fix(spawn): write launch command to a file instead of typing it whole - #95
Merged
Merged
Conversation
…e command A launch typed while the pane shell runs a pre-prompt hook (mise's, after each typed export) waits in the terminal's canonical input, which keeps only its first 1024 bytes on macOS and drops the rest with the Enter behind it. The shell is left holding a cut command, the agent never starts, and the in-place retry types into the same race. fm-spawn now writes the launch command to a private launch.sh in the per-task temp root and types only a short line sourcing it, so the pane shell evaluates the same command without ever receiving a long typed line. Every backend launched through that one typing site. The no-agent-started confirmation stays as the safety net. The launch-confirm regression gains a case whose pane shell waits on a slow prompt hook for every typed line, and the cut-launch shim now cuts the short line. Test fakes that log the typed launch read a sourced launch file as the command it holds.
…t fake's typed launch literal through the fixture helper fm_fake_sourced_launch (from tests/fixtures.sh), which resolves the ". '/path/to/launch-file'" source line fm-spawn.sh now types into the underlying launch command text a fake previously matched directly. This was needed because bin/fm-spawn.sh (base commit 4a9d372) changed from typing the full launch command literally to typing a short line that sources a launch file, which several test fakes still matched/logged as if it were the old literal. Files fixed (all test-only, no product code touched, no assertions weakened): - tests/fm-control-relaunch.test.sh: make_tmux_stub now resolves the -l payload through fm_fake_sourced_launch before logging/matching "encode launch-brief". - tests/fm-control.test.sh: same fix in its make_tmux_stub (the noted dead branch, kept in step). - tests/fm-secondmate-restart.test.sh: same fix in make_stub; now sources tests/fixtures.sh instead of tests/lib.sh. - tests/secondmate-helpers.sh: make_fake_tmux's send-keys logging now resolves the -l literal via fm_fake_sourced_launch before writing to FM_FAKE_TMUX_LOG; now sources tests/fixtures.sh. - tests/fm-backend-orca.test.sh: make_orca_fakebin resolves the argument following --text through fm_fake_sourced_launch before logging; now sources tests/fixtures.sh. - tests/remote-herdr-fixture.sh: install_remote_herdr_fixture resolves the `pane send-text` payload (4th arg) through fm_fake_sourced_launch before appending to the log; now sources tests/fixtures.sh. This fixture is shared by fm-remote-secondmate-parent-binding.test.sh (the exact CI-failing test, which greps this log for FM_PUBLIC_FOLLOWUP_PRIMARY_HOME) and fm-remote-secondmate-trace-context.test.sh (whose GOTMPDIR/TRACEPARENT/FM_TRACE_CONTEXT assertions match separate `pane send-keys` calls, unaffected by this change). Verified locally (via mutex-serialized runs): tests/fm-control.test.sh, tests/fm-control-relaunch.test.sh, tests/fm-secondmate-restart.test.sh, tests/fm-secondmate-lifecycle-e2e.test.sh, and tests/fm-backend-orca.test.sh all pass with zero failures. tests/fm-remote-secondmate-parent-binding.test.sh (the actual CI-failing script in shard 1) was still running via mutex-serialized background job when this response was finalized; I have not fabricated its result. All syntax-checked with bash -n. No product code was modified, consistent with the user's explicit instruction to change fakes only
marano
added a commit
that referenced
this pull request
Sep 24, 2026
…led (#115) Every crewmate, scout, and secondmate now starts with COMPACT_ADVISER_DISABLE=1, on a fresh spawn and on a relaunch alike, so an unattended session never activates the compact adviser. The value is unconditional. Two carriers deliver it: an export in the pane shell beside GOTMPDIR, and an export at the head of the launch command, which also covers a compound raw launch and survives the cleared launch environment. Also pins the launch-file typing from #95 with a case that spawns the longest real launch (long task id, effort flag, allowlist of thirty names) and bounds the longest typed line.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Intent
Observed 2026-09-23 on the tmux backend:
bin/fm-spawn.sh fm-nondeterministic-test-pair ... --harness claudeprinted "spawned ..." but the worker never started. The pane showed the launch command garbled at the end of the trust statement - "...absent from the brief.' -env -u CURSOR_AGENT -u CURSOR_INVOKED_AS -u GEnoenv -u CURSOR_AGENT -u CURSOR_INVOKED_AS -u GEMINI_CLI CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=fate" - and zsh answered "error: unknown option '-env'", leaving a bare shell prompt. Theenv -u ...prefix appears duplicated and interleaved, as if the typed command raced the shell's startup or was typed twice.Consequence: a spawn reports success for a worker that is not running; only a pane read caught it.
fm-control.sh relaunch --noterecovered it cleanly.Wanted: spawn verifies the agent process actually started (or the pane left the shell prompt) before reporting success, and the typed launch cannot interleave with shell startup. Reproduce first; one observation so far.
Group key: solo
SECOND OCCURRENCE 2026-09-23, about 20 minutes later, again a firstmate-repo spawn (fm-crew-state-hides-pipeline-custody, claude opus high, tmux). This time fm-spawn itself caught it - "no agent started in window ... after the launch command was typed (the endpoint reads 'dead')" - and rolled back: no window left, item still queued. An immediate retry succeeded. So the success-reporting half may already be covered by the check that caught this one; the garbling itself (two of four spawns that night, both firstmate-repo, bluejam spawns fine in between) is the open defect.
What Changed
bin/fm-spawn.shnow writes the full launch command to a privatelaunch.shin the per-task temp root and types only a short. <launch.sh>line into the pane, instead of typing the whole launch command; the temp root is validated to be a real, non-symlinked directory owned by the current user before the file is written, andspawn_type_launchsends the short sourcing line (LAUNCH_LINE) rather than the rawLAUNCHstring.tests/fm-spawn-launch-confirm.test.shadds a case reproducing the actual cause (a pane shell whose prompt runs a slow pre-prompt hook, keeping the shell out of its line editor while text is typed) and updates the tmux shim to garble the new sourcing line instead of the old literal;tests/fixtures.shgains a sharedfm_fake_sourced_launch/fm_test_fake_sourced_launch_fnhelper so faketmuxscripts resolve a sourced launch file back to the underlying command for logging, and the harness test fixtures (fm-agy,fm-kimi,fm-muse,fm-rovo,fm-secondmate,fm-trace-context-spawn) are updated to use it..agents/skills/harness-adapters/references/harness/rovo.mdanddocs/verification/rovo.mdto note that the described send-keys shape was the then-current one and thatfm-spawn.shnow types a short line sourcing a launch file instead.Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com
Risk Assessment
✅ Low: The change is well-scoped: it writes the launch command to a mode-600 file and types a short, fixed-length sourcing line instead of the full command, directly addressing the reported truncation cause, with sound new ownership/symlink guards, unchanged pre-existing confirmation/rollback logic, and behavioral (not string-matching) test coverage.
Testing
Drove the real fm-spawn-launch-confirm regression suite (spawns into actual tmux panes/shells), including the new adversarial case that reproduces the incident's root cause — a pane shell running a slow pre-prompt hook outside its line editor, with a launch file exceeding macOS's 1024-byte canonical-input limit — and it starts the agent correctly; also confirmed the existing cut-launch and continuation-prompt recovery cases still pass with the new sourcing-line typing path. All six harness/trace-context suites touched by the fixture change (agy, kimi, rovo, secondmate, trace-context-spawn) pass in full; muse's one unrelated failing assertion was independently reproduced on the pre-change base commit, so it is a pre-existing flake, not a regression from this change.
ok - a launch typed while the pane shell runs a slow prompt hook starts the agentfrombash tests/fm-spawn-launch-confirm.test.sh, which types a real launch into a real tmux pane whose zsh shell s…ok - a spawn clears a cut launch's continuation prompt and starts the agent(existing case, now exercising the short sourcing-line path via LAUNCH_LINE)ok - a relaunch clears a continuation prompt before typing its launchok - a spawn whose launch never started fails and closes its panebash tests/fm-*-harness.test.sh/fm-trace-context-spawn.test.shruns pass (agy: 20/20, kimi: 17/17, rovo: 11/11, secondmate: 19/19, trace-context-spawn: 12/12), each using the updated `fm_fak…Pipeline
Updates from git push no-mistakes
✅ **intent** - passed
✅ No issues found.
✅ **Rebase** - passed
✅ No issues found.
✅ **Review** - passed
✅ No issues found.
✅ **Test** - passed
✅ No issues found.
ok - a launch typed while the pane shell runs a slow prompt hook starts the agentfrombash tests/fm-spawn-launch-confirm.test.sh, which types a real launch into a real tmux pane whose zsh shell s…ok - a spawn clears a cut launch's continuation prompt and starts the agent(existing case, now exercising the short sourcing-line path via LAUNCH_LINE)ok - a relaunch clears a continuation prompt before typing its launchok - a spawn whose launch never started fails and closes its panebash tests/fm-*-harness.test.sh/fm-trace-context-spawn.test.shruns pass (agy: 20/20, kimi: 17/17, rovo: 11/11, secondmate: 19/19, trace-context-spawn: 12/12), each using the updated `fm_fak…bash tests/fm-spawn-launch-confirm.test.sh(real tmux server, real pane shells)bash tests/fm-agy-harness.test.shbash tests/fm-kimi-harness.test.shbash tests/fm-muse-harness.test.shbash tests/fm-rovo-harness.test.shbash tests/fm-secondmate-harness.test.shbash tests/fm-trace-context-spawn.test.shBaseline check: re-rantests/fm-muse-harness.test.shagainst base commit 083de4f in a scratch worktree to confirm its one failure predates this change✅ **Document** - passed
✅ No issues found.
✅ **Lint** - passed
✅ No issues found.
✅ **Push** - passed
✅ No issues found.