| name | thales-start |
| description | Entry point for a new Thales investigation. Initializes the ledger and runs the first orchestration cycle. Use when the user says "use Thales" or similar explicit invocation with a task. |
/thales-start
Initialize a fresh Thales run on the user's stated task.
Steps
-
Confirm invocation is explicit. User must have said "use Thales", "thales this", "run Thales" or similar. If not, do not run — explain Thales and ask if they want to invoke.
-
Verify scaffold. Check _scratch/thales/ exists. If not, copy from the Thales skill's assets/ledger_template/. If it exists and is non-empty, confirm with user whether to overwrite or resume.
-
Write task.md. Record: the task verbatim from user; acceptance criteria (ask user if not given); starting phase = explore.
-
Load the thales skill. Read skills/thales/SKILL.md fully before the first cycle plan. Also load skills/thales/references/ledger_format.md for the session.
-
Run plan_cycle state. Per the state machine in the thales skill. Cycle 1 spawn is typically Bridger + Prober for research tasks, Prober + Challenger for coding tasks — but follow the spawn decision tree, not this note.
-
Brief, spawn, collect, write to ledger. Standard cycle.
-
Report cycle 1 result to user in plain prose. Do not present a checkpoint yet — checkpoints come at cycle 3 (first Judge) or stall.
Guardrails
- Do not spawn more than 3 subagents in cycle 1.
- Do not skip writing
task.md before spawning.
- Do not do substantive work yourself — you are the orchestrator.
- Cycle 1 is always an evidence cycle (cycle_count starts at 1, not 0). Cycle 3 will be the first judge cycle.
- Final user-facing output comes from
thales-deliverer, not from you. When the time comes, spawn Deliverer and present its output verbatim.