johba/disinto
1
0
Fork 0
Code Issues 4 Pull requests Projects Releases Packages Wiki Activity Actions

chore: planner run 2026-03-26

Browse source
This commit is contained in:
openhands 2026-03-26 07:18:30 +00:00
parent 805fa69770
commit b6c2b611e3
3 changed files with 124 additions and 27 deletions
Download patch file Download diff file Expand all files Collapse all files

52
planner/MEMORY.md
View file

@ -1,27 +1,45 @@
<!-- summarized-through-run: 1 -->
<!-- summarized-through-run: 6 -->
# Planner Memory
## 2026-03-21 — Second planner run
## 2026-03-26 — Sixth planner run
### Milestone state
- **Foundation**: COMPLETE. Agent loop, supervisor, planner, multi-project all working.
- **Adoption**: IN PROGRESS. Bootstrap (#393), docs (#394), dashboard (#395), example project (#466) all in backlog. No work started yet — issues were unlabeled until this run fixed them.
- **Scale**: PARTIALLY started (multi-project works for 3 repos).
- **Foundation**: COMPLETE. All agent loops, supervisor, planner, multi-project, knowledge graph, predictor-planner feedback loop — all working.
- **Adoption**: 4/5 COMPLETE. Bootstrap (#393), docs (#394), dashboard (#395), landing page (#534) all done. Only #466 (example project) remains — stuck on human decision since 2026-03-23.
- **Ship (Fold 2)**: ENTERING SCOPE. Rent-a-human (#679) done. Exec agent (#699) done. Observable addressables (#718) filed. Deploy profiles and assumptions register not yet tracked.
- **Scale**: DEFERRED. No external users yet. Plugin system, community formulas, hosted option all premature.
### Completed since last summary (runs 2-6)
- Bootstrap fully hardened: init smoke test (#668), CI wiring (#661), Forgejo reachability (#660), 10+ bootstrap fixes
- Full stack containerized (#618, #619) with Forgejo, Woodpecker, Dendrite
- Autonomous merge pipeline (#568) — PRs auto-merge on CI pass + approval
- Unified escalation path (#510) — PHASE:escalate replaces needs_human
- Factory operational reliability — guard logging (#663), stale phase cleanup (#664)
- Prediction/backlog killed (#686) — planner now only ACTIONs or DISMISSes predictions
- Planner v2 — graph-driven formula (#667), tea CLI integration (#666)
- Exec agent (#699) — interactive assistant via Matrix
- Rent-a-human (#679) — formula-dispatchable human action drafts
- Tech-debt queue cleared (~30 items)
- Skill package initiative started (#710-#715) from research (#709)
### Patterns
- **Predictor over-signals on transient ops issues**: 4/6 predictions this run were orphaned tmux sessions or crashed reviews — things the supervisor handles automatically. Expected to continue until predictor learns to filter supervisor-handled issues.
- **Label loss on issue creation**: The 3 Adoption issues created last run had no labels when checked this run. Root cause unknown — could be a silent API failure in the planner's issue creation, or labels removed by another process. Watch for recurrence.
- **Long tech-debt backlog blocks features**: ~20 small backlog items (tech-debt, bug fixes) will be processed before Adoption features due to sequential pipeline and lower issue numbers. Not a problem per se — maintains factory health — but means Adoption work won't start for weeks unless manually prioritized.
- **needs_human is a silent pipeline killer**: When a dev-agent writes PHASE:needs_human and no human responds, the pipeline stalls silently. Supervisor doesn't escalate. Filed #465 to fix.
- **Label loss resolved**: #535 fixed the recurring label-loss pattern. Labels now persist reliably.
- **Predictor signal quality improved**: Later runs show 100% substantive predictions. Over-signaling on transient ops issues has stopped.
- **Human bottleneck is real**: #466 escalated 2026-03-23, still no response after 3 days. When the factory needs human input and doesn't get it, work halts on that branch entirely.
- **Factory throughput is extreme when unblocked**: 50+ issues cleared in ~5 days (2026-03-20 to 2026-03-25). Pipeline processes ~10 issues/day when backlog is stocked.
- **Duplicate issues from parallel creation**: #710/#714 and #711/#715 are duplicates — likely created in separate exec/research sessions. Gardener should catch these.
- **prediction/backlog migration**: All 4 legacy prediction/backlog items dismissed and closed in run 6. prediction/dismissed label created.
### Strategic direction
- Adoption remains the leverage multiplier. All 4 Adoption issues are now in backlog: #393 (init) → #394 (docs) → #395 (dashboard), plus #466 (example project, depends on #393).
- The critical path is: #393 (init) must land first — docs and example project both reference it.
- #465 (supervisor needs_human escalation) is operational leverage — prevents the kind of silent stall observed via #446.
- Ship milestone is the next frontier. Adoption is blocked only on #466 (human decision).
- Skill package distribution (#710→#711→#712) is the immediate pipeline work — packaging disinto for external discovery.
- Observable addressables (#718) bridges Fold 2 → Fold 3 — core vision item.
- The factory has the exec agent (#699) and rent-a-human (#679) — two vision capabilities now live.
- VISION.md updated with factory primitives (resources, addressables, observables) — formalizes the framework.
### Watch list
- Label persistence: verify #393/#394/#395 retain their backlog labels next run
- Tech-debt throughput: how fast is the dev-agent clearing the backlog queue?
- #357 (in-progress): action-agent runtime isolation — track completion
- #448 (prediction/backlog): disk at 75%, trend improving
- #446 (prediction/backlog): harb needs_human pattern — is #465 picked up?
- #466: human response overdue (3 days) — will it ever be unblocked?
- #710-#712: skill package pipeline — first new work direction since Adoption
- #714/#715: duplicate cleanup by gardener
- prediction/backlog label: should be deleted per #686, still exists
- Ship milestone gaps: deploy profiles, assumptions register, vault-gated folds — not yet filed