Harness Phase Upgrades Grill Status

Branch Dashboard

Removed Scope

• Strategy and ideation are out of this initiative.
• Product pulse is out of this initiative.

Parked Branches

• Distribution and pack placement, unless implementation planning needs it later.

Glossary

Terms

• TDD Hard Law: A delivery invariant that behavior-changing production code must have failing-test evidence before implementation and passing-test evidence before completion.
• Learning Artifact: A typed post-work knowledge record for a bug or reusable lesson, deduplicated before writing and refreshable over time.
• Phase Wrapper: A Harness workflow entrypoint that orchestrates skills and gates for a lifecycle phase.
• RED Evidence: Recorded proof that the intended public behavior failed before implementation.
• GREEN Evidence: Recorded proof that the same intended public behavior passed after implementation.
• Non-Testable Reason: An explicit rationale for why a task cannot produce meaningful RED/GREEN evidence, paired with alternative verification.
• Task Evidence: Short RED/GREEN proof recorded in the PLAN.md task log before a task can be marked complete.
• Intended-Result Test: A test that proves the public behavior the task is meant to deliver, independent of current implementation shape.
• TDD Status: A compact task state for TDD enforcement, using required, not_testable, recovered, or not_applicable.
• Knowledge Loop: The docs-ingest cycle that captures fresh lessons, deduplicates them, routes them into canonical wiki knowledge, and refreshes them when reality changes.
• Future-Use Hook: An inbound route from future work surfaces back to a learning artifact.
• Consumer Instruction: A phase-skill instruction requiring agents to scan relevant prior learnings during discovery, planning, debugging, or review.
• Learning Type: The artifact track for a learning, either bug for diagnosed failures or knowledge for reusable guidance.
• Learning Refresh Outcome: The docs-ingest decision for an existing learning artifact: keep, update, consolidate, replace, delete, or mark_stale.
• Memory Note: A compact cross-session gotcha or routing aid, not canonical knowledge.

Relationships

• Learning Artifact is produced through Docs Ingest Phase after implementation, debugging, or review reveals durable knowledge.
• Knowledge Loop produces, deduplicates, refreshes, and retires Learning Artifact pages inside the routed wiki/project docs model.
• Consumer Instruction tells phase skills when to look for Learning Artifact pages.
• Future-Use Hook points future agents from specs, notes, flows, concepts, runbooks, skill references, checklists, or memory notes back to a Learning Artifact.
• Learning Type controls which metadata fields a Learning Artifact must include.
• Docs Ingest Phase applies Learning Refresh Outcome during every relevant learning ingest.
• Memory Note may point to a Learning Artifact only when normal wiki/search discovery is likely to miss the gotcha.
• Phase Wrapper activates reusable skills when relevant.
• TDD Hard Law requires RED Evidence and GREEN Evidence unless a Non-Testable Reason is accepted.
• Task Evidence lives in PLAN.md; non-obvious narrative lives in IMPLEMENTATION-NOTES.md.
• Code-Before-RED Recovery uses an Intended-Result Test to recreate real RED evidence, then patches production code to GREEN.
• TDD Status summarizes whether Task Evidence is required, recovered, not testable, or not applicable.

Axioms

• Harness should steal schemas, gates, and artifact lifecycles, not external operating models wholesale.
• Harness should preserve repo-local wiki and scaffold authority.
• New global distribution should avoid splitting context away from the repository.
• TDD strictness should be artifact-backed, not only agent-compliance-backed.
• For behavior-changing implementation work, a task is incomplete until it has RED evidence and GREEN evidence, or an explicit accepted reason_not_testable.
• If behavior-changing production code was written before RED evidence, restart that implementation slice from the failing test or record why the slice is genuinely non-testable.
• Hard enforcement can begin as a phase contract before it becomes a parser-backed CLI validation gate.
• PLAN.md is the auditable execution board for task completion proof.
• IMPLEMENTATION-NOTES.md preserves judgment, surprises, and deviations, not a duplicate task database.
• Code-before-RED recovery should produce a real intended-behavior test and a passing implementation, not just a written apology.
• The agent must not fake RED evidence when existing production code already makes the new test pass.
• The final TDD task schema uses tdd_status, tdd_target, red_command, expected_red_failure, green_command, reason_not_testable, red_evidence, and green_evidence.
• Learning artifacts should compound inside the existing Harness docs authority, not create a parallel knowledge surface.
• A learning is not closed until future phases know to look for it and at least one artifact points back to it.
• Learning artifacts use learning_type: bug | knowledge plus Harness wiki/source metadata and future-use hooks.
• Compounding happens inside docs-ingest-phase by maintaining the learning set as part of every relevant ingest.
• Memory must not become a hidden second learning tree.

Flagged Ambiguities

• None currently active.