fix(pretty): coerce string offset and limit in readCall line range

[iamxeph]

Problem

When an LLM provides tool call arguments with string-typed numbers (e.g. offset: "1", limit: "2000", as observed with models like glm-5.3-flash), the read tool summary displays bizarre, inflated line ranges such as:

• lines 1–11999 (for offset: "1", limit: "2000")
• lines 1046–1046399 (for offset: "1046", limit: "400")

Root Cause

In src/summary.ts, readCall calculates the end of the line range as follows:

const range =
  args.offset || args.limit
    ? theme.fg("dim", ` · lines ${args.offset ?? 1}${args.limit ? `–${(args.offset ?? 1) + args.limit - 1}` : "+"}`)
    : "";

While args is typed as { offset?: number; limit?: number }, TypeScript types are erased at runtime. Tool call arguments deserialized from LLM responses may carry string values.
In JavaScript:

1. (args.offset ?? 1) evaluates to "1" (string).
2. ("1") + args.limit evaluates to "1" + "2000" = "12000" (string concatenation due to + associativity).
3. "12000" - 1 coerces to a number via subtraction, resulting in 11999.

Solution

Explicitly coerce args.offset and args.limit to numbers using Number() before calculating the range:

const offset = args.offset ? Number(args.offset) : undefined;
const limit = args.limit ? Number(args.limit) : undefined;
const range =
  offset || limit
    ? theme.fg("dim", ` · lines ${offset ?? 1}${limit ? `–${(offset ?? 1) + limit - 1}` : "+"}`)
    : "";

This prevents string concatenation while preserving the exact original logic and public type definition.

Verification

• Added regression tests in test/pretty.test.ts verifying string-typed arguments for offset and limit.
• Ran tsc --noEmit: clean (0 errors).
• Ran node --test test/pretty.test.ts: all 32 tests passed (0 failed).

[iamxeph]

@hypnguyen1209 could you take a look at the two small fix PRs I have open here when you get a chance? Both are one-liner changes with tests attached — details are in each description.

[hypnguyen1209]

Good catch and a correct diagnosis: "1" + "2000" concatenates before the - 1 coerces, hence lines 1–11999. The coercion is in the right place (the summary is the only consumer, the public type stays honest), and the regression tests pin the exact shapes you saw from glm-5.3-flash. Typecheck and the full suite (132 tests) pass on your branch locally, and I have approved the CI run.

One non-blocking nit, since the point of this change is that the arguments cannot be trusted: a non-numeric string such as offset: "abc" now renders as lines NaN–NaN. Guarding with Number.isFinite (inline suggestion) keeps a bad value out of the range the same way a missing one is, without changing anything for the cases you fixed. Happy to merge either way.

[hypnguyen1209]

Suggested change

	const offset = args.offset ? Number(args.offset) : undefined;
	const limit = args.limit ? Number(args.limit) : undefined;
	const offset = Number.isFinite(Number(args.offset)) && args.offset ? Number(args.offset) : undefined;
	const limit = Number.isFinite(Number(args.limit)) && args.limit ? Number(args.limit) : undefined;

[iamxeph]

Guarded with Number.isFinite and added a test for a non-numeric value. 132 tests + typecheck pass. The workflow run on the new SHA needs your approval.