/mcp kennet in Hum).
Tiers
Each person holds one tier. Tiers add up:l2 reaches everything l1 does.
| Tier | Who | Adds |
|---|---|---|
kennet:l1 | Task authors | Check and run bundles, read runs, adjudicate, follow your tasks. Runs only the models listed for l1 |
kennet:l2 | Trusted authors | Any catalog model, the authoring tools, review_run. run_repo_to_tasks at most $100 a session |
kennet:twin | A capability, not a tier | upload_trace, twin_trace, pull_trace, and get_task and cancel_task on your own jobs: five tools |
Authoring
| Tool | Title | What it does | Tier |
|---|---|---|---|
run_repo_to_tasks | Forge tasks from a repo | Run taskforge’s whole loop over a repository. | l2 |
run_task_proposals | Mine task ideas | Mine a repository for task candidates. | l2 |
score_candidate | Score a task idea | Score one task candidate and return its three measurements. | l2 |
generate_harbor_task | Write a Harbor task | Generate an eval-ready Harbor task bundle. | l2 |
run_sandbox_session | Run a sandbox session | Run an open-ended sandbox session on the environment the repository declares. | l2 |
Pre-run
| Tool | Title | What it does | Tier |
|---|---|---|---|
run_task_ci_checks | Pre-run checks | Run task CI on Harbor bundles: the screen, the advisory audits, and the validate run that decides eval_ready. | l1 |
run_redteam | Red-team a task’s tests | Red-team Harbor bundles’ tests and return a task whose result holds each bundle’s confirmed findings. | l2 |
prerun_validate | Validate before a run | Validate one submitted bundle’s construction before any run and return a task whose result says passed, what is missing and what warns, check by check. | l1 |
Conduct
| Tool | Title | What it does | Tier |
|---|---|---|---|
run_harbor_tasks | Run bundles on models | Run each named bundle and return one task covering every run. | l1 |
Post-run
| Tool | Title | What it does | Tier |
|---|---|---|---|
postrun_validate | Validate on a model | Run one submitted bundle k times on one model, judge each trial, and return a task whose result holds pass_at_k, the trials solved and the task flags its review raised (reward hacking, instructions sufficient). | l1 |
review_run | Review a run | Review a run’s trajectories. | l2 |
get_report | Read a review report | Read a trajectory report whole. | l1 |
submit_adjudication | Adjudicate a verdict | File one adjudication beside a trajectory report, as a suggestion under your name. | l1 |
list_adjudications | List adjudications | List the adjudications filed against a trajectory report, with its review timeline. | l1 |
resolve_suggestion | Resolve a suggestion | Resolve an adjudication’s suggestion under your name. | l1 |
harbor_runs | List a repo’s runs | List a repository’s runs newest first. | l1 |
harbor_run_report | Report one run | Report one run. | l1 |
harbor_rollouts | Rollouts and costs | Return one row per rollout across a repository’s eval runs (ids, model, harness, reward and the grade beside it, failure type, cost, tokens, phase times, model wait against tool time, sandbox CPU and peak memory) and their folds. | l1 |
pull_run | Pull a run locally | Pull a run, or one file or directory in it, into your working directory. | l1 |
Twin
| Tool | Title | What it does | Tier |
|---|---|---|---|
upload_trace | Upload a trace | Get a curl command that uploads one agent trace from your working directory. | twin |
twin_trace | Twin a trace | Make a synthetic twin of an uploaded trace. | twin |
pull_trace | Pull a trace | Get a curl command that downloads a twinned trace into your working directory. | twin |
Jobs
| Tool | Title | What it does | Tier |
|---|---|---|---|
get_task | Check a running job | Read a task a tool started. | l1 |
answer_task | Answer a running job | Answer the questions a running task asks you (its status is input_required and its inputRequests carry them); the task goes on with your answers. | l1 |
cancel_task | Stop a running job | Cancel a task a tool started. | l1 |
get_job | Track a job | Track a job you started in one read: the launch, its units, each unit’s attempts and the trials each attempt ran. | l1 |
list_jobs | List your jobs | List the jobs you started, newest first. | l1 |
Common refusals
A refusal comes back as a tool error with one sentence that names what to fix.| Refusal | Do |
|---|---|
<tool> needs kennet:l2 and you hold kennet:l1; ask Metaphi for kennet:l2 | Ask for the tier, or use a tool your tier lists |
<tool> on kennet:l1 runs <models>; ask Metaphi for kennet:l2 | Pick a model from the list it names |
run_repo_to_tasks on kennet:l2 spends at most $100 (cost_budget_usd) | Lower cost_budget_usd |
you do not push to <repo> on GitHub, and this call writes to it; ask its owner for write access | Get write access, or work in a repository you push to |
<you> does not read <repo> on GitHub: give <you> read on it | Check the name, and get read access to the repository |
<repo> has no ref <ref> | Push the branch first; Kennet clones GitHub, never your checkout |
name a bundle in <repo> by ref + path, or by staged for one already on its runs | Pass ref and path (or paths), not both forms |
mode conduct requires at least one model | Pass models |
reasoning_effort is required: name the effort this run is to use (for example medium) … | Pass reasoning_effort |
no run started: the model service did not answer … or every model was refused … | Nothing ran and nothing was spent. Try again later, or pick another model |
task <id> asks no open question <qid> | The question was answered or withdrawn; get_task shows what is open |