Codex can still overflow between a parallel tool batch and its continuation request

[justrach]

Symptom

A Codex turn requested six read_file calls in parallel and the immediate continuation failed with:

codex api error: Your input exceeds the context window of this model. Please adjust your input and try again.

The model was gpt-5.6-sol via the Codex provider.

Why the previous fixes do not fully cover it

• Compaction can fail or skip silently, leaving huge context #88 recovers an errored turn only when last_context_tokens >= compactAt().
• Codex/gpt-5.x sessions brick near the context cap: compaction WriteFails (sends full over-cap history) and emergency-trim can't recover a tool-churn history #163 makes compaction/emergency trim survive old tool-output churn, but that logic runs only once compaction is invoked.
• codex: context meter reads the WS-chained total, so full-history resends blow past the window at "35% ctx" and the session wedges (no auto-compact) #174 fixes the stale cross-turn WS/full-resend meter and recognizes the raw rejection, but it does not add an in-turn pre-sampling compaction gate.
• Responses tool outputs are individually capped at 16,384 chars, but a parallel batch is not checked as an aggregate before the next model request.

Agent.runTurn() currently does request -> stepResponses/runTools -> request with no context check between the tool results being appended and the continuation request. Normal auto-compaction is in mainloop.zig after a completed turn, which is too late for this path.

OpenAI Codex comparison

The upstream openai/codex turn loop checks active context after sampling/tool execution and, when another sampling iteration is needed, runs MidTurn compaction before the continuation request (codex-rs/core/src/session/turn.rs). It also combines server usage with local estimates for newly appended items.

Proposed fix

1. Port/retain codex: context meter reads the WS-chained total, so full-history resends blow past the window at "35% ctx" and the session wedges (no auto-compact) #174's full-history estimate and context-rejection recovery on the affected release branch.
2. Add a pre-request context gate inside Agent.runTurn(), including between tool-loop iterations.
3. Refresh the Responses estimate after normalizing newly appended tool outputs, so a parallel batch can trigger compaction before it is sent.
4. Add a regression test with several parallel-sized function_call_output items crossing compactAt().

[justrach]

Fixed in v0.0.226 (PR #328).

The mechanism was already released; the regression test used one 40KB output rather than the parallel batch the issue was actually about. It now appends several mid-sized function_call_outputs that individually pass the threshold and jointly cross it.

Tests went 449 -> 575 on this release; every Zig fix carries a pinning test that was falsified independently (revert the production change, confirm the test goes red naming itself, restore).