[jorgemanrubia]

The connector cared where work happened. It routed a project to a directory, then made a git worktree of that directory for every task, branched it, tracked it in the ledger, kept it when the task ended, and gave you two commands to list and prune what it had kept. That was about 5,400 lines in worktree-only files and references in twenty-odd more.

It did bound a worker's edits, and that is worth stating accurately: the permission policy refuses an edit whose paths resolve outside the working directory, symlinks followed, without asking anyone — and with worktrees on, that directory was the worktree. But it is policy, not containment. Nothing in the kernel enforces it, so anything that escapes the tool layer — a shell, a path the check cannot see — is unstopped, which is what the Codex driver means when it says the sandbox is Codex's and not the connector's. This change moves that bound from a per-task worktree to the route itself, which is where it already sat whenever worktrees were off. It is neutral on safety: the case for it is not that it makes anything safer, but that worktrees answered a concurrency question we are not keeping, at the price of 5,400 lines and two public commands.

Worktrees existed so two tasks on one repository could run side by side. That is a concurrency answer, and it is not one we are keeping: two workers in one directory is managed by the people running them until sandboxes are real, and sandboxing is being built separately. The connector runs where it is started, and the agent decides the rest — from its skills, its CLAUDE.md, its memory. A task that needs a clone or a directory of its own makes one.

Originally tracked in The connector runs where it is started: no routes to directories, no worktrees. This is the worktrees half only; the path half of a route — Route.Path, --route <id>=<dir>, directory resolution — is untouched and comes next. It closes rather than fixes Two connector settings that do not say what they cost, A route waiting for a worktree is visible nowhere and Worktrees and a Codex worker: accepted, and nothing could commit, all of which were about worktrees behaving badly.

What the shape of it is

Worktrees turned out to be the only implementation of the Workspaces seam, so the seam went with them — PerTaskWorkspaces and WaitingWorkspaces are worktree concepts by name, and RecoveringWorkspaces existed to reconcile worktrees on start. A task's working directory is its route, exactly as it already was whenever worktrees were off, and the dispatcher prepares nothing before a launch and releases nothing after one.

That cascaded once more. Nothing can raise ErrRouteUnusable any more — only a worktree that could not be made ever did — so the whole refused-start path is unreachable and goes too: the ledger's refusal, its hook, its blocked reason, and the holding reply it posted. Route waits go with it; they were in-memory in the worktree preparer, and nothing else waits on a route.

Two things worth a reviewer's eye

The ledger keeps migration 9 and adds migration 12. A migration that has been applied is never taken out of the list: removing it would change what every later one means, and a fresh ledger has to walk the same numbers to reach the newest. So migration 9 stays byte for byte, unedited, and migration 12 drops the table it made. The directories are not touched — a worktree the connector kept is still on disk and still git's, and git worktree remove still removes it. A test upgrades a ledger that really ran 9, with a retained row in it, and proves nothing named worktrees* is left.

connect.json keeps a read-and-discard field for the old worktrees key, and it must stay. The file is parsed with DisallowUnknownFields, which is what makes a misspelled setting a refusal rather than a silent nothing. Every connect.json ever written carries "worktrees", so deleting the field outright would have stopped every connector that has been set up. It is zeroed after the decode and kept out of anything written, so the next connect setup takes the key out of the file for good.

What happens now when a route's directory is gone

Removing the refusal removed one way a person found out their work could not start, so: the worktrees-off path never checked the route at dispatch — Prepare was if w.off { return route, nil }, no stat, no refusal — so that error could only ever arise with worktrees on. A vanished or unreadable directory has always surfaced the same way and still does: the spawn fails, is retried once, and the record is blocked spawn_failed, which posts "the worker could not be started. Needs a person: basecamp connect redispatch <n>" on the card. No regression.

One thing the tests caught

Taking renderRefusedStart out broke the retraction test that proves an upgraded ledger answers both of the asks it inherits, and it broke it quietly: the stand-in body I put in the fixture had no redispatch ask in it, and a notice with no ask is a notice postedAsks correctly passes over. The assertion caught it because it counts — two retractions, not "the call returned no error" — and a test written the other way would have gone green on one. The fixture now carries the message that build really posted, ask included.

Notes

connect worktrees list and prune are public commands, so .surface records their removal and .surface-breaking acknowledges the 73 entries that go.

Two pieces of route-path machinery crossed over: the route-is-a-repository check in setup and the helper behind it were both gated on worktrees, so they became unreachable the moment worktrees went rather than waiting for the follow-up.

The last commit is comments only, in driver/codex and driver/acp, where other work is in flight. It can be dropped whole without touching anything else.

[Copilot]

🔵 Needs a closer look

It removes a large stateful subsystem, and legacy refusal-intent compatibility still needs correction.

Pull request overview

Removes connector-managed Git worktrees so tasks run directly in their routed directories.

Changes:

• Removes worktree creation, recovery, status, pruning, and refusal paths.
• Preserves legacy configuration parsing and drops obsolete ledger data through migration 12.
• Updates tests, skills, command metadata, and public surface snapshots.

[!TIP]
If you aren't ready for review, convert to a draft PR.
Click "Convert to draft" or run gh pr ready --undo.
Click "Ready for review" or run gh pr ready to reengage.

File summaries

File	Description
skills/basecamp/SKILL.md	Documents route-directory execution.
skills/basecamp-connect/SKILL.md	Removes worktree setup guidance.
skill-evals/cases/basecamp-connect/route-quoting.yml	Updates mocked connector output.
skill-evals/cases/basecamp-connect/not-ready-agent-reads.yml	Updates mocked connector output.
internal/connector/worktrees_route_unix_test.go	Removes worktree route tests.
internal/connector/worktrees_route_test.go	Removes worktree route tests.
internal/connector/setup/route_repository_test.go	Removes repository checks.
internal/connector/setup/file.go	Accepts and discards legacy settings.
internal/connector/setup/file_test.go	Tests legacy configuration compatibility.
internal/connector/setup/checks.go	Removes worktree route validation.
internal/connector/setup/apply.go	Removes worktree configuration changes.
internal/connector/setup/apply_test.go	Updates setup application tests.
internal/connector/retraction_test.go	Retains legacy retraction fixtures.
internal/connector/recovery_harness_test.go	Removes workspace instrumentation.
internal/connector/recovery_dispatch_test.go	Updates recovery assertions.
internal/connector/recovery_connector_test.go	Removes harness workspaces.
internal/connector/outbox.go	Removes refused-start key handling.
internal/connector/outbox_run.go	Simplifies holding-reply claims.
internal/connector/operator_status_test.go	Removes worktree status tests.
internal/connector/lifecycle.go	Removes refused-start lifecycle handling.
internal/connector/ledger.go	Adds the table-removal migration.
internal/connector/ledger_worktrees.go	Removes worktree ledger operations.
internal/connector/ledger_tasks.go	Removes route-refusal state handling.
internal/connector/ledger_status.go	Removes worktree status fields.
internal/connector/ledger_migrations_test.go	Tests upgrading an old ledger.
internal/connector/ledger_dispatch.go	Updates transition documentation.
internal/connector/driver/worker.go	Updates ownership documentation.
internal/connector/driver/drivertest/drivertest.go	Updates test-package documentation.
internal/connector/driver/codex/codex.go	Clarifies Codex sandbox behavior.
internal/connector/driver/acp/adapters.go	Updates preflight documentation.
internal/connector/dispatcher.go	Runs tasks directly in routes.
internal/connector/dispatcher_test.go	Removes workspace-specific tests.
internal/connector/dispatcher_boundary_test.go	Narrows release-point checks.
internal/connector/dispatch_route_refusal_test.go	Removes worktree refusal tests.
internal/connector/admission/policy.go	Updates configuration documentation.
internal/commands/connect.go	Removes worktree commands and flags.
internal/commands/connect_worktrees.go	Removes list and prune commands.
internal/commands/connect_worktrees_test.go	Removes command tests.
internal/commands/connect_setup_test.go	Updates setup scenarios.
internal/commands/connect_run.go	Removes workspace initialization.
internal/commands/connect_operator.go	Removes worktree status output.
internal/commands/connect_operator_test.go	Updates operator and doctor tests.
internal/commands/connect_doctor.go	Checks routed directories directly.
internal/commands/commands.go	Removes worktrees from the catalog.
e2e/smoke/smoke_lifecycle.bats	Removes obsolete smoke entries.
.surface-breaking	Records removed public commands and flags.
.surface	Removes obsolete CLI surface entries.

Review details

• Files reviewed: 49/49 changed files
• Comments generated: 1
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🔵 Needs a closer look

The broad runtime removal, destructive schema migration, compatibility handling, and public CLI changes warrant final human validation.

Review details

• Files reviewed: 49/49 changed files
• Comments generated: 0 new
• Review effort level: Balanced