Skip to content

[CAPABILITY] Introspection pipeline blind to successful sessions — extractor reads only error-triggered request dumps #24

Description

@Da-Mikey

Problem (observed in real usage)

The introspection pipeline reads only error-capture request dumps — sessions that ended with a provider error. Successful sessions, which make up the vast majority of real user interactions, are invisible to the extractor.

Evidence (aggregated, anonymized)

  • Frequency: 1161 "sessions scanned" but all 50 source files are error-capture dumps (reason=max_retries_exhausted). Zero successful sessions represented.
  • Tool/area involved: scripts/introspection_extract.py, sessions/ ingestion path
  • Failure shape: The extractor iterates over request_dump_*.json files. These files are written ONLY when a session terminates with a provider error. Normal sessions that succeed produce no dump file and are never analyzed.

Impact on real tasks

Every introspection cycle since the pipeline was built has analyzed only what went wrong — never what went right, or what normal usage looks like. This means:

  1. The agent's actual strengths and effective patterns are invisible.
  2. All findings come from a biased sample (sessions already failing at the provider level).
  3. The 1514 "terminal failures" extracted may largely be cascading errors from provider disconnects, not standalone tool problems.
  4. True user-tool friction patterns in successful sessions are completely missed.

Proposed direction

Add a lightweight session capture mechanism for successful sessions — either:
(a) A compact digest written at session end (tool names used, turn count, outcome summary, no raw text), OR
(b) An opt-in transcript buffer that writes on any termination (success or failure), OR
(c) A real-time streaming digest written every N turns so mid-session state is captured even if a session does not end cleanly.

The key constraint: whatever captures successful sessions must be roughly 1KB or less per session (privacy-preserving, metadata only) to avoid the unbounded-bloat problem that the current dump-on-error design avoids.

Value

  • Impact: 0.8 — Every introspection cycle produces biased, incomplete analysis. Real user friction is invisible.
  • Effort: 0.3 — Adding a session-end callback that writes a compact JSON summary is a focused, self-contained change. The schema for aggregating tool-use statistics, turn counts, and outcome already exists in the extractor.
  • Priority Score: 5.33

Metadata

Metadata

Assignees

No one assigned

    Labels

    capabilityMissing ability users neededenhancementNew feature or requestintrospectionFound by session introspectionrejectedNot accepted by evolution — see closing comment

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions