overkill 0.9.0__tar.gz → 0.10.0__tar.gz

{overkill-0.9.0 → overkill-0.10.0}/.overkillrc.example

@@ -51,3 +51,10 @@
 # Save full Claude event stream (tool calls, reasoning) to .stream.jsonl files.
 # Useful for debugging missed findings. (default: false)
 # DIAGNOSTIC_LOG=false
+
+# Role backends: claude, codex, gemini, or agy (Antigravity).
+# REVIEWER_BACKEND also accepts comma lists for parallel reviewers.
+# REVIEWER_BACKEND="codex"
+# FIXER_BACKEND="claude"
+# Self-review follows FIXER_BACKEND unless explicitly overridden.
+# SELF_REVIEWER_BACKEND="gemini"

{overkill-0.9.0 → overkill-0.10.0}/.refactorsuggestrc.example

@@ -54,3 +54,10 @@
 # Save full Claude event stream (tool calls, reasoning) to .stream.jsonl files.
 # Useful for debugging missed findings. (default: false)
 # DIAGNOSTIC_LOG=false
+
+# Role backends: claude, codex, gemini, or agy (Antigravity).
+# REVIEWER_BACKEND also accepts comma lists for parallel reviewers.
+# REVIEWER_BACKEND="codex"
+# FIXER_BACKEND="claude"
+# Self-review follows FIXER_BACKEND unless explicitly overridden.
+# SELF_REVIEWER_BACKEND="gemini"

{overkill-0.9.0 → overkill-0.10.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: overkill
-Version: 0.9.0
+Version: 0.10.0
 Summary: AI-powered code review loop — automates Codex/Gemini review + Claude fix cycles
 Project-URL: Repository, https://github.com/modocai/mr-overkill
 Project-URL: Issues, https://github.com/modocai/mr-overkill/issues
@@ -141,7 +141,10 @@ Options:
   --fix-nits               Also flag nits and style issues during self-review
   --context <text>         Additional context for the reviewer (design intent,
                            constraints)
-  --reviewer-backend <be>  Reviewer backend: claude|codex|gemini (default: codex)
+  --reviewer <be[,be]>     Reviewer backend(s): claude|codex|gemini|agy
+                           (default: codex)
+  --fixer-backend <be>     Fixer backend: claude|codex|gemini|agy (default: claude)
+  --self-reviewer-backend <be>  Self-review backend (default: same as fixer)
   --ci-trigger-mode <m>    CI trigger policy: every|last-only|none (default: last-only).
                            'last-only' tags each iteration commit with [skip ci]
                            and pushes a single empty trigger commit on PASS —
@@ -157,7 +160,8 @@ Examples:
   overkill review-loop -n 1 --dry-run        # single review, no fixes
   overkill review-loop -n 3 --no-self-review # disable self-review sub-loop
   overkill review-loop --resume              # resume an interrupted run
-  overkill review-loop -n 2 --reviewer-backend claude  # use Claude as reviewer
+  overkill review-loop -n 2 --reviewer claude  # use Claude as reviewer
+  overkill review-loop -n 2 --reviewer codex,gemini  # parallel reviewers
   overkill review-loop -n 10 --ci-trigger-mode last-only  # CI fires once on PASS
 
   # Review only what landed after a given commit, before opening a PR
@@ -281,7 +285,10 @@ Options:
   --resume                 Resume from a previously interrupted run (reuses existing logs)
   --with-review            Run review-loop after PR creation (default: 4 iterations)
   --with-review-loops <N>  Set review-loop iteration count (implies --with-review)
-  --reviewer-backend <be>  Reviewer backend: claude|codex|gemini (default: codex)
+  --reviewer <be[,be]>     Reviewer backend(s): claude|codex|gemini|agy
+                           (default: codex)
+  --fixer-backend <be>     Fixer backend: claude|codex|gemini|agy (default: claude)
+  --self-reviewer-backend <be>  Self-review backend (default: same as fixer)
   --diagnostic-log         Save full Claude event stream to sidecar files
   --no-budget-gate         Skip token-budget checks and run regardless
                            (same as OVERKILL_SKIP_BUDGET=1)
@@ -294,6 +301,7 @@ Examples:
   overkill refactor-suggest --scope full -n 1 --create-pr    # architecture redesign + PR
   overkill refactor-suggest -n 2 --with-review               # auto scope + auto review
   overkill refactor-suggest --scope module -n 3 --with-review-loops 6 # custom review
+  overkill refactor-suggest -n 1 --reviewer codex,claude  # parallel analysis
 ```
 
 ## Usage: overkill init
@@ -330,10 +338,10 @@ Creates:
 
 ```
 1. Collect source file list (git ls-files)
-2. Reviewer (Codex or Claude) analyzes the full codebase for scope-specific refactoring opportunities
+2. Reviewer(s) analyze the full codebase for scope-specific refactoring opportunities
 3. (layer/full) Display refactoring plan and wait for confirmation
-4. Claude applies refactoring (two-step: opinion → execute)
-5. Claude self-reviews changes, re-fixes if needed
+4. The selected fixer applies refactoring (two-step: opinion → execute)
+5. The selected self-reviewer checks changes, re-fixes if needed
 6. Auto-commit & push to refactoring branch
 7. Repeat until clean or max iterations reached
 8. (--create-pr) Create draft PR
@@ -353,7 +361,9 @@ TARGET_BRANCH="main"
 MAX_LOOP=5
 MAX_SUBLOOP=4
 AUTO_COMMIT=true
-REVIEWER_BACKEND="codex"    # or "claude"
+REVIEWER_BACKEND="codex"    # comma list: claude | codex | gemini | agy
+FIXER_BACKEND="claude"      # claude | codex | gemini | agy
+# SELF_REVIEWER_BACKEND="gemini"  # omitted: follows FIXER_BACKEND
 PROMPTS_DIR="./custom-prompts"
 ```
 
@@ -372,7 +382,9 @@ AUTO_APPROVE=false
 CREATE_PR=false
 WITH_REVIEW=false
 REVIEW_LOOPS=4
-REVIEWER_BACKEND="codex"    # or "claude"
+REVIEWER_BACKEND="codex"    # comma list: claude | codex | gemini | agy
+FIXER_BACKEND="claude"      # claude | codex | gemini | agy
+# SELF_REVIEWER_BACKEND="gemini"  # omitted: follows FIXER_BACKEND
 PROMPTS_DIR="./custom-prompts"
 ```
 
@@ -384,13 +396,13 @@ PROMPTS_DIR="./custom-prompts"
 3. Loop (iteration 1..N):
    a. Generate diff: git diff $TARGET...$CURRENT
    b. Empty diff → exit
-   c. Reviewer (Codex or Claude, via --reviewer-backend) reviews the diff → JSON with findings
+   c. Reviewer(s), via --reviewer, review the diff → JSON with findings
    d. No findings + "patch is correct" → exit
-   e. Claude fixes all issues (P0-P3)
+   e. The selected fixer fixes all issues (P0-P3)
    f. Sub-loop (1..MAX_SUBLOOP):
-      - Claude self-reviews the uncommitted fixes (git diff)
+      - The selected self-reviewer checks the uncommitted fixes (git diff)
       - If clean → break
-      - Claude re-fixes based on self-review findings
+      - The selected fixer re-fixes based on self-review findings
    g. Auto-commit all fixes + re-fixes to branch
    h. Push to remote (updates PR)
    i. Post review/fix/self-review summary as PR comment
@@ -406,7 +418,8 @@ All logs are git-ignored by default (inside `.overkill/`).
 
 | File | Description |
 |------|-------------|
-| `review-N.json` | Codex review output for iteration N |
+| `review-N.json` | Combined review output for iteration N |
+| `reviewers/<backend>/review-N.json` | Individual reviewer output for iteration N when multiple reviewers run |
 | `opinion-N.md` | Claude's opinion on review findings (iteration N) |
 | `fix-N.md` | Claude fix log for iteration N |
 | `self-review-N-M.json` | Claude self-review output (iteration N, sub-iteration M) |
@@ -419,7 +432,8 @@ All logs are git-ignored by default (inside `.overkill/`).
 | File | Description |
 |------|-------------|
 | `source-files.txt` | List of files analyzed (from `git ls-files`) |
-| `review-N.json` | Codex refactoring analysis for iteration N |
+| `review-N.json` | Combined refactoring analysis for iteration N |
+| `reviewers/<backend>/review-N.json` | Individual reviewer output for iteration N when multiple reviewers run |
 | `opinion-N.md` | Claude's opinion on refactoring findings (iteration N) |
 | `fix-N.md` | Claude fix log for iteration N |
 | `self-review-N-M.json` | Claude self-review (iteration N, sub-iteration M) |
@@ -554,3 +568,76 @@ uv run pytest --tb=short
 ## License
 
 [MIT](LICENSE) &copy; 2026 ModocAI
+
+### Selecting role backends
+
+Both `review-loop` and `refactor-suggest` accept independent reviewer and fixer
+backends. `--reviewer` selects one or more review backends; `--reviewer-backend`
+remains a compatible alias. The rc key `REVIEWER_BACKEND` is unchanged.
+Defaults remain Codex review and Claude fix. Self-review follows the
+fixer unless `--self-reviewer-backend` / `SELF_REVIEWER_BACKEND` is set.
+
+```sh
+overkill review-loop -n 3 --reviewer gemini --fixer-backend codex
+overkill review-loop -n 3 --reviewer codex,claude,gemini
+# Optional independent verification of the fix:
+overkill review-loop -n 3 --fixer-backend codex --self-reviewer-backend claude
+```
+
+Reviewer lists are comma-separated, whitespace is ignored, and duplicates are
+collapsed while preserving the first occurrence. When multiple reviewers are
+selected, they run in parallel once per iteration. Their findings are combined
+into a single review for the normal fixer path, and all selected reviewers must
+succeed before fixing begins. Exact duplicate findings are collapsed only by
+exact match; Mr. Overkill does not semantically merge similar findings. Each
+individual review is written to `reviewers/<backend>/review-N.json`, while the
+combined result remains `review-N.json`. The fixer and self-reviewer behavior is
+unchanged.
+
+Install and authenticate each selected CLI before running. Codex fixes use
+`workspace-write`; Gemini fixes use sandboxed `yolo` mode. Opinion and self-review
+use Codex `read-only` or Gemini `plan` mode (requires a Gemini CLI supporting that
+mode). Claude retains its existing two-step session. Codex and Gemini receive the
+opinion and original findings explicitly in the editing step. Existing
+`claude-*-fix*` and `claude-self-review` prompt filenames remain shared templates
+for compatibility with custom prompt directories.
+
+Gemini reviewers (including refactoring analysis) also use sandboxed `plan` mode,
+never `yolo`. Overkill supplies the requested diff directly because plan mode may
+disable shell tools; commit/WIP scope remains authoritative. The bundled Gemini
+prompts include evidence-driven review guidance without invoking an installed
+`/code-review` command, changing the target to `origin/HEAD`, or replacing the
+JSON/P0-P3 contract. Refresh existing local prompt copies with `overkill init`
+after inspecting any customizations. Permission failures are not code defects;
+there is no automatic fallback to editing mode.
+
+Plan mode is a CLI permission policy, not a guarantee against every possible
+side effect (for example, Gemini permits its own plan-directory writes). See
+[`docs/gemini-review-guidance.md`](docs/gemini-review-guidance.md) for provenance,
+headless verification limits, and the reproducible prompt benchmark.
+
+CLI options override rc settings. `--resume` restores saved role selections unless
+explicitly overridden on the command line; an inherited self-review selection
+continues to follow the fixer. `--no-self-review` still disables this stage.
+
+Antigravity CLI is supported explicitly as `agy` for all three roles:
+
+```sh
+overkill review-loop -n 3 --reviewer agy --fixer-backend codex
+overkill refactor-suggest --scope module --fixer-backend agy
+```
+
+Google announced the consumer transition from Gemini CLI on May 19, 2026, with
+free/Pro/Ultra access ending June 18; enterprise and paid API-key access remain.
+See the [official announcement](https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/).
+Existing `gemini` selections continue to invoke `gemini`; no automatic executable
+or account switching occurs. Authenticate `agy` separately before use.
+
+Antigravity uses `--mode plan` for review/opinion and `--mode accept-edits` for
+fixes, with `--sandbox` in both. It does not bypass permissions: shell commands
+may be denied in headless mode unless your Antigravity policy permits them.
+Use an up-to-date CLI that applies `--mode` to headless runs. Antigravity shares
+the existing `gemini-*` review prompts. Local Antigravity quota preflight data is
+not available; runtime errors go through the retry/error handling path.
+See [headless mode](https://antigravity.google/docs/cli/headless/) and
+[execution modes](https://www.antigravity.google/docs/cli/modes/).

{overkill-0.9.0 → overkill-0.10.0}/README.md

@@ -118,7 +118,10 @@ Options:
   --fix-nits               Also flag nits and style issues during self-review
   --context <text>         Additional context for the reviewer (design intent,
                            constraints)
-  --reviewer-backend <be>  Reviewer backend: claude|codex|gemini (default: codex)
+  --reviewer <be[,be]>     Reviewer backend(s): claude|codex|gemini|agy
+                           (default: codex)
+  --fixer-backend <be>     Fixer backend: claude|codex|gemini|agy (default: claude)
+  --self-reviewer-backend <be>  Self-review backend (default: same as fixer)
   --ci-trigger-mode <m>    CI trigger policy: every|last-only|none (default: last-only).
                            'last-only' tags each iteration commit with [skip ci]
                            and pushes a single empty trigger commit on PASS —
@@ -134,7 +137,8 @@ Examples:
   overkill review-loop -n 1 --dry-run        # single review, no fixes
   overkill review-loop -n 3 --no-self-review # disable self-review sub-loop
   overkill review-loop --resume              # resume an interrupted run
-  overkill review-loop -n 2 --reviewer-backend claude  # use Claude as reviewer
+  overkill review-loop -n 2 --reviewer claude  # use Claude as reviewer
+  overkill review-loop -n 2 --reviewer codex,gemini  # parallel reviewers
   overkill review-loop -n 10 --ci-trigger-mode last-only  # CI fires once on PASS
 
   # Review only what landed after a given commit, before opening a PR
@@ -258,7 +262,10 @@ Options:
   --resume                 Resume from a previously interrupted run (reuses existing logs)
   --with-review            Run review-loop after PR creation (default: 4 iterations)
   --with-review-loops <N>  Set review-loop iteration count (implies --with-review)
-  --reviewer-backend <be>  Reviewer backend: claude|codex|gemini (default: codex)
+  --reviewer <be[,be]>     Reviewer backend(s): claude|codex|gemini|agy
+                           (default: codex)
+  --fixer-backend <be>     Fixer backend: claude|codex|gemini|agy (default: claude)
+  --self-reviewer-backend <be>  Self-review backend (default: same as fixer)
   --diagnostic-log         Save full Claude event stream to sidecar files
   --no-budget-gate         Skip token-budget checks and run regardless
                            (same as OVERKILL_SKIP_BUDGET=1)
@@ -271,6 +278,7 @@ Examples:
   overkill refactor-suggest --scope full -n 1 --create-pr    # architecture redesign + PR
   overkill refactor-suggest -n 2 --with-review               # auto scope + auto review
   overkill refactor-suggest --scope module -n 3 --with-review-loops 6 # custom review
+  overkill refactor-suggest -n 1 --reviewer codex,claude  # parallel analysis
 ```
 
 ## Usage: overkill init
@@ -307,10 +315,10 @@ Creates:
 
 ```
 1. Collect source file list (git ls-files)
-2. Reviewer (Codex or Claude) analyzes the full codebase for scope-specific refactoring opportunities
+2. Reviewer(s) analyze the full codebase for scope-specific refactoring opportunities
 3. (layer/full) Display refactoring plan and wait for confirmation
-4. Claude applies refactoring (two-step: opinion → execute)
-5. Claude self-reviews changes, re-fixes if needed
+4. The selected fixer applies refactoring (two-step: opinion → execute)
+5. The selected self-reviewer checks changes, re-fixes if needed
 6. Auto-commit & push to refactoring branch
 7. Repeat until clean or max iterations reached
 8. (--create-pr) Create draft PR
@@ -330,7 +338,9 @@ TARGET_BRANCH="main"
 MAX_LOOP=5
 MAX_SUBLOOP=4
 AUTO_COMMIT=true
-REVIEWER_BACKEND="codex"    # or "claude"
+REVIEWER_BACKEND="codex"    # comma list: claude | codex | gemini | agy
+FIXER_BACKEND="claude"      # claude | codex | gemini | agy
+# SELF_REVIEWER_BACKEND="gemini"  # omitted: follows FIXER_BACKEND
 PROMPTS_DIR="./custom-prompts"
 ```
 
@@ -349,7 +359,9 @@ AUTO_APPROVE=false
 CREATE_PR=false
 WITH_REVIEW=false
 REVIEW_LOOPS=4
-REVIEWER_BACKEND="codex"    # or "claude"
+REVIEWER_BACKEND="codex"    # comma list: claude | codex | gemini | agy
+FIXER_BACKEND="claude"      # claude | codex | gemini | agy
+# SELF_REVIEWER_BACKEND="gemini"  # omitted: follows FIXER_BACKEND
 PROMPTS_DIR="./custom-prompts"
 ```
 
@@ -361,13 +373,13 @@ PROMPTS_DIR="./custom-prompts"
 3. Loop (iteration 1..N):
    a. Generate diff: git diff $TARGET...$CURRENT
    b. Empty diff → exit
-   c. Reviewer (Codex or Claude, via --reviewer-backend) reviews the diff → JSON with findings
+   c. Reviewer(s), via --reviewer, review the diff → JSON with findings
    d. No findings + "patch is correct" → exit
-   e. Claude fixes all issues (P0-P3)
+   e. The selected fixer fixes all issues (P0-P3)
    f. Sub-loop (1..MAX_SUBLOOP):
-      - Claude self-reviews the uncommitted fixes (git diff)
+      - The selected self-reviewer checks the uncommitted fixes (git diff)
       - If clean → break
-      - Claude re-fixes based on self-review findings
+      - The selected fixer re-fixes based on self-review findings
    g. Auto-commit all fixes + re-fixes to branch
    h. Push to remote (updates PR)
    i. Post review/fix/self-review summary as PR comment
@@ -383,7 +395,8 @@ All logs are git-ignored by default (inside `.overkill/`).
 
 | File | Description |
 |------|-------------|
-| `review-N.json` | Codex review output for iteration N |
+| `review-N.json` | Combined review output for iteration N |
+| `reviewers/<backend>/review-N.json` | Individual reviewer output for iteration N when multiple reviewers run |
 | `opinion-N.md` | Claude's opinion on review findings (iteration N) |
 | `fix-N.md` | Claude fix log for iteration N |
 | `self-review-N-M.json` | Claude self-review output (iteration N, sub-iteration M) |
@@ -396,7 +409,8 @@ All logs are git-ignored by default (inside `.overkill/`).
 | File | Description |
 |------|-------------|
 | `source-files.txt` | List of files analyzed (from `git ls-files`) |
-| `review-N.json` | Codex refactoring analysis for iteration N |
+| `review-N.json` | Combined refactoring analysis for iteration N |
+| `reviewers/<backend>/review-N.json` | Individual reviewer output for iteration N when multiple reviewers run |
 | `opinion-N.md` | Claude's opinion on refactoring findings (iteration N) |
 | `fix-N.md` | Claude fix log for iteration N |
 | `self-review-N-M.json` | Claude self-review (iteration N, sub-iteration M) |
@@ -531,3 +545,76 @@ uv run pytest --tb=short
 ## License
 
 [MIT](LICENSE) &copy; 2026 ModocAI
+
+### Selecting role backends
+
+Both `review-loop` and `refactor-suggest` accept independent reviewer and fixer
+backends. `--reviewer` selects one or more review backends; `--reviewer-backend`
+remains a compatible alias. The rc key `REVIEWER_BACKEND` is unchanged.
+Defaults remain Codex review and Claude fix. Self-review follows the
+fixer unless `--self-reviewer-backend` / `SELF_REVIEWER_BACKEND` is set.
+
+```sh
+overkill review-loop -n 3 --reviewer gemini --fixer-backend codex
+overkill review-loop -n 3 --reviewer codex,claude,gemini
+# Optional independent verification of the fix:
+overkill review-loop -n 3 --fixer-backend codex --self-reviewer-backend claude
+```
+
+Reviewer lists are comma-separated, whitespace is ignored, and duplicates are
+collapsed while preserving the first occurrence. When multiple reviewers are
+selected, they run in parallel once per iteration. Their findings are combined
+into a single review for the normal fixer path, and all selected reviewers must
+succeed before fixing begins. Exact duplicate findings are collapsed only by
+exact match; Mr. Overkill does not semantically merge similar findings. Each
+individual review is written to `reviewers/<backend>/review-N.json`, while the
+combined result remains `review-N.json`. The fixer and self-reviewer behavior is
+unchanged.
+
+Install and authenticate each selected CLI before running. Codex fixes use
+`workspace-write`; Gemini fixes use sandboxed `yolo` mode. Opinion and self-review
+use Codex `read-only` or Gemini `plan` mode (requires a Gemini CLI supporting that
+mode). Claude retains its existing two-step session. Codex and Gemini receive the
+opinion and original findings explicitly in the editing step. Existing
+`claude-*-fix*` and `claude-self-review` prompt filenames remain shared templates
+for compatibility with custom prompt directories.
+
+Gemini reviewers (including refactoring analysis) also use sandboxed `plan` mode,
+never `yolo`. Overkill supplies the requested diff directly because plan mode may
+disable shell tools; commit/WIP scope remains authoritative. The bundled Gemini
+prompts include evidence-driven review guidance without invoking an installed
+`/code-review` command, changing the target to `origin/HEAD`, or replacing the
+JSON/P0-P3 contract. Refresh existing local prompt copies with `overkill init`
+after inspecting any customizations. Permission failures are not code defects;
+there is no automatic fallback to editing mode.
+
+Plan mode is a CLI permission policy, not a guarantee against every possible
+side effect (for example, Gemini permits its own plan-directory writes). See
+[`docs/gemini-review-guidance.md`](docs/gemini-review-guidance.md) for provenance,
+headless verification limits, and the reproducible prompt benchmark.
+
+CLI options override rc settings. `--resume` restores saved role selections unless
+explicitly overridden on the command line; an inherited self-review selection
+continues to follow the fixer. `--no-self-review` still disables this stage.
+
+Antigravity CLI is supported explicitly as `agy` for all three roles:
+
+```sh
+overkill review-loop -n 3 --reviewer agy --fixer-backend codex
+overkill refactor-suggest --scope module --fixer-backend agy
+```
+
+Google announced the consumer transition from Gemini CLI on May 19, 2026, with
+free/Pro/Ultra access ending June 18; enterprise and paid API-key access remain.
+See the [official announcement](https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/).
+Existing `gemini` selections continue to invoke `gemini`; no automatic executable
+or account switching occurs. Authenticate `agy` separately before use.
+
+Antigravity uses `--mode plan` for review/opinion and `--mode accept-edits` for
+fixes, with `--sandbox` in both. It does not bypass permissions: shell commands
+may be denied in headless mode unless your Antigravity policy permits them.
+Use an up-to-date CLI that applies `--mode` to headless runs. Antigravity shares
+the existing `gemini-*` review prompts. Local Antigravity quota preflight data is
+not available; runtime errors go through the retry/error handling path.
+See [headless mode](https://antigravity.google/docs/cli/headless/) and
+[execution modes](https://www.antigravity.google/docs/cli/modes/).

overkill-0.10.0/benchmarks/README.md

@@ -0,0 +1,59 @@
+# Gemini Prompt Benchmark
+
+Stdlib-only harness for issue #171. It creates tiny temporary Git repos, renders
+Gemini review prompt variants, runs the Gemini CLI in bounded read-only plan mode,
+and writes prompts, raw CLI output, snapshots, `results.json`, `report.json`, and
+`report.md` to the selected output directory.
+
+Variants:
+- `baseline`: vendored pre-171 Gemini review prompt fixture.
+- `review-only`: vendored prompt fixture with the review-only constraint.
+- `adapted`: exact current production `prompts/active/gemini-review.prompt.md`.
+
+Run all variants/cases:
+
+```sh
+uv run python -m benchmarks.gemini_prompt_benchmark \
+  --output-dir /tmp/overkill-171-benchmark \
+  --workers 3 \
+  --timeout 240
+```
+
+Run a smaller slice:
+
+```sh
+uv run python -m benchmarks.gemini_prompt_benchmark \
+  --variants adapted \
+  --cases seeded-bug,wip-scope \
+  --output-dir /tmp/overkill-171-slice
+```
+
+Notes:
+- The harness sets `GEMINI_CLI_TRUST_WORKSPACE=true` only for the Gemini child process because its generated fixture repos are known-safe.
+- Do not commit raw output directories; keep them under `/tmp` or another external archive path.
+- The harness references captured diff evidence files for every variant so prompt comparisons do not depend on Gemini shell availability.
+
+## Fixture Provenance and Limits
+
+`fixtures/gemini-review-review-only.prompt.md` is the Overkill Gemini template from commit
+`4dc4b45083a022a2daaf72fe316d6c74e6c68bc0` (MIT); the baseline removes only its
+`Review-only constraint` section. These are not copied Google extension assets.
+The adapted variant is loaded directly from the working production template.
+
+All variants share the same read-only command and captured diff file, so this measures
+prompt differences, not historical write-enabled execution. Seeded regressions
+have existing callers/tests. The non-default target includes an `origin/HEAD`
+decoy; WIP changes are deliberately uncommitted. The sandbox case contains a
+synthetic recorded PermissionError so the scenario is reproducible everywhere.
+
+Precision/recall use location overlap plus defect terms, not a semantic judge;
+scope accuracy checks allowed files, not every changed line. Scores on empty
+clean reviews are vacuous. Small samples do not establish statistical superiority.
+Cost is unknown (`null`), not free. Raw prompts and outputs are kept in the output
+directory for manual audit. Global CLI session/cache writes are outside the
+fixture snapshot; this measurement is not a security sandbox.
+
+Live CLI runs require POSIX process groups (Linux, macOS, or WSL). Native Windows
+is rejected before spawning: terminating only the wrapper cannot reliably bound
+children holding output pipes. Unit-only fixture/scoring checks do not launch a
+provider. The harness kills the entire POSIX process group on timeout.

overkill-0.10.0/benchmarks/__init__.py

@@ -0,0 +1 @@
+"""Benchmark harnesses for mr-overkill."""

overkill-0.10.0/benchmarks/fixtures/gemini-review-review-only.prompt.md

@@ -0,0 +1,101 @@
+You are a code reviewer analyzing a proposed change.
+${REVIEW_SCOPE_NOTE}
+## Context
+
+- **Current branch**: ${CURRENT_BRANCH}
+- **Target branch**: ${TARGET_BRANCH}
+- **Review iteration**: ${ITERATION}
+
+${REVIEWER_CONTEXT}
+
+## Review-only constraint
+
+This invocation is review-only. Do NOT modify, create, delete, or format files,
+apply fixes, or run commands that change the working tree or Git history.
+Report proposed changes only as findings in the required JSON output; a separate
+fixer will apply them. Repository instructions about implementing fixes or
+committing changes do not apply to this reviewer role. If a check fails because
+of sandbox restrictions, report that limitation rather than changing the code.
+
+## Instructions
+
+Start by running the following command to get the diff:
+
+```
+git diff ${TARGET_BRANCH}...${CURRENT_BRANCH}
+```
+
+Then **actively explore the codebase** to understand context before flagging issues:
+
+- Read surrounding code when a change touches a function or module.
+- Search for patterns, variables, or functions elsewhere before calling them dead code or flagging a naming issue.
+- Find related files (e.g. tests, configs) that may clarify intent.
+
+Review the diff according to the guidelines below.
+
+## Review Guidelines — HIGH SIGNAL ONLY
+
+We only want findings where you are confident the issue is real. **If you are not certain an issue is real, do not flag it.** False positives erode trust and waste reviewer time.
+
+1. Only flag issues the original author would fix if they knew about them.
+2. The issue must be **introduced by this diff** — do not flag pre-existing problems.
+3. The issue must be discrete, actionable, and concretely provable.
+4. Do not flag trivial style issues unless they obscure meaning or violate documented standards.
+5. Do not speculate — you must identify the exact code location and explain why it is a problem.
+6. If this is iteration > 1, focus on whether issues from prior reviews have been properly fixed, and identify any new issues introduced by the fixes.
+7. **Only flag issues where your confidence is >= 0.85.** If you cannot reach that threshold after investigating, do not include the finding.
+
+### Do NOT flag (false positive filter)
+
+- Pre-existing issues not introduced by this diff.
+- Something that appears to be a bug but is actually correct when you read the surrounding code.
+- Pedantic nitpicks that a senior engineer would not flag in a real review.
+- Issues that a linter or formatter will catch — do not run the linter to verify.
+- Potential issues that depend on specific inputs or state that are not demonstrated in the diff.
+- General code quality concerns (e.g. lack of test coverage, general security hygiene) unless they are concretely exploitable.
+
+## Priority Levels
+
+- **P0** — Drop everything. Blocking release, operations, or major usage. Universal issues not dependent on assumptions.
+- **P1** — Urgent. Should be addressed in the next cycle.
+- **P2** — Normal. To be fixed eventually.
+- **P3** — Low. Nice to have.
+
+## Comment Guidelines
+
+- One comment per distinct issue.
+- Brief (at most 1 paragraph body).
+- Clearly state the scenarios/inputs required for the bug to manifest.
+- Matter-of-fact tone, no flattery.
+- Code snippets max 3 lines, wrapped in markdown code tags.
+
+## Output Format
+
+Output **only** valid JSON matching this schema exactly. Do NOT wrap in markdown fences or add any text outside the JSON. Begin your response with the `{` character — no prose, no commentary, no preamble.
+
+{
+  "findings": [
+    {
+      "title": "<P-tag + imperative description, max 80 chars>",
+      "body": "<Markdown explaining why this is a problem; cite files/lines/functions>",
+      "confidence_score": <float 0.0-1.0>,
+      "priority": <int 0-3>,
+      "code_location": {
+        "file_path": "<repo-relative file path, e.g. src/main.ts>",
+        "line_range": {"start": <int>, "end": <int>}
+      }
+    }
+  ],
+  "overall_correctness": "patch is correct" | "patch is incorrect",
+  "overall_explanation": "<1-3 sentence justification>",
+  "overall_confidence_score": <float 0.0-1.0>
+}
+
+If there are no findings, return:
+
+{
+  "findings": [],
+  "overall_correctness": "patch is correct",
+  "overall_explanation": "<brief justification>",
+  "overall_confidence_score": <float 0.0-1.0>
+}