enallagi 发布单二进制编码智能体任务循环工具

The fractional harness that could.

An autonomous task loop for a coding agent, in one static binary.

Highlights

Reads a queue of tasks and works them one at a time.
Runs each task in a fresh agent process, never a long-lived session.
Splits the work across five roles that cannot grade each other.
Re-derives every done from the tree instead of trusting a claim.
Runs each lane in its own git worktree and fast-forwards on success.
Wraps any headless agent CLI, with thirteen presets shipped.
Pins every vendored skill by commit and hash.

Install

curl -fsSL https://raw.githubusercontent.com/thmsmtylr/enallagi/main/install.sh | sh

From source:

cargo install --locked --git https://github.com/thmsmtylr/enallagi enallagi

Releasing is one procedure, stated in .github/workflows/release.yml.

From install to a landed task

docs/setup.md walks the same path on a real repository, step by step.

enallagi init                  # one pass: detects the check, asks what it cannot decide, writes and commits everything
enallagi init --yes            # or takes every default, as a stdin that is not a terminal does

Turn an issue into a task:

enallagi issue owner/repo#12   # appends a proposed block to .enallagi/TASKS.md; `init --issue <url>` does the same

The block quotes the issue in notes:. The first run's adjudicator writes what done means and sets it ready; $EDITOR .enallagi/TASKS.md does the same by hand.

Or write a task into .enallagi/TASKS.md by hand:

## [T-001] the date parser drops a timezone
scope: src/date.ts, src/date.test.ts
blockedBy:
status: ready
rows: none — harness
criteria:
  - `npm test -- date` passes with a case for `+10:00`
notes:

Run it:

enallagi run --pipeline task --iterations 1   # one task, in this checkout
enallagi worktree 1                          # or one task, in its own worktree
enallagi watch                               # attach read-only to the live loop
enallagi pr T-001 --push                     # one pull request for the landed task
enallagi run --pr-per-task --iterations 3    # or one pull request per task, opened as each lands

--pr-per-task stacks a task on the branch of a blocker landed earlier in the same run.

How a run works

Each iteration takes the first ready task whose blockers are done.

The implementer edits only the paths on the task's scope: line.
It commits, writes its notes, and sets the task to review.
The verifier reads the criteria and returns done or a rejection.
Gates then re-run behind that verdict and can overturn it.

A done the gates disagree with goes back to ready. A question the criteria do not answer becomes needs-spec and halts the run.

enallagi run --pipeline task runs only the pipeline named. Repeat the flag to name more than one. With none named, every pipeline in .enallagi/enallagi.toml is eligible.

The pipeline, gate, probe, key and file tables are in docs/reference.md. Changing a skill, a pipeline or a stage is in docs/pipeline.md.

Configuration

.enallagi/enallagi.toml holds the whole configuration. Every key and its default is in docs/configuration.md.

enallagi init writes only the keys whose value differs from the embedded defaults. A key left out takes its default.

A re-run names every key that still equals one. enallagi init --prune-defaults deletes those keys.

A first enallagi init also detects the test runner from the files the repository already carries. Seven runners ship: cargo, node, bun, vitest, jest, pytest and go. Detection writes check.command, check.fail_name and the layout test keys.

Each detected value is printed with the file:line it came from. A tree that matches more than one runner keeps the defaults and names the candidates. A tree that matches none has the agent propose a check, written only once the check proves it.

agent.preset names the CLI to drive, or custom with your own agent.command.
agent.model and agent.effort set defaults. A role or a task can override both.
A claude lane runs sandboxed with edits accepted, so it needs no bypass flag. See docs/configuration.md.
agent.dangerously_skip_permissions adds the preset's bypass flag to every lane. Default false.
enallagi run --dangerously-skip-permissions does the same for one run. A preset that declares no bypass flag refuses the run.
run.start records the choice as permissions_skipped, and enallagi events prints it.
check.command is the single command that decides green.
[[skill]] declares a skill to vendor, with its rev and the gate that enforces it.

Budgets come from the environment. Set BUDGET_USD, BUDGET_SECONDS, or BUDGET_TOKENS. BUDGET_TOKENS counts every token lane [agent.usage] names, cached reads and cache writes included.

Commands

init installs into a repository. eject removes it and keeps the record outside it. Both commit one change to the product's .gitignore: init adds /.enallagi/, and eject removes it.
run drives the pipelines in place. worktree drives them in an isolated checkout.
run prints its version and build commit first, and binary predates HEAD when HEAD descends from that commit.
watch attaches to a live loop. events queries the log. Each event carries sha, the product HEAD, and run.start carries binary.
events --summary prints one row per task: stages, verify rounds, rejections, overturned verdicts, seconds, cost and token lanes.
tasks reads and edits the queue. base prints the commit a task was queued against.
tasks landed lists each done task, its commits, and no branch, built <branch> or pushed <url>. --built keeps the branches waiting on a push.
issue appends a GitHub issue, and review a pull request's review comments, as proposed tasks.
probe reports findings. gate runs one gate. hook is the agent's lifecycle entry point.
audit proposes a learning for each class of finding that recurs. skills resolves declared skills. eval runs the evals.
pr builds a pull request. It builds a task whose commits conflict with the default branch on the newest task/<id> branch, built this run or on the remote, that changed a conflicting file and takes the commits, and its description names that task. A blockedBy: task whose branch is on the remote is a base the same way. When no branch takes the commits, the refusal names the conflicting files.

enallagi pr --push reads the target's contribution guide first. A sentence there that refuses or conditions generated changes stops the push.

Pass --policy-read once you have read it. enallagi pr records either outcome in the description file, and commits it under <harness directory>/pr/.

Run enallagi <command> --help for the flags.

License

MIT