Summon a fleet,
token savings
you can't beat.

The model reads a result. The transcript stays on your machine.

Everything a worker produces — its narration, its tool calls, the files it opened — stops at the manager. What passes through is a fixed-shape result under a hard budget. Drawn to scale:

1. Spawned, and maybe queued

   worker_spawn returns an id in under a second. Past the concurrency cap a worker waits in line, costing you nothing; its session opens and its clock starts when it actually runs.

2. Running in your repository

   By default every worker works in your checkout, together, seeing each other's edits. The work stays uncommitted in your tree. Three watchdogs read the event stream — wall clock, idle, budget.

3. Blocked, if it needs you

   A worker that hits a question stops and says so. You answer with worker_message and it picks up in the same session, keeping everything it had already worked out.

4. Settled, then checked

   The manager snapshots the worktree, re-runs the repository's own test suite itself, and reconciles the worker's report against git diff --name-only. Its summary is a claim. The discrepancies are Dispatched Code's own finding.

5. Or it fails, and says which kind

   Timed out, over budget, cancelled, interrupted by a dead process — the result says which. A kill -9 mid-run is salvaged from the worktree into a real result on a mergeable branch.

This is the whole of what the model sees for one worker.

Worker: w-001 · model: opencode/muse-spark-1.2-contributor-free · mode: implement · status: completed · 15s · ~12,985 tok
Task: Add a `range` function to src/stats.js that returns the largest value minus the smallest.

Summary: Added `range(values)` to src/stats.js returning max-min with RangeError on empty input, and
added corresponding tests to test/checks.mjs. Verified range([3,1,4])===3 and range([]) throws
RangeError, and npm test passes.
Changes (2 files, +18/−1): src/stats.js, test/checks.mjs
Tests: 4 passed / 0 failed / 0 skipped
Discrepancies: none
Snapshot: a7c6376179 on the worker's branch

The client is a slot. Codex fills it.

Dispatched Code speaks stdio MCP and nothing else, so whatever holds the other end of the pipe is the thing doing the delegating. Codex is the second one that has actually held it: one entry in ~/.codex/config.toml, and it spawns, waits, answers a blocked worker, revises and merges through the same sixteen tools, reading the same verified result.

• The tool surface is the same sixteen. Same names, same arguments, same results. Nothing in it branches on who is calling.
• The dashboard is the same page. It is served over HTTP on 127.0.0.1:4180, and a browser does not care which client started the process behind it.
• The repository is the same repository. The server works in the directory its client was started in, so the rule does not change: start the client in the project.

The list you would find on your own, in week two.

• No container sandbox. Workers run with your permissions, in your repository. An implement worker can write anywhere this process can, and the diff is what tells you where it went. Your client's own approval and sandbox settings cover the client; they do not reach the workers behind the tool call.
• Cost is unverified on a paid provider. Every run so far has been on a free tier, where the number is always zero. Trust the token counts; treat the dollar figure as a placeholder.
• The queue is lost on restart. A restarted manager mints ids from w-001 again, so it can collide with rows a dead one left behind.
• Two clients means two managers. Each client launches its own copy of the server over stdio. Claude Code and Codex open at once are two independent managers minting ids from w-001 into one repository, and only one of them can have 127.0.0.1:4180.
• A question still costs a turn. Permission requests are answered in band and the worker carries straight on. An actual question is answered by picking one of the labels the worker offered, so the exchange moves one label at a time.
• Your test command is taken at face value. The manager re-runs it and reports the exit code honestly, but node --check file.js passes on anything that parses. Whether the command proves something is your call.