name: frb-fix-ci description: Use when CI fails in flutter_rust_bridge - before deep investigation
FRB Fix CI
Note: Check your user-level
remote-testingrules before running commands. Tests and codegen may require remote execution. Note: When reproducing FRB failures remotely, do not run multiple remote commands in parallel against the same workspace/container. Serialize them unless you have isolated workspaces.
Overview
CI failures in flutter_rust_bridge often have simple fixes. Try the appropriate approach below before deep investigation.
Core principle: Start with lazy fixes (re-run, copy diff, --fix) before expensive investigation.
If the CI failure or PR state is caused by merge/rebase conflicts, especially conflicts in generated files after master and a PR both added tests or APIs, read the /frb-fix-merge-conflict skill before resolving them.
Triage Order
Use this order before diving into individual failure types:
- Check the latest relevant run or job first. Do not reason from stale CI state.
- If the failure looks flaky, rerun only the failed jobs.
- Reproduce the exact failing
./frb_internal ...command from CI, but first check your user-levelremote-testingrules instead of assuming local execution is correct. If reproducing remotely, keep one remote workspace single-writer: do not run multiple FRB reproduction commands in parallel against that same workspace. Also do not trust a dirty remote workspace by default: FRB remote runs often leave stray generated or untracked files behind, so prefergit reset --hard HEAD && git clean -fdxbefore a meaningful reproduction. - Decide where the failure sits in the dependency graph: is it more likely a prerequisite cause (
Generate,Integrate, or a high-relevanceGenerate Internalstage) or a downstream symptom (Build :: Flutter, native tests)? - Only do deeper debugging after you have ruled out flakes, stale runs, and failure propagation.
Checking the Right Run
Do not answer from stale CI state. Read the latest relevant run or job information first.
Temporary CI Narrowing
When CI feedback is slow and only one job family matters for the current investigation, read frb-ci-filter before using ci_filter, the ci-manual-dispatch label, or related workflow configuration.
Quick Reference
| Symptom | Fix |
|---|---|
| Flaky test (passes sometimes) | gh run rerun --failed |
| Git diff shown in CI | git apply OR regenerate |
| Lint/format errors | Add --fix flag |
| Can't reproduce locally | Use same ./frb_internal command from CI, following remote-testing rules |
Dependency Order
When several related jobs are failing, use this dependency graph instead of treating all failures as peers:
Legend: rectangles are files or directories. Ovals are CI operations. In each oval, the first line is the CI job name and the second line in parentheses is the corresponding ./frb_internal ... command when there is one.
flowchart LR
CodegenSources["frb_codegen/src/** + codegen config"]
Versions["pinned Flutter / Dart / Rust versions"]
Templates["frb_codegen/assets/integration_template/**"]
Cargokit["cargokit"]
PureDart["frb_example/pure_dart/**"]
GeneratedOutputs["frb_example/**/frb_generated.*"]
PureDartGenerated["frb_example/pure_dart/**/frb_generated.*"]
ExampleOutputs["integrate outputs under frb_example/**"]
PureDartPde["frb_example/pure_dart_pde/**"]
PureDartPdeGenerated["frb_example/pure_dart_pde/**/frb_generated.*"]
NativeTests(["Test :: Flutter :: Native::*\n(./frb_internal test-flutter-native ...)"])
Generate(["Generate :: FRB Codegen :: Command Generate\n(./frb_internal generate-run-frb-codegen-command-generate)"])
Integrate(["Generate :: FRB Codegen :: Command Integrate\n(./frb_internal generate-run-frb-codegen-command-integrate)"])
Build(["Build :: Flutter\n(./frb_internal build-flutter ...)"])
subgraph GenerateInternal["Generate Internal stages"]
GenerateInternalPureDart(["Generate Internal :: frb_example pure_dart chain\n(./frb_internal generate-internal-frb-example-pure-dart)"])
GenerateInternalRust(["Generate Internal :: Rust\n(./frb_internal generate-internal-rust)"])
GenerateInternalOther(["Other internal stages\n(book-help / dart-source /\nbuild-runner / contributor / readme)"])
end
CodegenSources -->|used by| Generate
Versions -->|used by| Generate
Generate -->|writes| GeneratedOutputs
CodegenSources -->|used by| GenerateInternalPureDart
Versions -->|used by| GenerateInternalPureDart
PureDart -->|read by| GenerateInternalPureDart
GenerateInternalPureDart -->|rewrites| PureDart
GenerateInternalPureDart -->|derives| PureDartPde
GenerateInternalPureDart -->|writes| PureDartGenerated
GenerateInternalPureDart -->|writes| PureDartPdeGenerated
CodegenSources -->|used by| GenerateInternalRust
Versions -->|used by| GenerateInternalRust
Versions -->|used by| GenerateInternalOther
CodegenSources -->|used by| Integrate
Versions -->|used by| Integrate
Templates -->|used by| Integrate
Cargokit -->|used by| Integrate
Integrate -->|writes| ExampleOutputs
GeneratedOutputs -->|consumed by| Build
GeneratedOutputs -->|consumed by| NativeTests
PureDartGenerated -->|consumed by| Build
PureDartPdeGenerated -->|consumed by| Build
ExampleOutputs -->|consumed by| Build
ExampleOutputs -->|consumed by| NativeTests
How to Read
Read the graph as artifact and input dependencies, not as a literal GitHub Actions job graph.
Key Chains
frb_codegen/assets/integration_template/+cargokit-> integrate outputs underfrb_example/**If Flutter integrate examples, example platform files,Build :: Flutter, and native Flutter tests regress together, suspect these template inputs first. Do not hand-edit generated example outputs. If the bug is actually inside the embeddedcargokitsubmodule, it is acceptable to edit that submodule directly and push tofzyzcjy/cargokit, then update the submodule ref.- Apple scaffold under integrate examples:
when Linux-side raw
create/integratedoes not reproduce checked-in Apple files such as.metadataiOS stanzas,ios/**,example/ios/**,macos/Podfile, or Apple-specificpubspec.yamlfragments, do not treat that as proof the checked-in files are wrong. In this repo, checked-in mac-generated Apple scaffold may be the source of truth, and Linux integrate may explicitly apply it before diff comparison. Generate Internal+frb_example/pure_dart/**->frb_example/pure_dart_pde/**Ifpure_dart_pdeis failing, do not only refreshpure_dart_pde. First check whether./frb_internal generate-internal-frb-example-pure-dart --set-exit-if-changed ...is still changingfrb_example/pure_dart.
When to Consult
Use this graph when several nearby categories start failing together in the same run, especially when earlier nodes such as Generate, Integrate, or the high-relevance Generate Internal stages are already red and later failures look consistent with missing, stale, or mismatched generated files or platform files.
Rule
Prefer fixing prerequisite nodes before symptom nodes. If a prerequisite node is still unstable, treat later failures as propagated symptoms until proven otherwise. If this pattern keeps repeating across multiple commits or CI runs, jump to Whack-a-Mole Prevention.
Whack-a-Mole Prevention
This section is about history across multiple commits or CI runs, not a single failing job.
Use it when the same area keeps becoming green and then red again, especially with commits like refresh, regenerate, sync, and revert.
What this usually means:
- You may be chasing generated outputs instead of fixing their source inputs
- A prerequisite node on the dependency graph is still unstable
- A temporary green run may only mean one symptom was suppressed, not that the root cause was fixed
What to do:
- Stop adding more generated-output-only sync commits by default
- Go back to the dependency graph and identify the unstable source of truth first:
frb_codegen/src/**,frb_codegen/assets/integration_template/**, pinned Flutter/Dart/Rust versions, generation order, or package relationship - Only accept regenerated outputs after that source node is stable in a clean matching environment
- If you have already accepted two package-level
Generatesync commits with very similar generated Dart drift in one session or across adjacent CI runs, stop the package-by-package loop and escalate to a clean remote./frb_internal precommit-generate
Common FRB patterns:
- Flutter integrate examples:
suspect
frb_codegen/assets/integration_template/andcargokit; do not hand-edit generated example outputs. If the real fix belongs incargokit, patch the submodule and update its ref instead of only patching copied outputs. pure_dartandpure_dart_pde: if both are moving, stabilizefrb_example/pure_dartfirst and treatpure_dart_pdeas a dependent output- Repeated
Generatefailures across different example packages: if the sameGenerate :: FRB Codegen :: Command Generatesymptom keeps rotating across packages with similar generated Dart diffs, stop fixing one package at a time. Re-run clean remote./frb_internal precommit-generateand treat that full run as the source of truth for the wholeGeneratechain. Only after that full run stabilizes should you accept any remaining per-package tail diff. Practical cutoff: after two similar package-level sync fixes, the burden of proof switches. Do not add a third similar package sync until you have checked whether clean remoteprecommit-generatecollapses the whole diff surface.
Escalation Protocol for Repeated Generate Drift
Trigger this protocol if both are true:
- Two or more example packages fail in the same
Generate :: FRB Codegen :: Command Generatefamily across the current session or adjacent CI runs - The diffs are mostly generated Dart outputs such as
frb_generated.dart,frb_generated.io.dart,frb_generated.web.dart, generated API wrapper files, or nearby formatting/layout churn, with no evidence yet of a package-specific semantic bug or hand-written file change
Then do this, in order:
- Stop accepting more package-by-package sync commits
- Run clean remote
./frb_internal precommit-generate - Pull back the remote workspace
- Inspect the full diff surface
- If the full diff surface collapses to one small tail package, accept that tail from this same clean run
- If the diff surface is still broad, treat
precommit-generateas the authoritativeGeneratesource and debug the shared generation workflow before touching more package-level symptoms - Do not resume package-by-package syncs unless a clearly different failure class appears
Example:
- If
dart_minimal, thenrust_ui_counter--ui, thengalleryfail in separateGeneratejobs with similar generated Dart drift, stop after the second one and escalate to clean remoteprecommit-generate
Fixes by Failure Type
Flaky Test
Sometimes CI fails due to timing issues, not real bugs. Rerun only failed jobs:
gh run rerun --failed
If it passes on retry -> flaky, not your bug.
Git Diff Errors
When CI shows a diff, you have two options:
Option A: git apply (faster)
CI already ran the generator. Just apply what it computed:
# Copy the diff from CI, then:
pbpaste | git apply # macOS
Option B: Regenerate (slower but more "proper")
./frb_internal precommit-generate
After codegen: Check your user-level
remote-testingrules. If codegen was run remotely, pull changes back to local.
Both are correct. Option A is faster; Option B is more thorough.
Do not hand-edit generated files as the final fix.
You may use CI diffs only as a diagnosis aid to understand what changed, but the final accepted output should come from re-running the appropriate generation workflow in a clean matching environment.
Can't Reproduce Locally
CI shows the command it ran. Before running it, check your user-level remote-testing rules to determine whether this repo requires remote execution.
Before reproducing, make sure the toolchain versions match CI closely enough to be meaningful. In practice this usually means Flutter, Dart, Rust, cargo subcommands, and any pinned template or helper dependency should match the versions used by CI.
If reproducing remotely, clean the remote workspace first unless you have a specific reason not to. In FRB this often matters because old generated files and untracked outputs can survive earlier runs and create fake diffs:
git reset --hard HEAD && git clean -fdx
Then run the same command:
# CI shows: ./frb_internal test-dart --package frb_example/pure_dart
./frb_internal test-dart --package frb_example/pure_dart
Lint/Format Errors
For clippy, dart analyze, or format errors, use --fix:
./frb_internal lint --fix
This runs:
cargo clippy --fix- Rust lint fixescargo fmt- Rust formatdart format- Dart formatdart fix --apply- Dart auto fixes
When lint/format failures happen on generated files, do not default to hand-editing those files just to match formatter output.
Instead:
- Compare the formatter inputs across branch head, PR merge ref, and remote workspace
- Check whether generation order, hidden generation steps, or toolchain/environment drift changed the generated file before formatter ran
- Only accept formatter output after confirming the pre-format generated input is actually the correct one
Command Integrate Failures
When Generate :: FRB Codegen :: Command Integrate fails because integrated output is wrong, do not hand-edit the generated integrate example outputs.
Instead:
- Fix the source templates under
frb_codegen/assets/integration_template/ - If the bad logic is inside the embedded
cargokitsubmodule, fix and pushfzyzcjy/cargokit, then update the submodule ref - Re-run integrate generation after updating the templates
Special case: Apple scaffold
- If Linux-side raw
create/integratedrops checked-in Apple files, do not describe the current fix as "restoring old scaffold" unless it really is an accidental workaround. - First check whether the repo intentionally treats checked-in mac-generated Apple scaffold as source of truth.
- If that is the contract, prefer names and comments like "apply checked-in Apple scaffold source of truth" over vague restore/preserve wording.
Generate-caused Failures
If CI previously failed mainly in Generate while other jobs passed, and after accepting generated changes additional non-Generate jobs start failing, treat this as strong evidence that the accepted generated outputs are incorrect or incomplete.
In that situation:
- Do not continue fixing downstream failures one by one first
- First validate the generation logic or generation workflow
- Re-generate from a clean environment
- Only accept generated outputs after confirming they do not introduce new non-
Generateregressions
If the symptom is not one package but a sequence of packages failing with very similar generated Dart drift, prefer validating clean remote ./frb_internal precommit-generate before accepting more package-by-package sync commits.
If clean remote precommit-generate reduces the remaining diff surface to one small tail package, accept that tail from the same clean run and rerun CI. Do not go back into a long per-package loop unless a new failure class appears.
Dart Web Browser Startup Flakes
When Test :: Dart :: Web (...) fails after the web build and server startup already succeeded, and the failure is:
Exception: Websocket url not found.
treat it as a likely browser / puppeteer startup flake first, not an immediate code regression.
In that situation:
- Do not assume the generated code or test logic is broken just from this error
- Check whether wasm build,
dart compile js, and local web server startup already succeeded - Prefer rerunning only the failed job first
- Only start code investigation if the same job keeps failing with the same error repeatedly
Common Mistakes
- Investigating root cause when a simple re-run would work
- Not trying
git applyfirst when CI provides a diff - Fixing many new downstream test/build failures one by one after accepting generated changes, when CI previously failed mainly in
Generate - Hand-editing generated files to chase CI formatter output before checking whether CI, merge ref, and remote environments are formatting the same input
- Hand-editing integrate-generated example outputs instead of fixing
frb_codegen/assets/integration_template/ - Assuming
cargokitsubmodule changes are off-limits when the real bug is there - Chasing repeated
refresh/regenerate/syncdiffs without re-checking the upstream generation inputs - Continuing package-by-package
Generatesync commits after two similar generated Dart drifts, instead of escalating to clean remoteprecommit-generate - Fixing downstream build/test jobs before upstream generate/integrate/high-relevance generate-internal stages are stable
- Answering from stale CI state instead of reading the latest relevant run or job information first
- Forgetting to remove
ci-manual-dispatchand rerun normal CI before treating the PR as ready
Related Skills
frb-code-generation- Which generation commands to runfrb-debugging- Deep debugging when simple fixes don't work