[KongchangAI]
Product

moshcode work

A command-line workflow feature of moshcode that supports running multiple headless AI agents in parallel to execute tasks, using the exit code of a verify command as the completion criterion, and managing failed tasks through git worktree isolation, atomic task claiming and an inbox mechanism.

Timeline (last 90 days)

Oct 17

moshcode version 0.117.0 adds the `moshcode work` feature.

Unverified50%
Oct 17

In moshcode work, each task carries a `verify` command, and exit code 0 is the sole definition of task completion.

Unverified50%
Oct 17

The workers in moshcode work are headless processes, each running in its own git worktree.

Unverified50%
Oct 17

The engine's assessment of its own work is recorded but never accepted as the basis for completion.

Unverified50%
Oct 17

A failed attempt returns to the queue with the verification output and goes into the next prompt; after three failures, the task goes into the inbox along with its worktree.

Unverified50%
Oct 17

A task that has already passed requires only one verification and consumes no tokens.

Unverified50%
Oct 17

Task claiming in moshcode work is an atomic file rename, so multiple runs on the same queue can split the tasks.

Unverified50%
Oct 17

Each run of moshcode work is recorded as an OpenFleet swarm.

Unverified50%
Oct 17

In the first run, 50 Claude workers processed 19 tasks in parallel, all passed on the first attempt, the inbox remained empty, and local failures dropped from 139 to 0.

Unverified50%
Oct 17

The design philosophy of moshcode work is borrowed from the field of mathematical proofs: humans do not review each agent's reasoning process, but instead rely on an objective pass/fail signal (a Lean proof checker or a verify command).

Unverified50%

2 more timeline events

All Facts (12)

Unverified

The inbox mechanism ensures that human attention is focused on genuinely difficult problems: only tasks that fail after exhausting all attempts (three by default) enter the inbox, with their worktrees preserved for inspection.

50%
Unverified

In moshcode work, each task carries a `verify` command, and exit code 0 is the sole definition of task completion.

50%
Unverified

The workers in moshcode work are headless processes, each running in its own git worktree.

50%
Unverified

The engine's assessment of its own work is recorded but never accepted as the basis for completion.

50%
Unverified

A failed attempt returns to the queue with the verification output and goes into the next prompt; after three failures, the task goes into the inbox along with its worktree.

50%
Unverified

A task that has already passed requires only one verification and consumes no tokens.

50%
Unverified

Task claiming in moshcode work is an atomic file rename, so multiple runs on the same queue can split the tasks.

50%
Unverified

Each run of moshcode work is recorded as an OpenFleet swarm.

50%
Unverified

The design philosophy of moshcode work is borrowed from the field of mathematical proofs: humans do not review each agent's reasoning process, but instead rely on an objective pass/fail signal (a Lean proof checker or a verify command).

50%
Unverified

By using a git worktree to isolate each worker, moshcode work avoids file conflicts between agents and provides a foundation for horizontal scaling.

50%
Unverified

In the first run, 50 Claude workers processed 19 tasks in parallel, all passed on the first attempt, the inbox remained empty, and local failures dropped from 139 to 0.

50%
Unverified

moshcode version 0.117.0 adds the `moshcode work` feature.

50%

Source Articles