Summary
OpenOwl's learning loop is broken for the common exit-and-resume workflow. Insights generated during a session are not persisted to cerebrum.md because the session lifecycle hooks available in OpenCode don't cover the resume case.
Problem
When a user exits and resumes a session (the most common workflow), OpenOwl has no mechanism to capture and persist insights. The current approach relies on the model voluntarily writing to cerebrum.md, which nothing triggers or reminds it to do.
OpenCode Plugin Lifecycle Events
From the official docs at https://opencode.ai/docs/plugins/:
Available session events:
session.created — fires on new session creationsession.deleted — fires on session deletionsession.updated — fires on session update (generic)session.idle — fires when session goes idlesession.compacted — fires on context compactionsession.error — fires on session errorsession.status — fires on status changesession.diff — fires on diff events
Critically: there is no session.resumed event. When a user exits and resumes, it likely fires session.updated or session.status, not session.created.
Why This Matters
session.created only fires for fresh sessions — exit-and-resume never triggers itsession.idle fires when the session goes idle, but a user can exit without the session entering idle state, and resume without a new session.createdsession.updated is a generic event — using it as a proxy for resume would require inspecting event properties to detect the transition- The model is told via AGENTS.md to write learnings to cerebrum.md, but nothing in the plugin triggers this — the instruction is in a file the model is told not to read
Proposed Approaches
-
Use session.idle for learn-back — On idle, generate a summary of session insights and persist to cerebrum.md. This covers the case where the user steps away (even if they resume later).
-
Use session.updated as resume proxy — Detect resume transitions in session.updated event properties and trigger learn-back or re-initialization.
-
Tool-based debounce (no lifecycle dependency) — Track insights during tool.execute.after and periodically flush to cerebrum.md based on accumulated changes, not session lifecycle.
-
Compaction hook injection — Use experimental.session.compacting to inject a "persist your learnings" prompt before compaction, leveraging the model's own summary capability.
-
SDK SSE stream — Use event.subscribe() to detect session transitions with more granularity than individual hook events provide.
Related
memory.md auto-logs tool actions (reads, writes) — this works because it writes on every tool.execute.aftercerebrum.md has no auto-logging — the learning loop is one-way (injection in, no persistence out)- The daemon's cron tasks (cerebrum-reflection) could consolidate entries but can't create them from scratch
Reference
Summary
OpenOwl's learning loop is broken for the common exit-and-resume workflow. Insights generated during a session are not persisted to
cerebrum.mdbecause the session lifecycle hooks available in OpenCode don't cover the resume case.Problem
When a user exits and resumes a session (the most common workflow), OpenOwl has no mechanism to capture and persist insights. The current approach relies on the model voluntarily writing to cerebrum.md, which nothing triggers or reminds it to do.
OpenCode Plugin Lifecycle Events
From the official docs at https://opencode.ai/docs/plugins/:
Available session events:
session.created— fires on new session creationsession.deleted— fires on session deletionsession.updated— fires on session update (generic)session.idle— fires when session goes idlesession.compacted— fires on context compactionsession.error— fires on session errorsession.status— fires on status changesession.diff— fires on diff eventsCritically: there is no
session.resumedevent. When a user exits and resumes, it likely firessession.updatedorsession.status, notsession.created.Why This Matters
session.createdonly fires for fresh sessions — exit-and-resume never triggers itsession.idlefires when the session goes idle, but a user can exit without the session entering idle state, and resume without a newsession.createdsession.updatedis a generic event — using it as a proxy for resume would require inspecting event properties to detect the transitionProposed Approaches
Use
session.idlefor learn-back — On idle, generate a summary of session insights and persist to cerebrum.md. This covers the case where the user steps away (even if they resume later).Use
session.updatedas resume proxy — Detect resume transitions insession.updatedevent properties and trigger learn-back or re-initialization.Tool-based debounce (no lifecycle dependency) — Track insights during
tool.execute.afterand periodically flush to cerebrum.md based on accumulated changes, not session lifecycle.Compaction hook injection — Use
experimental.session.compactingto inject a "persist your learnings" prompt before compaction, leveraging the model's own summary capability.SDK SSE stream — Use
event.subscribe()to detect session transitions with more granularity than individual hook events provide.Related
memory.mdauto-logs tool actions (reads, writes) — this works because it writes on everytool.execute.aftercerebrum.mdhas no auto-logging — the learning loop is one-way (injection in, no persistence out)Reference