| name | run |
| description | Drive a task through the generic gated workflow — intake, design panel, implementation, hardening and review, wrap-up — with explicit human gates, resumable state, and every gate decision captured into the project knowledge base. Use for non-trivial tasks that deserve staged delivery; input is a task description or spec file, no ticket system required. |
| invocation | user |
| disable-model-invocation | true |
run — the generic gated workflow
One stage per invocation; freeze at every human gate; never infer approval
from silence. Helpers live in scripts/ next to this file; state lives at
.autodidact-workflow/runs/<KEY>/ (KEY = TASK-<n> or any ABC-123-style id
you pick for the run).
Pipeline
INTAKE → DESIGN → IMPLEMENT → HARDEN+REVIEW → WRAP-UP
Use scripts/run-state.sh for init/status/record at every stage;
--status reports and performs no work.
- Intake (with bounded interrogation). Read the task description or
spec file. Extract requirements, acceptance criteria, affected
components, risks. Then interrogate the gaps — ask the user the
questions whose answers would change the design (max ~5, one round;
contradictory or materially incomplete spec → stop and say so). Propose
a profile via
scripts/profile-select.sh from spec signals (points/ACs/
components/type/risk) — never from estimated lines of code.
- Profile gate. Explicit
--approve or --profile quick|standard|deep.
- Design. For profiles that call for it, spawn
panel-role agents
per scripts/panel-members.sh, then synthesizer merges into
design.md. Surface conflicts with proposed resolutions; the user
decides.
- Design gate. Show the artifact; require explicit approval or a
decline reason.
- Implement. Delegate production edits to
implementer task by task.
- Implementation gate. Show the cumulative diff; require explicit
approval or revision instructions.
- Harden + review. Run
test-hardener, then auto-reviewer per
profile. Critical findings block until fixed or explicitly escalated.
- Wrap-up. Summarize; branch/PR per project convention (
gh pr create where available — ticket/PR-system specifics belong to a
downstream overlay, not here); then close the learning loop (below).
Learning-loop wiring (every gate — this is not optional)
After EACH explicit gate decision, capture it with the autodidact-learning
capture CLI (deterministic protocol — never hand-write entries). Locate it
at <autodidact-plugins>/plugins/autodidact-learning/skills/capture/scripts/new_entry.py
(installed plugin cache or repo checkout):
- gate approved and the stage taught something reusable →
--outcome approve (what worked, principle-level);
- gate returned with corrections →
--outcome correct (the corrected
rule — this is the highest-value signal);
- gate declined an approach →
--outcome decline (what was rejected
and why);
- nothing learned (routine pass) → capture nothing; do not manufacture
entries.
Always --source gate --kb-root docs/knowledge, --ref the touched
files. Exit 10 (duplicate) is fine; exit 2 (secret refusal) → redact and
retry, never bypass.
At wrap-up, run the thresholds check and act on it:
python3 <autodidact-plugins>/plugins/autodidact-learning/skills/distill/scripts/mechanical.py \
thresholds --kb-root docs/knowledge
Exit 4 → tell the user thresholds are hit and offer to run the distill
skill now (workflow boundary = the designed distillation moment); exit 0 →
nothing to do; exit 1 → the project has no KB (suggest adopting the autodidact
skeleton).
