Builder sessions go idle with queued messages unprocessed; Tower registration can silently vanish from a live session

[waleedkadous]

Two related reliability problems, both recurring (3+ occurrences over 2026-07-25/26, shannon workspace)

1. Idle-wedge: builders stop processing queued messages

Builder sessions repeatedly go idle with messages sitting unprocessed in their queue, after any of: an in-session /clear, an API-500 session death, or simply overnight. The builder process is alive, afx status shows it, sends return [ok] — but nothing is consumed until the terminal is poked. One overnight occurrence cost 15 idle hours on a hot lane.

Working wake recipe (reproducible): a normal afx send <builder> "<msg>" followed by a separate afx send <builder> --raw "status?" — the raw send with Enter is what actually unwedges the prompt.

Ask: Tower-side detection (message queued + no consumption within N minutes → nudge or flag in afx status), or at minimum a documented afx wake <builder>.

2. Registration-loss: a live session silently loses its Tower registry row

A builder worktree's session can lose its Tower registration while its shellper process keeps running (observed sequence: wedge → wake → registration gone). Effects observed:

• All address forms for the agent return NOT_FOUND (afx send <id>, affinity architect, explicit architect:<name>), including agents it had successfully messaged minutes earlier.
• Its own outbound sends fall through to the workspace-qualified fallback and land on the WRONG recipient, stamped with a WRONG sender name — the injection surface already filed as Nested task-spawned builders are send-only ghosts: no registry row, affinity NOT_FOUND, sends fall through to MAIN — an instruction-injection surface #1254.
• From outside, pgrep shows the session alive; no afx surface shows the discrepancy (status says nothing, sends just NOT_FOUND).

Ask: (a) persist/re-establish registry rows on Tower or session hiccups rather than losing them while the PTY lives; (b) an afx doctor-style check that diffs live shellper processes against registry rows and reports orphans, so operators can see 'alive but unaddressable' instead of inferring it from NOT_FOUND archaeology.

Related: #1254.

[waleedkadous]

Frequency data point for prioritization: SEVEN occurrences of the idle-wedge on a single builder session in one working day (2026-07-26, shannon workspace, builder-spir-3330 — a long-running SPIR lane). Signature each time: the session completes a unit of work (a commit, a merge, receiving consultation results), then goes idle with the next step unstarted and any queued messages unprocessed; process alive; the wake recipe (normal send + separate --raw "status?") recovers it every time within a minute. The cost is architect attention: each wedge needs detection (an external monitor), diagnosis (passive checks to distinguish wedge from long-running work), and a wake round-trip. Across the day this consumed more supervision time than any other single cause.