Quickstart

1. Connect

If your account is granted Taskforge, the kennet server connects when you sign in to Hum. Otherwise add it:

hum mcp add kennet https://kennet.metaphi.ai/mcp/
hum mcp login kennet      # the browser opens; sign in with GitHub

Sign in with GitHub: every call is checked against your own GitHub access. /mcp kennet shows the tools your tier reaches, and /tools lists them by group. See MCP servers. Any MCP client works the same way: point it at the URL above and sign in when it answers 401.

2. Push a bundle

A bundle is a directory holding task.toml, with instruction.md, environment/, solution/solve.sh and tests/test.sh. Push it first: Kennet clones GitHub, never your checkout. The steps below use acme/tasks at main, bundle tasks/fix-rounding. Its runs land under fix-rounding/. You can ask Hum in plain words (“run pre-run checks on tasks/fix-rounding”); the calls are shown so you know what it sends.

3. Pre-run checks

run_task_ci_checks {"repo": "acme/tasks", "ref": "main", "path": "tasks/fix-rounding",
                    "run_oracle": true, "run_nop": true}

It returns a task. Poll get_task until it ends. Read tasks[<bundle>].eval_ready; when false, each failing check names what to fix. Edit, push, call again. The report lands at fix-rounding/ci/<run id>/.

4. Validate

run_harbor_tasks {"repo": "acme/tasks", "ref": "main", "paths": ["tasks/fix-rounding"],
                  "mode": "validate"}

One trial per leg: the reference solution must score 1.0, the untouched code below it, and a cheat probe must not pass. It lands at fix-rounding/validate/<run id>/.

5. Conduct

run_harbor_tasks {"repo": "acme/tasks", "ref": "main", "paths": ["tasks/fix-rounding"],
                  "mode": "conduct", "models": ["google/gemini-3.8-flash"],
                  "harness": "terminus-2", "k": 8, "reasoning_effort": "medium"}

Eight trials, each judged as it finishes. reasoning_effort is required; Kennet has no default. The task’s progress counts calls by phase while it works. The run lands at fix-rounding/eval/<run id>/<model slug>/.

6. Read the run

harbor_rollouts   {"repo": "acme/tasks", "task": "fix-rounding"}
harbor_run_report {"repo": "acme/tasks", "run_id": "<run id>"}

harbor_rollouts gives one row per trial (reward, failure type, cost, tokens, phase times) and per task the pass@k and its band. harbor_run_report lists one run’s trials and files; name a trial for its judged report.

7. Review the trajectories

review_run {"repo": "acme/tasks", "run_path": "fix-rounding/eval/<run id>/<model slug>"}

Every trial gets trajectory_report.json: whether the reward was earned, whether the agent got stuck, and recommended fixes to the task. Disagree with a verdict? See Post-run.

8. Pull it

pull_run {"repo": "acme/tasks", "run_id": "<run id>"}

It returns a curl command. Run it; the run lands in ./jobs/<run id>/ in Harbor’s layout, and harbor view jobs/<run id> browses it.

Then decide

Too easy (pass@k near 1)? Harden the task. Unsolved and the reviews say the instruction is short of information? Fix the instruction. Then push and go round again. Keeping, revising and rejecting are yours.

Getting started

The loop

Traces

Reference

1. Connect

2. Push a bundle

3. Pre-run checks

4. Validate

5. Conduct

6. Read the run

7. Review the trajectories

8. Pull it

Then decide

​1. Connect

​2. Push a bundle

​3. Pre-run checks

​4. Validate

​5. Conduct

​6. Read the run

​7. Review the trajectories

​8. Pull it

​Then decide

1. Connect

2. Push a bundle

3. Pre-run checks

4. Validate

5. Conduct

6. Read the run

7. Review the trajectories

8. Pull it

Then decide