Repository navigation
feat(discord): distinguish worker, SSH and conversation failures - #289
Conversation
|
Merged PR #289 and deployed Validation: 343 focused tests passed, typecheck/lint CI passed. Broad CI had 20,948 passing tests and the same 3 existing failures tracked in #287 (CLI-id allowlist and two stale Discord glue expectations); no new failing test names. Separately, the user requested dead-pane recovery during deployment. Two dead local SSH wrappers (w29-buildmesh and w3-infinidysk) were moved by startup cleanup into Recovery Queue; their old local pane/tab IDs were removed. The remote worker was reachable again, but their original remote tmux targets were absent. Verified each exact transcript ID and working directory before launching strict native resumes into its original remote target, with no fresh-conversation fallback. Verified the Claude runtime resume ID and actual Codex resume argv. Reconnected via the supported Recovery Queue API. Recovered w29-buildmesh: Claude conversation Final audits: 16 local panes, zero dead; all 14 previously healthy local pane PIDs/dead-state snapshots unchanged. Active mapped remote panes: ubuntu-worker 4/4 alive, mac-mini 5/5 alive. Discord REST shows existing recovered titles Idle for w29 and Working for w3. No new model prompt was injected or prior prompt replayed. |
The same red “Connection unverified” title hid several unrelated conditions: an unreachable worker, SSH failures, a lost attachment, and a reachable terminal whose exact conversation could not be verified. w30-buildmesh reproduced this with
No route to host, while the activity updater was healthy.Add bounded, whitelisted SSH failure classification and carry a typed health reason through thread titles, session/recovery cards, Activity and status replies. Labels distinguish Worker unreachable, SSH authentication failed, SSH trust changed, Worker check failed, Terminal disconnected, Agent exited, Conversation unverified, Terminal check failed and Session updating. Unexpected failures have a separate Status check failed label. Unknown probes are never guessed to be network failures or agent exits.
Cards explain the evidence and appropriate next step. Network/auth/trust failures retain Refresh status/Open in Codeman, without offering a new session or automatic resume. Exact-conversation/input safety checks, existing recovery gates, automatic retries and legacy probe consumers are preserved. Successful verification clears the reason. Raw stderr and credentials are never displayed or persisted as a reason.
Fixes #288.
Validation: 343 tests passed across classification, connection health, activity, native UI, recovery controls, native services and Full-mode integration. Includes exact “No route to host” classification, non-SSH failure refusal, automatic recovery, label propagation and absence of unsafe recovery actions. Typecheck, changed-source ESLint, formatting and lockfile checks passed. Existing broad-CI baseline failures are tracked separately in #287.
No production deployment. The image below is a labelled design preview, not a capture of the live Discord client.