Skill: AI session report (context → report)

Turns a session context — the ordered user prompts and agent responses of one AI session — into a markdown report on local disk. This skill covers only that transformation. It does not know or care where the report goes afterwards; the destination, credentials, and the policy for when to report are supplied by the caller (a knowledge note, a playbook, or the user).

session_report.py beside this file does the work. Python 3 stdlib only.

1. Report format

One report per session, at $AI_SESSION_REPORT_DIR/<YYYY-MM-DD>-<id>.md (default directory ~/ai-session-reports; <id> is the session id without any devin- prefix, derived from an id or a session URL).

---
date: 2026-09-15
session: devin-<id>
user: <username>
<key>: <value>          # any --meta pairs given on first write
---

# Session <id>

## [2026-09-15T18:20:00Z] prompt — <username>

<verbatim user message>

## [2026-09-15T18:21:30Z] response

<verbatim agent message>

Roles are prompt (user → agent), response (agent → user, i.e. each message_user), and note (anything else worth keeping, e.g. a redaction marker). Text is verbatim; the only permitted edit is replacing a secret value with [REDACTED].

2. Live reporting (one exchange at a time)

S=.github/skills/ai-session-report/session_report.py

# each incoming user message
python3 $S append --context "$SESSION_ID" --role prompt --author "$USERNAME" \
  --meta url="$SESSION_URL" --meta repo="$OWNER_REPO" <<'EOF'
<user message>
EOF

# each outgoing response (same tool batch as the message_user call)
python3 $S append --context "$SESSION_ID" --role response <<'EOF'
<exact text passed to message_user>
EOF

python3 $S path --context "$SESSION_ID"    # where the report is

--meta pairs are written only when the report is first created; it is harmless to pass them every time. Append calls print the report path so the caller can hand the file to its delivery mechanism.

3. Backfilling a whole session

When a session was not reported live, rebuild the report from its event history:

1. List message events for the session (for Devin: devin_mcp → devin_session_events with {"action": "list", "session_id": "devin-<id>", "categories": ["message"], "first": 100}, paging with after), then fetch full text with {"action": "details", "event_ids": [...]} (≤ 20 ids per call).
2. For each event in chronological order run append with --timestamp <event time> and --date <session start date>; incoming events are --role prompt --author <username>, outgoing events are --role response.
3. If a report already exists for the session, read its last ## [...] timestamp and append only newer events.

4. Rules

• Verbatim. Never summarize, reword, or drop exchanges.
• Never write secret values into a report.
• One writer per session; sequence append calls, do not parallelize them.
• Reporting is best-effort: a failure to append must not block the session's actual task.