AGENTS.md

Repository invariants

• README.md is the sole entry point. Current product, roadmap, and delivery authority lives only in docs/current/PRODUCT.md, docs/current/ROADMAP.md, and docs/current/DELIVERY.md; proposed or conflicting plans are HOLD, not permission to expand scope.
• Keep one writer per root. Declare ownership, exact base SHA, immutable target, and mechanical guide before fan-out; serialize migrations, lockfiles, generated files, OpenAPI, CI, and authority records.
• Never use destructive shared-workspace Git operations or overwrite another lane's work. Preserve historical evidence.
• Do not skip, delete, quarantine, or weaken tests without an approved receipt and independent review. Record exact invocations and discovered/executed counts.
• Keep facts, inferences, hypotheses, and legal conclusions distinct. Production exposure and legal/compliance claims require separate authority and evidence.
• Every lane records pre-mortem, blast radius, detection, rollback, stop conditions, review identities, head SHA, and remaining HOLDs.

Detailed current method: docs/current/DELIVERY.md. The program playbook is retained as historical method and reusable reference only.

Task-selected reasoning lenses

All substantive reasoning, planning, implementation, review, and verification must use the smallest task-appropriate subset. Select at least two lenses before nontrivial work, re-evaluate the set when evidence or risk changes, and do not mechanically apply all lenses.

1. Cartesian doubt — challenge assumptions and separate evidence, inference, and uncertainty.
2. Essentialism / YAGNI — pursue the smallest sufficient outcome and avoid speculative scope.
3. Chesterton's Fence — understand why an existing constraint or mechanism exists before removing it.
4. Contrarian / outside-the-box — test non-obvious alternatives when the default framing may be wrong.
5. Socratic — expose hidden premises with focused questions; ask the user only when the answer materially blocks safe progress.
6. Pragmatism — optimize for the real-world outcome under actual constraints.
7. Red Team — model misuse, adversaries, hostile inputs, and ways the plan can fail.
8. Systems Thinking — trace dependencies, feedback loops, second-order effects, and system boundaries.
9. Operability / Day-2 — account for deployment, diagnosis, maintenance, recovery, and ownership after launch.
10. Opportunity Cost — compare the chosen work against the best alternatives in time, complexity, and value.
11. Blast-radius / cell-based — contain changes and failures; prefer independently recoverable boundaries.
12. Constant-work / anti-fragility — avoid input-dependent blowups, degrade predictably, and use stress to improve the system.
13. Shared-nothing / eventual consistency — minimize coordination and make convergence, conflicts, and stale-state behavior explicit.
14. FinOps / unit-cost — reason about cost per useful outcome, including operational and scaling costs.
15. Telemetry-first — make important state, decisions, failures, and success criteria observable.
16. Zero-trust / defense-in-depth — verify every boundary, minimize privilege, and layer independent safeguards.

High-risk authz, migration, contracts, approval, HR/payroll, release, production, and compliance-sensitive work should include Red Team, Operability / Day-2, Blast-radius / cell-based, and Zero-trust / defense-in-depth. Report concise conclusions, evidence, decisions, and tradeoffs rather than private chain-of-thought.

This is guidance, not a filing requirement. A machine cannot verify that a lens was applied, only that a document claims it was, so no gate demands a per-record evidence block. The one lens artifact still enforced is the identifier-only manifest projected into CLAUDE.md, which must not drift from the canonical list above -- two lists either match or they do not.