DATUM — Agentic Production Line

Commands

/datum go          Run from current phase through merge. Halt at gates.
/datum yolo        Skip optional gates. Hard stops still halt.
/datum <phase>     Run one phase: refine, plan, act, validate, review, closeout, etc.
/datum resume      Resume from .datum/state.json after interruption.
/datum status      Print phase, RUN_ID, lane progress, last failure.
/datum init        Bootstrap repo: hooks, linter, AGENTS.md, CURRENT_STATE.md, ROADMAP.md.
/datum classify    Auto-classify epic complexity (Patch/Feature/System)
/datum landscape   Generate docs/LANDSCAPE.md from filesystem analysis
/datum mermaid     Generate Mermaid diagrams
/datum dream       Memory consolidation — staleness audit + transcript extraction + pruning

Rule: Determinism

Orchestration is deterministic. State, transitions, routing, gates — all enforced by Python scripts and TypeScript workflow pipelines. No improvisation. The LLM works within one phase. It does not decide pipeline structure or skip steps.

Dispatcher

Execute in order before any phase work:

0. Branch Guard — If on main/master, auto-create a feature branch and switch. Slugify the brief/TICKET title into datum/<slug>.

0.5. Self-check — datum doctor. If it fails, halt.

1. Load Config — .datum/config.toml, falling back to assets/config.toml.default.

2. Read State — datum status --json. If no state, detect entry:

Epic artifacts always live at docs/epics/<branch>/.

3. Dispatch Phase — Each phase is a TypeScript workflow. datum-go chains them all.

After each phase: datum gate <phase> [--approve]

Act Phase — TDD Workflow Pipeline

Act is handled by the datum-tdd-act TypeScript workflow (skills/src/datum-tdd-act.ts).

Invocation:

Workflow({ name: "datum-tdd-act", args: "yolo" })

Yolo mode auto-detects the current branch and generates a run ID. Or pass explicit args:

Workflow({ name: "datum-tdd-act", args: { epicBranch: "datum/epic-17", runId: "20260614-010000" } })

Pipeline stages per lane:

1. RED — write failing tests (sonnet), structural assertion check (haiku)
2. REFLECT — score test quality 0-10 (haiku), gate at <4
3. GREEN — make tests pass (sonnet, escalates to opus on retry)
4. SKEPTIC — adversarial verification panel (3 lenses: edge/error/contract)
5. REFACTOR — optional cleanup if haiku pre-check finds improvements
6. File ownership — verify each commit only touches allowed files

Source: skills/src/ (TypeScript) -> skills/*.js (generated via bash scripts/build-workflows.sh)

Prompt templates: skills/src/prompts/*.md with {{placeholder}} syntax

Model tiers: haiku (evaluators), sonnet (writers), sonnet->opus (GREEN retry)

Gates

Hard stops never bypass: tests_red_after_3x_retry, hook_blocked_write, merge_conflict, schema_validation_failed, file_ownership_violation.

Error Recovery

• ENVIRONMENTAL -> fix in place, same tier, counter not incremented
• REASONING -> retry ladder: standard -> reasoning
• Self-healing: datum bugfile <module> "<description>" --trace "<traceback>"

Artifacts

Cross-Tool

SKILL.md is tool-agnostic. Model tiers resolve per-tool via [models] in config. See references/model-tiers.md.