Authoring

You choose what to build from: a repository’s history, a directory of source material, a draft. Taskforge drafts. Every authoring tool reads your repository as you and writes to its runs, so each needs your push.

The forge: `run_repo_to_tasks`

Runs the whole loop over a repository: find candidates, queue them, author each as a bundle, verify it (pre-run checks, oracle, nop and probe, pass@1 on target_model, trajectory review), revise, and debrief. A task is kept when its pass@1 is below the threshold.

run_repo_to_tasks {"repo": "acme/service", "ref": "main",
                   "target": 5, "pass_at_1_threshold": 0.2,
                   "target_model": "google/gemini-3.8-flash", "k_rollouts": 1,
                   "reasoning_effort": "medium", "cost_budget_usd": 50,
                   "environment": "Dockerfile"}

Argument	Meaning
`target`	Kept tasks the session aims for. Default 10
`pass_at_1_threshold`	Keep a task when pass@1 is below this. Default 0.2
`target_model`, `k_rollouts`	What pass@1 is measured on, and how many rollouts
`cost_budget_usd`	What the session may spend, sessions and rollouts together. Default 100; your tier caps it
`batch`, `max_attempts`	Candidates authored at once (three at most), and tried before it stops
`environment`	The file at `ref` that declares the toolchain (a Dockerfile, a requirements `.txt` or a `pyproject.toml`); every bundle’s Dockerfile opens with it
`seed`	A `run_task_proposals` run id; its candidates seed the queue
`resume`	A session id to continue

Kept bundles stand at <task>/harbor/ with their checks and runs beside them. The session’s worklog, queue and memory land at agent-sessions/<session id>/. While it runs it may ask you a question; see Consult. A session caps at

100 on `kennet:l2` and

500 on kennet:l3.

Mining: `run_task_proposals`

Mines a repository for candidates: its commits, and its merged pull requests with the issues they close. Deduplicated, ranked and scored. A candidate is an idea with its base and final commits, not a bundle.

run_task_proposals {"repo": "acme/service", "ref": "main", "miners": ["prs"], "max_candidates": 20}

proposals.json lands at proposals/<run id>/. Pass the run id to run_repo_to_tasks as seed.

Scoring: `score_candidate`

Scores one candidate and returns three measures: idea_difficulty, idea_quality, and oracle_quality_score when the record proposes test cases. No overall score and no bar: what a score is worth is your call. Stateless; nothing is written.

score_candidate {"record": {"change": "...", "evidence": "...", "cases": []}}

One bundle: `generate_harbor_task`

Writes one bundle from source material and pushes it to a branch. Use it rather than writing a bundle by hand.

generate_harbor_task {"repo": "acme/tasks", "ref": "main", "path": "tasks/fix-rounding",
                      "source_path": "material/rounding", "examples_paths": ["tasks/gold-1"],
                      "run_ci": true, "brief": "a maintenance fix, no refactor"}

The result carries bundle_path, the conversion report, the branch it landed on, and eval_ready when run_ci is set. generate_harbor_task is the only tool that writes to your repository.

Also

run_sandbox_session: an open-ended session on the environment the repository declares, toward a goal you write (a toolchain spike, fixtures). It writes SANDBOX_REPORT.md and lands at sandbox/<run id>/.
run_skill_eval: evaluates one of Taskforge’s own skills against a fixed case. For the team that maintains the skills.

Getting started

The loop

Traces

Reference

The forge: `run_repo_to_tasks`

Mining: `run_task_proposals`

Scoring: `score_candidate`

One bundle: `generate_harbor_task`

Also

​The forge: run_repo_to_tasks

​Mining: run_task_proposals

​Scoring: score_candidate

​One bundle: generate_harbor_task

​Also

The forge: `run_repo_to_tasks`

Mining: `run_task_proposals`

Scoring: `score_candidate`

One bundle: `generate_harbor_task`

Also