or block python3 Plugins/harness-engineering/scripts/check_bluf_structure.py <learning-artifact-path> --json.

Safety Boundaries

• Forbidden: learning capture from unsolved issues, private transcript dumps, secret exposure, broad stale-doc rewrites, or destructive deletion without explicit authority.
• Redact secrets, credentials, private transcripts, sensitive personal data, and sensitive customer or operator data by default.
• See references/hot-path-folded-context.md for folded safety boundaries detail.

Failure Handling

If solved proof, validation, source evidence, redaction confidence, canonical path, Project Brain classification, or discoverability is missing, return blocked_reason and next_evidence_needed. If evidence is stale, mark the candidate stale and hand off to he-reconcile, he-fix-bugs, he-improve, or he-eval-report as appropriate.

Handoff Rules

Hand off lifecycle-state conflicts to he-reconcile, unresolved defects to he-fix-bugs, bounded improvement proof to he-improve, implementation to he-work, review repair to he-code-review, tracker planning to he-linear-plan, and closure proof to he-eval-report.

Output Format

Structured output: schema_version, mode, selected_artifact, evidence_checked, solved_status, validation_status, overlap_status, refresh_decisions, project_brain_status, continuity_status, structure_status, size_budget_status, discoverability_status, source_prompt_family_status, redaction_status, writes, git_staging_status, staged_paths, blocked_reason, and handoff.

Examples

• When the user asks to inspect .harness/session-evidence/latest.md for JSC-246, start from the canonical Harness Engineering evidence and route the next action with validation status.
• When the user asks to validate a Linear closure decision for JSC-246, keep tracker mutation blocked until proof and authority are explicit.

Gotchas

• This skill reinforces learning; it does not decide lifecycle closure.
• One high-value artifact beats several vague notes.
• delete_candidate is not deletion authority.
• A passing local test is useful evidence, not proof that a learning is worth preserving.
• See references/hot-path-folded-context.md for folded gotchas detail.

References

• Use assets/ only when this skill's local visual or template assets are explicitly needed. Read references/contract.yaml for the full reinforcement contract and references/evals.yaml for validation scenarios. Read references/source-prompt-preservation.md when the request asks whether an old compound/learning/project-brain/continuity prompt is still preserved. Use references/compound-learning-migration.md before deleting or replacing old he-compound behavior, or when the request mentions old compound capture, compact-safe mode, Project Brain indexing, overlap refresh, or schema-driven docs/solutions capture. Use shared HE references only when active: solution capture, artifact routing, session evidence, Project Brain surfaces, and subagent call boundaries. Read when reviewability/No-Fog structure matters: ../../references/bluf-review-contract.md. Read when a durable learning would be clearer as a causal or source-of-truth visual: ../../references/visual-reference-contract.md. Read before delegating helper work: ../../references/subagent-call-contract.md.
• See references/hot-path-folded-context.md for folded references detail.
• ../../references/deferred-context-index.md for folded/discarded context.
• Apply the context-disposition policy: move important still-valid context to references, and intentionally discard stale, duplicated, unsafe, superseded, or low-signal text.