Self-review: show the configured remote agents to the model

The self-review node's proof. Policy (the-loop critic policy): selfReviewCount: 2, stop on no new findings, escalate on a repeated finding.

Review cycles

Round	Reviewer	Outcome	Findings → disposition	Link
1	[claude/claude-opus-5-5]	new findings	F1: an agent registered at two versions was listed twice in the roster, possibly with a card the delegation does not reach → will-fix, fixed in 399f6b7 (registered_agents: one per id at version="*", red→green unit test)	comment
2	[claude/claude-opus-5-5]	new findings	F2: stale context.py module docstring (static sections omit agents) → will-fix, fixed in 3fc7011	comment

Two rounds is the operator's cap, so the loop stops here, not on convergence. Round 2 found only a docstring nit and no repeat of F1. The critic round follows.

Review of R5–R7 (after PR review)

The rounds above covered R1–R4, at 7cb2103. R5 ([card]), R6 (the example) and R7 (every child over A2A) came later from PR review, with 16 commits and about 650 lines of production code. Round 3 reviewed git diff 7cb2103..HEAD -- tiny_harness.

Round	Reviewer	Outcome	Findings → disposition
3	[claude/claude-opus-5-5]	new findings	SR-1 — replay determinism: an in-flight parent that spawned a child under the old code would, on replay, meet start_agent_task where its history has StartChildWorkflowExecution (the same for /sub-task and the new completion shortcut), and the workflow task would fail. The codebase has never used workflow.patched. → escalated: patch the changed sequences or document "finish or cancel in-flight tasks with sub-tasks before upgrading"; a decision for the reviewer. SR-2 — a briefed model skipped the selectable-models check that /model applies → will-fix, fixed in fe0587d (check_model). SR-3 — behaviour change for outside clients: a first message that is only a task envelope now starts the task without a model turn (it used to run a turn on a synthetic "[task updated…]" text) → accepted, recorded in the a2a-server capability doc. SR-4 — a parent cancelled while start_agent_task is in flight can leave the created child without a stand-in, so the cancel does not reach it → accepted as a narrow residual; noted for a follow-up. SR-5 — the completion shortcut that waits for an unread result skips the task.complete hook for that one turn → accepted: the task does not complete in that turn, and the hook runs at the next completion.

Found and fixed during R7's verification and CI healing, before this round:

• a follow read replayed history as change (2dba276);
• agent streams left open (2dba276);
• a delegated parent completing without reading a result (2dba276);
• a stuck new stand-in (74ff83b);
• a cancel mid-delivery failing the workflow (92ffab1);
• a child's signal before a continued run's load (3104ea1).