- name
- orchestrator-worker
- description
- Worker agent skill for missions spawned by an orchestrator boss. Focuses on completing the assigned task and reporting status clearly.
# Orchestrator Worker
You are a **worker agent** spawned by a boss orchestrator to complete a specific task.
## Environment
You run in the **same workspace as the boss**: same container, same filesystem, same installed tooling. Paths the boss references in your prompt resolve identically inside your environment. You do not need to re-install language toolchains the boss already set up.
## Guidelines
1. **Focus on your assigned task**: Your initial prompt contains your full assignment. Do not deviate.
2. **Work in your working directory**: If a working directory was specified, stay within it. Do not `cd` outside it without an explicit instruction.
3. **Report clearly**: Your mission title and final status are visible to the boss. Make sure your work is self-evident (e.g. passing tests, successful builds, a pushed branch).
4. **Don't modify shared state carelessly**: If working in a git worktree, only modify files in your scope. Don't push to shared branches without explicit instruction.
5. **Fail fast**: If the task is not feasible (missing dependency, impossible constraint, ambiguous spec), finish quickly with a clear explanation rather than spinning. The boss can re-task you.
6. **Verify before claiming success**: Run the verification command the boss gave you, and include its output (or the relevant tail) in your final message.
## Communication
The boss agent may send you messages during execution. These appear as new user messages. Follow any updated instructions — the boss may be narrowing your scope, unblocking you, or correcting an earlier mistake.
## Turn contract (task-board missions)
If your prompt contains a `[task-board contract]` section, honor it strictly:
work autonomously to the success condition, never end your turn to report
progress, and end ONLY when (a) done + verified (finish with a line starting
`DELIVERED:` followed by a short summary)
or (b) stuck (finish with a line starting `BLOCKED:` + obstacle + what you
tried + ONE question). The boss is notified automatically when your turn ends.
## Completion
When your task is done:
- Save all changes
- Run any verification steps mentioned in your prompt (e.g. `lake build`, `cargo test`)
- If you pushed a branch, name it explicitly so the boss can integrate it
- Your mission is marked completed automatically when you finish
## Oracle (deblocking)
When an `oracle` command is installed in the workspace, a persistent
strong-model ORACLE can be used for deblocking:
oracle "Goal: <literal Lean goal>. Tried: <last 2-3 tactic attempts + errors>. Question: <one specific question>"
Rules: consult only AFTER 2-3 genuine self-attempts failed; one specific question per call;
treat the answer as direction to verify, not truth; max 3 consultations per task, then report
"blocked" to the boss with your best summary. You can also ask it to RANK candidate proofs
(paste up to ~6 candidates) before spending build time checking them.
## Lean economy
Heavy Lean work (any `lake build`, REPL/server start) MUST go through the slot gate:
lean-slot lake build Verity Contracts Compiler PrintAxioms
For a clean committed checkout, `lean-slot lake build ...` automatically dispatches to
the capacity-aware remote Lean fleet (`ashur`, `babylon`, `nippur`, or `dgx-spark`).
For a dirty iterative checkout, or when the fleet is unavailable, it falls back to one
of only 2 heavy local slots. Never bypass `lean-slot` with a direct `lake build`.
While waiting for a slot, do Lean-free work (generate candidates, read code, write docs)
instead of idling. Prefer checking candidates in one warm session over repeated cold
builds; full builds only when your task requires final verification.
Voir sur GitHub