extract-skeleton.py silently returns empty for all current Claude sessions (10-line detection cap)

[anthonygiuliano]

Summary

ce-sessions' extract-skeleton.py silently returns an empty skeleton for every current Claude Code session, because its platform auto-detection only inspects the first 10 lines and current Claude session files front-load more than 10 non-message metadata records. This disables session-history synthesis across every skill that depends on it (ce-sessions, ce-compound's optional session pass, etc.) — and it fails silently (empty output, parse_errors: 0), so it reads as "no relevant prior sessions" rather than a bug.

Root cause

plugins/compound-engineering/skills/ce-sessions/scripts/extract-skeleton.py, platform auto-detection loop:

for line in sys.stdin:
    ...
    if not detected and len(buffer) <= 10:   # <-- caps detection to first 10 lines
        obj = json.loads(line)
        if obj.get("type") in ("user", "assistant"):
            detected = "claude"
        elif obj.get("type") in ("session_meta", "turn_context", "response_item", "event_msg"):
            detected = "codex"
        elif obj.get("role") in ("user", "assistant") and "type" not in obj:
            detected = "cursor"
...
handler = handlers.get(detected, handle_codex)   # <-- None falls through to codex

Current Claude Code session JSONL front-loads a metadata preamble before the first message. Observed first 10 record types in a real session:

1: last-prompt      6: attachment
2: custom-title     7: attachment
3: agent-name       8: attachment
4: mode             9: attachment
5: permission-mode 10: attachment

The first user/assistant record is at line 13. Because detection is capped at line 10, it never sees a message record, detected stays None, and handlers.get(detected, handle_codex) falls through to the codex handler. Every Claude record then fails the codex checks → user: 0, assistant: 0, tool: 0, parse_errors: 0 → empty skeleton.

Reproduces on a session that has 141 user + 300 assistant records — the parser reads all 1212 lines without error and extracts none.

Fix

Drop the 10-line cap so detection scans until it finds a decisive record (the if not detected: guard already short-circuits once set):

-    if not detected and len(buffer) <= 10:
+    if not detected:

Verified locally: the same session goes from user:0 assistant:0 tool:0 (88-byte empty skeleton) to user:26 assistant:106 tool:114 (59 KB skeleton with real content).

Optional hardening

The deeper trap is that an undetected platform falls through to handle_codex and produces empty output silently. Consider emitting a one-line stderr note when detected is None after the full scan, so a future format change surfaces as a visible warning instead of a silent "no sessions found." (Not required for the fix above.)

Affected

extract-skeleton.py is the shared extractor for ce-sessions; any skill routing session history through it (ce-sessions, ce-compound opt-in session pass) silently loses Claude session context until this lands. Found in v1.0.2 (e74e298).