WS4 — Telaio research rail: campaign interview → skills injection → campaign execution

Acceptance Criteria

• A campaign interview runs end-to-end in the terminal client: domain → direction → budget → warrant request, each stage a recorded entity
• Skills inject on demand during the authoring turn
• Campaign execution runs researcher and corrector as separately qualified tasks; an unqualified endpoint is refused with a route event carrying the reason
• Every artifact write goes through the amico CLI boundary — no harness-internal writes to run dirs, vaults, or catalogs
• The full suite stays green with no network access

Testing Decisions

Extend the mock-transport suite with campaign-loop tests (sticky last-step semantics script a full campaign); add a qualification-gate test (unqualified endpoint refused); reuse the golden tiered-solve test as the authoring-turn fixture.

Key Decisions

• The interview rail is the campaign rail — telaio never grows a pulse-specific interview; pack content arrives as skills.
• CONTRACT-region immutability extends to campaign artifacts (fill-in-aware diffs, per the repo's invariant 5).

Constraints & Invariants

• The telaio repo invariants apply verbatim: events are the only state; the kernel names tasks never models; qualification is measured; no silent routing; keys never touch events; zero runtime dependencies.

Prior Art

• The telaio spike itself (kernel, policy engine, golden solve); the pulse-designer score's stage contract (the interview shape to generalize); the QEC hunt loop (the campaign shape).

Source

Part of #368 · Blocked by #369 · design-of-record: vault spec note spec-20260813-162300-autoresearch-studio