fix: make multi-writer artifacts readable, and stop destroying codex error causes

[saucam]

Two defects found by the first live end-to-end run of a collaborative session with real models: a mixed fleet (claude opus 4.8 / sonnet 5, codex gpt-5.6-sol) where an orchestrator drove a cross-vendor panel.

Neither was reachable from the test suite as written. One of them a test was actively enforcing.

---

1. blackboard_read could not address a multi-writer artifact

findings is a MULTI_WRITER kind — each reviewer writes its own slot. So blackboard_index correctly reported findings in slot review, and blackboard_read had no slot parameter, read the singleton slot, and answered "No findings has been written on this goal yet."

The index advertised something the reader could not fetch.

Live: the orchestrator saw the artifact, tried six times, was told it didn't exist, concluded "Blackboard still unreadable on a third attempt", and abandoned the board — routing around the entire typed-handoff mechanism the feature exists for.

read now takes an optional slot, and a slotless read that finds nothing points at the slots that do exist rather than claiming absence. The asymmetry with write is unchanged and deliberate: §4 gives write no slot so a writer cannot name a peer's and overwrite it. The service always supported read(kind, slot) — only the tool schema was missing it.

A test was enforcing the bug

blackboard-mcp.test.ts asserted that no tool accepts a slot — stricter than the design, which forbids it on write only. That over-assertion is why reads were slotless, and it would have rejected this fix. It now states the invariant per tool: no tool takes a goal/session id (scope rides the token), write never takes a slot, read does.

2. Codex errors arrived as [object Object]

The error notification carries an object; the handler did (params.message ?? params.error) as string then String(...). Every codex failure reported the literal [object Object] — the cause destroyed at the one place whose job was to report it.

Live: a codex role-child failed twice. Nothing in the transcript, audit log, or daemon log said why, and the orchestrator could only report "identical bare worker turn ended in error, no diagnostic detail."

describeCodexError now handles a direct string, a bare string error, a JSON-RPC-shaped {code, message}, and falls back to bounded JSON — tolerating shape rather than asserting one, since a wrong assumption costs the whole diagnostic, which is what happened.

With it, the same run reported:

codex: {"type":"error","status":400,"error":{"type":"invalid_request_error",
  "message":"The 'gpt-5.6-sol' model requires a newer version of Codex.
             Please upgrade to the latest app or CLI and try again."}}

The codex failure itself is environmental — the installed CLI (client 0.145.0) predates the model its own cache advertises. Not a codeoid bug, and now diagnosable in one read instead of not at all.

---

Verification

Daemon suite 2172 → 2191. Both fixes mutation-checked: removing slot from the read schema fails the multi-writer read test; reverting the error extraction fails three of the five codex cases.

The codex error path had no test at all before this — which is how [object Object] shipped. It has five now, including one asserting that string never appears for any input shape.

What the live run confirmed working

Not part of this fix, but it's the first time any of it ran against real models:

• an orchestrator on opus chose fleet_panel itself and fanned one brief to two children on different vendors
• the R3 gate fired per dispatch with the cost roll-up attached, accumulating across dispatches: $0 → $0.68 → $1.39
• the barrier behaved exactly as designed, twice — group_waiting done=1/2 then group_joined members=2 failed=1: it absorbed the first completion, did not fire early, and joined on all-terminal with the failed member reported rather than hanging on it
• after the slot fix, the reviewer's findings read back through the orchestrator's mount, closing the write-then-read handoff

🤖 Generated with Claude Code

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.