Transcript Retrieval¶
Use transcript tools when the answer lives in prior agent conversations: who said a rule, what a task did, where a failed attempt happened, or what context a handoff lost.
Use graph tools when the answer depends on connected entities. Use transcript tools when you need the conversation itself.
Retrieval Ladder¶
Start broad:
Open context around a hit:
Inspect the session:
bbox_session(session_id="<session-id>")
bbox_messages(session_id="<session-id>", from_end=true, limit=40)
If you need the origin of a standing claim, use cite:
bbox_cite returns oldest-first so the origin comes before later repetitions.
Search¶
bbox_search is for topics when you do not know the session:
bbox_search(
query="atom binding workflow",
project="/repo/x",
role="assistant",
exclude_self=true,
limit=10
)
Default smart mode broadens adjacent terms for recall. Use quoted phrases for
exact text and -term for exclusions. Use mode="fulltext" only when you want
raw Tantivy/Lucene syntax.
Filter early:
| Filter | Use |
|---|---|
project |
Keep results scoped to the repo you care about. |
role |
Find user directives, assistant summaries, tool results, or thinking blocks. |
account |
Separate multiple Claude/Codex accounts. |
include_subagents |
Include or exclude subagent transcript blocks. |
exclude_self |
Avoid echoing the current turn. |
Sessions And Messages¶
Use bbox_sessions_list when recency matters more than text:
Use bbox_session for metadata: first prompt, project, duration, tool counts,
and transcript file path.
Use bbox_messages when you already know the session and need chronological
flow. Tail mode is useful for takeover:
Topics¶
bbox_topics is a cheap "what was this session about?" read. It is term
frequency, not summarization:
Reindex And Health¶
Background reindex runs periodically. Reach for manual tools only when you are diagnosing freshness:
Use full=true only after schema changes, corruption, or an explicit reason to
throw away incremental assumptions.
When To Use Graph Instead¶
Use Internals and the agentic opening sequence when the question needs provenance across entities:
- why a line exists
- which decision superseded another
- what thread/session/note produced a commit
- which docs and symbols relate to an artifact
Transcript search finds text. The graph finds paths.