SkillPedia Skills Encyclopedia
中文 EN
Today's updates 2026-11-18 UTC

No new commits yet today

Open the update feed

Skills 2026-10-11 02:21 UTC

cc-haha

Committed by NanmiCoder. Update Category: Bug Fixes & Stability Improvements. Commit details: fix(ci): make the PR quality gate trustworthy and add a mock-LLM agent lane (#1511). This update resolves known runtime errors and enhances system reliability.

NanmiCoder f38deae NanmiCoder/cc-haha

CommitRaw commit message

fix(ci): make the PR quality gate trustworthy and add a mock-LLM agent lane (#1511)

* fix(ci): make the PR quality gate trustworthy and add a mock-LLM agent lane

The PR gate passed 4 of its last ~100 runs. Almost every failure came from
the same eight test files that were already red on main, so every pull
request inherited failures it did not cause and a red gate meant nothing.

Make main green:
- cron launcher tests: fake CLIs now answer before waiting for EOF, matching
  the stream-json protocol the scheduler has used since #1433 (deadlock).
- cron migration test: mocked child env lacked PATH, so bun never spawned.
- connector test: pin the platform instead of borrowing the host's.
- workspace: resolve NFC/NFD filename mismatches inside the workspace
  (product bug, visible on ext4/NTFS; APFS hid it).
- data migration: moving a restored read-only directory failed on POSIX and
  rollback left an undeletable staging dir (product bug); this also caused
  the coverage job's EACCES cleanup failure.
- /help panel count, office zip timing assertion, sidecar test isolation.
- desktop-checks installs adapter deps the sidecar bundle needs.

Restructure the gate:
- Run PR Quality again after each merge to main, scoped to the pushed range.
- Coverage is reported on PRs but no longer blocks merges.
- Full Quality runs daily in addition to manual dispatch.

Add check:agent-e2e: the real server and the real CLI agent loop against a
loopback Anthropic-compatible endpoint backed by a scripted model. It shares
the live lane's scenarios and adds upstream-side assertions (tool output
returned to the model, denial as an error result, stop closing the upstream
stream, recovery after an upstream error). It exposed that the live interrupt
scenario sent a message type the server ignores.

* fix(ci): fix the Linux-only failures the first gate run exposed

- agent-e2e: the lane stands in for a desktop machine, so it clears CI. With
  CI set, the CLI demands ANTHROPIC_API_KEY (src/utils/auth.ts) and exits 0 at
  startup on the desktop provider path. Same fix for the claudeBetas
  integration test, which drives that path.
- print.sessionMessage: feed stdin through a pipe. On Linux a Blob stdin does
  not survive bin/claude-haha re-spawning the CLI, which then reads nothing.
- workspaceWatch: assert the dirty-directory contract the desktop relies on;
  Bun's Linux watcher names only the source of a rename.
- connectors: pin a supported host instead of borrowing the Linux runner's.
- conversations: the deferred-permission test settles only after the
  inspection it asserts on, with enough stream delay for a busy runner.
- team supervisor: wait for the crash notice, which follows the terminated flag.
- swift-checks shares coverage's best-effort sandbox cleanup: a leftover the
  macOS runner refuses to delete (EROFS) no longer fails a passing lane.