feat: per-role model defaults + surveyor arsenal#3
Conversation
The installer now emits the model frontmatter field so the harness launches each agent at its expected tier without per-dispatch specification: council seats opus (architect-tier judgment survives a cheaper composer), workers and reviewers sonnet, the surveyor haiku - quick and bounded, predicated on the composer supplying well-defined boundaries. A campaign IBC may elevate a specific dispatch; these are ambient defaults. Companion conditioning: the surveyor runs the outward-environment arsenal step before its first search (semantic code search such as grepai beats grep for intent-shaped questions), and the composer names the project's search arsenal in every survey dispatch - the surveyor checks regardless, but the composer holds the project context.
|
Added per nrd: explicit CLI-merge semantics. The forge skill's §4 previously permitted CLI merges while listing |
|
Correction on the record: commit a9a59a4 truncated Two durable takeaways worth predicate's process-feedback ledger: (1) the link auditor passes empty files — a non-emptiness check on audited docs would have blocked a9a59a4 at the gate; (2) patch-script writes should assert the post-state, not just the anchors. Both are gate-hardening candidates. The substance stands as described in the previous comment: |
Summary
Restores (properly, on master) the
model:frontmatter the installer never emitted — the field that lets the harness launch each agent at its expected tier without the composer specifying it per dispatch — plus the surveyor's arsenal discipline.Defaults
*-seat)survey-workerA campaign IBC may elevate a specific dispatch; these are ambient defaults, deliberately encoding the routine posture.
Arsenal discipline (companion conditioning)
Verification
Conditioning suite all-pass; installer re-run against the live harness confirms
model: opuson seats,model: haikuon the surveyor,model: sonnetelsewhere.