Skip to content

ci: six-cell OS/arch matrix, coverage gate, VT fuzz test, fix two stale swarm tests - #71

Merged
bsommers merged 21 commits into
mainfrom
chore/ci-matrix-and-coverage
Nov 11, 2026
Merged

bsommers merged 21 commits into
mainfrom
chore/ci-matrix-and-coverage

Conversation

@bsommers

@bsommers bsommers commented Nov 10, 2026 •

Copy link
Copy Markdown
Owner

What

  • CI matrix: Linux, macOS and Windows, each on x86_64 and ARM64 (ubuntu-latest, ubuntu-24.04-arm, macos-latest, macos-15-intel, windows-latest, windows-11-arm). quick (locked build, clippy, unit tests, binary smoke) runs on all six for every event; integration (every tests/*.rs, 3 shards per cell) runs on linux-x86_64 and macos-arm64 for PRs and on all six for main, nightly and manual runs.
  • Required check names are kept on purpose. Branch protection on main requires clippy + quick verify (ubuntu-latest), clippy + quick verify (macos-latest) and cargo-deny. The linux-x86_64 and macos-arm64 quick cells keep exactly those names so the required checks keep reporting; rename them only together with the protection rule.
  • Coverage: scripts/coverage.sh with a per-module ratchet (.github/coverage-floor.txt) and a coverage job. Line coverage on main before this branch 83.1%; this branch 84.7% locally, 84.x% on the runner. The stt floor is calibrated on the runner (72.4%): a dev box with an audio recorder on PATH reaches about 75.6%.
  • New tests: deterministic property fuzz of the VT parser, per-OS binary smoke test, :clipboard grammar end to end, malformed-input robustness, config dotted keys, CLI argument parsing.
  • scripts/verify.sh: shards, parallel pre-build, skip report, panic messages kept in fail.txt, a short TMPDIR and a private XDG_RUNTIME_DIR for every run; scripts/verify.ps1 for Windows.
  • Merged main (2efb58d). fix: harden pane output handling and app lifecycle #70 fixed the two VT char-boundary panics this branch's fuzz test found, so the KNOWN_BUG_PANIC tolerance, the four #[ignore]s and the canary are gone; the reproducers now run as regression tests.

Product bugs found by running the suites in clean CI cells (fixed here)

  • :font add <zip|url> failed on a fresh profile with "Cannot create fonts directory": the directory was created non-recursively (src/pane/font_fetch.rs).
  • macOS: a non-socket file left at a daemon socket path was never cleaned up, because connect() reports ENOTSOCK there and ECONNREFUSED on Linux (src/daemon/process.rs, src/daemon/mod.rs).

Test hermeticity fixes

  • navigator_two_line_rows: depended on the length of the machine's user@host.
  • phase66_67_live_agent_control: socket directory under the long macOS temp dir overflowed sun_path; a third test that sets the process-global live socket was missing the lock (the earlier flake fix was incomplete).
  • phase51_1_tracing, phase50_1_e2e: depended on a short temp/runtime path; the gate now provides one.
  • Two tests that failed on unmodified main (phase68_69_surface_parity, swarm_wiring) were stale and are fixed test-side only.

Quarantined, visibly

  • daemon_resilience_lifecycle::zombie_registry_entry_is_not_listed_active is #[ignore]d with a reason off Linux. Product gap: daemon::liveness::pid_is_zombie reads /proc, so on macOS a zombie daemon entry is still listed as active.
  • Windows test steps are non-blocking (WIN-TESTS); build and clippy are hard gates and pass on both Windows targets. Two tests fail there today and are shown in each Windows job summary:
    • wm::session_manager::tests::session_persists_and_restores_shell_running_commands_and_cwd_per_pane: expects /tmp/frontend, gets C:\tmp\frontend.
    • smoke_binary_per_os::headless_daemon_scripted_session_over_loopback_tcp: bilterm --headless --listen 127.0.0.1:PORT exits with code 1 on Windows. Not diagnosed further.
  • Seven tests that drive a /bin/sh child are ignored, <reason> on Windows; tests/tmux_link_live.rs is Unix-only.

Other findings, not changed here

  • macOS without XDG_RUNTIME_DIR: the default runtime dir is longer than the 40-byte ssh ControlPath limit, so ssh ControlMaster is always disabled with a warning.
  • An unknown top-level switch (bilterm --bogus) is not rejected; it falls through to the interactive launch.

Verification

  • Local: ./scripts/verify.sh → PASS 183/183 65s; scripts/check_classic_goldens.sh → GOLDENS OK 11; cross cargo clippy --all-targets -- -D warnings clean for the five non-host targets.
  • CI run 42811677754 on 83bf55b: all 15 jobs green (6 quick cells, 6 integration shards, coverage, cargo-deny, plan). The Windows quick jobs are green with the two non-blocking test failures above.
  • Not yet exercised: integration on linux-arm64, macos-x86_64 and Windows runs only on main, nightly and manual dispatch, so its first run happens after merge.

🤖 Generated with Claude Code

bsommers and others added 12 commits November 10, 2026 00:58
- swarm_wiring: use apply_automation (returns command errors as Err); plain
  apply only reports them on the status bar, so is_err() never held.
- phase68_69_surface_parity: swarm.run / :swarm run shipped; keep the
  not-yet-wired guard for review/heal and assert run is wired.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…fix)

spawned_child_pty_receives_bilterm_env raced client_socket_resolution_order
(LiveRpc::start sets the same global), failing 1 in 5 runs under load.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
tests/common: isolated HOME/XDG/runtime env, deadline polling, portable
shell helper, visible runtime skips, deterministic PRNG.
smoke_binary_per_os: --version/--help, generators, CLI/ctl error paths,
scripted MCP stdio session, scripted RPC session against a headless daemon
on loopback TCP, and a real PTY child.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Random, grammar-shaped and mutated escape sequences with fixed seeds.
Found two char-boundary panics (src/vt/scan.rs:29, src/vt/codec.rs:77);
src/vt belongs to another branch, so the harness tolerates exactly that
panic message visibly, with ignored minimal reproducers and a canary that
fails once the parser is fixed.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
InstanceGuard.runtime is only read by the Unix default-link publisher; three
test files import Unix-only daemon/PTY helpers. Scoped cfg_attr allows only.
Also tidy clippy lints in the new fuzz test.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
… parsing

malformed_input_robustness: JSON (round trip, nesting bombs, mutations),
JSON-RPC dispatch and MCP line handler with wrongly typed params, colon
commands for every registered head, selectors/coordinates/chords, static
path traversal, bind/CORS validation, tmux control parsers, config /
playbook / swarm plan / theme loaders.
config_keys_and_cli_args: dotted-key round trips and atomic failures,
load/save error paths, CLI option syntax and endpoint resolution.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…form)

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…ime skips

Seven tests drive a /bin/sh child; on Windows they are reported as
'ignored, <reason>' instead of failing or silently passing.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
… point

verify.sh: --integration and --shard K/N for CI fan-out; one parallel
cargo pre-build of the selected test binaries; target/verify/skips.txt
lists ignored tests and runtime SKIPPED lines; drop MSYS's SHELL so tests
see a native Windows environment.
verify.ps1: thin PowerShell wrapper (Git for Windows bash) with the same
one-line PASS/FAIL contract. ci_report.sh: job-summary + fail.txt groups.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
cargo-llvm-cov over unit + integration tests; per-module table, largest
uncovered regions, lcov + HTML; fails when a module drops below
.github/coverage-floor.txt. No external coverage service.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…rage gate

quick (all cells, every event): locked build, clippy, unit tests, binary
smoke. integration (3 shards/cell): PRs on linux-x86_64 + macos-arm64,
main/nightly/manual on all six. coverage: llvm-cov ratchet + artifact.
Windows test steps are continue-on-error (WIN-TESTS) with their real
outcome in the job summary; build + clippy are hard gates there.
Actions pinned by SHA, contents:read only, superseded runs cancelled.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
bsommers and others added 9 commits November 10, 2026 06:34
…y panics are fixed

Merged main (#70): removes KNOWN_BUG_PANIC, the guarded() tolerance, the
four #[ignore]s and the canary; the reproducers now run as regressions.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The fonts directory was created non-recursively, so the command failed with
'Cannot create fonts directory' until something else had created
<config>/bilterm. Found by running pane_typography_coverage in a clean CI
shard. The directory itself keeps the private-dir checks.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
- navigator_two_line_rows: user@host is as long as the machine makes it;
  assert the row starts the detail text instead of fitting its first word.
- phase66_67: socket dirs under /tmp (macOS temp dir overflows sun_path);
  the third test that sets the global live socket now takes the lock too.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…he runner

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…h on macOS

connect() to a regular file is ECONNREFUSED on Linux but ENOTSOCK on macOS,
so stale-lock recovery, registry pruning and stop left the file behind
there (phase55 daemon tests, macOS CI).

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…tmux live test to Unix

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…the /proc zombie test off Linux

macOS's long per-user temp dir overflowed socket paths, disabled the ssh
ControlPath (extra WARN in the tracing test) and clipped a path in a
rendered frame. zombie_registry_entry_is_not_listed_active needs /proc:
ignored with the reason on other OSes (product gap, listed in the PR).

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@bsommers
bsommers merged commit 98e2776 into main Nov 11, 2026
15 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant