Problem
Problem
The OpenCode L3 memory_transcript path can only retrieve a tail window of a session (--limit) or the last turn (--last-turn). It cannot jump to an arbitrary turn in a long session,
so transcript drill-down becomes coarse and inefficient as sessions grow.
Current behavior
plugins/opencode/scripts/parse-transcript.py loads messages by session_id ordered by time_created.- The script exposes
--limit and --last-turn, but no stable turn/message cursor.
- The OpenCode capture flow stores daily markdown summaries, but does not persist a reusable turn anchor for later drill-down.
Why this matters
For long sessions, L3 recall often needs to land on the exact turn that introduced a decision, error, or code snippet. Without a turn-level cursor, the tool has to scan too much
surrounding text and can return more irrelevant context than needed.
Reference
The Claude Code plugin already writes turn-level anchors and supports drilling into a specific turn with surrounding context via memsearch transcript ... --turn ... --context ....
Proposed solution
Proposed change
Add a stable cursor to the OpenCode capture/transcript flow, for example:
- store a turn/message identifier in the anchor
- allow
memory_transcript to accept --turn or --message-id
- add a
--context option so the tool can return the target turn plus surrounding messages
The exact cursor field can match OpenCode's schema, as long as it reliably identifies a turn and its nearby context.
Expected result
memory_transcript can jump directly to the relevant turn instead of only reading a session tail.
Alternatives considered
No response
Problem
Problem
The OpenCode L3
memory_transcriptpath can only retrieve a tail window of a session (--limit) or the last turn (--last-turn). It cannot jump to an arbitrary turn in a long session,so transcript drill-down becomes coarse and inefficient as sessions grow.
Current behavior
plugins/opencode/scripts/parse-transcript.pyloads messages bysession_idordered bytime_created.--limitand--last-turn, but no stable turn/message cursor.Why this matters
For long sessions, L3 recall often needs to land on the exact turn that introduced a decision, error, or code snippet. Without a turn-level cursor, the tool has to scan too much
surrounding text and can return more irrelevant context than needed.
Reference
The Claude Code plugin already writes turn-level anchors and supports drilling into a specific turn with surrounding context via
memsearch transcript ... --turn ... --context ....Proposed solution
Proposed change
Add a stable cursor to the OpenCode capture/transcript flow, for example:
memory_transcriptto accept--turnor--message-id--contextoption so the tool can return the target turn plus surrounding messagesThe exact cursor field can match OpenCode's schema, as long as it reliably identifies a turn and its nearby context.
Expected result
memory_transcriptcan jump directly to the relevant turn instead of only reading a session tail.Alternatives considered
No response