Finding AgentsView History

Use AgentsView as an evidence source for past agent behavior: find relevant
recorded sessions, inspect the surrounding conversation, and extract
decisions, instructions, or patterns. A snippet is a lead, not an answer.

Core Workflow

Delegate to a search subagent if your harness supports one; otherwise run the bounded probes yourself in order. The coordinator decides what evidence it wants, delegates
the archive crawl, then synthesizes and verifies.

1. Translate the ask into a target behavior and desired evidence: likely
   projects, agents, time windows, vocabulary.

2. Concept questions ("why did we…", "how did we handle…") start with
   hybrid search. Exact identifiers (IDs, paths, commands, error strings,
   tool names) start with plain substring search over tool inputs and results;
   FTS searches messages only and cannot find identifiers stored only in tool
   calls:

   agentsview session search "<concept query>" --hybrid --context 2 --json --limit 8
   agentsview session search "<exact string>" --in tool_input,tool_result --json --limit 8

3. Narrow by recency when the ask implies it: --since 14d, --since 3m
   (h/d/w/m/y units; m is months, not minutes). Widen or drop --since
   if results are thin.

4. Triage from the inline context_before/context_after in the results.
   Every hit, in every mode, carries an ordinal_range conversation-unit
   span ([start, end]; a run of assistant messages when end > start) with
   ordinal as the anchor message; center the window on the anchor.
   Deep-dive only the strongest candidates:

   agentsview session messages <session-id> --around <ordinal> --before 8 --after 8 --role user,assistant --json

5. Give the search a mechanical budget: 4-6 probes, top 2-4 sessions, one
   window per session, then stop and report.

6. Answer from evidence, citing session IDs and ordinals. Cite the session
   plus the hit's ordinal_range with the @ordinal anchor (e.g.
   #12-40 @19), in every mode — not just the single anchor ordinal.

7. A hit marked subordinate (sidechain or subagent/fork content) is
   supporting evidence only: corroborate it from its parent session
   (parent_session_id) before treating it as a decision or instruction.

Mode Fallbacks

• "semantic search not available" (HTTP 501): embeddings are not set up on
  this archive. Fall back to FTS probes — several short queries with
  synonyms beat one long phrase:

  agentsview session search "<two or three words>" --fts --json --limit 8

• "temporarily unavailable" (HTTP 503): the embeddings endpoint is down.
  Retry once; if it persists, say so and continue with FTS — do not
  silently downgrade, the user should know their embeddings are broken.

Reconstructing a Decision

1. Find the earliest substantive mention: hybrid query for the decision's
   subject with --scope top (delegated subagent sessions echo their
   parent's instructions and bury the conversation where the decision was
   made), starting --since 3m and widening (6m, 1y, none) until the
   origin appears.
2. Walk forward from the origin with message windows. User messages carry
   intent and constraints; assistant messages carry rationale, options
   considered, and tradeoffs.
3. Watch for durable artifacts referenced in the conversation (spec or
   plan documents, ADRs, PR descriptions) and read those files if they
   still exist.
4. Produce a decision record: what was decided, when, by whom, the stated
   reasons, alternatives that were rejected, and citations
   (session-id + ordinals) for each claim.

Rationalization Table

Output Shape

## Searches
- `query` (mode) -> why it was useful or not

## Strong Matches
- `<session-id>` (`project`, `agent`, ordinals N-M): finding and evidence

## Synthesis
- Decision/pattern grounded in the matches, with citations

## Gaps / Follow-ups
- What was not found and the next narrower probe