Agent Session Review (learning tail)

The final edge of every job. A job is not done with a text answer — it is done when the tail is written.
Karpathy method: synthesize on write, not store raw. Short, precise, survives context compression.

Runs at the end of every meaningful agent session before handoff / deliver gate.
Mandated by the final-learning-tail gate in handoff-graph.yaml.

Trigger

• After completing a multi-step job (any provider: Claude Code, Codex).
• Before writing to vault, before handoff, before the deliver gate fires.
• Gate edge final-learning-tail checks all four blocks are present.

Four required blocks

All four must be present. Missing any block → gate fails.

brief_md

3–5 bullets. What was actually done — artifacts produced, decisions made.
Describe what changed, not what was asked.

## Session Brief
- [artifact]: [one concrete line]
- [decision]: [one concrete line]

agent_trace

Chronology of key tool calls and decision points.
Not every tool — only inflection points where direction was chosen or changed.

## Agent Trace
1. [tool / action] → [outcome or decision]
2. ...

self_reflection

One short paragraph each.

## Self-Reflection
**Went well:** ...
**Went wrong / could be better:** ...
**Next time:** ...

most_important

Single most important thing not to lose: a constraint, a lesson, a warning, a pattern.
Must fit in one sentence — this is what survives context compression.

## Most Important
[one sentence]

Output

Write the tail to ~/.agents/reports/<run-id>-session-tail.md
or append to the job outbox note.

Optionally pass to the postrun report:

node ~/.agents/registry/scripts/athena-postrun-report.mjs \
  --results-json <results.json> \
  --run-id <run-id>

Gate check (smoke/agent-contract.sh)

Smoke verifies this file exists and contains all four block names:
brief_md, agent_trace, self_reflection, most_important.