fix: send the Agent SDK's initialize request to every relaunched stream-json child

[anaclumos]

Closes #322.

Problem

The Agent SDK sends its initialize control request once per query. The request carries appendSystemPrompt, the hook callback ids, the SDK MCP servers, agents, and the other init settings. The supervisor's stdin relay forwarded each harness line to the live child once, so after a move that resumed with --resume, the relaunched child never received initialize. A T3 Code thread lost its appended runtime instructions after its first move, and an SDK consumer lost its hooks and in-process MCP servers.

Change

• StdinRelay keeps the latest line of each control request in REPLAYED_REQUESTS (initialize, set_model, set_permission_mode, set_max_thinking_tokens) for the life of the session, and attach sends those lines to every child ahead of the resume prompt and the queued input. A first launch has none, so it is unchanged.
• The --session-id replay from fix: replay a stream-json session's lost lines on a move before its first transcript row #324 keeps the other lines the killed child received (user lines), and still answers every other control request with an error control_response. Setters now go before the replayed user lines instead of in arrival order, so a replayed first turn runs under the latest model and permission mode.
• AGENTS.md, DESIGN.md, and docs/content/docs/architecture.mdx describe the relaunch.

Decisions

• Replaying initialize is safe for the SDK. In @anthropic-ai/claude-agent-sdk 0.3.276, Query.readMessages files a control_response with no pending request in a bounded unmatchedControlResponses map, and hookCallbacks lives on the Query object for the life of the query, so the hook callback ids in the replayed request still resolve. This answers the issue's open question about hook callback ids.
• Keeping the latest line per subtype bounds the state to four lines, instead of the full control history.

Verification

Hermetic run under a throwaway HOME and TOKENMAXXING_HOME, with a stub claude that records its stdin and writes a respawn marker after the first user line:

• --resume move, on main: the relaunched child receives only the resume prompt and later input. On this branch: initialize, set_model, the resume prompt, then later input.
• --session-id move (no transcript) with an mcp_call request before it: the relaunched child receives initialize, set_model, the replayed user line, then later input, and stdout carries the error control_response for the mcp_call.

Not run: a live T3 Code session through a real move, because that spends pooled quota.

[vercel]

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

Project	Deployment	Actions	Updated
tokenmaxxing	Ready	Preview	Oct 6, 2026 2:32pm UTC

[anaclumos]

Adversarial review (OpenCode on MiMo v2.6 Pro, because the Codex review plugin is not installed on the host) returned three findings:

1. Map order with delete then set lets setters precede a new initialize. Not a defect. The Agent SDK sends one initialize per Query, and each query spawns its own process, so one supervisor's stdin carries one initialize. initialize is the first line, so it stays first. The follow-up commit drops the delete, and the saved lines keep first-arrival order.
2. A setup request that arrives while no child is attached (compaction, depleted wait) went to the queue behind the resume prompt, so that turn ran on the old model or permission mode. Fixed in the follow-up commit: such a line updates the saved state, which attach sends before the resume prompt. Verified hermetically: a set_model sent during a 3-second wait reaches the relaunched child before the resume prompt, and the stale value is no longer sent.
3. A new initialize does not clear the previous query's setters. Not a defect, for the same reason as finding 1.