Countzero Windows Llama Cpp Windows Llama Cpp

Multi-Pass PR Code Review

This skill performs 3 iterative review passes over a pull request diff with escalating focus: general defects, cross-file interactions, then absent behavior. After all passes, a final summary deduplicates, validates, assigns severity, and filters to Critical/High/Medium. The output includes clickable deep links to the relevant code on GitHub so the developer or reviewer can easily navigate to each finding.

PR and Diff Resolution

1. Run gh pr view --json number,baseRefName,state,isDraft. This single call determines whether a PR exists and, if so, its state.
2. If a PR exists:
   • If the PR is closed or a draft, inform the user and stop.
   • Otherwise, use the PR's baseRefName as the base for the diff.
3. If no PR exists (the command fails):
   • Use main as the default base branch (the repo's mainline; feature work and develop are PR'd into it).
   • Inform the user: "No PR found. Diffing against main."
4. Get the diff via git diff <base>...HEAD.
5. Resolve the full HEAD SHA via git rev-parse HEAD.
6. Resolve the repository owner/name via gh repo view --json owner,name (fall back to parsing git remote get-url origin if gh is unavailable). Store the owner, repo name, full SHA, and PR number (or null when no PR exists) for constructing deep links in the output.

Context Gathering

Before starting the review passes, gather context to understand the intent behind the changes:

1. Branch and commits: Read the current branch name via git branch --show-current and the commit subjects in range via git log --oneline <base>..HEAD to understand the stated intent.
2. Full file reads: Do NOT read full files upfront. Instead, during each pass, if a diff hunk lacks sufficient surrounding context to assess a potential defect, read the full file at that point using the Read tool.

Design Review (before passes)

After gathering context and the diff, but before starting the defect passes, evaluate the overall design of the change:

1. Scope: Does the PR implement more or less than its stated intent (branch name, commit subjects)? Flag scope creep or missing pieces.
2. Placement: Do the changes live in the correct layer? This repo is a PowerShell wrapper: build/orchestration logic belongs in *.ps1 at the repo root or examples/; model configuration belongs in presets/*.ini; no original C/C++/Python source belongs here, and nothing under vendor/llama.cpp/ is expected to persist (each rebuild wipes it).
3. Complexity: Is any part of the change more complex than necessary? Flag over-engineering, unnecessary abstractions, or functionality that is not required by the change's intent.

Record design-level findings separately under ### Design in the output, before the severity tables. Design findings do not go through the severity classification; they are qualitative observations for the author.

Workflow and User Communication

Use the TodoWrite tool to create and track these steps:

1. Resolve PR, base branch, and gather context
2. Design review
3. Pass 1
4. Pass 2
5. Pass 3
6. Build final summary (deduplicate, re-examine, severity, filter); internally only
7. Print the report in the terminal

Mark each todo as in_progress when starting and completed when done. After each pass, tell the user in one line how many new defects were found (e.g., "Pass 1 complete; found 6 defects.").

After Pass 3, build the final summary internally; do not print it yet. Once the summary is complete, print the report body in the conversation as the turn's final output, following the Report Output section below. Nothing (no recap, no "Let me know if…", no "Review complete" line, no follow-up commentary) may be printed after the report. The report is the terminal output of the turn; any follow-up question opens a new turn.

Review Passes

Three passes over the full diff, each with a different focus. All passes review the full diff. Findings are recorded without severity: just File, Line(s), and Description.

Only flag defects in lines that are added or modified in this PR. Do not flag issues in unchanged context lines, even if they appear in diff hunks.

Skip any finding that:

• Matches a finding from a previous pass (same file and overlapping line range).
• Describes the same logical issue as an existing finding but references a different location (e.g., a function definition vs. its call site). Two findings about the same root cause count as one; keep the one closest to the root cause.

Pass 1: General scan Review the diff. Report all defects: bugs, logic errors, security issues, bad practices, missing validation, incorrect error handling. Also check the project-specific concerns listed in the Project-Specific Review Checklist section below. Only flag defects in lines that are added or modified in this PR.

Pass 2: What was missed Review the diff again, assuming defects were missed on the first pass. Focus on interactions between changed files, subtle logic errors, and implicit assumptions in the code. Only flag defects in lines that are added or modified in this PR.

Pass 3: What the code does NOT do Assume there are still undiscovered defects. Focus on what is absent: missing error handling, missing edge cases, missing input validation, missing null checks, race conditions, resource leaks, and incorrect assumptions about state. Only flag defects in lines that are added or modified in this PR.

Track findings internally across passes (in conversation context). The format for each finding is: File, Line(s), Description.

False Positive Exclusion List

Do NOT flag any of the following:

• Pre-existing issues not introduced in this PR's changes.
• Code that appears to be a bug but is actually correct.
• Pedantic nitpicks that a senior engineer would not flag.
• General code quality concerns unless explicitly required in AGENTS.md.
• Issues explicitly silenced in code (e.g., via a lint ignore comment).
• Pure code style or formatting preferences.
• Potential issues that depend on specific inputs or runtime state.
• Findings inside vendor/ (the submodule is wiped and re-checked-out each rebuild; only the documented idempotent CMakeLists shim is sanctioned).

Project-Specific Review Checklist

These are architectural and safety concerns specific to this PowerShell wrapper around llama.cpp; there is no linter to catch them. Check for these during all passes, in addition to general defect scanning. They are NOT style issues; they are correctness and safety rules. Each maps to a rule in AGENTS.md ("Non-obvious behavior", "Presets", "Changelog style").

• Ephemeral submodule edits: Changes under vendor/llama.cpp/ that expect to persist. Each rebuild_llama.cpp.ps1 resets the submodule to origin/master and re-checks-out; the only sanctioned edit is the idempotent OpenBLAS CMakeLists.txt shim the build re-applies.
• Submodule "cleanup": Removing .gitmodules ignore = dirty, or resetting the vendor/llama.cpp pointer casually. The dirty state is by design.
• PowerShell argument passing: Values that can contain spaces routed through an Invoke-Expression-built command string or the -additionalArguments whitespace-split/re-pair parser (they will not survive); unquoted space-bearing paths in llama-server/python invocations.
• CWD-relative path assumptions: chat-template-file and read_file() resolve against the launch CWD, not the INI directory. Scripts that shell out to vendored Python (gguf_dump.py, speed_bench.py) must guard the path against upstream relocation.
• Build-detection invariants: CUDA is selected only when both nvidia-smi and nvcc are present; do not drop -DCMAKE_ASM_COMPILER=ml64 (CMP0194 ASM breakage); do not defeat the SMT-aware --parallel cap or the running-build-tree-process abort.
• Python requirements layering: Edits to requirements_override.txt that break the documented torch (cu126), transformers, numpy<2.3, or tiktoken pins.
• Preset INI semantics: mmproj-offload = true silently OOMing CLIP warmup on a saturated GPU; dropping a required chat-template-file pin (it replaces the GGUF-embedded template, not redundant with jinja = true); stale spec-type flag names.
• CHANGELOG style: Entries that are not one physical line, omit the [Component] <verb> <thing> form, or carry rationale/paths/line numbers.

Final Summary

After Pass 3:

1. Deduplicate: Two findings are duplicates if they reference the same file and overlapping line ranges, or if they describe the same root cause at different locations. Keep the more detailed description.
2. Re-examine: For each remaining finding, verify it against the full file context. If the file was not already read during a pass, read it now using the Read tool. Remove any finding that cannot be confirmed after re-examination.
3. Confidence filter: Remove any finding where confidence is below 80% that it is a real defect. When in doubt, exclude.
4. Assign severity using these definitions:
   • Critical: Data loss, security breach, corruption of state, a build that destroys a working tree, or a crash on a common path.
   • High: Incorrect behavior under normal usage, unhandled error paths that will be hit, significant logic flaw.
   • Medium: Bad practice that could lead to bugs, missing validation for unlikely but possible inputs, minor logic issue.
5. Filter: Keep only Critical, High, and Medium.
6. Print: Print the deduped, severity-filtered findings directly in the terminal, following the Report Output section below. When the filtered set is empty, print a single line No Critical/High/Medium findings. in place of the severity tables; the metadata table and Design section still print normally. When the set is non-empty, print one table per severity level (omitting empty levels), following the link format rules in the Link Format section below.

No data is written to GitHub. The developer or reviewer uses the output to manually create PR comments.

Link Format

This section defines how links appear in the output tables. Follow these rules exactly.

Columns: File, Description. The File column contains a markdown link. The link label is the file path relative to the repository root with a leading /, plus the line range in :{start}-{end} format (e.g., /examples/speed-bench.ps1:41-43). The link URL is the full GitHub URL constructed as described below.

Line range: Include 1 line of context before and after the finding. A finding on line 42 links to lines 41-43. A finding spanning lines 42-45 links to lines 41-46. The label always uses :{start}-{end} regardless of whether the URL uses L or R anchors.

SHA-256 hash for PR links: Compute the SHA-256 hex digest of each unique file path via Node.js (cross-platform): node -e "process.stdout.write(require('crypto').createHash('sha256').update('{path}').digest('hex'))".

When a PR exists: Use PR deep links. The URL format is https://github.com/{owner}/{repo}/pull/{pr-number}/files#diff-{sha256}R{start}-R{end}.

When no PR exists (fallback): Use blob links with the full HEAD SHA: https://github.com/{owner}/{repo}/blob/{full-sha}/{path}#L{start}-L{end}. Note: blob links use L (not R) for line anchors.

Correct: markdown links with file path labels.

| File                                                                                                            | Description                  |
| --------------------------------------------------------------------------------------------------------------- | ---------------------------- |
| [/examples/speed-bench.ps1:41-43](https://github.com/owner/repo/pull/42/files#diff-a1b2c3d4e5f6a7b8c9d0R41-R43) | Unquoted path breaks on space |
| [/examples/speed-bench.ps1:41-43](https://github.com/owner/repo/blob/4a7c9e1f/examples/speed-bench.ps1#L41-L43) | Unquoted path breaks on space |

Wrong: do NOT use any of these formats.

| https://github.com/owner/repo/pull/42/files#diff-a1b2c3d4e5f6a7b8c9d0R41-R43 | ...  |
| [https://github.com/...](https://github.com/...)                             | ...  |
| [speed-bench.ps1:41-43](https://github.com/...)                              | ...  |

The first is wrong because it uses a raw URL instead of a markdown link. The second is wrong because the label is a URL instead of a file path. The third is wrong because it uses a filename only; the label must be the full path from the repository root with a leading /.

Report Output

The report is printed directly in the terminal as the turn's final output; no files are written and no trailer follows it. Print the body in this fixed order: (1) the title (# PR Review then ## <branch> (vs <base>)), (2) the metadata table, (3) the ### Design section, (4) the severity tables (or the empty-findings line). Nothing may be printed after the report.

All tables (the metadata table and every severity table) must have every cell (header, separator, body) padded with spaces (or dashes for the separator row) so all cells in the same column occupy the same character width between pipes. This keeps the report scannable in the monospace TUI.

The metadata block is a 2-column GFM table with explicit Field / Value headers. Severity glyphs (🔴 🟠 🟡) keep the Findings row scannable without custom styling.

Metadata-row formatting rules:

• Branch: link both branches to their GitHub tree URLs: https://github.com/<owner>/<repo>/tree/<branch-name>. Multi-segment branch names (e.g. feature/foo) work without encoding because the tree/ path accepts literal slashes; percent-encode any branch name that contains spaces or other reserved characters.
• HEAD: use the 7-char short SHA in both the link label and the URL (/commit/<head7>). The short SHA keeps the row narrow enough that the OpenCode TUI does not wrap the Field column. GitHub redirects /commit/<short> to the full commit, so the link resolves identically.
• PR: the PR number linked to its GitHub URL when one exists ([#42](https://github.com/<owner>/<repo>/pull/42)), none otherwise.
• Diff: plain text, no colour, no emoji. Format: <files> files · +<additions> · −<deletions>, with middle dots (·, U+00B7) as separators. Use the Unicode minus sign (U+2212, −) for the deletions count, not the ASCII hyphen-minus, for typographic correctness.

When the severity-filtered set is empty, print No Critical/High/Medium findings. in place of the severity tables; the metadata table and Design section still print.

Full report template:

# PR Review

## develop (vs main)

| Field     | Value                                                                        |
| --------- | ---------------------------------------------------------------------------- |
| Branch    | [`develop`](https://github.com/owner/repo/tree/develop) → [`main`](https://github.com/owner/repo/tree/main) |
| HEAD      | [`head7`](https://github.com/owner/repo/commit/head7)                        |
| PR        | [#42](https://github.com/owner/repo/pull/42) (or `none` when no PR is open)  |
| Generated | YYYY-MM-DD HH:MM                                                              |
| Diff      | 6 files · +390 · −0                                                          |
| Findings  | 🔴 0 Critical · 🟠 1 High · 🟡 6 Medium                                      |

### Design

- <Qualitative observation about scope, placement, or complexity>

### Critical

| File                                                                                                          | Description                |
| ------------------------------------------------------------------------------------------------------------- | -------------------------- |
| [/rebuild_llama.cpp.ps1:41-43](https://github.com/owner/repo/pull/42/files#diff-a1b2c3d4e5f6a7b8c9d0R41-R43)  | Destroys tree on running proc |

### High

| File                                                                                                       | Description               |
| ---------------------------------------------------------------------------------------------------------- | ------------------------- |
| [/examples/server.ps1:14-19](https://github.com/owner/repo/pull/42/files#diff-f6a7b8c9d0a1b2c3d4e5R14-R19) | Unhandled startup failure |

Constraints

• Do NOT modify source code.
• Do NOT post comments, reviews, or any data to GitHub. This skill is advisory with respect to GitHub; it never writes there.
• This skill is read-only: do NOT write any files. The report is printed in the terminal only. Do not touch .claude/, the repo root, or vendor/.
• Do NOT report pure style or formatting preferences.
• Do NOT report issues in generated files, lock files, or changelog entries.
• Do NOT report findings inside vendor/.
• File cells in output tables must use markdown link syntax with the file path as label; see the Link Format section.

Source: countzero/windows_llama.cpp — distributed by TomeVault.