[getdownandcode]

Summary

Fixes all 8 findings from the code review (bugs 1\u20138), each with regression tests where meaningful.

High

1. Free-tier users billed at paid rates \u2014 commit_flow now passes free_tier=cfg.free_tier to both telemetry calls (initial generation and regenerate), so rows are stored with the correct flag and get_invocations no longer reprices them at paid rates. Free-tier usage can no longer trip budget_cap_usd.

2. Rich markup ate text in square brackets \u2014 commit preview panel, git stdout/stderr, and the a/e/r/c menus now render as literal text (rich.text.Text / rich.markup.escape / \[…\]). fix(cli): handle list[str] input [skip ci] displays verbatim; git's [main 1a2b3c4] … line keeps the branch and hash; invalid input containing [/x] no longer crashes with MarkupError.

Medium

3. Truncated files lost their filename \u2014 patches replaced by [patch omitted for size: …] are now prefixed with File <path> (<status>): in the prompt (new _prompt_diff_text, OMITTED_NOTE_PREFIX).

4. Binary changes never reached the model \u2014 binary files are now included in the prompt as Binary file <path> (<status>).

5. pr-summary / changelog ignored ignore_globs \u2014 both now pass extra_ignores=cfg.ignore_globs to DiffExtractor, matching commit and debug-diff.

6. Cached chunk summaries were billed \u2014 CachedProvider exposes last_hit; summarize_omitted_diffs excludes cache hits from billed tokens, and the orchestrator uses last_hit (instead of probing cache internals) to classify the main call. A fully cached rerun now costs $0.

Low

7. Budget warning printed twice \u2014 new check_budget flag (default True) on generate_commit_message/run_generation_pipeline; callers that already ran check_budget_cap pass check_budget=False.

8. Misleading fallback message \u2014 token-counting failures now print "Token budgeting failed, using template fallback" instead of "API unavailable".

Testing

• 304 tests pass (uv run pytest -q) \u2014 includes new regressions: free-tier $0 spend, literal brackets end-to-end, prompt labels for truncated/binary files, cached-chunk $0 billing, single budget check, ignore_globs in pr-summary
• ruff check . and ruff format . clean
• mypy (strict) clean
• One stale test that pinned the old bug-6 behaviour was updated (test_generate_commit_message_fully_cached_rerun_costs_nothing)

@@ tests/test_orchestrator.py
-    # Final call cost was 0 (cache hit), but chunk_cost MUST still be preserved!
-    assert res2.estimated_cost_usd == chunk_cost
+    # No API tokens were billed on the rerun: cached chunk summaries are excluded.
+    assert res2.estimated_cost_usd == 0.0

Note: this PR is based on feat (not main) so it contains only the fixes \u2014 they'll flow into main via the open #14.

[getdownandcode]

🤖 Automated PR Review: #15 - fix: address review findings on billing, markup, and prompt fidelity

1. Summary of Changes

• Scope: Addresses 8 defect findings from previous code review across billing metrics, terminal UI rendering, prompt fidelity, and chunk cache accounting.
• Key Files Modified: src/gitmate/committer.py, src/gitmate/orchestrator.py, src/gitmate/chunking.py, src/gitmate/providers/cache.py, src/gitmate/changelog.py, src/gitmate/pr_summary.py, src/gitmate/fallback.py, src/gitmate/token_budget.py, tests/test_committer.py, tests/test_orchestrator.py, tests/test_pr_summary.py

2. Analysis & Key Highlights

• Strengths:
  • Accurate Telemetry & Spend Isolation: Properly threads free_tier=active_cfg.free_tier into commit_flow telemetry calls, preventing free-tier users from having usage repriced at commercial API rates.
  • Crash-Resistant Rich UI: Wraps user message text, git command stdout/stderr, and prompt options with Text(...) and escape(...), neutralizing MarkupError exceptions when commit messages contain unescaped brackets like [skip ci] or [feat].
  • O(1) Cache State Tracking: Introduces last_hit attribute on CachedProvider, eliminating intrusive cache key probing and ensuring cached chunk summaries are not double-billed on subsequent runs.
  • Prompt Fidelity: Restores file path and change status context for omitted/truncated patches (File <path> (<status>): ...) and binary files (Binary file <path> (<status>)).
• Observations:
  • This PR targets feat (PR feat: Phase 9 — ship it (packaging, release workflow, docs) #14) directly rather than main. Merging this PR into feat first ensures that these bugfixes flow into main cleanly when PR feat: Phase 9 — ship it (packaging, release workflow, docs) #14 is merged.

3. Actionable Suggestions & Concerns

• [Severity: Medium] [tests/test_committer.py:270, 295, 470]: Ambient GIT_EDITOR shadows EDITOR in interactive edit unit tests. In test_interactive_edit_compound_editor_and_comment_stripping, test_interactive_edit_aborts_on_empty_commit_message, and test_interactive_edit_trailing_whitespace_stripped, monkeypatch.setenv("EDITOR", ...) is called without clearing higher-precedence editor variables. When GIT_EDITOR is present in the ambient environment (common in CI/subshells, e.g. GIT_EDITOR=false), _resolve_editor() selects GIT_EDITOR, causing editor execution to fail with exit code 1.

  # Ensure clean isolation across all interactive edit tests:
  monkeypatch.delenv("GIT_EDITOR", raising=False)
  monkeypatch.delenv("VISUAL", raising=False)
  monkeypatch.setenv("EDITOR", f"{sys.executable} -c \"{editor_script}\"")

• [Severity: Low] [src/gitmate/orchestrator.py:266]: Inconsistent provider duck-typing. chunking.py:68 checks getattr(provider, "last_hit", False), whereas orchestrator.py:266 uses isinstance(active_provider, CachedProvider) and active_provider.last_hit.

  # Suggestion: Use consistent duck-typing in orchestrator.py:
  cache_hit = bool(getattr(active_provider, "last_hit", False))

• [Severity: Low] [tests/test_orchestrator.py:446-447, 461-462]: Unpack typing suppression. The tests use # type: ignore[arg-type] due to dictionary unpacking (**kwargs / **common). Calling generate_commit_message(diffs=_sample_diff(), cfg=cfg, provider=provider, counter=FakeCounter()) with explicit keyword arguments preserves type checking and avoids suppression comments.

4. Merge Status

Verdict: READY TO MERGE
Reasoning: Resolves all 8 target defects with rigorous regression tests, clean type hints, and zero breaking changes.