[anilcancakir]

Summary

Two defects from a field test of app-local connected commands (Watchools player:*, which call ext.watchools.player from bin/dispatcher.dart):

1. A connected command run while the app was still booting printed Connected command player:state failed: ext.watchools.player: (-32601) Method not found plus a vm_service stack (exit 3), then worked a few seconds later. A consumer cannot catch this itself without importing package:vm_service, which trips depend_on_referenced_packages.
2. start --device macos on a fresh worktree: the first native build (CocoaPods, SwiftPM) outlasted the 90 s scrape, the timeout's StateError escaped as an unexpected error whose visible tail was <asynchronous suspension>, and the session was left booting: true with no URI while the app kept building.

Changes:

• VmServiceClient.callServiceExtension maps JSON-RPC -32601 to a new public ExtensionNotRegisteredException, on the first call and on both retries (collected promise, refreshed isolate after a hot restart). ArtisanApplication prints it as one line and exits 75.
• ArtisanCommand.exitTryAgain = 75 (EX_TEMPFAIL): "not ready yet, retry".
• start, plain branch: the scrape throws TimeoutException. With flutter run alive, start keeps the booting session, prints flutter run (pid <n>) is still building after <n>s with the status / retry / --timeout hints and exits 75. With it gone, it stops the FIFO holder, removes the session (the log stays), names the log and exits 1.
• A connected command against a booting session with no URI yet exits 75 while the recorded pid is alive (was 1), so a caller can retry on one code from start to the first command that connects.
• --cdp-port branch: behaviour unchanged (fail + reap, exit 1); the message drops the TimeoutException after <duration> prefix.
• Docs: doc/commands/start.md, skills/.../state-and-recovery.md, mcp-tools.md, SKILL.md, CHANGELOG under [Unreleased].

Left as is: a permanent -32601 (a typo in the method, or a plugin newer than the app's copy) also exits 75 with "retry"; telling it apart would need an extensionRPCs lookup per failure. A web runtime that reports a missing extension with a code other than -32601 (older DWDS) keeps the old output.

Verification

dart test: 1432 passed; dart analyze lib/ test/ bin/ and dart format clean; line coverage 87.78%; dart pub publish --dry-run 0 warnings. Each new test failed first. The extension tests run against a fake JSON-RPC WebSocket server; a real Dart VM with no extension registered printed the one line and exited 75. The cold-build path is covered through the start seams; it has not been run against a real multi-minute macOS build.

[kodizm]

Note

Kodizm (AI-generated). May contain mistakes; verify before acting.

The change looks correct to me, and I have no findings; the one gap is that the repo's own CI workflow has not reported on this commit.

Why I think it is correct:

• -32601 is mapped inside _invokeExtension (lib/src/vm/vm_service_client.dart:121), so the first call and both retries all raise ExtensionNotRegisteredException. The exception is not a vm.RPCError, so it does not wrongly trigger the -32603 retry.
• Callers keep working. The MCP plugin dispatch (lib/src/mcp/mcp_server.dart:296) catches everything and prints $e, which now gives the one-line message. dusk_evaluate goes through evaluate, not this path, so it is unaffected.
• When flutter run has died (lib/src/commands/start_command.dart:574), StateFile.delete() removes only the session's state.json and the legacy pointer. flutter-dev.log stays, as the message promises.
• Docs, skill references and CHANGELOG [Unreleased] are updated, as Golden Rules 1, 2 and 5 in CLAUDE.md require.
• The new type is a final class, matching the conventions in CLAUDE.md.

Tests

New tests cover every new path:

• The -32601 mapping on the first call and on the retries, against a fake WebSocket server.
• The dispatcher printing one line and exiting 75, for both an unregistered extension and a live booting session.
• start timing out with flutter run alive (exit 75) and gone (exit 1, session reaped), through the startIsAlive seam.
• The --cdp-port message no longer carrying the TimeoutException after prefix.

CI

• Auto-merge low-risk Dependabot PRs: skipped. This is the only check reported for 3b50620, so the overall: success it produces is not a real pass.
• CI / Lint & Test (format, analyze, tests, 80% coverage floor, publish dry-run): no result reported for this commit, although .github/workflows/ci.yml triggers on pull requests into master. Treat this review as static until that run shows up and passes.

[codecov]

Codecov Report

❌ Patch coverage is 94.11765% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
lib/src/vm/extension_not_registered_exception.dart	50.00%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[kodizm]

Note

Kodizm (AI-generated). May contain mistakes; verify before acting.

The only new commit merges master into this branch, and the merge looks correct. I have no new findings.

The new commit is 100cdd0, a merge of origin/master. It brings in the 0.0.19 release (#66), the bump to xml 7 (#64) and cupertino_icons 2.0.0 for the example app (#65). Those changes were reviewed in their own pull requests, so I did not review them again here. The one hand-resolved conflict was in CHANGELOG.md, and it is resolved correctly: this PR's Added and Fixed entries stay under [Unreleased], above ## [0.0.19] - 2026-10-09. The merge leaves this PR's own code unchanged, and that code does not touch the XML helpers that #64 changed.

Tests

The tests for this PR are the same as in my last review. The merge only adjusts the existing XML and installer tests that came from #64.

CI

• Lint & Test: success. This answers the gap I raised last time, because this check had not reported then.
• codecov/patch: success (94.11% of the diff covered, target 80%).
• Auto-merge low-risk Dependabot PRs: skipped. This is expected, since this is not a Dependabot PR.