v0.2.2 — contract delivered to the child

The output contract was being written into the parent-facing instruction, so the child never saw it. That was the real cause of the 0/3 result previously attributed to a model ceiling.nnWith the contract inside the child task block: 8/8 compliant on the cheapest bulk-tier model (3 items + 5 items, two different schemas). --expect-json stays, now guarding genuine model faults instead of a prompt-plumbing bug.nnCumulative: 20/20 fan-out items delegated to the right role; 860/1000 test calls. See docs/test-report.md §5 and docs/review.md (ten defects).

v0.2.1 — deterministic fan-out entry point

bin/roster-fanout.mjs removes the "should I delegate?" decision from the model: N items in, N role-children out, answers read from the child tool_result event. 12/12 items delegated correctly across four runs.nnContract adherence remains a model ceiling (--expect-json: 0/3 on both free tiers), so the entry point now fails loudly rather than returning prose.nnAlso: operations page, 20 roles, and nine fixed defects from the review round. See docs/test-report.md and docs/review.md.

v0.2.0 — operations page, 20 roles, negative result

One delegate tool with a role enum replaces twelve sibling tools (parent tool surface 36 → 25) with no loss of constraint: persona, tool filter, and model tier are pinned per call and verified from gateway request logs.nnAlso adds a loopback operations page for editing the roster without a restart, eight new roles (20 total), and a review round covering seven defects.nnHonest headline: consolidating tools did NOT fix free-choice routing (2/12 → 0/6). Controlled delegation stayed at 14/15. See docs/test-report.md.

v0.1.0

First release: 12 constrained role tools, generated composition, 18 unit tests, and real-run evidence (42/42 controlled delegation).