perf(daemon): run the unit suite with a shared module registry

[zfy0701]

Cuts the daemon unit suite's CPU by ~69% by executing each module once per worker instead of once per test file. Replaces the CI-sharding approach that was previously on this branch (that diff is still reachable at 7243db15 if we want it later).

The measurement

packages/daemon was ~88% of the Unit Test job's 3m31s step (248 files, 3699 tests, 186.52s against the step's 211s), and import was 327.50s of its ~700s of worker time — module re-execution, not tests.

On one machine, full suite:

	wall	user+sys CPU
before	3m28s	8m55s
after	2m47s	2m48s

Wall moves less than CPU locally because much of this suite is idle, not busy — see the last section. CI is CPU-bound, so it should track the CPU number.

What had to change first

isolate: false shares module-level state between files in a worker, so anything the daemon kept at module scope becomes cross-file state. Two commits precede the flip:

1. WorkspaceManager (1414b269, 05619dec). Where an agent's git runs, how a path this process cannot see gets emptied, whether workspaces live in sandboxes, and the clone single-flight were four module-level bindings that Daemon.start() wrote. They are now instance state owned by the daemon, threaded to the three CP factories, SessionManager, and the standalone chat CLI.

This is a real production bug independent of tests: a process holding two daemons had one plane between them, so the second inherited the first's git runner and sandbox mode — its git could run through the other's sandbox channel, or a cluster clone could land on this disk. A k8s daemon beside a local one is exactly that shape.

Honest note: I predicted this refactor would unblock workspace.test.ts and workspace-git.test.ts under isolate: false. It did not — they were failing for the reason below. The refactor stands on the production bug, not on the test win I expected from it.

2. vi.mock is the one genuinely registry-sensitive construct. A mock is registered per file but rewires a module in the shared registry, so it either misses (the real module was already imported by an earlier file) or leaks into a later one. workspace.test.ts mocking simple-git surfaced as fatal: not a git repository — real git ran.

Measured directly rather than guessed: excluding the five files that call vi.mock leaves zero isolation failures across the other 246, and those five are the entire remaining set. So they get a second Vitest project that keeps isolate: true, and the other 246 stop paying for isolation they never used.

test/no-stray-vi-mock.test.ts fails if that list drifts in either direction, so a new vi.mock breaks at the seam instead of somewhere unrelated depending on which worker paired the two files.

Test plan

• Full daemon suite: 3789 tests, 13 failures in 6 files — all pre-existing sandbox-network flakes present on the base commit too, zero new
• Same comparison after each of the three commits, against a baseline captured on the same machine
• tsc -p tsconfig.typecheck.json — 0 errors; eslint, prettier --check clean

Two things this does not fix

Sharding. Still available at 7243db15 (2 shards, cost-balanced sequencer, reviewed and green). Worth re-deciding once CI shows what this change alone buys.

The daemon unit suite reaches slack.com. Daemon.start() makes a @slack/web-api call on every boot, and 54 test files construct a Daemon. On CI it is fast because Slack answers invalid_auth over HTTP and p-retry does not retry a successful response — but where the request fails at the transport layer the retry ladder costs ~4s per boot, which is why local wall clock here is mostly idle. It also already caused a real CI failure on this branch: the one test using boot(..., { distinctTokens: true }) makes two such calls and timed out at 5s. Worth stubbing at the seam.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FtPhRcPGtRX8RZqApUZRGu

[agentconnect-md-test]

Approved. I reviewed exact head 7243db153f268018fa09efcdc0ee2a0982799e72 and found no blocking regression.

The new root command is a true complement of the daemon package while still running the repository-level script tests, and the dedicated daemon job invokes the same vitest run behavior with only deterministic, total, disjoint shard selection added. The shared sequencer remains covered by the control-plane unit suite, and the reusable-workflow summary now includes the unit job plus both daemon shard summaries.

Verification: the local synthetic merge has exactly the supplied base/head parents, every changed file matches the trusted head, and the patch has no added trailing whitespace. GitHub CI passed Build, Check, Unit Test, Daemon Unit Test 2/2, Sandbox, and both integration shards. Daemon Unit Test 1/2 currently has one 5-second timeout in the unchanged daemon-agent-mention-routing.test.ts test at line 547; because that test and behavior predate this patch, I do not consider it a review blocker.

_{sent by review-bot (Codex · gpt-5.6-sol) · open in session}

[agentconnect-md-test]

Approved for current head 1414b269edf5b7fc0b5ebdbb2a47a0030b60ccd4; I found no blocking regression in this revision.

This synchronized PR diff no longer contains the CI-sharding changes from the earlier head. Against base 385d1f1e1a78ceb3bd7d333403eceb503539a91d, it contains only the WorkspaceManager staging refactor in packages/daemon/src/workspace/workspace-manager.ts. The existing resolver, path clearer, sandbox-mode flag, and clone single-flight now live behind a manager object, while all current module-level call paths continue through a single default instance, preserving current behavior.

Scope note: this commit does not yet provide per-daemon workspace isolation because no WorkspaceManager instance is owned or passed by Daemon; the source explicitly identifies that threading as step 2, consistent with the PR being parked. I treated the pre-existing process-global behavior as non-blocking because this patch does not amplify it.

Verification: the checkout is exactly the trusted head, the changed file matches that commit's raw content, the current PR head/base remain the supplied SHAs, and the patch has no added trailing whitespace. CI was still running at review time.

_{sent by review-bot (Codex · gpt-5.6-sol) · open in session}

[agentconnect-md-test]

I found one blocking regression in current head cc52a85e4d8ea2a739fc415e7b38a04b84531e07.

The workspace execution plane is now correctly owned by Daemon and passed to the console workspace/git/skills seams, but it is not passed into the production SessionManager. That manager therefore creates a separate local-mode WorkspaceManager; for Kubernetes git workspaces with an agentDir, session creation/load resolves a pod-side cwd and then tries to canonicalize it on the daemon filesystem when calculating additional workspace directories. Pass the owning manager in this constructor (workspaces: this.workspaces) so all session workspace calculations use the same sandbox-aware plane.

Verification: the checkout is exactly the trusted head and all changed files match that revision; the patch has no added trailing whitespace. GitHub CI passed Build, Check, Sandbox, Daemon Store, Daemon Image, Runtime Sandbox Image, and both integration shards. Unit Test is currently failing under the new shared-registry configuration; GitHub exposed only the step-level exit-code annotation, so I could not attribute that failure to a specific test from the public logs.

_{sent by review-bot (Codex · gpt-5.6-sol) · open in session}

[agentconnect-md-test]

Pass workspaces: this.workspaces into this SessionManager construction. Its new getter falls back to new WorkspaceManager() when the dependency is omitted, so runtimeWorkspaceDirectories() uses a local-mode plane even though this resolver returns pod coordinates. For a --k8s git workspace with agentDir, additionalWorkspaceDirectories() then takes its local branch and calls realpathSync on paths that exist only in the sandbox, causing session creation/load to fail before the runtime call.

[agentconnect-md-test]

Reviewed at 0d1f110ba0d033656cd796a83e9b7a64bad3cccc.

The previous blocking issue is fixed: Daemon now passes its owning WorkspaceManager into SessionManager, so k8s session creation and loading resolve additional workspace directories through the sandbox execution plane instead of a fresh local-mode manager.

I also checked the shared-registry split and its supporting refactor. The non-isolated and isolated projects are disjoint, every current vi.mock caller is routed to the isolated project, and the guard test will catch drift in either direction. Workspace git/read/skills consumers consistently receive the owning manager, and the shim shutdown change clears the in-flight handshake timer without altering normal bind/reconnect handling.

No blocking findings on this revision. I could not rerun the suite in this isolated checkout because dependencies are not installed; the review therefore relies on exact-revision source inspection and the submitted test results.

_{sent by review-bot (Codex · gpt-5.6-sol) · open in session}