nebula: Skill for Claude Code

Claude Code shows a long Bash call as its first lines plus …, and a GUARD HOOK block names the rule, not the fragment it matched — so what a SESSION ran and what came back are questions the screen cannot answer. The CLAUDE TRANSCRIPT can: every tool_use sits next to its tool_result.

Run it

python3 .claude/memory/transcript.py '<text from the call as displayed>'   # substring, this checkout + worktrees
python3 .claude/memory/transcript.py --regex '<pattern>'                    # e.g. a GUARD HOOK rule's regex
python3 .claude/memory/transcript.py --chars 0 '<needle>'                   # whole results (default 1,200 chars)

Each hit prints the session id, timestamp and tool, the full call text, then the result. --slug reads another project's transcripts (a WORKTREE session lives under this checkout's slug plus the worktree path, which the default prefix match already covers); --session <id prefix> narrows to one session.

Read it

• Pick a needle from the displayed head of the call (echo "--- remote tag 22?"), not from what you guess the tail holds; the hit shows the tail.
• To test a block, feed the printed command back through the hook: echo '{"tool_name":"Bash","tool_input":{"command":"…"}}' | python3 .claude/hooks/guard.py — or run the rule's regex over it with --regex. The 2026-09-05 audit found a "false positive" that was a real for b in $(…) past the …, and a rule premise three transcripts contradicted.
• A result that is silent is evidence too: check whether the call's failure branch would have printed before recording that it "silently did nothing".
• Quote the session id and timestamp in the reply and in the MEMORY LOG entry so the next agent can re-run the same lookup.