feat: write truncated tool outputs to files

[rekram1-node]

closes: #7168
closes: #4560

Combines ideas from:
#6234 by @ben-vargas
#6048 by @shantur

[rekram1-node]

/review

[rekram1-node]

/review

[rekram1-node]

/review

[github-actions]

lgtm

[rekram1-node]

/review

[github-actions]

Overall the PR looks good! The main suggestion is a minor style preference:

• The else if/else chain in read.ts (line 130-134) could potentially use early returns, but the current code is readable and this is not a violation per se.

The code follows good practices, uses Bun APIs appropriately, and the implementation is solid. The let statements used are appropriate for loop contexts.

[rekram1-node]

/review

[github-actions]

lgtm

[jogoossens]

This seem to have adverse effects, only 50kb of agents.MD files is read :/

[rekram1-node]

agents.md files are parsed independent of this

[rekram1-node]

also this changed 0 about reading of files

[jogoossens]

The cap is OpenCode‑side, not model‑side:

The stored tool result is 53,592 chars and ends with: (Output capped at 50 KB. Showing lines 1-526. Use offset=527 to continue.)
Where: packages/opencode/src/tool/read.ts line 18, MAX_BYTES = 50 * 1024. This runs in the OpenCode process, before tool output ever reaches the LLM. It applies identically whether the agent is talking to Qwen 3.5, Qwen 3.6, gpt‑5.4, or anything else.
First introduced: upstream commit 1b82511 on 2026‑01‑07 (PR #7239).
First tagged release containing it: v1.2.0, published 2026‑02‑14.

It is not /AGENTS.md, that one does read 100% ,it are subfolders like apps/AGENTS.md that seem to have the cap.

We added some subfolder AGENTS.md to instructions and the used tokens are suddenly way higher. We also saw that rules on top of apps/AGENTS.md were used by the model, rules on the bottom were completely ignored.

I would say at least something is going on ,manually apps/agents.MD to instructions made the file be read 100%