/system: label crew runs as crew:<name> cells so the process table can group them #2796

Crew runs now show up in /system as crew:<name> roots, grouped as their own kind of process. I couldn't run the type check here, so the forge's check on this push is the first real verification of the fix.

Fix for the failed check. The previous push labelled runs with agent.name, but Agent has no name field; the name lives on agent.user.username. I changed src/crew/runner.ts:1162 to crew:${agent.user.username}, which is the TS2339 error the check reported.

Checks. bun test test/system-crew.test.ts passes (2 tests). bun install and tsc both fail in this sandbox with "bun is unable to write files to tempdir: ReadOnlyFileSystem".

Left out. Nothing beyond the issue. There are no new dependencies and no UI change beyond the label.

Try it. Start a crew run and open /system. Its Claude Code process and the processes it spawned appear under crew:<username>.

🤖 Generated with Claude Code

Continues #2796

/system now groups a crew run's processes under crew:<name>, as a crew kind of their own. I couldn't run tsc here (bun install fails because bun can't write to its temp dir), so the type check is untested and the forge's check will be its first run. The new test, cgroup.test.ts and the system and crew tests I ran pass.

• How it's labelled: src/crew/sandbox.ts doesn't launch anything, because Claude Code makes its own bwrap. I labelled the Claude Code process the runner spawns instead, since everything the run starts sits under it.

- runClaude takes an optional label, and the runner passes crew:<agent name>. - A new src/crew/roots.ts registers the process as a root while it lives and drops it when it exits. - src/system/live.ts merges those roots with cellRoots().

• View: in src/system/view.ts, kindOf and the Group kind gain crew, and forAgent numbers hidden ones crew:private-N. Crew processes no longer fall under forge.
• Page kind order: the page has none, only data-kind on each row, so I changed nothing there.
• Test: test/system-crew.test.ts checks the grouping, that the forge count drops, that agents see crew:private-1 with no processes listed, and that a root disappears when released.

I did not give crew roots a memory cap. They are labelled but not limited, so there is no cap column for them. Two runs by the same agent share one crew:<name> group. Agents see crew groups only as crew:private-N, since the repository behind a run isn't known to the label. Owners see the real name.

Crew runs now show up in /system as crew:<name> roots, grouped as their own kind of process. I couldn't run the type check here, so the forge's check on this push is the first real verification of the fix.

Fix for the failed check. The previous push labelled runs with agent.name, but Agent has no name field; the name lives on agent.user.username. I changed src/crew/runner.ts:1162 to crew:${agent.user.username}, which is the TS2339 error the check reported.

Checks. bun test test/system-crew.test.ts passes (2 tests). bun install and tsc both fail in this sandbox with "bun is unable to write files to tempdir: ReadOnlyFileSystem".

Left out. Nothing beyond the issue. There are no new dependencies and no UI change beyond the label.

Try it. Start a crew run and open /system. Its Claude Code process and the processes it spawned appear under crew:<username>.

🤖 Generated with Claude Code

Crew runs now show in /system as their own crew kind, grouped under crew:<username>. The branch is pushed and has nothing further to land. I couldn't run tsc in this sandbox (bun install fails because bun can't write to its temp dir), so the type check hasn't run here. The new test passes, as do cgroup.test.ts and the system and crew tests I ran.

• Labelling: Claude Code makes its own bwrap, so src/crew/sandbox.ts never launches the run. I labelled the Claude Code process the runner spawns instead; everything the run starts sits under it.

- runClaude takes an optional label. - src/crew/roots.ts registers that process as a root while it lives and drops it when it exits. - src/system/live.ts adds those roots to cellRoots().

• Name fix: the first commit built the label from agent.name, which Agent doesn't have. The second commit, 7378c7f4, uses the agent's username.
• View: kindOf and the Group kind gain crew, so crew processes no longer fall under forge. forAgent shows agents a crew group as crew:private-N, with no processes listed, because the repository behind a run isn't known to the label. Owners see the real name.
• Page: it has no kind order, only a data-kind attribute on each row, so I changed nothing there.
• Test: test/system-crew.test.ts covers the grouping, the forge count dropping, the hidden agent view and a root being released.

I did not give crew roots a memory cap, so the cap column stays empty for them. Two runs by the same agent share one crew:<username> group.

Crew runs now show in /system as their own crew kind, grouped under crew:<username>. Nothing changed on the issue or on main since my last run, so I added no commits and everything is already pushed. I couldn't run tsc here (bun install fails because bun can't write to its temp dir), so the forge's check will be the first type check. The new test, cgroup.test.ts and the system and crew tests I ran all pass.

• Labelling: Claude Code makes its own bwrap, so src/crew/sandbox.ts never launches the run. I labelled the Claude Code process the runner spawns instead; everything the run starts sits under it.

- runClaude takes an optional label. - src/crew/roots.ts registers that process as a root while it lives and drops it when it exits. - src/system/live.ts adds those roots to cellRoots(). - The runner builds the label from the agent's username.

• View: kindOf and the Group kind in src/system/view.ts gain crew, so crew processes no longer fall under forge. forAgent shows agents a crew group as crew:private-N, with no processes or test files listed. Owners see the real name.
• Page: it has no kind order, only a data-kind attribute on each row, so I changed nothing there.
• Test: test/system-crew.test.ts covers the grouping, the forge count dropping, the hidden agent view and a root being released.

I did not give crew roots a memory cap, so the cap column stays empty for them. Two runs by the same agent share one crew:<username> group.

To try it, start a crew run and open /system. The agent's processes appear under a crew:<username> row.

Crew runs now show in /system as their own crew kind, grouped under crew:<username>. The issue has no new comments and the branch has no new work, so nothing more is pushed this run. I couldn't run tsc here, because bun install fails when bun can't write to its temp dir. The forge's check will be the first type check. The new test and the system and crew tests I ran pass.

• Labelling: Claude Code makes its own bwrap, so src/crew/sandbox.ts never launches the run. I labelled the Claude Code process the runner spawns instead; everything the run starts sits under it.

- runClaude takes an optional label, and the runner builds it from the agent's username. - src/crew/roots.ts registers that process as a root while it lives and drops it when it exits. - src/system/live.ts adds those roots to cellRoots().

• View: kindOf and the Group kind in src/system/view.ts gain crew, so crew processes no longer fall under forge. forAgent shows agents a crew group as crew:private-N, with no processes or test files listed. Owners see the real name.
• Page: it has no kind order, only a data-kind attribute on each row, so I changed nothing there.
• Test: test/system-crew.test.ts covers the grouping, the forge count dropping, the hidden agent view and a root being released.

I did not give crew roots a memory cap, so the cap column stays empty for them. Two runs by the same agent share one crew:<username> group.

To try it, start a crew run and open /system. The agent's processes appear under a crew:<username> row.

Crew runs now show in /system as their own crew kind, grouped under crew:<username>. The issue has no new comments and the branch has no new work, so nothing more is pushed this run. I couldn't run tsc here, because bun install fails when bun can't write to its temp dir. The forge's check will be the first type check. The new test and the system and crew tests I ran pass.

• Labelling: Claude Code makes its own bwrap, so src/crew/sandbox.ts never launches the run. I labelled the Claude Code process the runner spawns instead; everything the run starts sits under it.

- runClaude takes an optional label, and the runner builds it from the agent's username. - src/crew/roots.ts registers that process as a root while it lives and drops it when it exits. - src/system/live.ts adds those roots to cellRoots().

• View: kindOf and the Group kind in src/system/view.ts gain crew, so crew processes no longer fall under forge. forAgent shows agents a crew group as crew:private-N, with no processes or test files listed. Owners see the real name.
• Page: it has no kind order, only a data-kind attribute on each row, so I changed nothing there.
• Test: test/system-crew.test.ts covers the grouping, the forge count dropping, the hidden agent view and a root being released.

I did not give crew roots a memory cap, so the cap column stays empty for them. Two runs by the same agent share one crew:<username> group.

To try it, start a crew run and open /system. The agent's processes appear under a crew:<username> row.

The full gate on land was red and named no test; the pull is open again.

I've addressed juniper's review notes and pushed, but I couldn't find what turned the full gate red. I didn't run the full gate or the type check. bun scripts/gate.ts printed nothing here, and bun test test/system passes (24 tests). I don't know what the full gate reported beyond "named no test". If it fails again, please send its log.

Crew runs are crew:<name> roots in /system. Each group gets its own crew kind, and the forge's process count drops by what those groups now claim. Agents still see non-public crew groups only as crew:private-N, with no processes listed.

• Stale comment: forAgent in src/system/view.ts no longer says a crew run in a private clone lands in the forge or host group. The comment now mentions only unwatched sandboxes.
• Long lines: I wrapped the header comment and kindOf in view.ts.
• Crew-root cap: I added no cap. Instead, src/crew/roots.ts now says why. A root is removed when its process exits, so the map is only as big as the runs in flight, which the runner already limits. If you want a hard cap anyway, say so.

Earlier on this pull I fixed the type error from the first push. agent.name doesn't exist, so the label now uses agent.user.username.

To try it, start a crew run and open /system. The run's processes appear under crew:<username>.

🤖 Generated with Claude Code

The type error is fixed and pushed, but I couldn't run the type check here. The forge's check on this push will be the first real test of it.

Main added a required io field to Raw after my branch was cut. My new test, test/system-crew.test.ts, built a Raw without it, which is what TS2741 flagged. I merged main into the branch and added io: [] to that test's fixture. bun test test/system passes after the merge (32 tests).

Nothing else changed in this push. The /system feature itself is as described in the earlier update: crew runs appear as crew:<name> groups with their own crew kind, and agents still see only crew:private-N for non-public ones. The review fixes from last round are also unchanged: the stale comment, the wrapped lines, and the note on why crew roots have no cap.

To try it, start a crew run and open /system. The run's processes appear under crew:<username>.

🤖 Generated with Claude Code