fix(opencode): share tool response truncation

[MagMueller]

Summary

• use the existing tool-response truncator for successful output, persisted tool errors, workflow-tool errors, and malformed tool-call repairs
• reduce the shared default limit from 50 KiB to 40 KiB while retaining the existing 2,000-line limit and configuration overrides
• save complete truncated content in the existing seven-day tool-output store and return the same head preview plus file path
• keep a bounded preview if the archive write fails, for both successful responses and errors

Why

A production malformed browser_execute call contained roughly 118,000 newline characters. AI SDK InvalidToolInputError copied the malformed tool input into its error, and JSON reserialization persisted a roughly 445,000-character assistant message. A later tiny user turn then exceeded model context. Normal tool-output truncation did not apply because the oversized content crossed the boundary as tool input and error rather than successful output.

Validation

• bun test test/tool/truncation.test.ts test/session/session.test.ts: 28 passed
• production-shaped 118,000-newline malformed-input regression: passed and full error archived
• shell truncation regressions: 4 passed
• packages/opencode bun typecheck: passed
• root filtered typecheck: 16 packages passed
• targeted oxlint: 0 errors
• Prettier and git diff --check: passed
• repository pre-push hook with Bun 1.3.14: passed

Scope

The final diff is four files. Three source integration points remain necessary: the shared truncator, the central persisted-tool-error boundary, and the AI SDK malformed-call repair boundary.

[cubic-dev-ai]

All reported issues were addressed across 6 files

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 6 files (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Fix all with cubic | Re-trigger cubic</sub>

[MagMueller]

Pre-merge Cloud/explore verification:

• The Cloud v4 worker pins bcode 0.1.15. That tag contains the native task tool and explore subagent; generated bcode.json only denies websearch, so task/explore remain enabled.
• Live current-source smoke with an isolated data root completed end to end: build -> task(subagent_type=explore) -> child Grep on the archived tool-output -> child Read(offset=12002, limit=1) -> exact RECOVERY_SENTINEL_7F3A91C2 -> exact parent answer.
• Exporting the child session confirmed agent=explore and both tool calls completed.
• Relevant local checks: truncation 21/21, native-agent 2/2, Cloud config 2/2. The broader Task file has four existing config-fixture failures; the same 16-pass/4-fail result reproduced on untouched origin/main.
• A first macOS /tmp run encountered the /tmp -> /private/tmp alias permission boundary; rerunning with the canonical path passed. Cloud workers are Linux, so that alias is not present there.