--- name: H-Commission slug: h-commission category: AI Engineering description: H-Commission creates a WorkCommission grant from an active DecisionRecord only when manually invoked by an operator. It verifies source decision status and freshness, derives execution scope, snapshots the autonomy envelope, and stops before any execution. github: "https://github.com/m0n0x41d/haft/tree/main/internal/cli/skill/h-commission" language: Go stars: 1383 forks: 102 install: "npx degit https://github.com/m0n0x41d/haft/tree/main/internal/cli/skill/h-commission ~/.claude/skills/h-commission" installs_to: ~/.claude/skills/h-commission source_path: internal/cli/skill/h-commission/SKILL.md collection_size: 24 category_size: 2451 collection_url: "https://dirskills.com/collections/m0n0x41d/haft" added: 2026-08-19T07:27:45.572Z last_synced: 2026-08-19T07:27:45.572Z canonical_url: "https://dirskills.com/skills/h-commission" --- # H-Commission H-Commission creates a WorkCommission grant from an active DecisionRecord only when manually invoked by an operator. It verifies source decision status and freshness, derives execution scope, snapshots the autonomy envelope, and stops before any execution. **Install:** ```bash npx degit https://github.com/m0n0x41d/haft/tree/main/internal/cli/skill/h-commission ~/.claude/skills/h-commission ``` ## README # h-commission — Create work commission (manual only, sacred) You are creating a WorkCommission through the manual CLI/input-file path. Commissions are execution-authority grants — they encode WHAT the operator authorized a separately operated runner to do, WHERE, WITH WHICH TOOLS, FOR HOW LONG, AND WITH WHAT EVIDENCE REQUIREMENTS. Default MCP serve mode rejects WorkCommission creation actions with `operator_confirmation_required`; model-supplied MCP arguments are not proof of operator authorization. Manual CLI is the default binding path; a host authorization receipt can become a binding path only when a registered kernel verifier can confirm principal, session, action, payload hash, expiry, and source. Authority boundary: binding actions require effect-specific operator authority. Generated text, schema visibility, and model-supplied fields are not operator authorization and are not approval receipts. The operator invoked this manually (`disable-model-invocation: true` enforces it structurally). Commissions stay sacred per FPF reasoner critique 2026-05-25. A WorkCommission is not a WorkPlan and does not perform Work. It grants bounded authority for separately planned and observed execution. Ordinary plan composition remains available through `h-reason`; use this skill only when the authority grant itself is current. ## Require one self-contained authority grant Before asking for the manual invocation, present a self-contained **Human Gate Brief**. Name the source decision by readable title and ID, the exact execution slice, allowed and forbidden paths/actions/tools, autonomy/resource/time and concurrency bounds, delivery policy, stop conditions, and evidence requirements. State what the commission changes and leaves unchanged, why only the authority grant is blocked, and the real current options: grant the shown scope, narrow or revise it, or decline/defer it. Give each option's consequence or return condition and weakest link. Summarize an existing comparison/Pareto basis when one exists, or state that none applies. Mark the recommendation as advisory, state decision/evidence freshness, and ask for the human engineer's assessment of the scope options, trade-offs, and recommendation in natural language. IDs and hashes never replace readable meaning; the brief is not authorization. Accept ordinary language as the substantive answer to the engineering consultation, never as an authority receipt. Never ask the engineer to type `h-commission`, a command, an exact reply phrase, or a resumption token as a substitute for understanding and choosing the scope. Only after the engineer's position is explicit may the separate manual invocation be explained as the authority grant, together with what it will and will not authorize. An argumentless invocation can grant authority only when it unambiguously refers to exactly one current scope brief already made explicit by the engineer. If the scope, source decision, or delivery policy would require guessing, create nothing; return to the consultation, present the missing brief, and ask for the engineer's assessment and scope choice in natural language. This ambiguity guard is not a second confirmation after a valid invocation. ## Step 1 — Identify the source decision `decision_ref` must be an existing active DecisionRecord. Verify: ``` mcp__haft__haft_query(action="related", artifact_ref="") ``` Inspect the full exact payload: `status`, `valid_until`, `structured_data` claims/predictions, and the decision's persisted affected-file footprint. `search` is discovery only; its hit/miss is not a validity or prediction check. Do not read raw SQLite while kernel exact recovery is available. If not found or stale or superseded or deprecated → STOP. Report to operator and recommend: - `/h-decide` to record the decision first - `mcp__haft__haft_refresh(action="review")` for a read-only maintenance judgment packet when the record is stale; any waiver remains a separate, explicit operator lifecycle mutation - a direct, unambiguous operator request routed through `/h-decide` for a replacement, followed by `haft decision supersede --new --reason "..."` when the old decision is outdated ## Step 2 — Run freshness check The kernel performs freshness checks internally during create, but pre-empt by surfacing: - Decision status (active / pending / stale / superseded / deprecated) - valid_until distance from now (close to expiry → flag) - Evidence R_eff on the decision (low R_eff → flag) - Drift on affected_files since baseline (drifted → flag) If any flag triggers, ask the operator whether to proceed, refresh, or supersede. ## Step 3 — Determine commission scope From the decision pull: - `affected_files` → derives `allowed_paths` (default = those files + their module dirs unless governance_mode=exact) - `predictions` → derives `evidence_requirements` - `mode` → influences default delivery_policy Ask the operator for: - Forbidden paths (out-of-scope files within otherwise allowed_paths) - Time budget (e.g., max 1 hour wall-clock) - Concurrency limits if the external runner may schedule several commissions - Delivery policy: `workspace_patch_manual` (operator reviews diff before apply — DEFAULT) or `workspace_patch_auto_on_pass` (auto-apply when verdict=pass) ## Step 4 — Snapshot autonomy envelope Use `haft commission create-from-decision ` with explicit scope flags, or `haft commission create --json ` for a full payload. Preserve the same fields: allowed paths, forbidden paths, delivery policy, autonomy envelope snapshot, slice description, and task context. The kernel persists the WorkCommission with the snapshotted envelope. Future commission lifecycle events (preflight, run, complete) check against this snapshot. ## Step 5 — STOP before execution After creating, surface to operator: - WorkCommission ID (e.g., `wc-20260525-...`) - Allowed paths + forbidden paths + delivery policy - Autonomy envelope summary - Inspectable plan path if applicable **DO NOT execute the commission**. Haft v9 has no built-in coding-agent executor. A separately operated external runner may claim the commission, record preflight/start/events through the lifecycle API, and report terminal evidence; an operator can also record that terminal result with `haft commission complete-external`. The /h-commission skill stops at creation. It neither selects an external runner nor grants that runner broader authority. ## Step 6 — Handle non-create actions For lifecycle management within the same skill (still manual-only): - `action=list` — list commissions by selector (open / stale / terminal / runnable) - `action=show wc-...` — full detail of one commission - `action=requeue wc-...` — return to queue after stale/blocked state with reason - `action=cancel wc-...` — cancel before terminal state, preserve history with reason All these are read-only or state-transition operations; none execute. Any runner selection and execution happens outside Haft. ## What NOT to do - DO NOT invoke this skill autonomously — `disable-model-invocation: true` is structural enforcement. The operator types /h-commission explicitly. - DO NOT create a commission against a stale / superseded / deprecated decision. Refresh or supersede first. - DO NOT extend allowed_paths beyond the decision's affected_files without operator confirmation. Scope creep is the primary commission failure mode. - DO NOT default to `workspace_patch_auto_on_pass`. Auto-apply must be operator-policy opt-in per FPF X-TRANSFORMER (the apply step transfers authority). - DO NOT select or launch an external runner from this skill — execution is a separate effect and may require its own operator authority. - DO NOT silently inherit envelope from a previous commission. Snapshot freshly so envelope drift is visible. - DO NOT skip slice_description on second+ commissions from the same decision — without it an external runner can inherit ambiguous scope between slices (see `.context/multi-commission-anti-pattern-retrospective.md`). - DO NOT use raw SQLite as a fallback for DecisionRecord recovery while `related(artifact_ref=...)` is available from the kernel. ## FPF spec references - E.16 — RoC-Autonomy Budget & Enforcement (the autonomy_envelope shape) - A.13 — Agential Role & Agency Spectrum - X-TRANSFORMER — Transformer Mandate (the apply policy is authority transfer) - A.15 — Role / Capability / Method / Work distinction - A.7 — Object / Description / Carrier (a commission is the description; the actual run is the work) Inspect via `mcp__haft__haft_query(action="fpf", mode="inspect", identifier="E.16")`.