jason.woltje
83c4e9851e
Collaboration record (conductor loop, docs/plans/CONDUCTOR.md): - round 1 (worker session worker-1, 2m28s): retry implemented per spec - conductor live test exposed spec gap: direct invocation lacked launcher env exports - round 2 (same worker session, 59s): spawnEnv made self-sufficient, but used PI_* where compose interpolates MOSAIC_* - conductor hotfix: 3-line rename to MOSAIC_PROVIDER/MOSAIC_MODEL/ MOSAIC_DATA_ROOT Final: node scripts/mosaic-task.mjs retry <runId> re-executes a run's task snapshot as a new run; live retry replied REMEMBERED; all suites green (24/32/14 + verify). Known limitation: retrying a run whose task used a RELATIVE mission path resolves it against the temp dir; lineage tracking deferred. Closes #25, closes #26, closes #27
2.3 KiB
2.3 KiB
Conductor protocol — poor-man orchestration loop
How the stack orchestrates headless pi workers to do work on itself.
Roles
| Role | Runs where | Powers | Never has |
|---|---|---|---|
| Conductor | host (assistant or owner) | git (clone/commit/push), task dispatch, review, verification suites, Gitea | nothing new |
| Worker | container (headless pi via scripts/run-task.sh) |
read/write/edit/bash inside its workspace; persistent session on request | git credentials, docker socket, host filesystem |
The loop
- Decompose: conductor turns a goal into worker tasks small enough to
specify completely in one prompt (file paths, acceptance criteria, style
constraints, verification the worker can run itself, e.g.
node --check). - Mirror: conductor maintains the repo clone at
<dataRoot>/workspaces/stack-repo(host-side git; workers see it read-write through their workspace mount). - Dispatch:
scripts/run-task.sh run <worker-task.json>— worker edits the clone. Session nameworker-<n>keeps continuity across refinement rounds. - Extract:
git -C <workspace> diff > patch— the worker's entire output is a reviewable diff. Run record (result.json, stderr.txt) is the receipt. - Review: conductor reads the diff line by line. Bad output → refine the prompt, re-dispatch (same session: "your patch had these problems…").
- Integrate: conductor applies the patch to the real repo, runs the full suites, commits and pushes. Suites failing → revert apply, back to step 5.
- Record: update CURRENT.md, BUILD-LOG, close the Gitea issue.
Guardrails
- Workers never receive credentials; they never run git; they never leave the workspace (container is the boundary; tools allowlist is the gate).
- Every worker diff is reviewed by the conductor before integration. No auto-apply. (Auto-apply would be a capability-policy decision for later.)
- Verification is mechanical: suites +
node --check/bash -ngates. - Recursive decomposition = "fail → smaller task", never "hope."
Worker task template
{
"taskVersion": 1,
"id": "t-worker-<name>",
"prompt": "<full spec: goal, files, constraints, acceptance, self-checks>",
"workspace": "stack-repo",
"capabilities": { "tools": ["read", "write", "edit", "bash"] },
"session": "worker-1",
"timeoutSeconds": 600
}