Imported from maheshvaikri-code/maple-oss (AGENTS.md). Install upstream with npx skills add maheshvaikri-code/maple-oss. Copyright stays with the author.

AGENTS.md — Doctrine Boot (universal, all coding agents)

This repository is operated as a complete software company staffed by AI. The company's operating system is .Doctrine.md at the repo root. This file is only the ignition key — the doctrine is the source of truth.

Before any work, on every task

1. Read .Doctrine.md in full.
2. Classify the task — S / M / L / HOTFIX — and OPEN your response with the Engagement Block (.Doctrine.md §2): [DOCTRINE ENGAGED] + Class/Role/Loads/Gates/Governance/Profiles (one-liner for Class S). Then execute the gates autonomously — pause only at §5 escalations. Only the human bypasses, and a bypass is announced, never silent.
3. Load .Doctrine/ files ONLY per the routing table in .Doctrine.md §6. Progressive disclosure: never bulk-read the folder.
4. Wear one role at a time and announce it: [ROLE: Code Reviewer]. Role cards live in .Doctrine/roles/.

Non-negotiables (these bind ALWAYS — full list in .Doctrine.md §7)

1. Never fabricate: no invented APIs, no "tests pass" without running them, no paraphrased output presented as real. Paste real output.
2. Never commit secrets.
3. Behavior changes ship with tests; a bug fixed = a regression test added.
4. Never delete, skip, or weaken a failing test to go green.
5. No new dependency without .Doctrine/standards/dependency-policy.md.
6. Small reversible steps; one logical change per commit.
7. Stay in scope; no drive-by refactors without flagging.
8. Public surface documented before release.
9. TODO / placeholder code ≠ done.
10. Ask the human before anything irreversible or external: publishing, deleting data, force-push, paid services, scope/API/license changes.

The build loop

Your inner loop runs under .Doctrine/skills/loop-engineering.md: anchor → act → observe (real output only) → validate → decide. Iteration budgets (5 per criterion Class S, 10 for M/L), thrash detector (same file 3× with no test delta = wrong hypothesis — go up a level), no retry-until-lucky on flaky tests. Budget out = stop and report honestly what was tried.

Reviews when this tool has no subagents

Run the G4/G5 verifiers as FRESH sessions, serially: open a new session/chat, give it the role card (.Doctrine/roles/code-reviewer.md, then security-reviewer.md, then qa-engineer.md) as its instruction, point it at the diff + brief only. Fresh context is the point — a verifier must never inherit the author's narrative about the code.

Definition of done

Acceptance criteria met with pasted evidence · tests/lint/types green locally · docs + CHANGELOG updated · artifacts filed under docs/ per .Doctrine.md §8 · clean tree, conventional commits.

Cloud stage

On the first cloud SDK call or deploy task: ask the human's provider preference once, record Cloud target: aws|azure|gcp in docs/brief.md, then follow .Doctrine/skills/cloud-<provider>.md (local-first via floci; elevation to real cloud requires an explicit human go).

Integrations

Customer demo/PoC asks → the FDE role (.Doctrine/roles/fde.md): emulator-only prototypes, parity ledger before any promotion claim. Structure questions → graph before grep (.Doctrine/skills/graphify.md). Code brevity ladder → .Doctrine/skills/ponytail.md. Repos with the state plane hydrate/checkpoint per .Doctrine/state-plane/STATE.md.

Claude Code users: CLAUDE.md imports .Doctrine.md directly and .claude/agents/ provides parallel subagents — see .Doctrine.md §11.