fix(bash): treat linter/test-runner exit code 1 as non-error (#1436)

[laurentftech]

Closes #1436

Problem

Linters and test runners use exit code 1 to mean "issues found" — not a command crash. commandSemantics.ts falls back to DEFAULT_SEMANTIC for unknown commands, which treats any non-zero exit as an error.

Result: the model is told isError: true, the system prompt tells it to retry on tool errors, and it retries the same ruff/eslint/pytest command 2-3 times before giving up — even though the lint output it needs is right there in the result.

Reported on v0.15.0 (Windows) with uvx ruff check --fix:

Bash(uvx ruff check --fix app/foo.py 2>&1)
  └  Error: Exit code 1
E501 Line too long (105 > 100)

Fix

src/tools/BashTool/commandSemantics.ts:

1. Add tools to COMMAND_SEMANTICS — exit 1 = informational, 2+ = real error:
   • Linters: ruff, eslint, flake8, biome
   • Type checkers: mypy, pyright, tsc
   • Test runners: pytest, jest, vitest
2. pylint — uses an OR-ed bitfield (1=fatal, 2=error msg, 4=warn, 8=refactor, 16=convention, 32=usage). Only fatal (1) or usage error (32) is a real failure; message bits are findings the model should read.
3. Runner-aware base command extraction — heuristicallyExtractBaseCommand now looks past package/module runners and resolves path/version-pinned invocations:
   • uvx ruff check → ruff
   • npx --yes eslint@8 src → eslint
   • python -m pytest → pytest
   • ./node_modules/.bin/tsc → tsc

Deliberately conservative

• pnpm/yarn/poetry are not unwrapped: their subcommand is usually a package.json/script name, not a tool, so exit-code semantics are ambiguous.
• black --check (exit 1 = would reformat) is not added — it falls through to default semantics. Can follow up if wanted.
• Anything not in the list keeps today's exact behavior.

Test plan

• bun test src/tools/BashTool/commandSemantics.test.ts — 58/58 pass
• bun test src/tools/BashTool/ — 126/126 pass (no regressions)
• tsc --noEmit — no new errors in changed file

New tests cover each tool family, the pylint bitfield (including OR-ed codes like 17 and 20), runner unwrapping, version pins, path-based invocations, compound chains, and the negative case (python script.py exit 1 stays a real error).

🤖 Generated with Claude Code

[laurentftech]

Update: e2e testing caught a real bug

Added commandSemantics.e2e.test.ts that spawns real binaries and feeds their actual exit codes to interpretCommandResult (skips gracefully when a binary is absent).

This revealed tsc is inverted vs every other linter (verified against TypeScript 5.9):

Exit	tsc meaning	Classification
0	clean	not error
1	CLI/usage error	real error
2	type/syntax diagnostics found	not error (read them)
3+	config/internal error	real error

The original exitOneInformational('tsc') mapping was backwards — it would have flagged real type errors (exit 2) as failures and retried, the exact bug this PR fixes. Now corrected with a dedicated semantic.

e2e coverage (real tsc, and real ruff via uvx):

• tsc clean→0, type-error→2 (non-error), bad-flag→1 (error)
• ruff clean→0, violations→1 (non-error, uvx prefix unwraps to ruff), bad-flag→2 (error)

bun test src/tools/BashTool/ → 135 pass.