XQUIC PR Pre-Review
Review the published pull request's exact base-to-head diff and write the
result to
~/build/harness/pr-review-<number>/pr-review-<number>.md. Require a GitHub
PR number and URL; do not substitute a local branch candidate. Never stage or
commit the report or other local review artifacts.
Local Review Workspace
- Expand
~to the current user's home directory and use exactly~/build/harness/pr-review-<number>/as the review directory. Do not place the workspace under the repository checkout. - Use
pr-review-<number>.mdas the single current retrospective. Before rerunning a changed PR head, read the existing retrospective and local artifacts as hypotheses and reproduction inputs, then rewrite the current retrospective for the new head. - Keep every reviewer-created abnormal-case source file, patch, build or run script, peer input, packet, compiler output, and log inside the same review directory. Use descriptive filenames and preserve reusable abnormal-case code across iterations.
- Do not edit the submitted PR diff to construct a review case. Build or run reviewer-created code from the local review directory against the exact published PR head.
- Index every retained artifact in the retrospective with its relative path, purpose, reviewed head, exact reproduction command, and result. Mark stale artifacts explicitly; never treat a prior-head result as current evidence.
- Keep the entire review directory local and uncommitted. It is both the output of the current review and an input to the next PR iteration.
Inputs and Gate
- Fetch the pull request from GitHub. Record its number, URL, base commit, published head commit, changed files, draft state, and review time.
- Confirm the fetched head matches the PR's published head, then read the changed implementation, callers, tests, architecture map, and adjacent module documentation. Treat the PR and issue descriptions as context, not evidence.
- For protocol behavior, retrieve the authoritative text from RFC Editor or the IETF Datatracker. Pin an Internet-Draft to the revision implemented by the repository. Record the document URL, exact section, normative keyword, and a source excerpt of no more than 25 words. Do not rely on a quotation copied into the PR or issue.
- Use
pass,fail,inconclusive, ornot-applicablefor each section. Usenot-applicableonly with a concrete scope reason. - Set
pre_review_result: trueonly when every section ispassor justifiednot-applicable. Afailorinconclusiveresult is a closed gate. - Invalidate and rerun the report whenever the published PR head changes.
If the official specification, reviewed source, or current-head validation
evidence cannot be inspected, return inconclusive; never infer a pass.
1. RFC Conformance
- Derive the affected protocol mechanism from the code and wire behavior.
- Compare every changed state transition, limit, role restriction, error code, and peer-visible result with the authoritative section.
- Trace peer input from parsing through validation, state mutation, and error handling. Cite current file and line evidence.
- Check interactions with related RFC requirements and negotiated extensions, not only the paragraph cited by the PR.
- Check verified RFC errata or pinned-draft changes that alter the cited mechanism.
- Fail on a semantic deviation unless the RFC permits the implementation choice and the report explains that allowance.
2. Validation Coverage
- Inspect committed tests rather than trusting the PR summary.
- Require a positive and negative CUnit case tied to each changed production path.
- For endpoint-visible behavior, require positive and abnormal
client-to-server coverage or an explicit case-test gap. When
scripts/case_test.shcases are added, require distinct IDs and matching selectors intests/test_client.candtests/test_server.c. - Confirm any executed end-to-end cases prove the endpoint-visible result, not merely a process exit or generic failure.
- Verify current-head local evidence shows the complete CUnit suite passed and any accepted targeted case evidence or gaps are current. Fail when required unit coverage or current-head evidence is missing.
3. Adversarial Bad Case
- Construct the smallest peer-controlled input or event sequence that could bypass the RFC guard. Consider boundary values, duplicate or reordered frames, wrong endpoint roles, invalid state transitions, truncated or non-canonical encodings, retransmission, 0-RTT, and callback reentrancy when relevant.
- Trace the bad case through the actual parser and state machine. Attempt an executable focused unit or client-to-server reproduction when the local environment permits it.
- Keep all abnormal-case code and supporting artifacts in the fixed local review directory; do not modify the submitted diff.
- Fail when the bad case bypasses the intended restriction or exposes an
unhandled state. Return
inconclusivewhen a material bad case cannot be executed or ruled out.
4. Performance and Compiler Optimization
- Identify whether the change is on a packet, frame, stream, request, timer, or connection hot path and estimate its invocation frequency.
- Compare the time complexity and critical-path work with viable alternatives. Check repeated scans, avoidable parsing, allocations, copies, branches, cache-unfriendly access, locking, and contention.
- Inspect whether types, aliasing, control flow, data layout, and function boundaries allow the configured compiler to inline, fold constants, eliminate dead work, and vectorize where useful.
- Read the actual release build flags before assuming an optimization is enabled.
- Use optimized-build assembly, compiler remarks, profiles, or benchmarks for material performance claims when practical. Do not call an implementation optimal without evidence.
- Fail on an avoidable asymptotic regression or a material hot-path slowdown without a correctness requirement or measured tradeoff.
5. Memory Safety and Footprint
- Build an ownership and lifetime map for every touched allocation, buffer, pointer, list node, callback context, and borrowed reference.
- Check success, error, timeout, close, retry, and reentrant callback paths for leaks, double frees, use-after-free, stale aliases, null dereferences, integer overflow, out-of-bounds access, and length/capacity mismatches.
- Quantify new long-lived bytes per connection, stream, request, and path. Project server cost at realistic concurrency and check for unbounded or peer-controlled retention.
- Use sanitizer, static-analysis, or focused stress evidence when the changed ownership or bounds cannot be proved by inspection.
- Prefer execution efficiency when memory footprint and speed conflict, but record the alternatives, concurrency cost, measured or reasoned speed benefit, and why the retained memory is bounded and acceptable.
Report Format
Write exactly five numbered review sections after the metadata:
---
reviewed_base: <commit>
reviewed_head: <commit>
review_target: <PR URL>
review_time_utc: <timestamp>
review_artifact_dir: <expanded absolute review directory>
review_artifacts: <relative paths and reviewed heads, or none>
pre_review_result: <true|false>
---
# XQUIC PR Pre-Review
## 1. RFC Conformance
Status: <pass|fail|inconclusive|not-applicable>
Sources: <official document, section, URL>
Evidence: <source/tests and short authoritative excerpt>
Findings: <blocking and non-blocking findings, or none>
Conclusion: <reasoned conclusion>
## 2. Validation Coverage
Status: <...>
Evidence: <positive/negative unit and end-to-end case IDs/results>
Findings: <...>
Conclusion: <...>
## 3. Adversarial Bad Case
Status: <...>
Attempt: <constructed input or sequence and execution result>
Artifacts: <relative paths, reviewed heads, commands, and results, or none>
Findings: <...>
Conclusion: <...>
## 4. Performance and Compiler Optimization
Status: <...>
Evidence: <hot path, complexity, compiler and measurement evidence>
Findings: <...>
Conclusion: <...>
## 5. Memory Safety and Footprint
Status: <...>
Evidence: <ownership, bounds, per-scope footprint and concurrency estimate>
Tradeoff: <memory versus execution-efficiency decision>
Findings: <...>
Conclusion: <...>
For every blocking finding, cite the reviewed file and line, explain the failure mode, and state the minimum condition for the gate to pass. Do not fix the PR while performing pre-review unless the user separately requests it.