Skills, CLI & MCP
You do not need to memorize this page. In ordinary use,
talk to the coding agent and use h-reason when the route is
unclear. This is the reference for scripting, debugging, and exact
integration work.
Skills help the host agent choose a method. The CLI gives a human direct
access. The haft serve MCP server exposes structured operations
to the host. All three use the same local project state.
The 12 v9 skills
| Skill | Use it when |
|---|---|
h-reason | The question is ambiguous and the right route is not yet clear |
h-frame | The problem itself is under-articulated |
h-diagnose | A concrete failure has rival causes |
h-explore | You need genuinely different approaches |
h-compare | Existing options need the same comparison basis |
h-decide | You directly choose one option for one clear subject and scope |
h-verify | A recorded decision or claim must meet current evidence |
h-status | You want the current attention view: work, drift, evidence, or coverage |
h-onboard | Project setup or scope detection needs attention |
h-spec | The current product or software contract is in question |
h-note | You explicitly want to save one non-binding observation |
h-commission | You manually create durable, bounded authority for a runner |
These are independent entry points. “Auto” routing means the host may use a
matching skill; it does not mean every answer is saved or that the agent
can invent a human choice. h-commission is manual-only.
CLI families
Run haft --help on the installed binary for the complete tree.
| CLI | Purpose |
|---|---|
haft init | Initialize project memory and selected host integrations |
haft serve | Start the MCP server for this project |
haft fpf query|lookup|inspect | Find and open the bundled FPF source |
haft graph explore | Find code and related project records |
haft spec ... | Draft, inspect, check, exchange, and manage specifications |
haft memory ... | Resolve and inspect exact project memory |
haft onboard ... | Inspect or complete project setup |
haft decision ... | Inspect decisions and reconcile their links |
haft commission ... | Manage durable records for external runner work |
haft artifact create ... | Create one supported project record from an input file |
haft sync | Import the supported subset of readable artifact projections |
haft check | Run deterministic governance checks suitable for CI |
haft doctor | Inspect installation and project-binding problems |
haft interface [capability] --json | Read exact machine-facing contracts |
Check the version your host loaded
Documentation describes a release. An already running host may still have
an older binary and tool catalog in memory. After init or upgrade, restart
the host and use haft version, haft doctor, and
haft interface when exact behavior matters.
The 12 MCP tools
| Tool | Purpose |
|---|---|
haft_query | Read status, FPF source, code context, specifications, decisions, and project memory |
haft_problem | Frame or characterize a problem |
haft_solution | Explore or compare options |
haft_decision | Work with decisions, claims, evidence, and baselines |
haft_refresh | Scan and maintain stale or drifted records |
haft_note | Save one non-binding note |
haft_method | Open and close task-local engineering guidance |
haft_commission | Manage runner-neutral commissions and run records |
haft_spec_section | Read and update one specification section |
haft_onboard | Read setup status and prepare a scope review when needed |
haft_entity | Establish one exact, non-binding project identity |
haft_memory | Low-level memory validation and admission for expert use |
MCP validates arguments in the server and returns structured errors. A tool being visible does not grant permission to create a human decision, commission work, or change another binding state.
Removed from v9
haft run,haft harness, and bundled Open-Sleigh execution- the Elixir/OTP/BEAM runtime dependency
/h-search,/h-problems, and/h-viewhaft fpf search; usequery,lookup, orinspect- the old five-stage workflow as a required product model