v0.9.14: cap every child tool result at capture time — 1 MiB bytes / 10k tokens into context, pacing checked at result insertion

[Hmbown]

Why this matters

Tonight a worker reading a 542KB source file in chunks burned 638k input tokens and died with zero workspace changes — pure read starvation. The step cap cannot help (one step has no cost ceiling), and the 75%-consumed pacing notice is evaluated between steps, so one huge read jumps 60%→dead without ever crossing the warning band. Budget awareness already shipped (#6194) and did not help.

Design (from the harness comparison, codex-rs citations)

Adopt codex-rs's placement rule: truncation at capture time, not accounting after insertion (refs/codex/codex-rs/core/src/exec.rs:719-728; default 1 MiB DEFAULT_OUTPUT_BYTES_CAP at utils/pty/src/lib.rs:14):

1. In crates/tui/src/tools/subagent/, intercept each child tool result before it is appended to the child's history; hard-truncate to a byte cap (1 MiB default) with a truncated: true marker the model can see.
2. Add a model-facing max_output_tokens param defaulting to 10k (codex unified_exec/mod.rs:79, tools/handlers/shell_spec.rs:61).
3. Evaluate the budget pacing check against the pending result size pre-insertion, so a single call can never jump the budget silently.

Full comparison with citations: codewhale-ops/designs/HARNESS-COMPARISON-20260916.md (parallel doc; will land with the slice).

Acceptance criteria

• A child issuing one tool call whose raw result is >1 MiB receives a truncated result with truncated: true and its context grows by ≤ the cap
• The pacing/budget check runs at result-insertion time; a test proves a child at 60% cannot die mid-read without having crossed the warning threshold first
• Focused subagent tests green; cargo test -p codewhale-tui --lib for the touched modules quotes real counts

Related

• v0.9.14: a worker's reserved report turn is spent by its own descendants — budget death returns nothing at all #6277 (reserve/report-turn sizing — complementary)
• v0.9.14: write-claim contention forbids N workers writing disjoint files under one shared root — the natural fan-out pattern is impossible #6278 (fan-out economics)