`read` tool call rendering in TUI breaks if line numbers are strings

[moritzwilksch]

What happened?

I was trying out openrouter:xiaomi/mimo-v2.6-flash, which seemingly likes to generate offset and limit values for the read tool as strings. The TUI then incorrectly concatenates strings instead of adding numbers when rendering the result.

The model requested lines 25-37, corresponding to offset: 25 and limit: 13, but the rendered tool call shows dummy.txt:25-2512.

The issue is in packages/coding-agent/src/core/tools/read.ts:

const startLine = args.offset ?? 1;
const endLine = args.limit !== undefined ? startLine + args.limit - 1 : "";

With offset: 25 and limit: "13", this becomes 25 + "13" - 1 → "2513" - 1 → 2512.

Steps to reproduce

Use openrouter:xiaomi/mimo-v2.6-flash and ask it to read a certain line range in a file. Read tool renders string-concatenated numbers.

Expected behavior

Performs addition, i see 25:37 instead of 25:2512.

Version

v0.87.0

[github-actions]

This issue was auto-closed. All issues from new contributors are auto-closed by default.

Maintainers review auto-closed issues daily and reopen worthwhile ones. Issues that do not meet the quality bar in CONTRIBUTING.md will not be reopened or receive a reply.

If a maintainer replies lgtmi on one of your issues, your future issues will stay open. If a maintainer replies lgtm, your future issues and PRs will stay open. The command must be at the start of the reply (optionally after one or more @username mentions) or at the end.

See CONTRIBUTING.md.

[sxh313]

Reproduced on main (a328aa89a). The transcript label is built from the raw streamed tool arguments, not the converted ones, so a model that emits offset: "25", limit: "13" renders:

read notes.txt:25-2512

instead of read notes.txt:25-37 — startLine + args.limit - 1 concatenates the two strings before the subtraction.

execute() is not affected, because validateToolArguments converts a clone of the arguments (packages/ai/src/utils/validation.ts), which is why only the display lies. The session tree label has the same arithmetic with two as number casts on toolCall.arguments, and since raw arguments are persisted in the session file it misrenders again after resume.

Fix: parse both values through one shared num() helper in core/tools/render-utils.ts, and omit the range when a value isn't numeric. Two surfaces touched (renderers/read.ts, components/tree-selector.ts), plus regression tests for each, including the mixed number/string cases.

I have this on a fork branch: https://github.com/sxh313/pi/tree/fix/read-line-range-string-args — npm run check is clean and the two affected test files pass locally. Per CONTRIBUTING I won't open a PR unasked; if you'd like it as one, reply lgtm @sxh313.

[Demonese]

I believe the issue lies with the AI large model inference provider. We should provide feedback to openrouter and request a fix.

Some open-weights models has bugs when handling non-string tool call parameters.