# Review Fixes Native Image Run Fail > Review pull requests with the `fixes-native-image-run-fail` label in graalvm-reachability-metadata. Use when asked to review or triage a PR that fixes a native-image runtime failure for an existing library version update. Focus on validating the runtime/native fix, keeping the diff scoped, and applying the repair coverage gate to overall and breakdown dynamic-access reports. - Skill: `oracle/review-fixes-native-image-run-fail` (Agent Skill) - Install (CLI): `npx skillmds@latest add oracle/review-fixes-native-image-run-fail` - Raw SKILL.md: https://api.skillmd.com/api/skills/oracle/review-fixes-native-image-run-fail/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: Oracle (https://skillmd.com/u/oracle) - Updated: 2026-09-10 - Page: https://skillmd.com/skills/oracle/review-fixes-native-image-run-fail --- # Review `fixes-native-image-run-fail` PRs These PRs repair metadata or tests after an existing library version compiles and builds a native image, but the native image fails when the test runs. Review them more lightly than `library-new-request` PRs: the library is already supported, so the goal is to restore native-image behavior for a newer version while preserving existing dynamic-access coverage. The rules below implement the contribution contract: block only on a concrete violation of a must — the shape limits (§FS-contribution-contract.2), the coverage gates (§FS-contribution-contract.3), the cheating patterns (§FS-contribution-contract.4), and the test contract's musts (§FS-test-contract) — while shoulds stay advisory in review (§FS-contribution-contract.1). ## Chunked dynamic-access PRs When the PR has the `chunked-dynamic-access` label, skip every repair coverage comparison in this skill, including for the final chunk. Request changes for coverage only when the reported `dynamicAccess.coveredCalls` is `0`. Keep every non-coverage rule. §FS-contribution-contract.3 The PR number or URL can be passed as an optional argument (for example, `1234`, `https://github.com/oracle/graalvm-reachability-metadata/pull/1234`). If the user says "review this PR" without an argument, infer the PR from the surrounding conversation or `gh pr status`; only ask the user when it cannot be inferred. Use `gh pr view `, `gh pr diff `, and `gh pr checks ` against the resolved PR throughout the workflow below. ## Review Principles - Confirm the PR has label `fixes-native-image-run-fail`. - Expect targeted metadata, resource, proxy, serialization, JNI, initialization, or test adjustments that make the native executable run successfully for the new version. - Be more relaxed than `library-new-request`: do not reject only because a test is inherited from older support, uses compatibility branches, or keeps an existing package layout. - Do not accept fixes that hide the failing native path by skipping assertions, skipping native-image runtime execution, or weakening the test until the failure disappears. - Treat dynamic-access coverage preservation as the main quality gate. Apply the ordered repair comparison in the workflow below independently to the overall report and every breakdown present in either version. - Use the reported stats evidence as-is. Do not inspect generation filters, agent configuration, or metadata contents to second-guess the reported `totalCalls`, `coveredCalls`, or coverage ratios. - Treat metadata entry counts as telemetry only. Never request changes or apply `human-intervention` because the new version reports fewer metadata entries. - A new version that reports zero total dynamic-access calls has no comparable call surface and passes the numeric coverage gate. The native-image runtime fix itself must still pass and must not be made green by skipping the native path or disabling native-image behavior. - Accept only `reachability-metadata.json` files as metadata files. Reject legacy native-image metadata config files such as `reflect-config.json`, `resource-config.json`, `proxy-config.json`, `serialization-config.json`, `jni-config.json`, or `predefined-classes-config.json`. - Prefer concrete evidence from native run output, generated metadata, stats, and CI over style objections. ## Workflow 1. Inspect the PR summary. - Resolve the target PR from the optional argument, or infer it from context when possible. - Confirm the PR has label `fixes-native-image-run-fail`. - Identify the target coordinate, the previous tested version, and the new tested version from the PR body, title, changed `index.json`, metadata path, and test path. - Gather files, reviews, inline comments, and CI checks. 2. Validate the diff scope. - Expected files are usually limited to: - `metadata////reachability-metadata.json` - `metadata///index.json` - `stats////stats.json` - `tests/src////**` - allowed Docker image entries or test resources only when the native test requires them - Treat generated test project files such as `.gitignore`, `build.gradle`, `gradle.properties`, `settings.gradle`, and `user-code-filter.json` as normal when they live under the target version's test directory. - Accept metadata additions that are necessary for the new upstream version. - Accept narrow test edits that keep the same behavior covered across old and new versions. - Be suspicious of unrelated build logic, workflows, generated sources, other libraries, or broad refactors. - Reject legacy native-image metadata config files. Metadata for generated support and test-only metadata must use `reachability-metadata.json`. - Reject or request changes if the PR fixes the native run by disabling the failing behavior. 3. Review the native-image fix. - Confirm the metadata or test change matches the observed native runtime failure, such as missing reflection, resources, proxies, serialization constructors, JNI access, or class initialization behavior. - Metadata additions should be specific enough to the target library behavior; do not require hand-minimized entries when generated metadata is coherent and validation passes. - Test changes are acceptable when the upstream API or runtime behavior changed, but they must still exercise the native path that previously failed. - Do not require the stricter `library-new-request` rules about scaffold-only tests or test package placement unless the PR is also adding a new library. 4. Check dynamic-access coverage across versions. - Compare `stats////stats.json` and `stats////stats.json` when stats are present in the PR or available on the branch. - Apply the following ordered comparison independently to the overall `dynamicAccess` report and every `dynamicAccess.breakdown` scope present in either version. Treat `N/A` on either side as `0/0`: zero `totalCalls` and zero `coveredCalls`. 1. Pass when the new scope reports `totalCalls == 0`. 2. Fail when the new scope reports `totalCalls > 0` and `coveredCalls == 0`. 3. Pass when new `coveredCalls` is at least old `coveredCalls`, or when the new coverage percentage is at least the old percentage. 4. When both scopes report fewer than 10 `totalCalls`, fail only when old `coveredCalls - new coveredCalls > 2`. 5. Otherwise, fail only when coverage drops by more than 20 percentage points. - Reach step 4 only after both covered calls and percentage decreased, so the subtraction is always positive. A failing scope is blocking unless the PR gives a concrete, credible explanation of the changed upstream surface. - For every failing scope, inspect the old and new stats and the test diff, then relevant upstream API/runtime or CI evidence, before deciding. State the evidence-backed cause. If the evidence does not establish one, say `Cause not established from the available evidence` and ask for an explanation. Never guess or present speculation as the cause. - For reference: `8/8 -> 7/9` and `8/8 -> 6/9` pass the small-report rule; `8/8 -> 5/9` fails it; `10/10 -> 15/20` passes because covered calls grew; and `20/20 -> 15/20` fails the large-report rule. - Do not compare metadata entry counts; they are telemetry, not a review gate. - If required old or new stats are missing or stale, ask for `generateLibraryStats` or the relevant CI stats job before approving. 5. Check CI before deciding. - Expected minimum: native-image compile and native test execution are green for the target coordinate. - Metadata validation and Java tests should also pass for the changed coordinate. - If current-defaults and future-defaults lanes both run, both should pass unless the PR clearly targets only one failing lane and the other failure is unrelated infrastructure noise. - If CI is flaky but the diff and coverage comparison are sound, ask for a rerun instead of blocking on speculation. ## Disposition Finding a violation is half the review. **Read §FS-contribution-contract.5 and take the disposition it assigns** — `grund FS-contribution-contract.5 --full`, or section 5 of `docs/functional-spec/contribution-contract.md` in the checkout. That section is authoritative and this table is only a dispatch: it tells you which case you are in, not what the case permits. You are the party that acts, because the PR was generated and no author is waiting to answer a review comment. | Case | When | What you do | |---|---|---| | 5.1 | The violation is in the PR's own `metadata/`, `stats/`, and `tests/src/` files | Repair it on the PR head, push, re-run the affected checks, and say what you changed and why | | 5.2 | The only possible fix is build logic, harness code, workflows, or another coordinate | Do not repair there; revert such a change if the PR already carries one, then take 5.3 | | 5.3 | The cause is a defect in shared repository code | Find or open one issue for the defect, comment that the PR is blocked on it rather than on its own content, then take 5.5 | | 5.4 | The library's dynamic access is reachable only through behavior Native Image cannot support | Close the PR, label the linked issue `library-unsupported-version`, and close that issue | | 5.5 | Anything else, uncertainty included | Label the PR `human-intervention` and comment with the rule at issue, what you tried, and what you could not decide | Never close a PR under any case but 5.4. Read §FS-contribution-contract.5.1 before you repair: it bounds what a repair may do, and it carries the one exception that lets you drop a test scenario rather than fix it. ## Decision Rules These decide the verdict. **Disposition** above decides what you then do with it: a violation you can correct inside the PR's own files is repaired and pushed, not requested from an author who is not there. Approve when all of these are true: - The PR is scoped to the target existing library and the native-image runtime failure it fixes. - The native-image runtime path still executes meaningful library behavior. - The overall report and every breakdown pass the ordered repair coverage gate, or any failing scope is convincingly explained by a changed upstream surface. - Required metadata, Java, native-image compile, and native-image run checks are green. Request changes when any of these are true: - The fix makes native execution pass by skipping the failing native path, disabling assertions, or removing coverage. - The overall report or a breakdown fails the ordered repair coverage gate without a credible explanation and replacement coverage. - Metadata additions are unrelated to the failure or affect other libraries without justification. - CI failures indicate the native-image runtime problem is not actually fixed. Ask for follow-up instead of rejecting when: - Stats needed for the old/new version comparison are missing or stale. - CI failed in a way that looks like infrastructure noise. A transient external failure is waited out; a reproducible defect in the repository's own code is case 5.3 of **Disposition** (§FS-contribution-contract.5.3). - A failing repair-coverage scope may reflect a plausible upstream API or runtime change, but the PR does not explain it. ## Output Style Keep comments short and factual: - For coverage failures: name the failing overall or breakdown scope, report old and new `coveredCalls`, `totalCalls`, and percentages, and identify whether it failed the zero-covered, small-report, or large-report rule. Ask for restored coverage or a concrete explanation. Explain why the regression happened using evidence from the stats, test diff, upstream changes, or CI. If the cause cannot be established, say so explicitly and request an explanation; never guess. Do not report a change that an earlier step of the ordered gate accepts. - For native skips: say that the PR avoids the failing native path instead of fixing it, so it does not demonstrate native-image runtime coverage. - For unrelated changes: say the PR should stay scoped to the `fixes-native-image-run-fail` repair and remove unrelated files. - For legacy metadata files: say that metadata must use `reachability-metadata.json` and ask for old config files such as `reflect-config.json` or `resource-config.json` to be replaced. - For missing stats: ask for regenerated library stats or CI evidence before approval.