Integrate Anthropic harness-engineering observations into the methodology

[EsatanGW]

Summary

Closes seven canonical-methodology gaps surfaced by cross-referencing Anthropic's Harness Design for Long-Running Agentic Applications (Mar 2026) and the existing OpenAI Harness engineering: leveraging Codex in an agent-first world (Feb 2026) against this repo.

All seven entries land in [Unreleased]; no version bump. +331 / −1 across 13 files (1 new file).

What changed (seven gaps closed)

Critical — P1 / P2 / P3

	Where	What
P1	docs/ai-operating-contract.md §12	New Context anxiety failure mode — premature task-shrink under perceived context pressure, parallel in structure to §11 verbal-completion illusion. Explicit prefer Context Reset over Compaction remedy. Glossary entry + cross-refs in resumption-protocol.md §Step 2b (outgoing-session symmetric rule) and ai-project-memory.md §Pre-compression protection list.
P2	docs/multi-agent-handoff.md §Acceptance criteria as a Sprint Contract	Planner-side time-axis discipline catching unverifiable AC at write-time, not at Implementer's egress self-check. Two named rules — Reviewer-anticipation (Planner imagines Reviewer's audit) and Reverse-shape (AC text reverse-shapes Implementer choices). Three pre-handoff self-check questions.
P3	docs/harness-evolution-discipline.md (new file) + mode-decision-tree.md row + index registrations	Per-material-model-release re-evaluation of canonical methodology components whose load-bearing-ness depends on a specific model-class failure mode. Sibling to anti-entropy-discipline.md (which excludes canonical methodology by design). Four-step procedure (map / re-test / classify / record). Companion mode-decision-tree.md row makes sweep-backed canonical retirement Lean-eligible — first canonical-component weight-shedding path.

Enrichments — 1d / 1e / B2 / 1f

	Where	What
1d	multi-agent-handoff.md §Single-agent anti-collusion rule §Why this rule exists	Preamble naming self-evaluation bias as the underlying behavioural failure the structural rule enforces against. Explicit "this is not prompt-engineerable away" with the Tool-permission matrix's no-write row as the load-bearing form.
1e	folded into P2 §Reverse-shape rule	AC text reverse-shapes Implementer choices (parallel to agent-persona-discipline.md's observation that medium reverse-shapes persona).
B2	multi-agent-handoff.md §Capability gating by risk level	"Risk is one axis; capability frontier is another" callout. Risk-axis is encoded mechanical enforcement boundary; capability-frontier is the Planner-judgement signal alongside it. Matrix's gating column is a floor, not a ceiling.
1f	mechanical-enforcement-discipline.md §Boundary with non-mechanical evaluation	First canonical three-evaluator comparison: Mechanical / Application-driven / Agentic Reviewer audit, with layering (floor / bridge / ceiling) and a Planner-side allocation rule at Phase 3 (each AC to the cheapest evaluator that catches its failure shape). New anti-pattern: routing by familiarity rather than by failure shape. Companion: multi-agent-handoff.md §Reviewer §Must not do gains a row pointing back.

Discipline

• Tool-agnostic. No vendor / model names in normative content (per CLAUDE.md §2).
• No schema impact. No new manifest fields, no new evidence enums, no new role definitions.
• SoT-consistent. All seven entries cite existing fields and rules; new normative content stays inside the file-role-map.md SoT discipline. New file registered in docs/README.md Tier-3 and docs/file-role-map.md.
• Forced Full mode per CLAUDE.md §Mode implication (canonical methodology edits at L1+).

Test plan

Local CI validators (mirrors of .github/workflows/validate.yml jobs):

• internal-links — no broken relative links across 189 markdown files
• legacy-terms — no legacy terms outside allow-list
• summary-drift — 53 docs, no TL;DR-vs-body drift
• role-consistency — role contract consistent across SoT + 9 mirrors (8 invariants)
• schema-syntax — 4 schemas valid
• changelog-drift — CHANGELOG.json ↔ CHANGELOG.md in sync
• template-conformance — 7 manifest examples valid
• cluster-disjointness — self-test passes
• schema-drift — generated .json matches .yaml
• CHANGELOG.json — valid JSON syntax

Out of scope (deferred)

Two lower-priority items from the original cross-reference were deliberately not addressed:

• 2b (feature-at-a-time sprint cadence) — Anthropic's sprint construct is time-bounded; the repo's design philosophy is scope-driven (surfaces, clusters, modes). Adding time-axis cadence would create tension with existing decomposition philosophy without proportional value.
• 2e (Sprint Contract as persistent artifact) — Task Prompt is intentionally non-persisted (multi-agent-handoff.md §Task Prompt structure); persisting it would split the SoT against the Manifest-as-state-snapshot discipline. Already a "design choice, not gap" in the original analysis.

🤖 Generated with Claude Code