Repository navigation
Long-running goals: zero mid-run progress events (milestone RFC not landed); operator blind for the entire run #5886
Description
Activity
I’m taking a bounded, read-only CLI slice for this: add an opt-in watch mode to the existing Turn journal inspection command so operators can see persisted phase transitions while
run-onceis still executing.At canonical main
3ed5d6b,periodic_reportdoes implementbounded_segment_milestone, but it is emitted for validated, durably settled stage closures; it does not surface ordinary phase progress during a long host Turn.turn run-oncealso renders one CLI payload after the synchronous executor and post-writeback hooks return. The existingturn inspect-journalcommand reads the canonical fenced journal without executing effects, so I’ll build on that boundary and show only safe journal status/phase fields, with no host output or session content.I’ll keep this optional and read-only, verify polling against concurrent journal checkpoints, and exit when the Turn reaches a terminal state. Please flag any conflict with the issue’s desired progress contract while I work.
- added 2 commits that reference this issue
on Oct 8, 2026 Opened PR #5998 with the bounded journal-watch implementation. It exposes persisted journal status and phase checkpoints through turn inspect-journal --watch, with allowlisted JSONL/Markdown output and no host/session content. The PR includes a concurrent-writer regression test and the English/Chinese integration recipe. It does not add transient host-step progress before a journal checkpoint is persisted.
From the filer side: no conflict — this matches what the issue was asking for. Two data points from our orchestrated runs that the design covers well:
- The 83-minute blind wait cited in the issue: persisted phase checkpoints would have shown the turn was alive and roughly where it was, which is all an operator needs to decide "keep waiting vs. investigate".
- The stuck-turn case we filed separately as turn run-once: quota_spend CAS failure leaves turn permanently in_progress (completed work unauditable) #5885 (quota CAS failure leaves the turn permanently
in_progress): under "exit on terminal state", a stuck turn shows zero transitions and the watch never exits — which is itself the liveness signal an operator needs to intervene. The current surface gives nothing at all in that scenario.
Read-only scope +
journal_consistentgating + no host/session content is the right boundary. Following the review on #5998.
Environment
turn run-onceWhat happened (measured)
A single research goal runs 40–85 minutes wall-clock (8-9 agent turns, measured across 5 production runs). During the entire run there are zero mid-run progress events an operator or dashboard could consume:
turn run-onceemits nothing until final settle;drainonly fires at the end.periodic_report'sbounded_segment_milestonetrigger class exists in the RFC/design docs but is not implemented (source for the trigger is byte-identical between v1.2.4 and v1.3.1 — no code landed).Impact
status/todo listshow coarse state, but "what is the agent doing / how far along" is invisible mid-run.Suggestions (directions)
--on-phasehook).loopx status --watchrecipe if preferred.Five fully-audited production runs behind these observations — happy to share timelines.