[matt2e]

Problem

• Library-only agents disappeared on every library read. This covers hosted agents, agents created on another device, and every agent on web. Mentions, author names, agent icons, typing labels and the Solo default-agent chip all dropped them while a read was in flight. The gap lasted exactly as long as the read (2.04 s with a 2 s read in the live take).
• A failed read hid them until the next successful read. A dropped relay connection hid them for the whole outage, and after reconnecting too (≥ 11.6 s in the live take, still missing 3.7 s after "Reconnected").
• Four DM paths refused agents during launch. New message, search's person action and the agent's Message button refused an agent the native list had not returned yet. The RecipientPicker showed "No matching people." for it. In the live take a restored New message draft sent at launch was refused every time.

Why it was built this way

Git history shows no security or ghost-agent reason for either behavior:

• The disconnect clear() came from the initial import (Import plugin core #2) as a read fence. Fix startup inventory duplication and late panel scroll shifts #160 kept it so that reconnecting would reread the library.
• Add channel-backed Sessions for agent conversations #77 deliberately kept the rows and gated only admission on ready.
• Add optional channel templates, teams and personal group defaults #181 moved that gate into the shared agentChoices projection, which turned the admission gate into a display gate.

Archives already moved to stale() in #1665. The library was the last store at that call site that still emptied on disconnect.

Fix

• Rows survive refreshes. The library and name directory keep the last read through loading and error. pending/complete still mark the read as partial, so inference and templates keep refusing until a ready read.
• Disconnect marks the library stale instead of clearing it. stale() replaces clear(). In-flight reads are still aborted, and reconnecting still rereads. Sign-out, cache clear and dispose still empty the library.
• One shared admission rule. admissible means managed, or the library read is ready. Channel addAgents, Solo start and tagged sends use it, so stale rows are display-only. addAgents rereads once before refusing an unknown key.
• DM paths wait for the native list instead of refusing. The four DM paths gate on the new managedPending flag. The three that act on a click await refresh(false), which joins the native list already in flight and never waits on the relay library.
• docs/agents.md is updated.

Net +83 production lines; no new timers or abstractions.

Before / after

Video evidence

All four takes are side by side: BEFORE (a3a6223d7) on the left, AFTER on the right, with the clocks synchronized. They were recorded in the Playwright browser fixture with library or native reads held for exactly 2 s.

A1, Refresh member data. Watch the message author: BEFORE shows a raw key with no agent icon for 2 s, while AFTER keeps "Owned Honey".

A1.mp4

A2, relay disconnect. Watch the author: BEFORE drops it from the disconnect until after "Reconnected", while AFTER keeps it and still offers it in @Hon while offline.

A2.mp4

B1, sending a restored New message draft during launch. Watch the line above the composer: BEFORE refuses at once and stays refused, while AFTER waits and creates the DM once the list lands.

B1.mp4

B2, the recipient picker during launch. Watch under "To: Owned": BEFORE shows "No matching people." for about 1 s, while AFTER lists the agent as soon as people load.

B2.mp4

Take	BEFORE	AFTER
A1: author name/icon gap during a 2 s read	2.04 s	none
A2: author name/icon gap across a disconnect	≥ 11.6 s (to end of take)	none
B1: send at launch	refused	DM create issued 10 ms after the list lands
B2: "No matching people." at launch	0.96 s	never

Cost / risk

• Rendering is unchanged. Publishes and renders per refresh are the same as before, and rows are never remounted.
• Waiting is new only where the app used to refuse instantly. That means the DM paths while the native list is loading. An offline addAgents refusal of a library-only agent now takes 3.5–10 s (one reread) instead of being instant.
• Stale rows stay after reconnect until a reread. They are display-only, and admission refuses them.
• Admission safety depends on callers using admissible. All three admission paths I know of (channel add, Solo start, tagged send) use it and are tested.
• Not fixed here (pre-existing). On desktop, a partial inventory read (relay half fails, local half succeeds) still drops relay-only agents until the next full read. This is left as a follow-up.

Testing

• Unit tests fail on the parent and pass on this branch. With the branch's tests copied onto the parent sources, 18 of 249 tests in the 10 affected files fail; the tests added by the later commits also fail on their parent commits (the bridge test fails with the mapping inverted). Covered: admission across loading, error and stale libraries; DM waits; the picker; Solo start; tagged sends; admissible; joining the in-flight native list; and the local-agents bridge mapping.
• ProfilePanel.test.tsx now expects the library-configured name ("Configured Carl") to stay through a relay drop instead of falling back to the public name; it fails on the parent's clear-on-disconnect.
• Typecheck, lint and format pass.
• The four Playwright fixture takes above were run live.
• Not run:
  • the native desktop app;
  • a real relay;
  • DM creation end to end (the fixture has no DM-create endpoint).
• Unit tests only: the agent Message button and search person-action paths. They cannot be reached at launch in the fixture.

🤖 Generated with Claude Code

[kalvinnchau]

🤖 Two reproducible issues in recipient selection and native-read waiting.

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, integrating independent systems/lifecycle, correctness/security and product/accessibility passes.

One low-priority finding is recorded inline. The stale-display/admission separation and launch-time wait paths were reviewed against their callers and lifecycle guards.

Reviewed head fa94a50 against base a65463a. Source-only: no tests, builds, app execution or independent CI assessment. Native DM creation, real-relay behavior, performance, rendered focus and assistive-technology behavior remain unverified. Inspected the public description, commit messages, diagram and one-second video frame samples; the fixture recordings show an earlier revision and do not establish current-head runtime correctness.

[wesbillman]

🤖 Star Lord’s automated source follow-up via Wes’s account, integrating independent systems/lifecycle, correctness/security and product/accessibility passes.

No new actionable findings in the rebase integration. The existing P3 recipient-picker loading-status finding remains unresolved at this head: the native-list wait can still leave an empty picker without loading feedback. Please retain that existing finding; no duplicate inline comment is added.

Compared the earlier reviewed fa94a50 with this revision: four patches are unchanged; the first adapts to the current name-fallback and reconnect code. Checked retained library rows, freshness/admission separation, native-list joins, cancellation and error/retry callers.

Reviewed head 0c8eb04 against base 4bcd920 (merge base 561265d). Source-only: no tests, builds or app execution. Hosted checks were still running in the single snapshot. Native DM creation, real-relay recovery, performance, rendered focus and assistive-technology behavior remain unverified. Inspected the public description, five commit messages, diagram and sampled frames from all four unchanged fixture videos; those recordings show an earlier revision and do not establish current-head runtime correctness.

[wesbillman]

🤖 Star Lord’s automated source follow-up via Wes’s account, integrating independent systems/lifecycle, correctness/security and product/accessibility passes.

No new actionable findings. The prior missing loading-status finding is repaired: the picker announces loading while the native agent list is pending. The follow-up also keeps uncontrolled agents out of selection and makes joined refresh callers wait for superseding native reads. Checked the five-file delta from 0c8eb04, its regression-test source, and the affected callers and lifecycle guards.

Reviewed head 5831134 against base 4bcd920 (merge base 561265d). Source-only: no tests, builds, app execution or CI assessment. Native DM creation, real-relay behavior, performance, rendered focus and assistive-technology behavior remain unverified. Inspected the public description, six commit messages, diagram and sampled frames from all four unchanged fixture videos; unsampled footage was not reviewed, and those earlier-revision recordings do not establish current-head runtime correctness.

[wesbillman]

🤖 Star Lord’s automated follow-up source review via Wes’s account, integrating independent Rocket, Mantis and Groot review lanes.

No new actionable findings. The prior recipient-filtering, loading-status and superseding-native-read fixes remain intact. Reviewed the rebase integration with hosted-agent canMessageAgent eligibility, existing-DM reopening, refresh/retry and cancellation boundaries. No changes requested from this source review.

Reviewed head 36397014e43a29ecf2d8a2212cb9e63bf76e6490, base e7ae7b18cff03f371382ae60fc1064b60a667ed2 (merge-base ca836a3109be30a7fd46db97bdcceaf32ebabf65). Source-only: no tests, builds or app execution. One hosted-check snapshot showed 39 passed and two skipped (Windows validation and flake check); this does not establish native/real-relay behavior, performance, rendered focus or assistive-technology behavior. The public description, six commit messages, diagram and 16 sampled frames across all four videos were inspected; unsampled footage remains unreviewed, and those earlier fixture recordings do not validate this head.

[wesbillman]

🤖 Star Lord’s automated source follow-up via Wes’s account, integrating independent Rocket, Mantis and Groot review lanes.

No new actionable findings in the rebase integration. The prior recipient-filtering, loading-status and superseding-native-read fixes remain intact. Compared all six feature patches with previously reviewed 36397014e43a29ecf2d8a2212cb9e63bf76e6490, including the short-query picker state, keyed search cancellation, native refresh bridge and retained-display/admission boundaries. No changes requested from this source review.

Reviewed head 823454777007257fbc374827e8fd68b06f3c71e6 against base 90ce5e37a851909cc0e5fd4da2f469010e4eb7bf. Source-only: no tests, builds or app execution. One hosted-check snapshot showed 39 successful checks and two skipped (Windows native validation and flake check); native/real-relay behavior, performance, rendered focus and assistive-technology behavior remain unverified. Inspected the public description, six commit messages and attached diagram. The four historical fixture videos were not viewed in this cycle and do not establish current-head runtime correctness.

[wesbillman]

🤖 Star Lord’s automated source follow-up via Wes’s account, integrating independent Rocket, Mantis and Groot review lanes.

No new actionable findings in the rebase integration. The prior recipient-filtering, loading-status and superseding-native-read fixes remain intact. Compared the six feature patches with previously reviewed 823454777007257fbc374827e8fd68b06f3c71e6, focusing on verified-name recipient selection, existing-DM visibility, remote-agent labels, native-list lifecycle and Solo/channel admission. The documented distinction between hosted DM eligibility and channel-admission evidence is preserved. No changes requested from this source review.

Reviewed head 1438e6593a3f375a41b339bfd2ca251f5517f77c against base fe3dd1947181de66dc6e4fd8ca5d3b61573c0ca2 (merge-base 1c3047558547e76bd006c2dc76951b3f1d601c5d). Source-only: no tests, builds, app execution or CI assessment. Native/real-relay behavior, performance, rendered focus and assistive-technology behavior remain unverified. Inspected the public description, six feature commit messages and attached diagram. The four historical fixture videos were not viewed in this cycle and do not validate this head.

[wesbillman]

🤖 Star Lord automated source review via Wes’s account, integrating independent Rocket, Mantis and Groot review lanes.

No new actionable findings in this focused rebase follow-up at 4a0e3d4f85f4a782f90f69f0293fd6d02a0f0d20, base/merge-base cae3922574c5c1f0ce5d32231810229e94b1bb40, compared with reviewed 1438e6593a3f375a41b339bfd2ca251f5517f77c.

The six feature patches retain the reviewed behavior: five are patch-identical and the first differs only in documentation context. We checked their integration with the changed native-agent service, relay recovery, DM/picker and Solo/composer consumers. Retained library rows remain display evidence, admission still requires managed-or-ready-library evidence, and native waits preserve joining/supersession and cancellation. The prior picker filtering/loading and superseding-read repairs remain intact; their regression coverage was inspected as source.

This was a focused integration review, not a full review of unrelated changes in the new base. No tests, builds or app execution were performed. Native/real-relay behavior, performance, rendering, focus and assistive-technology behavior remain unverified. I inspected the current PR description and attached diagram; the four historical fixture videos were not reviewed in this round and do not establish current-head validation.

[tellaho]

🤖 Design review: no design blockers found at 4a0e3d4f85f4a782f90f69f0293fd6d02a0f0d20. Agents keep their familiar names during refresh/disconnection, and starting a chat waits for the agent list. The recipient picker distinguishes loading from an empty result while preserving eligibility checks.

Reviewed source, the diagram and sampled frames from all four attached fixture recordings. Those recordings show an older revision (011465070), not this head. Tests were not run; native/real-relay DM creation, live failure/retry, keyboard/focus, dark mode, narrow/wide layouts, enlarged text and many-agent rendering remain unverified.

[wesbillman]

🤖 Star Lord’s automated source follow-up via Wes’s account, integrating independent Rocket, Mantis and Groot review lanes.

No new actionable findings at f1c3cfe773d2174e24f990078b59499d35076294, base/merge-base c9239f737f4985979eb5939e84c33ab3da2b1321, compared with reviewed 4a0e3d4f85f4a782f90f69f0293fd6d02a0f0d20.

Five feature patches are unchanged; the first adapts to the base’s partial-library recovery contract. We checked retained display rows versus admission, reconnect failure recovery and generation fencing, native-list joining/supersession, and the changed picker/retry consumers. The prior recipient filtering, loading-status and superseding-read repairs remain intact. Regression tests were inspected as source.

This is a focused integration review; unrelated changes in the new base were not fully reviewed. No tests, builds, app execution or CI assessment were performed. Native/real-relay behavior, performance, rendering, focus and assistive-technology behavior remain unverified. The current public description and attached diagram were inspected; the four historical fixture videos were not viewed this round and do not validate this head.