fix(pair): refuse a Codex lead, name the pair tools, and end paired mode when the pair is off

[rynfar]

What

A live end-to-end run of pair mode found that a Codex lead does not pair, while a Claude lead does. This makes the product say so, and tightens what a paired lead is told.

• Codex cannot lead a pair for now. pair_start refuses it and the composer's Pair control shows it as unavailable with a reason, the same way Antigravity already was. Both still work as the executor.
• The pair protocol names its tools. It now says the executor is reached only through pair_handoff, pair_await and pair_stop on the t3-code MCP server, tells the lead to search for pair_handoff if it is not listed, and says that an agent started with the harness's own tools (Agent, Task, spawn_agent, a collaboration namespace) is not the executor.
• Turning a pair off now really turns it off. The pair capability was granted whenever the executor thread existed, and turning off a pair with history archives the executor instead of deleting it. That lead would have kept the Agent tool denied and the pair protocol in its instructions. The capability now needs an executor that is not archived, which is what the internals doc already said. Found by reading the code while testing, covered by a new case in ProviderService.test.ts.
• A paired Codex app-server also gets features.multi_agent_v2=false. Kept for Codex threads that already have an executor; it is not enough on its own (below).

Evidence

Run against an isolated server with real providers, task: "brief your executor to create a file, wait, verify".

• Codex lead (codex-cli 0.153.4, GPT-6-Astra), four runs, two of them on a fresh thread with both flags off and the new protocol text: every time the lead called Codex's own collaboration.spawn_agent / followup_task / wait_agent, the work ran on the lead's model, and the pair executor never got a session.
• codex exec -c features.multi_agent=false -c features.multi_agent_v2=false still lists all six collaboration.* tools, so the flags flip in codex features list without removing the tools. agents.max_depth=0 does not remove them either.
• Asked directly, the Codex lead reported no <pylon_pair> block and no t3-code tools, although the provider log shows t3-code reaching ready and the protocol inside the turn's developer instructions. Not yet explained; tracked in feat: pair mode, a linked lead/executor thread pair replacing fan-out delegation #622.
• Claude lead (Claude provider, Sonnet 5) with a Claude executor: the lead called pair_handoff and pair_await, the executor's own session wrote the file, the lead verified it, the control read "Finished", and the wake message rendered as the compact notice.

Verification

• vp test run on the pair toolkit, RuntimeInstructions.test.ts, CodexAdapter.test.ts and packages/client-runtime/src/state: 710 passed. ProviderService.test.ts: 138 passed.
• Typecheck t3, @t3tools/client-runtime, @t3tools/web: 0 errors. Lint on the changed files: clean.

Known limits

• The lead is woken once more for a completion it already read through pair_await. Follow-up in feat: pair mode, a linked lead/executor thread pair replacing fan-out delegation #622.
• Prime Agent, Cursor and OpenCode leads are allowed but have not been run end to end.

Part of #622.

Model: Claude Fable 5.1, in Pylon via the Claude Code harness.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
pylon-marketing	Ready	Preview	Sep 18, 2026 5:14pm UTC

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ The exact PR base did not have a successful artifact. Baseline uses the latest successful main measurement shown below.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.9 KiB	13.9 KiB	+49 B (+0.3%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.2 KiB	7.2 KiB	+9 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.7 KiB	6.7 KiB	+40 B (+0.6%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.0 KiB	58.0 KiB	+44 B (+0.1%)	66.4 KiB	✅
Codex	Live turn messages	8	9	+1 (+12.5%)	21	✅
Claude	Total thread wire	13.9 KiB	14.0 KiB	+45 B (+0.3%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.2 KiB	7.2 KiB	+16 B (+0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.7 KiB	6.7 KiB	+29 B (+0.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.8 KiB	58.9 KiB	+44 B (+0.1%)	66.4 KiB	✅
Claude	Live turn messages	8	9	+1 (+12.5%)	21	✅

Baseline: 93931e7 · PR result: 7bbe9ac · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 115.7 KiB
• Claude decoded thread snapshot: 116.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.