F1: one stale registry `done` state silently voids FOUR ratified contracts on the default path (U6, U8/#404, P11 closure, #378 guard)

[EtanHey]

Source: P12 adversarial evaluation, finding F1 · HIGH — docs.local/plan/stability-v2/phase-12/evaluation.md:60-83
Evaluated against the deployed build: brew list --versions cmuxlayer → 0.4.46, daemon from /opt/homebrew/Cellar/cmuxlayer/0.4.46/libexec/dist/, git log HEAD = 4a88b43.

Related: #408 (registry marks live idle agents "done") is the upstream cause. This issue tracks the composition consequence: that one stale field silently voids four separate ratified contracts at once, on the default path.

Evidence (verbatim from the evaluation, evaluation.md:60-83)

F1 · HIGH · One stale registry state silently disables four contracts

TERMINAL_AGENT_STATES = new Set(["done","error"]) (server.ts:497) and
resolveCurrentCallerAgent filters those records out (server.ts:3541-3556). The registry marks live
agents done within minutes (#408, open, not on the residual list). Consequences, all observed in
one list_agents response for my own live seat:

• state = {value:"working", source:"screen"} but detail.state = "done" (registry).
• U6 violated: the child I spawned got parent_agent_id: null, spawn_depth: 0 — because I was
  filtered out as a caller. My own record's parent is the malformed "cmuxlayerClaude", not an agent id.
• U8/fix: make agent delivery receipts trustworthy #404 regression: send_to returned delivery:"failed", terminal:true, delivered:false
  against an agent whose screen read status:"idle", control_state:"ready" at a live prompt. send_to's
  own published description promises the opposite: "Omit it to receive a nonterminal queued receipt."
  This is ledger row docs: proper README — 221 tests, 1,423x speedup, 10 MCP tools #4, still live on 0.4.46 after the entire P1/fix: make agent delivery receipts trustworthy #404/P5b delivery program.
• P11 Contract C violated: closure:"artifact_missing" on a working agent, because both call sites
  pass the registry agent.state (agent-engine.ts:1723,1847) while U7/P4 established truth is derived
  live. PR feat(p11): engine-issued coordination paths + closure state (U10) #454's own table reads artifact_missing as "route a reviewer NOW", and promised it
  "actionable on its own, no cross-referencing required" at default detail — yet the disqualifying
  caveat lives in health (detail:full only).
• Silently disables the spawn_agent: role is inferred from the CLI/launcher, so every Claude reviewer lands in the orchestrator column #378 guard: callerIsWorker (server.ts:11562) can never be true for a
  stale-done caller, so the force-child-to-worker/record-parent guard — the fix for the 3×-recurring
  ledger fix: model extraction for narrow panes and no-cost status lines #13/fix: recover read_screen parser fallbacks #15/fix: infer context for truncated claude panes #16 "reviewer in the lead column" class — no-ops.

Fix direction: resolve callers and closure from the live-derived state (P4 already computes it),
not the registry record. One change closes four contract holes.

The four consequences, restated

1. U6 identity triple non-null — VIOLATED on the default path. Spawned children get parent_agent_id: null, spawn_depth: 0.
2. U8 / fix: make agent delivery receipts trustworthy #404 delivery regression. send_to returns a terminal delivery:"failed" against an idle-ready agent at a live prompt — ledger row docs: proper README — 221 tests, 1,423x speedup, 10 MCP tools #4, still live on 0.4.46.
3. P11 Contract C closure invariant — VIOLATED. closure:"artifact_missing" on a working agent, at default detail, with the disqualifying caveat only visible at health detail:"full".
4. spawn_agent: role is inferred from the CLI/launcher, so every Claude reviewer lands in the orchestrator column #378 placement guard silently dead. callerIsWorker can never be true for a stale-done caller, so the fix for the 3×-recurring ledger fix: model extraction for narrow panes and no-cost status lines #13/fix: recover read_screen parser fallbacks #15/fix: infer context for truncated claude panes #16 "reviewer in the lead column" class no-ops.

Fix direction

Resolve callers and closure from the live-derived state (P4/U7 already computes it), not the registry record. Per the evaluation's verdict (evaluation.md:196-200): "Fix F1 first: it is one change to the state source, and it closes U6, the U8 false-failure class, P11's closure invariant, and the #378 guard together."

Evaluation: docs.local/plan/stability-v2/phase-12/evaluation.md

---

— cmuxlayer-issue-filer (worker) · claude-code/claude-opus-5

[EtanHey]

Lane T1 — "#457-rest" DEFERRED, with what T1 did touch

Lane T1's phase file lists "#457-rest" (whatever survives PR #466's live-state resolution) as part
of its scope. T1 did not find a concrete residual to fix and did not run the audit that would prove
there is none, so this is a deferral, not a closure:

• What T1 did touch: the caller resolver fix(f1): resolve callers, delivery and closure from live state, not the stale registry record #466 rewrote. Caller resolution can attribute a call to a dead record on a recycled surface_id #468 (a terminal record claiming a
  recycled surface_id) was the one known hole left in it; T1 closes that with an
  observer-ownership guard on the ref-only tier. That is one of F1: one stale registry done state silently voids FOUR ratified contracts on the default path (U6, U8/#404, P11 closure, #378 guard) #457's four consequences (the spawn_agent: role is inferred from the CLI/launcher, so every Claude reviewer lands in the orchestrator column #378
  placement guard firing off a corpse) hardened further.
• Not audited here: whether any other consumer still reads agent.state as primary truth. That
  is a repo-wide sweep, not a registry/eviction change, and it wants the adversarial evaluator's
  eyes rather than the lane that would be marking its own homework.

Recommend the T6 evaluator pass own the residual sweep and either close this or name the remaining
call sites with file:line.

— cmuxlayerClaude-72284e3f (worker) · claude-code/opus-5