Repository navigation
Conversation
Pins the new contract for cmux claude-teams's tmux-compat env: TERM defaults to xterm-ghostty (was screen-256color) and TERM_PROGRAM is preserved (was unconditionally unset). Without the matching Swift CLI fix, this test fails — exposing the Claude Code v2.1.112 crash from issue manaflow-ai#2947 during permission escalation. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…#2947) Mirrors PR manaflow-ai#2960's Go-daemon fix into the local-Mac Swift CLI path, which is what `cmux claude-teams` actually goes through on the user's machine. Without this, the Go-only fix doesn't reach the reporter's crash. - Add `unsetTermProgram: Bool` to configureTmuxCompatEnvironment. - claude-teams: pass false (Claude Code v2.1.112+ needs TERM_PROGRAM set during permission escalation; missing it triggers the cli.js crash in manaflow-ai#2947). - omo / omx / omc: pass true (opencode flips to a light theme on TERM_PROGRAM=ghostty; preserves prior behavior). - Default TERM is now xterm-ghostty instead of screen-256color so agents render with full Ghostty fidelity, and COLORTERM=truecolor is set when the parent env didn't already supply one. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Fixes cmux crash when using claude-teams with teamCreate (issue manaflow-ai#2947). Root cause: Claude Code v2.1.112 changed how it handles permission escalation based on terminal environment variables. The previous code unconditionally unset TERM_PROGRAM which caused Claude Code's cli.js to crash during permissionExplainerEnabled checks. Changes: - Add unsetTermProgram flag to agentConfig to control TERM_PROGRAM handling per agent type - claude-teams: preserve TERM_PROGRAM (needed for Ghostty detection) - omo/omx/omc: unset TERM_PROGRAM (opencode switches to light theme) - Change default TERM from screen-256color to xterm-ghostty - Add COLORTERM=truecolor for proper truecolor support (cherry picked from commit 4c6f9b2)
…Swift Polish on top of the cherry-picked Go fix (manaflow-ai#2960) and our matching Swift fix, addressing CodeRabbit + adversarial-review feedback. Changes: - Rename agentConfig.unsetTermProgram (Go) and the matching Swift parameter to preserveTermProgram. The zero/default value now matches the historically common case (TERM_PROGRAM unset, the legacy behavior), so any future agent type defaults to the safe choice and only claude-teams opts in to preservation. CodeRabbit flagged the original naming as zero-value-unsafe; both adversarial reviewers concurred. - Add a resolveDefaultTerm() helper in both surfaces that probes `infocmp xterm-ghostty` and falls back to `xterm-256color` when the terminfo entry is missing. cmux's terminfo overlay installs asynchronously during shell bootstrap on remote SSH hosts (see tests_v2/test_ssh_remote_shell_integration.py for the documented race), so a child process exec'd before the install completes would otherwise inherit TERM=xterm-ghostty without the matching terminfo entry — landing it at a dumb terminal. xterm-256color is universally available. - Read COLORTERM from the live process env (getenv) instead of the Swift-side processEnvironment snapshot, mirroring the Go daemon's behavior in agent_launch.go. Functionally equivalent today (nothing mutates COLORTERM between snapshot and check) but more robust to future re-orderings and keeps the two surfaces aligned. - Tighten the asymmetry comment in both files. opencode reacts to any non-empty TERM_PROGRAM (treats it as an outer-host marker), not specifically TERM_PROGRAM=ghostty as the previous comment implied. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add comprehensive coverage for configureAgentEnvironment's behavior matrix in cmuxd-remote, plus a sibling Python test for the local Swift CLI's `cmux omo` path. Go (daemon/remote/cmd/cmuxd-remote/agent_launch_test.go): - TestConfigureAgentEnvironment_PreservesTermProgramWhenFlagSet — locks in claude-teams's preserve contract that prevents the Claude Code v2.1.112+ permission-escalation crash from manaflow-ai#2947. - TestConfigureAgentEnvironment_OpencodeFamilyUnsetsTermProgram — guards against silent regression of manaflow-ai#2516 (opencode flips to a light theme on any non-empty TERM_PROGRAM); without this assertion a future PR could remove the unset side of the asymmetry undetected. - TestConfigureAgentEnvironment_TermEnvVarOverride — asserts the per-agent TERM override env var (e.g. CMUX_CLAUDE_TEAMS_TERM) wins over resolveDefaultTerm's fallback. - TestConfigureAgentEnvironment_COLORTERMPreservedWhenSet — asserts the truecolor fallback never downgrades a caller-provided COLORTERM. - TestConfigureAgentEnvironment_AppliesExtraEnv — sanity check that the extraEnv map applies (regression guard for the field). - TestResolveDefaultTerm_FallsBackWhenInfocmpAbsent — simulates the remote-SSH-bootstrap race (PATH stripped of infocmp) and asserts resolveDefaultTerm returns "xterm-256color" instead of leaving a child process with an unresolvable TERM. - TestResolveDefaultTerm_PrefersXtermGhosttyWhenInfocmpSucceeds — happy-path check, gracefully skips on hosts without the terminfo entry installed. The existing TestConfigureAgentEnvironment in tmux_compat_test.go now also asserts TERM_PROGRAM is unset by default (zero-value preserveTermProgram = legacy behavior), filling a previously-empty assertion gap in the existing test. Python (tests/test_cli_omo_env.py): - New file mirroring tests/test_cli_claude_teams_env.py's fake-shim pattern but for `cmux omo`. Asserts TERM_PROGRAM is unset (the manaflow-ai#2516 guard), TERM defaults to xterm-ghostty, and COLORTERM=truecolor is set when the parent had no COLORTERM. Uses the cmux DEV CLI from /tmp/cmux-last-cli-path so it runs against the user's most recent reload.sh build. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
|
Someone is attempting to deploy a commit to the Manaflow Team on Vercel. A member of the Team first needs to authorize it. |
📝 WalkthroughWalkthroughThis PR fixes a cmux claude-teams crash by refactoring terminal environment handling: it adds dynamic TERM resolution via Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
Warning There were issues while running some tools. Please review the errors and either fix the tool's configuration or disable the tool if it's a critical failure. 🔧 golangci-lint (2.11.4)level=error msg="[linters_context] typechecking error: pattern ./...: directory prefix . does not contain main module or its selected dependencies" Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Greptile SummaryThis PR fixes
Confidence Score: 4/5Core fix is correct and well-tested in Go; two Python tests will reliably fail on CI hosts without xterm-ghostty terminfo and need a one-liner fix before merge. The implementation logic across both Go and Swift surfaces is sound, the Go unit tests are thorough and correct, and the tests/test_cli_claude_teams_env.py (line 205) and tests/test_cli_omo_env.py (line 100) — both need the term override env var pinned before the subprocess call. Important Files Changed
Flowchart%%{init: {'theme': 'neutral'}}%%
flowchart TD
A[cmux agent launch] --> B{agent type?}
B -->|claude-teams| C[preserveTermProgram = true]
B -->|omo / omx / omc| D[preserveTermProgram = false]
C --> E[configureAgentEnvironment]
D --> E
E --> F[Set PATH / TMUX / TMUX_PANE]
F --> G{termEnvVar set?}
G -->|yes| H[TERM = override value]
G -->|no| I[resolveDefaultTerm]
I --> J{infocmp xterm-ghostty exits 0?}
J -->|yes| K[TERM = xterm-ghostty]
J -->|no| L[TERM = xterm-256color]
H --> M[COLORTERM fallback]
K --> M
L --> M
M --> N{preserveTermProgram?}
N -->|true| O[TERM_PROGRAM preserved]
N -->|false| P[unsetenv TERM_PROGRAM]
O --> Q[exec agent]
P --> Q
|
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| term_value = read_text(term_log) | ||
| if term_value != "screen-256color": | ||
| print(f"FAIL: expected TERM=screen-256color, got {term_value!r}") | ||
| if term_value != "xterm-ghostty": | ||
| print(f"FAIL: expected TERM=xterm-ghostty, got {term_value!r}") | ||
| raise SystemExit(1) |
There was a problem hiding this comment.
TERM=xterm-ghostty assertion is non-deterministic
The test unconditionally expects xterm-ghostty, but resolveDefaultTerm() returns xterm-256color whenever infocmp xterm-ghostty fails — which is the case on any Linux CI runner or Mac without Ghostty installed. The test provides no fake infocmp in real_bin, so whether the assertion passes depends entirely on test-host config rather than code correctness.
The simplest fix is to pin the override env var so the test bypasses resolveDefaultTerm() and stays deterministic:
| term_value = read_text(term_log) | |
| if term_value != "screen-256color": | |
| print(f"FAIL: expected TERM=screen-256color, got {term_value!r}") | |
| if term_value != "xterm-ghostty": | |
| print(f"FAIL: expected TERM=xterm-ghostty, got {term_value!r}") | |
| raise SystemExit(1) | |
| env["CMUX_CLAUDE_TEAMS_TERM"] = "xterm-ghostty" | |
| env["TERM"] = "xterm-256color" |
Alternatively, accept both valid values:
if term_value not in ("xterm-ghostty", "xterm-256color"):
print(f"FAIL: expected TERM=xterm-ghostty or xterm-256color, got {term_value!r}")
raise SystemExit(1)| term_value = read_text(term_log) | ||
| if term_value != "xterm-ghostty": | ||
| print( | ||
| f"FAIL: expected TERM=xterm-ghostty for cmux omo, got {term_value!r}" | ||
| ) | ||
| raise SystemExit(1) |
There was a problem hiding this comment.
TERM=xterm-ghostty assertion will fail on hosts without Ghostty terminfo
Same issue as the sibling test: the check hard-asserts xterm-ghostty, but resolveDefaultTerm() falls back to xterm-256color on hosts where infocmp xterm-ghostty exits non-zero (any Linux CI runner, any Mac without Ghostty). No fake infocmp is placed in real_bin, so the result is host-dependent.
Pin CMUX_OMO_TERM in the test env to force the override path and make the assertion deterministic:
| term_value = read_text(term_log) | |
| if term_value != "xterm-ghostty": | |
| print( | |
| f"FAIL: expected TERM=xterm-ghostty for cmux omo, got {term_value!r}" | |
| ) | |
| raise SystemExit(1) | |
| env["CMUX_OMO_TERM"] = "xterm-ghostty" | |
| env["TERM"] = "xterm-256color" |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 4
🧹 Nitpick comments (2)
daemon/remote/cmd/cmuxd-remote/agent_launch.go (1)
441-449: Consider cachingresolveDefaultTerm()across a single process lifetime.The
infocmpprobe spawns a subprocess on every agent launch. It's only called whenCMUX_*_TERMisn't set, so the cost is bounded (one launch → one spawn), but if this function ends up being called from other code paths later it would be cheap insurance to memoize viasync.Once. Not needed today — just flagging for the next caller.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@daemon/remote/cmd/cmuxd-remote/agent_launch.go` around lines 441 - 449, The resolveDefaultTerm function currently runs infocmp on every call; memoize its result for the process lifetime by adding a package-level cached variable (e.g., defaultTerm string) and a sync.Once (e.g., defaultTermOnce) and change resolveDefaultTerm to call defaultTermOnce.Do(func(){ ...compute and set defaultTerm... }) and return defaultTerm; keep the same probe logic (exec.Command("infocmp", "xterm-ghostty") etc.) inside the once body so the subprocess is spawned only once per process.daemon/remote/cmd/cmuxd-remote/agent_launch_test.go (1)
160-177: Test doesn't actually verify the "applied LAST" contract in its docstring.The comment states
extraEnv"is applied LAST (after all other env mutations) so callers can override anything the function sets", but the assertion only checks thatCLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1is set — which would pass regardless of ordering since nothing else writes that variable. To actually pin the ordering contract, haveextraEnvcollide with a value the function sets earlier (e.g., overrideTERMorTMUX) and assert theextraEnvvalue wins.♻️ Proposed strengthening
func TestConfigureAgentEnvironment_AppliesExtraEnv(t *testing.T) { t.Cleanup(envSnapshot(t, "COLORTERM", "TERM", "PATH", "TMUX", "TMUX_PANE", "CMUX_CLAUDE_TEAMS_CMUX_BIN", "CMUX_SOCKET_PATH", "CMUX_SOCKET", "CMUX_CLAUDE_TEAMS_TERM", "TERM_PROGRAM", "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS", )) - configureAgentEnvironment(claudeTeamsConfig()) + cfg := claudeTeamsConfig() + // Collide with a value set earlier in configureAgentEnvironment to + // prove extraEnv wins (i.e., is applied LAST). + cfg.extraEnv["TERM"] = "extraenv-override" + configureAgentEnvironment(cfg) if got := os.Getenv("CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS"); got != "1" { t.Errorf("CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS = %q, want 1", got) } + if got := os.Getenv("TERM"); got != "extraenv-override" { + t.Errorf("TERM = %q, want %q (extraEnv must be applied LAST and override earlier writes)", got, "extraenv-override") + } }🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@daemon/remote/cmd/cmuxd-remote/agent_launch_test.go` around lines 160 - 177, The test TestConfigureAgentEnvironment_AppliesExtraEnv documents that extraEnv is applied last but doesn't verify ordering; update the test to make extraEnv collide with a variable that configureAgentEnvironment mutates (for example TERM or TMUX) by using claudeTeamsConfig() or constructing an extraEnv map that sets that variable to a distinct sentinel value, call configureAgentEnvironment, and then assert the environment variable equals the sentinel (verifying extraEnv wins). Ensure references to configureAgentEnvironment, claudeTeamsConfig (or the test's extraEnv construction), and the chosen env var (e.g., "TERM" or "TMUX") are used so the test actually checks the "applied LAST" contract.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@CLI/cmux.swift`:
- Around line 10381-10385: The code currently sets process.standardOutput and
process.standardError to Pipe() then calls try process.run() and
process.waitUntilExit(), which can block because the pipes are never read;
change those assignments to use FileHandle.nullDevice for both standardOutput
and standardError (replace the Pipe() uses), so output is discarded instead of
creating unread pipes before calling process.run() and process.waitUntilExit();
update references around process.standardOutput, process.standardError,
process.run(), and process.waitUntilExit() accordingly.
In `@tests/test_cli_claude_teams_env.py`:
- Around line 205-207: The test currently asserts term_value == "xterm-ghostty"
unconditionally; change it to accept the documented fallback by mirroring the
production resolveDefaultTerm logic: determine whether the runtime prefers
"xterm-ghostty" or falls back to "xterm-256color" (either by calling/importing
the existing resolveDefaultTerm helper or implementing the same detection in the
test) and then assert term_value equals the resolved default instead of
hard-coding "xterm-ghostty" (refer to term_log/term_value and resolveDefaultTerm
in the codebase).
In `@tests/test_cli_omo_env.py`:
- Around line 99-104: The test unconditionally asserts TERM == "xterm-ghostty"
which flakes when the terminfo entry is missing; change the test to either set
CMUX_OMO_TERM (or CMUX_CLAUDE_TEAMS_TERM) to a known value before launching the
agent so the produced term file (read_text -> term_value) is deterministic, or
probe for the terminfo entry and skip the assertion if it's absent (mirror Go's
resolveDefaultTerm probe-and-skip). Locate the block that reads term_value
(term_value = read_text(term_log)) and implement one of these fixes: export
os.environ["CMUX_OMO_TERM"]="xterm-ghostty" for the test run, or run a quick
check (e.g., subprocess call to infocmp or use shutil.which) to detect the
xterm-ghostty terminfo and only perform the equality assertion when present.
Ensure references remain to term_value, read_text, and CMUX_OMO_TERM so the
change is easy to find.
- Around line 33-60: The test currently allows real plugin installation because
run_omo (and the Swift runOMO path) calls omoEnsurePlugin unconditionally and
PATH still contains the host's npm/bun; fix by ensuring the plugin setup is
short-circuited during tests: either pre-seed a fake plugin installation
directory under the test's fake_home (create fake_home/.config/opencode with
expected files) before invoking run_omo, or export an opt-out env var (e.g.,
CMUX_OMO_SKIP_PLUGIN=1) in base_env so omoEnsurePlugin returns early; locate
calls to run_omo/runOMO and omoEnsurePlugin to implement the pre-seed or opt-out
behavior and ensure fake_home and base_env are used consistently so the test
cannot reach the network to install plugins.
---
Nitpick comments:
In `@daemon/remote/cmd/cmuxd-remote/agent_launch_test.go`:
- Around line 160-177: The test TestConfigureAgentEnvironment_AppliesExtraEnv
documents that extraEnv is applied last but doesn't verify ordering; update the
test to make extraEnv collide with a variable that configureAgentEnvironment
mutates (for example TERM or TMUX) by using claudeTeamsConfig() or constructing
an extraEnv map that sets that variable to a distinct sentinel value, call
configureAgentEnvironment, and then assert the environment variable equals the
sentinel (verifying extraEnv wins). Ensure references to
configureAgentEnvironment, claudeTeamsConfig (or the test's extraEnv
construction), and the chosen env var (e.g., "TERM" or "TMUX") are used so the
test actually checks the "applied LAST" contract.
In `@daemon/remote/cmd/cmuxd-remote/agent_launch.go`:
- Around line 441-449: The resolveDefaultTerm function currently runs infocmp on
every call; memoize its result for the process lifetime by adding a
package-level cached variable (e.g., defaultTerm string) and a sync.Once (e.g.,
defaultTermOnce) and change resolveDefaultTerm to call
defaultTermOnce.Do(func(){ ...compute and set defaultTerm... }) and return
defaultTerm; keep the same probe logic (exec.Command("infocmp", "xterm-ghostty")
etc.) inside the once body so the subprocess is spawned only once per process.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 44e634fa-aadc-464e-a9f0-09af9fd289b2
📒 Files selected for processing (6)
CLI/cmux.swiftdaemon/remote/cmd/cmuxd-remote/agent_launch.godaemon/remote/cmd/cmuxd-remote/agent_launch_test.godaemon/remote/cmd/cmuxd-remote/tmux_compat_test.gotests/test_cli_claude_teams_env.pytests/test_cli_omo_env.py
| process.standardOutput = Pipe() | ||
| process.standardError = Pipe() | ||
| do { | ||
| try process.run() | ||
| process.waitUntilExit() |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
python3 - <<'PY'
from pathlib import Path
for path in Path(".").rglob("cmux.swift"):
text = path.read_text(errors="ignore")
marker = "private static func resolveDefaultTerm()"
start = text.find(marker)
if start == -1:
continue
end = text.find("\n }\n", start)
body = text[start:end if end != -1 else start + 1200]
print(path)
print("uses Pipe():", "Pipe()" in body)
print("drains pipe data:", "readDataToEndOfFile" in body or ".read(" in body)
break
else:
raise SystemExit("cmux.swift with resolveDefaultTerm() not found")
PYRepository: manaflow-ai/cmux
Length of output: 116
🏁 Script executed:
#!/bin/bash
# View the exact code context at lines 10381-10385 and surrounding
sed -n '10369,10395p' CLI/cmux.swift
echo "---"
echo "Searching for FileHandle.nullDevice usage patterns in Swift files:"
rg "FileHandle\.nullDevice" -A 2 -B 2 CLI/cmux.swift | head -30
echo "---"
echo "Checking for other Pipe() + waitUntilExit patterns without draining:"
rg -A 5 'process\.standardOutput = Pipe\(\)' CLI/cmux.swiftRepository: manaflow-ai/cmux
Length of output: 1878
Discard probe output instead of piping it.
Lines 10382–10383 create pipes that are never drained before waitUntilExit(). Use FileHandle.nullDevice to discard output and avoid potential blocking, matching both the Go-side implementation and the pattern used elsewhere in this file.
🐛 Proposed fix
- process.standardOutput = Pipe()
- process.standardError = Pipe()
+ process.standardOutput = FileHandle.nullDevice
+ process.standardError = FileHandle.nullDevice📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| process.standardOutput = Pipe() | |
| process.standardError = Pipe() | |
| do { | |
| try process.run() | |
| process.waitUntilExit() | |
| process.standardOutput = FileHandle.nullDevice | |
| process.standardError = FileHandle.nullDevice | |
| do { | |
| try process.run() | |
| process.waitUntilExit() |
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@CLI/cmux.swift` around lines 10381 - 10385, The code currently sets
process.standardOutput and process.standardError to Pipe() then calls try
process.run() and process.waitUntilExit(), which can block because the pipes are
never read; change those assignments to use FileHandle.nullDevice for both
standardOutput and standardError (replace the Pipe() uses), so output is
discarded instead of creating unread pipes before calling process.run() and
process.waitUntilExit(); update references around process.standardOutput,
process.standardError, process.run(), and process.waitUntilExit() accordingly.
| term_value = read_text(term_log) | ||
| if term_value != "screen-256color": | ||
| print(f"FAIL: expected TERM=screen-256color, got {term_value!r}") | ||
| if term_value != "xterm-ghostty": | ||
| print(f"FAIL: expected TERM=xterm-ghostty, got {term_value!r}") |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
# Description: Check whether this host should expect xterm-ghostty or the fallback.
if command -v infocmp >/dev/null 2>&1 && infocmp xterm-ghostty >/dev/null 2>&1; then
printf 'expected TERM=xterm-ghostty\n'
else
printf 'expected TERM=xterm-256color\n'
fiRepository: manaflow-ai/cmux
Length of output: 88
🏁 Script executed:
cat -n tests/test_cli_claude_teams_env.py | sed -n '200,215p'Repository: manaflow-ai/cmux
Length of output: 949
🏁 Script executed:
# Search for TERM-related code to understand the fallback logic
rg -n "xterm-ghostty|xterm-256color" --type pyRepository: manaflow-ai/cmux
Length of output: 2524
🏁 Script executed:
# Search for infocmp usage in the codebase
rg -n "infocmp" --type pyRepository: manaflow-ai/cmux
Length of output: 903
🏁 Script executed:
# Search for where TERM default is set (likely in main claude-teams launch code)
rg -n "TERM.*=|DEFAULT.*TERM|fallback.*TERM" --type py | head -20Repository: manaflow-ai/cmux
Length of output: 2597
🏁 Script executed:
# Look for claude_teams or run_claude_teams function definition to understand the env setup
rg -n "def.*claude_teams|def.*run_claude_teams" --type pyRepository: manaflow-ai/cmux
Length of output: 119
🏁 Script executed:
# Check what the test file imports and setup to see if it mocks infocmp
rg -n "import|mock|patch|monkeypatch" tests/test_cli_claude_teams_env.py | head -20Repository: manaflow-ai/cmux
Length of output: 231
🏁 Script executed:
# Find the claude-teams main script/implementation
fd "claude[_-]teams" --type f | grep -v test | head -10Repository: manaflow-ai/cmux
Length of output: 102
🏁 Script executed:
# Look for TERM fallback logic in non-test code
rg -n "xterm-ghostty|xterm-256color" --type py | grep -v testRepository: manaflow-ai/cmux
Length of output: 42
🏁 Script executed:
# Read more of the test to understand the full context
cat -n tests/test_cli_claude_teams_env.py | sed -n '1,50p'Repository: manaflow-ai/cmux
Length of output: 1943
🏁 Script executed:
# Search for claude-teams implementation (might be shell script, Python, or compiled)
find . -name "*claude*teams*" -type f ! -path "./.git/*" ! -path "*/node_modules/*" ! -path "*/__pycache__/*" 2>/dev/nullRepository: manaflow-ai/cmux
Length of output: 457
🏁 Script executed:
# Check if there's a cmux binary or shell script
fd "cmux" --type f ! -path "./.git" ! -path "*/node_modules" | head -20Repository: manaflow-ai/cmux
Length of output: 228
🏁 Script executed:
# Look for environment variable handling in the codebase
rg -n "TERM.*=" --type sh | head -20Repository: manaflow-ai/cmux
Length of output: 1456
🏁 Script executed:
# Check the claude_teams_test_utils to understand test setup
cat tests/claude_teams_test_utils.pyRepository: manaflow-ai/cmux
Length of output: 771
🏁 Script executed:
# Look for where the claude-teams subprocess's TERM is actually set
# Maybe in the cmux CLI code itself
find . -name "*.py" -type f ! -path "./.git/*" ! -path "*/node_modules/*" ! -path "*/__pycache__/*" | xargs grep -l "claude.teams" | grep -v testRepository: manaflow-ai/cmux
Length of output: 42
🏁 Script executed:
# Search for TERM setting in shell scripts or config
rg -n "TERM.*ghostty" --type shRepository: manaflow-ai/cmux
Length of output: 142
🏁 Script executed:
# The test calls a cmux CLI binary. Let's see what the test is actually testing
cat -n tests/test_cli_claude_teams_env.py | sed -n '100,180p'Repository: manaflow-ai/cmux
Length of output: 3695
🏁 Script executed:
# Look for any implementation that sets TERM in the claude-teams context
rg -n "infocmp.*xterm-ghostty|ghostty.*terminfo" -A 3 -B 3Repository: manaflow-ai/cmux
Length of output: 14991
🏁 Script executed:
# Check if there's any reference to fallback behavior in the PR or comments
rg -n "fallback|xterm-256color" tests/test_cli_claude_teams_env.pyRepository: manaflow-ai/cmux
Length of output: 102
Account for the documented xterm-256color fallback.
This assertion hard-codes xterm-ghostty, but the runtime implementation (in both Swift and Go) prefers xterm-ghostty only when its terminfo entry is available, otherwise falls back to xterm-256color. On CI images without that terminfo entry (like this sandbox), this test will fail even when the fallback works correctly.
The proposed fix adds a helper function that mirrors the actual resolveDefaultTerm() logic already implemented in the codebase:
🐛 Proposed fix
def read_text(path: Path) -> str:
if not path.exists():
return ""
return path.read_text(encoding="utf-8").strip()
+def expected_default_term() -> str:
+ try:
+ proc = subprocess.run(
+ ["infocmp", "xterm-ghostty"],
+ stdout=subprocess.DEVNULL,
+ stderr=subprocess.DEVNULL,
+ check=False,
+ )
+ except FileNotFoundError:
+ return "xterm-256color"
+ return "xterm-ghostty" if proc.returncode == 0 else "xterm-256color"
+
+
def run_claude_teams( term_value = read_text(term_log)
- if term_value != "xterm-ghostty":
- print(f"FAIL: expected TERM=xterm-ghostty, got {term_value!r}")
+ expected_term = expected_default_term()
+ if term_value != expected_term:
+ print(f"FAIL: expected TERM={expected_term}, got {term_value!r}")
raise SystemExit(1)🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@tests/test_cli_claude_teams_env.py` around lines 205 - 207, The test
currently asserts term_value == "xterm-ghostty" unconditionally; change it to
accept the documented fallback by mirroring the production resolveDefaultTerm
logic: determine whether the runtime prefers "xterm-ghostty" or falls back to
"xterm-256color" (either by calling/importing the existing resolveDefaultTerm
helper or implementing the same detection in the test) and then assert
term_value equals the resolved default instead of hard-coding "xterm-ghostty"
(refer to term_log/term_value and resolveDefaultTerm in the codebase).
| def run_omo(cli_path: str, base_env: dict[str, str]) -> subprocess.CompletedProcess[str]: | ||
| with tempfile.TemporaryDirectory(prefix="cmux-omo-env-") as td: | ||
| tmp = Path(td) | ||
| real_bin = tmp / "real-bin" | ||
| real_bin.mkdir(parents=True, exist_ok=True) | ||
|
|
||
| term_log = tmp / "term.log" | ||
| term_program_log = tmp / "term-program.log" | ||
| colorterm_log = tmp / "colorterm.log" | ||
| cmux_bin_log = tmp / "cmux-bin.log" | ||
| argv_log = tmp / "argv.log" | ||
| fake_home = tmp / "home" | ||
| fake_home.mkdir(parents=True, exist_ok=True) | ||
|
|
||
| # Fake opencode shim — logs env on every invocation (the final `--version` | ||
| # call wins, overwriting earlier writes from any plugin-install probes). | ||
| make_executable( | ||
| real_bin / "opencode", | ||
| """#!/usr/bin/env bash | ||
| set -euo pipefail | ||
| printf '%s\\n' "${TERM-__UNSET__}" > "$FAKE_TERM_LOG" | ||
| printf '%s\\n' "${TERM_PROGRAM-__UNSET__}" > "$FAKE_TERM_PROGRAM_LOG" | ||
| printf '%s\\n' "${COLORTERM-__UNSET__}" > "$FAKE_COLORTERM_LOG" | ||
| printf '%s\\n' "${CMUX_OMO_CMUX_BIN-__UNSET__}" > "$FAKE_CMUX_BIN_LOG" | ||
| printf '%s\\n' "$@" > "$FAKE_ARGV_LOG" | ||
| exit 0 | ||
| """, | ||
| ) |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
# Look at the omo code path in CLI/cmux.swift to see whether --version bypasses plugin setup.
rg -nP --type=swift -C4 'runOMO|omoEnsurePlugin|oh-my-opencode'Repository: manaflow-ai/cmux
Length of output: 8754
🏁 Script executed:
#!/bin/bash
# Check if there's early --version handling before runOMO() is called
# Look around line 1783 where runOMO is called
sed -n '1770,1790p' CLI/cmux.swift | cat -nRepository: manaflow-ai/cmux
Length of output: 733
🏁 Script executed:
#!/bin/bash
# Check the beginning of runOMO function and whether it handles --version
sed -n '11073,11120p' CLI/cmux.swift | cat -nRepository: manaflow-ai/cmux
Length of output: 2605
🏁 Script executed:
#!/bin/bash
# Check the test to see if it pre-seeds the plugin or sets env vars to prevent install
sed -n '1,100p' tests/test_cli_omo_env.py | cat -nRepository: manaflow-ai/cmux
Length of output: 4382
🏁 Script executed:
#!/bin/bash
# Check omoEnsurePlugin signature and if it receives commandArgs or checks arguments
sed -n '10896,10910p' CLI/cmux.swift | cat -nRepository: manaflow-ai/cmux
Length of output: 979
🏁 Script executed:
#!/bin/bash
# Check if commandArgs (containing --version) are used anywhere to short-circuit plugin setup
rg -n 'commandArgs' CLI/cmux.swift | grep -i 'omo\|11073' | head -20Repository: manaflow-ai/cmux
Length of output: 277
🏁 Script executed:
#!/bin/bash
# Check if there's any env var that controls plugin installation
rg -n 'OPENCODE\|omoEnsure' CLI/cmux.swift | head -30Repository: manaflow-ai/cmux
Length of output: 42
The test will trigger a real plugin installation whenever bun or npm are available on the test host.
The Swift cmux omo path does not gate off or short-circuit plugin setup based on --version. The runOMO() function calls omoEnsurePlugin() unconditionally at line 11103, before any argument parsing. The function receives commandArgs: [String] but does not pass it to omoEnsurePlugin(), so there is no mechanism to bypass plugin installation when --version is requested.
Since the test sets PATH to prepend a fake bin directory but still includes the original PATH, any bun or npm installed on the test host will be found and used by omoEnsurePlugin() (line 10956–10989) to install oh-my-opencode into fake_home/.config/opencode/. This makes the test slow, network-dependent, and flaky on CI.
Pre-seeding a fake plugin directory in fake_home or exporting an opt-out environment variable would harden the test.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@tests/test_cli_omo_env.py` around lines 33 - 60, The test currently allows
real plugin installation because run_omo (and the Swift runOMO path) calls
omoEnsurePlugin unconditionally and PATH still contains the host's npm/bun; fix
by ensuring the plugin setup is short-circuited during tests: either pre-seed a
fake plugin installation directory under the test's fake_home (create
fake_home/.config/opencode with expected files) before invoking run_omo, or
export an opt-out env var (e.g., CMUX_OMO_SKIP_PLUGIN=1) in base_env so
omoEnsurePlugin returns early; locate calls to run_omo/runOMO and
omoEnsurePlugin to implement the pre-seed or opt-out behavior and ensure
fake_home and base_env are used consistently so the test cannot reach the
network to install plugins.
| term_value = read_text(term_log) | ||
| if term_value != "xterm-ghostty": | ||
| print( | ||
| f"FAIL: expected TERM=xterm-ghostty for cmux omo, got {term_value!r}" | ||
| ) | ||
| raise SystemExit(1) |
There was a problem hiding this comment.
Host-dependent assertion may cause CI flakiness.
The Go tests (agent_launch_test.go:217-235) deliberately skip the xterm-ghostty expectation when the terminfo entry isn't installed, because resolveDefaultTerm() falls back to xterm-256color. This Python test asserts TERM == "xterm-ghostty" unconditionally and will fail on any host where the xterm-ghostty terminfo entry isn't present (fresh Linux CI, minimal containers, etc.).
Consider mirroring the Go test's probe-and-skip pattern, or set CMUX_CLAUDE_TEAMS_TERM/CMUX_OMO_TERM explicitly so the test doesn't depend on the host's terminfo database.
🔧 Probe-and-skip sketch
+import shutil
+
+def has_xterm_ghostty_terminfo() -> bool:
+ if shutil.which("infocmp") is None:
+ return False
+ return subprocess.run(
+ ["infocmp", "xterm-ghostty"],
+ stdout=subprocess.DEVNULL,
+ stderr=subprocess.DEVNULL,
+ ).returncode == 0
+
@@
- term_value = read_text(term_log)
- if term_value != "xterm-ghostty":
- print(
- f"FAIL: expected TERM=xterm-ghostty for cmux omo, got {term_value!r}"
- )
- raise SystemExit(1)
+ term_value = read_text(term_log)
+ expected_term = "xterm-ghostty" if has_xterm_ghostty_terminfo() else "xterm-256color"
+ if term_value != expected_term:
+ print(
+ f"FAIL: expected TERM={expected_term} for cmux omo, got {term_value!r}"
+ )
+ raise SystemExit(1)🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@tests/test_cli_omo_env.py` around lines 99 - 104, The test unconditionally
asserts TERM == "xterm-ghostty" which flakes when the terminfo entry is missing;
change the test to either set CMUX_OMO_TERM (or CMUX_CLAUDE_TEAMS_TERM) to a
known value before launching the agent so the produced term file (read_text ->
term_value) is deterministic, or probe for the terminfo entry and skip the
assertion if it's absent (mirror Go's resolveDefaultTerm probe-and-skip). Locate
the block that reads term_value (term_value = read_text(term_log)) and implement
one of these fixes: export os.environ["CMUX_OMO_TERM"]="xterm-ghostty" for the
test run, or run a quick check (e.g., subprocess call to infocmp or use
shutil.which) to detect the xterm-ghostty terminfo and only perform the equality
assertion when present. Ensure references remain to term_value, read_text, and
CMUX_OMO_TERM so the change is easy to find.
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
There was a problem hiding this comment.
2 issues found across 6 files
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="CLI/cmux.swift">
<violation number="1" location="CLI/cmux.swift:10381">
P2: Use `FileHandle.nullDevice` instead of `Pipe()` for stdout/stderr. These pipes are never drained before `waitUntilExit()`, which can deadlock if the subprocess writes enough to fill the pipe buffer (~64KB). Since the output of `infocmp` is discarded anyway, `FileHandle.nullDevice` is the correct choice — and matches the Go-side pattern (`cmd.Stdout = nil`).</violation>
</file>
<file name="tests/test_cli_claude_teams_env.py">
<violation number="1" location="tests/test_cli_claude_teams_env.py:206">
P2: This assertion is host-dependent and will fail on CI runners (or any host) without the `xterm-ghostty` terminfo entry installed. The runtime's `resolveDefaultTerm()` falls back to `xterm-256color` when `infocmp xterm-ghostty` fails, but this test doesn't account for the fallback. The Go-side tests already handle this correctly by probing and skipping. Either probe `infocmp xterm-ghostty` to determine the expected value, or pin `CMUX_CLAUDE_TEAMS_TERM` in the test env to make the assertion deterministic.</violation>
</file>
Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.
| process.standardOutput = Pipe() | ||
| process.standardError = Pipe() |
There was a problem hiding this comment.
P2: Use FileHandle.nullDevice instead of Pipe() for stdout/stderr. These pipes are never drained before waitUntilExit(), which can deadlock if the subprocess writes enough to fill the pipe buffer (~64KB). Since the output of infocmp is discarded anyway, FileHandle.nullDevice is the correct choice — and matches the Go-side pattern (cmd.Stdout = nil).
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At CLI/cmux.swift, line 10381:
<comment>Use `FileHandle.nullDevice` instead of `Pipe()` for stdout/stderr. These pipes are never drained before `waitUntilExit()`, which can deadlock if the subprocess writes enough to fill the pipe buffer (~64KB). Since the output of `infocmp` is discarded anyway, `FileHandle.nullDevice` is the correct choice — and matches the Go-side pattern (`cmd.Stdout = nil`).</comment>
<file context>
@@ -10366,6 +10366,29 @@ struct CMUXCLI {
+ let process = Process()
+ process.executableURL = URL(fileURLWithPath: "/usr/bin/env")
+ process.arguments = ["infocmp", "xterm-ghostty"]
+ process.standardOutput = Pipe()
+ process.standardError = Pipe()
+ do {
</file context>
| process.standardOutput = Pipe() | |
| process.standardError = Pipe() | |
| process.standardOutput = FileHandle.nullDevice | |
| process.standardError = FileHandle.nullDevice |
| term_value = read_text(term_log) | ||
| if term_value != "screen-256color": | ||
| print(f"FAIL: expected TERM=screen-256color, got {term_value!r}") | ||
| if term_value != "xterm-ghostty": |
There was a problem hiding this comment.
P2: This assertion is host-dependent and will fail on CI runners (or any host) without the xterm-ghostty terminfo entry installed. The runtime's resolveDefaultTerm() falls back to xterm-256color when infocmp xterm-ghostty fails, but this test doesn't account for the fallback. The Go-side tests already handle this correctly by probing and skipping. Either probe infocmp xterm-ghostty to determine the expected value, or pin CMUX_CLAUDE_TEAMS_TERM in the test env to make the assertion deterministic.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At tests/test_cli_claude_teams_env.py, line 206:
<comment>This assertion is host-dependent and will fail on CI runners (or any host) without the `xterm-ghostty` terminfo entry installed. The runtime's `resolveDefaultTerm()` falls back to `xterm-256color` when `infocmp xterm-ghostty` fails, but this test doesn't account for the fallback. The Go-side tests already handle this correctly by probing and skipping. Either probe `infocmp xterm-ghostty` to determine the expected value, or pin `CMUX_CLAUDE_TEAMS_TERM` in the test env to make the assertion deterministic.</comment>
<file context>
@@ -203,13 +203,16 @@ def run_claude_teams(
term_value = read_text(term_log)
- if term_value != "screen-256color":
- print(f"FAIL: expected TERM=screen-256color, got {term_value!r}")
+ if term_value != "xterm-ghostty":
+ print(f"FAIL: expected TERM=xterm-ghostty, got {term_value!r}")
raise SystemExit(1)
</file context>
2 participants
Summary
Fixes #2947 —
cmux claude-teamscrashes during teammate→team-lead permission escalation on Claude Code v2.1.112+. Root cause: cmux's agent-launch shims unconditionallyunsetenv("TERM_PROGRAM")for every agent type, but Claude Code's new permission flow requiresTERM_PROGRAMto be set or itscli.jscrashes mid-prompt.This PR fixes both surfaces — the Go remote daemon (
cmuxd-remote) and the Swift local CLI — and adds the test coverage that was missing from the original report.Why this supersedes #2960
PR #2960 (by @bobo-xxx) correctly identified the bug and patched the Go daemon. However, the original repro in #2947 is local-Mac (
cmux claude-teamstyped directly into a terminal pane), which goes throughCLI/cmux.swift, notcmuxd-remote. So #2960 alone doesn't reach the reporter's crash path.Rather than open a Swift-only companion PR (and leave maintainers to coordinate two reviews / two test plans), this PR includes:
39527116)Happy to defer to whatever the maintainers prefer — see the comment I'll leave on #2960.
What's in this PR
Bug fix (verified end-to-end)
daemon/remote/cmd/cmuxd-remote/agent_launch.go) — gatesos.Unsetenv("TERM_PROGRAM")on a per-agent flag. claude-teams preserves; opencode-family (omo/omx/omc) unsets.CLI/cmux.swift) — same asymmetry on the local CLI path. Without this, the original reporter's local-Mac crash still reproduces.Adversarial-review improvements
preserveTermProgram(zero-value-safe naming) — CodeRabbit flagged the originalunsetTermProgramfield as having an unsafe zero value (false= preserve = the rare exception, not the historical default). Renamed across both surfaces; the zero value now matches the legacy "unset" behavior, so any new agent type defaults to the safe choice and only claude-teams opts in.infocmp xterm-ghosttyprobe — defaultingTERMtoxterm-ghosttyis the right call, but cmux's terminfo overlay installs asynchronously during shell bootstrap on remote SSH hosts (seetests_v2/test_ssh_remote_shell_integration.py:536-615which documents that fresh containers don't have the entry preinstalled). If an agent process exec's before that install completes, it inheritsTERM=xterm-ghosttywithout the matching terminfo entry → ncurses falls back to a dumb terminal. Both surfaces now probeinfocmp xterm-ghosttyand fall back toxterm-256color(universally supported).COLORTERMread alignment — the Swift implementation read from aprocessEnvironmentsnapshot taken before any env mutations; Go readsos.Getenv(live). Functionally equivalent today but fragile to re-orderings. Both surfaces now read live for parity.Comment accuracy — opencode reacts to any non-empty
TERM_PROGRAM(treats it as an outer-host marker), not specifically=ghosttyas the original comment implied. Updated in both files.Tests
tests/test_cli_claude_teams_env.pyTERM=xterm-ghostty+ preservedTERM_PROGRAMcontract for claude-teams (was: asserted the old contract; would have broken on this PR). Updated, not new.tests/test_cli_omo_env.pycmux omoassertingTERM_PROGRAMis unset (regression guard for #2516 — opencode's light-theme switch on any non-emptyTERM_PROGRAM),TERM=xterm-ghostty, andCOLORTERM=truecolorfallback.daemon/remote/cmd/cmuxd-remote/agent_launch_test.goconfigureAgentEnvironmentmatrix (preserve / unset / TERM override / COLORTERM preservation / extraEnv) plus 2 tests forresolveDefaultTerm(terminfo race success + fallback).daemon/remote/cmd/cmuxd-remote/tmux_compat_test.goTERM_PROGRAMunset assertion to the existingTestConfigureAgentEnvironment(the test setTERM_PROGRAM=should-be-removedbut never asserted the result).Per CLAUDE.md's regression-test policy, the original test/fix pair (commits
8369e3c2→3b2b37b2) is a two-commit failing-test-then-fix sequence so CI on the test commit alone goes red — proving the test catches the bug.Test plan
I've verified the GUI repro locally on macOS 26.4.1 / Apple Silicon. The original crash is gone; opencode's dark theme is intact.
Out of scope (flagged for follow-up)
tests/test_cli_claude_teams_env.py(the--max-old-space-size 2048 --trace-warningsheap-flag scenario) fails on Node 24 because Node 24 rejects the space-separated form (--max-old-space-size 2048) and now requires--max-old-space-size=2048. Pre-existing; unrelated to this PR. CI's macOS VM probably runs an older Node where it still passes.Related
🤖 Generated with Claude Code
Summary by cubic
Preserves TERM_PROGRAM for
claude-teamsacross the Go daemon and Swift CLI to fix the permission-escalation crash in Claude Code v2.1.112+ (Fixes #2947). Also hardens terminal env defaults and adds coverage to prevent regressions.Bug Fixes
preserveTermProgram;claude-teamspreserves, opencode‑family (omo/omx/omc) unsets.cmux claude-teams.xterm-ghosttywheninfocmp xterm-ghosttysucceeds, else fall back toxterm-256color.COLORTERM=truecolorwhen absent to ensure truecolor rendering.tests/test_cli_omo_env.py; updatedtests/test_cli_claude_teams_env.py.Refactors
preserveTermProgram(safe default: unset); readCOLORTERMfrom live env; clarify comments about opencode reacting to any non‑empty TERM_PROGRAM.Written for commit 4f02421. Summary will update on new commits.
Summary by CodeRabbit
Bug Fixes
Tests