Repository navigation
investigate(codex): receive input before the first turn through the owning App Server #306
Description
Problem
Ordinary session-peer refuses a live Codex thread before it has a saved thread record (thread_not_yet_persisted). A recent isolated startup test found the SQLite file already present but the target row absent. This does not establish that the owning App Server cannot accept input.
Investigation
Use a dedicated owned TUI and the same App Server, not a new server pretending to own an existing target. Inspect thread/loaded/list and test thread/queue/add before any user/model turn. If queueing cannot work, separately assess turn/start and its start-versus-steering semantics. Record CLI version, persisted-row state, submission evidence, and actual TUI receipt separately.
Acceptance and safety
- Preserve home, owner, generation, direct-input and version guards; no database writes or permission bypass.
- Bound deadlines and owned-process cleanup; never retry or fall back after an uncertain effect.
- Do not read unrelated transcripts, copy credentials, change global configuration, or resume another writer.
- Distinguish normal/shared-daemon launch from an opt-in dedicated App Server topology.
- This is an isolated feasibility experiment, not stable runtime support or release authorization.
Related: #166 (diagnostic already implemented), #245 (ephemeral side sessions), #247 (steering). This issue concerns ordinary fresh sessions and is not a duplicate of those scopes.
Activity
abruption commentedon Oct 13, 2026
Isolated feasibility result — 2026-10-13
PASS in a dedicated macOS arm64 Codex 0.162.1 App Server/TUI topology.
thread/queue/addreturnedqueuedSubmission.ACK CODEX-FIRST-INPUT-306-QUEUE.turn/start, resume, database mutation, fallback, or resend was used. Queue acceptance and observed reply are separate evidence.This proves a feasible first-input transport; it does not add support to the released
session-peer send. The server inherited normal authentication without copying credentials. No global settings or existing daemon were changed.Remaining gates
Safe discovery and generation/owner binding for ordinary daemon, Orca and
--no-daemonconfigurations; version/capability checks; bounded body-free observation; Linux/Windows/SSH validation. A post-reply metadata inspection stopped at the exclusive single-loaded-thread assertion; the changed inventory was not investigated, and no post-reply database-count claim is made.The dedicated server/TUI exited cleanly and their sockets were removed. The test conversation remains in normal native history. Sanitized harness and results are retained in the isolated local experiment worktree. Syntax checks passed; no runtime modification, PR, merge, release or deployment was performed.
abruption commentedon Oct 30, 2026
Milestone housekeeping — 2026-10-30: moved remaining acceptance from the already published v1.2.0 milestone to Long-term, following the owner's research/backlog organization request. Shipped implementation and merged evidence remain unchanged. This issue stays OPEN; existing priority and safety/validation gates are preserved. No next release date, implementation completion, new live-test permission or operational change is implied.