fix: save engine stderr to temp log and show --exclude hint on failure

[takai]

Closes #19

Problem

When the engine subprocess fails on a large diff, the error shown to the user is not actionable:

• Claude: exits 1 with empty stderr → engine command failed: exit status 1: (no information)
• Codex: exits 1 with the entire prompt echoed to stderr, burying the real error at the end → unreadable wall of text

Both cases left the user with no idea what went wrong or how to fix it.

Solution

engine.EngineError (new type in internal/engine/cli.go)

CLI.Generate() now returns a structured *EngineError on non-zero exit instead of embedding stderr into the error string. The Stderr field preserves the full output separately, accessible via errors.As.

Actionable error message (internal/app/app.go)

When the engine fails, buildEngineFailureError():

1. Saves the full stderr to a temp file (/tmp/git-ai-commit-stderr-*.log) — even when stderr is empty. The file is kept for the user to inspect.
2. Shows a --exclude hint when the diff had truncated or pattern-excluded files, suggesting the likely culprits for a context window overflow.

Example output:

engine command failed: exit status 1
Full engine output saved to: /tmp/git-ai-commit-stderr-3728164819.log
Hint: the following files were truncated or excluded and may have caused
a context window overflow. Re-run with --exclude to skip them:
  git ai-commit --exclude 1m_lines.txt

Changes

• internal/engine/cli.go — EngineError type; Generate() returns *EngineError on failure
• internal/engine/cli_test.go — unit tests for EngineError
• internal/app/app.go — buildEngineFailureError, writeTempLog, buildExcludeCandidates helpers; commitDiff now also returns git.Result
• internal/app/app_test.go — unit tests for the new helpers

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 309efbbd3c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Check temp log writes before reporting saved output

writeTempLog unconditionally returns the temp file path even when WriteString fails, so buildEngineFailureError can tell users “Full engine output saved to …” when the log is actually empty or truncated. This is reproducible on ENOSPC or other write errors and removes the only place stderr is preserved in this flow, making failures harder to debug in exactly the environments where diagnostics are most needed.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Exclude hint must avoid already pattern-excluded files

buildExcludeCandidates appends filterResult.ExcludedFiles directly, but those files were already excluded by git.Filter and therefore cannot be further reduced by adding --exclude again. When engine execution fails, this produces no-op remediation steps (for example default-excluded lockfiles), which can generate long misleading hints and distract from the actionable files that are actually still in the prompt.

Useful? React with 👍 / 👎.