speckit.boundedrelay.route

Route Approved Work

Read the active artifacts and pending routing.json named in $ARGUMENTS. This operation writes only that run-local evidence file.

1. Read the engine-generated taskManifest first. It is derived from the exact committed tasks.md checkbox IDs at the routing revision and is already content-addressed. Build the normalized codex_worker_sdd_route request for every ID in taskManifest.pendingTaskIds, in that exact canonical order, exactly once. Never omit a pending ID, invent an ID, or route a completed ID. Add integer effort points, dependency IDs, authority, kind, bounded repository-relative write scopes, lane eligibility, and risk for each task.
2. Resolve capability, safety, dependency, and path ownership constraints before balancing. Never assign overlapping writer scopes.
3. Call codex_worker_sdd_route with neutralCodexShareBps (default 5000). This is a soft tie-break for fit-neutral tasks, never a provider quota. Persist its full schema-versioned response, including routingPolicyVersion and planFingerprint. Do not invent, manually alter, or fall back from the returned assignments. The engine independently verifies that fingerprint and exact manifest coverage. The target is estimated effort, never a cost or token claim.
4. When pending evidence has a non-null projectProfile, read its exact sealed path and include the parsed JSON unchanged as request.projectProfile. Persist only the authoritative schema-v2 / sdd-routing-v3 result. Exact match executorId, capability requirements and eligibility, requiredCheckProfiles, and codexPolicy into each workflow assignment, and project the plan-level crossReviewPolicy exactly once. Never execute, interpolate, or shell-evaluate any check-profile argv; it is inert policy data. The profile may only restrict the routed write scopes, and its own path must not overlap a writer lease.
5. A Claude assignment uses { "source": "host-selected", "model": null }. Do not override the Claude Code host model.
6. A profiled Codex implementation or cross-review uses the exact model and reasoning effort returned in codexPolicy; a non-profiled Codex assignment uses server-allowlisted policy. For a profiled critical task, require the exact explicit non-null model and effort from codexPolicy.byRisk.critical and the authoritative plan-level cross-review policy; both remain subject to the server allowlist. A legacy no-profile critical route keeps the fixed gpt-5.6-sol / ultra lane. If the required lane is unavailable, leave routing incomplete and report the blocker; never silently downgrade it.
7. Project the returned lanes into the workflow assignments, add only coordinator policy metadata not owned by the router, set totals from the real assignments, and atomically replace the evidence with state: complete. Each write task may require at most 64 checks; reject the route before execution if all write assignments together require more than 256 receipts.

Do not start workers, edit implementation files, install dependencies, commit, push, or deploy.

$ARGUMENTS