eng

Platform-agnostic engineering agent with four modes — --plan, --build, --review, --bug. This file routes. The shared spine for every mode is refs/protocol-core.md (Steps 0–6); it never runs a mode's work itself.

Routing — the one routing table

Read the invocation flag and load exactly one mode file:

Natural-language triggers for --review: "review this change", "review my diff", "code review this branch", "adversarial review", "look for bugs in what I just wrote". Anything about style, naming or standards is /cook, not this mode.

Progress messages (--build) are always on, every 5 minutes (../shared/refs/status-heartbeat.md).

Exactly one mode flag must be present. Zero, several, or --todo hard-fail with the messages in refs/protocol-core.md § Step 0. Then read the active mode file fully and run the spine end to end.

The shared spine (refs/protocol-core.md)

• Step 1 — Input validation. The PRD/exec-table source (mode flag + prd-path + rows + agent) or the report=<path> alternate. --review skips Steps 1–4.
• Step 2 — Pre-flight, one pass. Injected scoped excerpts on an orchestrated run; otherwise the PRD (the digest on --build), devkit and the row-relevant code. Row ownership is checked by script-prd-digest.py --verify-rows.
• Step 3 — Summary gate (--plan only). --build asks no opening question; a manual build asks one question at the end, "Commit?".
• Step 4 — Coding standards (--build only). An injected payload wins; else script-eng-cook-flags.py + the standards cache, then /cook on a miss.
• Step 5 — Run the mode's work, then the closing message (../shared/refs/closing-message.md).
• Step 6 — Scope enforcement, continuous. Only the assigned rows.

References

• refs/plan/protocol.md — --plan: summary, output contract, the ## Todos — <Agent> ticket spec, the closing shape check. refs/plan/template-todo.md — the ticket schema the build reads. refs/plan/template-eng-plan.md — the plan section shapes. refs/plan/fix-plan.md — planning fixes from a report.
• refs/build/protocol.md — the manual build's solo-only steps (branch, ticket check, standards, progress messages, "Commit?", sub-branch PR, summary and run report). refs/build/protocol-packet.md — the build loop, written once: tickets, TDD, full-suite gate, review and re-test, db-touch pause, one commit per ticket, output. refs/build/protocol-build-debug.md — debug mode. refs/build/fix-build.md — the report source. refs/build/fix-build-orchestrated.md — the fix coordinator.
• refs/build/bug-report.md — --bug: a bug found by hand becomes a new report-prd-<N>-<K> pair, then the normal fix flow.
• refs/review/protocol.md — --review: change set, adversarial charter, severity, return contract, artifact, spawn rules. One review, whole change, by an agent that did not write the code.
• .claude/scripts/ — script-eng-db-touch.sh (production/data guardrail over this run's changes; a trip pauses for sign-off) · script-eng-cook-flags.py (the /cook flag set from the stack) · script-eng-comment-scan.sh (A4 comment presence) · script-eng-commit-cap.sh (A5 commit size, advisory; a missing Oversize-reason: trailer exits 3) · script-eng-plan-shape.py (the --plan closing check) · script-project-findings.py (the one finding→issue-ticket projection + issues-file validator) · script-eng-fix-grade.py (fix complexity + tier→model) · script-eng-close-loop.py (the only write to an issues file) · script-eng-bug-report.py (writes a hand-found bug as a new report pair)
• ../shared/refs/status-heartbeat.md — the progress-message contract
• Contract: the ## Engineering — <Agent> and ## Todos — <Agent> headings written by --plan are how plan-em detects a finished section and how --build finds its tickets. Do not rename them.