[Zeus-Deus]

Problem

Several sources of instability, slowness or security exposure were found while comparing Codemux against known failure modes in agent-driven coding environments:

• Background git polling took .git/index.lock. An agent's own git add / git commit could fail with index.lock: File exists. The 60 s background fetch could also break a concurrent git pull (unlocked FETCH_HEAD) and start auto-gc.
• Stuck agent turns.
  • Claude Stop could hang forever when the CLI never acknowledged it.
  • A Codex turn stuck in "retrying" never failed.
  • Cursor/Grok requests waited up to 24 h after the agent closed stdout.
  • ACP cancelled tool calls kept spinning.
• Remote connections that don't heal.
  • A short network drop permanently broke a remote workspace's SSH tunnel.
  • Tunnel ssh processes outlived the app.
  • The web/phone remote could sit in "reconnecting" for minutes after the network came back.
• Security.
  • An SVG or HTML file opened from the web remote's /api/assets ran as a same-origin page, able to mint a WebSocket ticket from the session cookie.
  • Automations synced origin URLs with embedded tokens to the cloud API.
• Integrations and UI.
  • GitHub repos cloned through an SSH host alias (git@github-work:…) lost all PR features.
  • PR watch reported green while workflow runs were still queued.
  • A huge file diff (lockfile, bundle) froze the diff pane.
  • The right panel redrew on every streamed frame.

Solution

One focused commit per fix:

Git

• GIT_OPTIONAL_LOCKS=0 on every git command Codemux runs. Status line counts now come from plumbing (diff-files, diff-index --cached), which never writes the index.
• The sweep counts changed files without re-running the diffs.
• The background fetch adds --no-write-fetch-head --no-auto-gc. On timeout it sends SIGTERM before kill, so git can remove its ref locks.
• New worktrees get one update-index --refresh.
• Per-file diffs are capped at 5,000 lines / 512 KiB, with a "Diff too large" notice. Git's stderr is drained concurrently, so a noisy external diff can't deadlock the read.

Agent runtimes

• Claude Stop: waits at most 10 s for the CLI's acknowledgement, then ends the CLI. It only ever touches the query it interrupted, and concurrent Stops share one stop.
• Pending requests: fail 1 s after a child closes stdout but keeps running.
• ACP: cancelled settles tool calls (as errors) and plan steps.
• Codex: a turn fails if a retry makes no progress for 10 min. Repeated retry notices don't extend that deadline.

Remote

• SSH tunnel:
  • After the circuit breaker opens, it keeps retrying every 60 s. The pill now reads "Connection lost — retrying".
  • The stale socket file is removed before each respawn.
  • ssh gets kill_on_drop, and tunnels are shut down on GUI exit and in headless serve teardown.
• Web/phone remote transport:
  • The ticket fetch has an 8 s deadline and socket open a 15 s deadline. Wake events restart a stalled attempt.
  • Abandoned grant requests are cancelled and can no longer overwrite a newer grant. Each socket keeps its own grant.

Security

• /api/assets always sends X-Content-Type-Options: nosniff. HTML, SVG and XML also get Content-Security-Policy: default-src 'none'; style-src 'unsafe-inline'; sandbox. Inline <img> rendering is unaffected.
• Automation project remotes have their credentials stripped (user, password, query, fragment) before they are stored or synced.

Integrations and UI

• SSH aliases: dot-less aliases containing "github", and github.com-<label> aliases, resolve to github.com. Explicit host mappings still win, and real dotted hosts (GitHub Enterprise) are untouched.
• PR watch: when checks look green, it makes one actions/runs?head_sha= call. Queued, waiting or in-progress runs keep it pending. A transient failure of that call holds green back and retries next pass; a permission or "not found" failure falls back to checks only.
• Right panel: subscribes only to derived subagent counts.
• Transcript: a press in its left gutter no longer cancels auto-follow.

Screenshots

The diff pane opening a 20,000-line lockfile-style diff (dev mock):

Before: everything rendered	After: capped with a notice

Verification

• cargo check: clean. cargo test --no-run compiles every test target.

• Focused cargo test --lib runs, all passing:

  Filter	Passed
  git::	231
  tunnel_supervisor	13
  jobs::	34
  pr_watch	35
  json_rpc_child	20
  web_remote::assets	5
  project_identity	16
  git_provider / _provider::	1234
  github_budget	35
  agent_provider::codex	238
  agent_provider::custom_acp	102

  The only failures were tests that need prebuilt fixture binaries (fake_cursor_acp, fake_claude_sidecar); CI builds those.

• cargo test --features test-fixtures --test integration git_operations: 65 passed.

• Sidecar: bun run typecheck clean, bun run test 245 passed.

• npm run check: clean. Affected vitest files: 469 passed (src/remote, src/environments, right panel, MessageList, diff).

• Each new regression test was confirmed to fail against the old code: transport stalls, right-panel redraw, left-gutter press, stale socket, the stalled Claude Stop, and the repeated-retry stall fixture.

• Independent code review over three rounds: 8 findings in total, all fixed; final verdict approve.

Caveats

• Not exercised live against a real SSH host, iroh relay, or real Claude/Codex/Cursor CLIs. Those paths are covered by unit and integration tests with fake processes and servers.
• Windows was not compiled locally. Unix-only signal and tunnel code is behind cfg(unix), and CI covers Windows.
• Automations saved before this change keep their stored remote URL until they are next edited.
• --no-write-fetch-head needs git ≥ 2.29, which Codemux already relies on elsewhere.

---

Implemented by Claude Opus 5.5 via Claude Code; reviewed by GPT 6.1 Sol (extra-high reasoning) via Codex.

🤖 Generated with Claude Code

[mario-reviewer]

Review of 03bc1d6ec089887994a3df4bb0635925a97e8241

1. Preserve real enterprise hosts when resolving SSH aliases

• File:Line: src-tauri/src/git_provider/detect.rs:287–292
• Severity: Warning
• What's wrong: let dotless_alias = !host.contains('.') && host.contains("github"); treats an actual internal hostname such as github or github-enterprise as a public-GitHub SSH alias, silently changing the API destination.
• Impact: With git@github:team/repo.git pointing to an enterprise instance through corporate DNS or SSH configuration, and no Codemux custom-host override, the base revision preserves github; this revision rewrites it to github.com. github_budget::remote_repository shares this resolver, so PR operations lose the enterprise login and target github.com/team/repo instead, potentially selecting an unrelated repository if that name exists publicly.
• Fix: Resolve the effective SSH HostName or require an explicit alias-to-host mapping before changing the destination. Keep the original host when its public-GitHub destination has not been established. Add coverage for a dotless enterprise hostname without a custom-host override.

2. Check every workflow-run page before declaring the head green

• File:Line: src-tauri/src/pr_watch.rs:769–774
• Severity: Warning
• What's wrong: "repos/{}/actions/runs?head_sha={}&per_page=100" reads only the first page, so an unfinished run on a later page does not hold back green.
• Impact: When the same SHA has more than 100 workflow runs and an older run remains queued behind concurrency, the first page can contain only completed runs while that queued run still has no jobs in statusCheckRollup. parse_runs_pending ignores total_count, returns false, and evaluate can send the green wake, refund attempts, and record green:<sha> before the queued workflow finishes.
• Fix: Page through the workflow-run envelope until all runs are checked, stopping early on an unfinished run; alternatively, query unfinished statuses directly. Keep green unconfirmed if the read is incomplete. Add a regression case with a completed first page and an unfinished run on the next page.

Validation: Full first-pass static review of the PR diff and exact-head repository context. No repository code was built, installed, or executed. No prior review findings were present.

CI, reported separately: changes and web-checks (ubuntu-latest) passed. web-checks (windows-latest), rust (ubuntu-latest), and rust (windows-latest) were in progress. iroh-wasm was skipped. Pending CI is not a code finding.

Verdict: REQUEST_CHANGES

🤖 Automated review by Mario (AI). Findings may be wrong; a human makes the final call.

[Zeus-Deus]

Thanks. Both findings are valid and both are fixed.

1. Enterprise hosts and SSH aliases (153dcfd)

• Alias-looking hosts are no longer redirected by name alone. Codemux now asks ssh -G -- <alias> for the real HostName.
  • The result is cached for 5 minutes.
  • The lookup has a 2 s timeout.
  • Only hosts made of [A-Za-z0-9._-] that don't start with - are passed to ssh.
• When ssh names no HostName (it echoes the alias back) or isn't available, the original host is kept. So git@github:team/repo reached through corporate DNS stays github.
• When ssh does name a HostName, that host is used. github-corp → github.corp.example now resolves to the enterprise instance instead of being forced to github.com, and ssh.github.com maps to github.com.
• Explicit custom-host mappings still win before any lookup happens.
• Tests use a per-thread fake ssh config. They cover dotless github / github-enterprise with no HostName (kept as they are) and an alias that points at an enterprise host.

2. Workflow-run paging (d13c4a0)

• runs_hold_green now reads &page=N and stops at the first unfinished run.
• It treats the head as finished once a short page arrives or total_count runs have been read.
• A transient failure on any page holds green back, exactly like a failure on page 1.
• Beyond 10 pages (1,000 runs) green stays unconfirmed.
• New regression test: a fully completed first page plus a queued run on page 2 holds green back; an all-completed second page goes green without reading a third page.

Verification:

• cargo check passes.
• cargo test --lib: pr_watch 36 passed, _provider:: 1238 passed, github_budget 35 passed.
• Checked real ssh -G output: with no config entry it prints hostname github-work, and with HostName github.com configured it prints hostname github.com.

[mario-reviewer]

Re-review of d13c4a000da90d0b4cc2e9cc11c8507aaa7b1652

No new blocking findings. Both findings from the previous review are resolved:

• SSH alias resolution: src-tauri/src/git_provider/detect.rs:203–218,265–294 now consults ssh -G rather than assuming a dotless GitHub-like hostname means github.com. Unresolved or unchanged hostnames remain intact, and explicit custom-host mappings still take precedence. Regression coverage includes dotless enterprise hosts and an alias targeting an enterprise instance.
• Workflow-run pagination: src-tauri/src/pr_watch.rs:230–255,801–809 now requests successive pages, stops on an unfinished run, and leaves green unconfirmed when the page limit is reached. Regression coverage includes a queued run on page two, completed pagination, a later-page timeout, and the page limit.

Validation: Reviewed the two commits since 03bc1d6ec089887994a3df4bb0635925a97e8241, their exact-head repository context, and both prior findings. Static review only; no repository code was built, installed, or executed. Regression tests were inspected, not run.

CI, reported separately: changes passed. web-checks (ubuntu-latest), web-checks (windows-latest), rust (ubuntu-latest), and rust (windows-latest) are in progress. iroh-wasm was skipped. Pending CI is not a code defect.

Verdict: APPROVE

🤖 Automated review by Mario (AI). Findings may be wrong; a human makes the final call.

[Zeus-Deus]

The Rust CI failures on d13c4a0 were not caused by this branch. CI picked up Rust 1.100.0 (released 2026-11-09); the last Rust run on main used 1.99.1. The new Cargo builds unit-test executables under target/debug/build/<pkg>/<hash>/out/ instead of target/debug/deps/, which broke two things:

• Linux: the 37 lib tests that locate fixtures two levels above the test binary (fake_cursor_acp, fake_codex_app_server, …) could not find them. 1e1fc6d adds test_isolation::fixture_bin, which searches every ancestor directory. Verified locally under the old layout, and by running a copy of the test binary from a build/codemux/<hash>/out/ directory.
• Windows: the Common Controls manifest step found no target/debug/deps directory. d8f9435 makes it search target/debug recursively.

main will hit the same failures on its next Rust run.

[mario-reviewer]

Re-review of d8f94356154f58ec785934344e3ecb67de7528b3

No new blocking findings. The two commits since the previous review look fine:

• src-tauri/src/test_isolation.rs:10–16 searches executable ancestors for fixture binaries, preserves the missing-fixture fallback, and is used by the updated unit tests. Existing explicit fixture overrides remain intact. The helper is in a test-only module.
• .github/workflows/ci.yml:548–551 recursively places Common Controls manifests beside Windows executables after test compilation and before test execution.

Both earlier findings remain resolved: SSH aliases are resolved through SSH configuration without guessing that enterprise hostnames mean public GitHub, and workflow-run checks paginate before confirming green. No unresolved prior blocking findings remain.

Validation: Inspected gh pr diff, both commits since d13c4a000da90d0b4cc2e9cc11c8507aaa7b1652, exact-head source and workflow context, and the prior reviews. Static review only. No repository code was built, installed, or executed; tests were inspected, not run.

CI, reported separately: changes and iroh-wasm passed. web-checks (ubuntu-latest), web-checks (windows-latest), rust (ubuntu-latest), and rust (windows-latest) are in progress. Pending CI is not a code defect.

Verdict: APPROVE

🤖 Automated review by Mario (AI). Findings may be wrong; a human makes the final call.