ADR 0027: task run pre-flight orchestrator

Summary

aidokit task run <id|path> runs aidokit's own deterministic pre-flight chain for a task in one command — score the brief, draft the goal contract, emit the run-plan — and then stops before the model step. It composes the existing runTaskScore, runGoalDraft, and runGoalPlan functions; it adds no new logic and makes zero model calls. The command prints the produced artifacts plus the ordered next steps (run your adapter → aidokit eval → aidokit verify --all-facets → aidokit pr packet).

This is the "Tier 2" orchestrator. It is explicitly not the autonomous runner (ROI #2's executing form) — that requires LLM calls and auto-progression, both permanent non-goals (CLAUDE.md §11, ROADMAP §Non-goals).

Context

ROI #2 asked for a one-command task pipeline. Three tiers exist:

1. Tier 1 — aidokit goal plan (ADR-0023): emits a run-plan script the human executes. Already shipped.
2. Tier 2 — this ADR: a command that runs the deterministic, model-free steps in sequence and halts at the boundary where a model (the user's own coding CLI) must act.
3. Tier 3 — autonomous runner: generate → drive a model → verify → loop. Permanent non-goal; not built.

The legitimate, in-charter slice is Tier 2: aidokit already owns several deterministic per-task steps (task score, goal draft, goal plan). Running them by hand, in order, is busywork. Orchestrating aidokit's own commands — not a model — is consistent with the deterministic-scaffolder charter.

Decision

Command

aidokit task run <id|path> [--brief <path>] [--json], registered as a subcommand of the existing task group. Brief resolution matches task score / goal / eval.

Behaviour (in order, all deterministic, no model)

1. runTaskScore — advisory quality score (does not gate).
2. runGoalDraft — write agent-artifacts/goals/<id>-goal.md.
3. runGoalPlan — write agent-artifacts/goals/<id>-plan.sh.
4. Print the score, the two artifact paths, and the ordered next steps: review contract → run YOUR adapter → aidokit eval → aidokit verify --all-facets → aidokit pr packet.

A missing brief raises AidokitError (BAD_INVOCATION). No new finding codes.

Hard boundary

task run never invokes a model, never auto-progresses past the human/adapter step, and runs no daemon. Step 2 of the printed chain (run your adapter) is a manual placeholder, exactly as in the goal plan script.

Consequences

• Positive: one command for the model-free pre-flight; less manual command stacking; composition-only, so it inherits the behaviour (and tests) of the commands it calls.
• Positive: the model boundary is explicit and visible in the output, reinforcing the no-autonomy charter.
• Negative / deferred: it does not orchestrate the post-model steps (eval/verify/pr packet) — those need the model to have run first, so they remain separate commands the user runs after. A future task verify resume step could chain the post-model half; deferred until there is demand.
• Alternative rejected: shelling out to the CLI binary per step. Calling the exported functions directly is hermetic, faster, and keeps error handling in one process.