transcript

Transcript

Use session_transcript; never read events.jsonl directly.

Invoking this skill is a request to retrieve the transcript, not to explain how to retrieve it. Call session_transcript in this turn using the arguments below. Do not stop after loading the skill, present a proposed tool call, or ask "shall I retrieve it?" The user's request already authorizes this read. Only ask for ID clarification after the tool reports ambiguity; honor any actual tool permission gate or explicit user restriction.

Sensitive content: stored captures are replayed verbatim and can contain sensitive content. The logging hook's JSON sanitization is not redaction.

Interpret $ARGUMENTS as follows:

• ID or ID -- <intent>: a bare full session ID or a hex ID prefix of at least 8 characters (for example, /transcript abcdef12) selects that session. Pass it unchanged in session_ids; the tool resolves it. Do not ask for the full ID first or treat it as an instruction about the current session. With no intent, replay verbatim; with intent, retrieve first, then perform that intent.
• No arguments: call session_transcript with no session_ids. Return the complete role-marked transcript pages verbatim. Do not summarize or add commentary.
• Intent only: call session_transcript with no session_ids, page through the whole transcript, then perform the requested intent using it as source material. State that the answer was based on the native transcript.
• --session ID[,ID...]: retrieve those sessions in the listed order. With no further intent, replay each transcript verbatim under its own session heading.
• --session ID[,ID...] -- <intent>: retrieve all named sessions first, then perform the intent using those transcripts.

A bare token selects a session only when it actually looks like a capture ID: eight or more hex characters (abcdef12, a full UUID), or the zero-padded sub-agent form 0000000000000000-<hex>[_<agent>]. Anything else is intent about the current session, even when it contains a hyphen, underscore, or digit — chapter-2 and summarize-day-3 are intents, not IDs. When a token could be read either way, treat it as intent: a wrong intent is visible and recoverable, a wrong lookup just fails. For an opaque ID that does not match those shapes, use explicit --session, which accepts unique prefixes and exact opaque IDs. If natural language explicitly names a session, pass that ID or prefix to the tool as well.

On an ambiguous prefix, show the tool's candidates and pass on the length it reports as needed — it names the character position where the candidates diverge, so relay that rather than asking vaguely for "a longer prefix". On a missing session, report that failure. Never choose a candidate, retry against the current session, or delegate to an agent to guess an ID.

A session whose capture directory exists but has no readable metadata.json is reported as capture_unavailable, and an ID whose only captures are derived <id>_<agent> sub-agent sessions is reported as session_not_found listing those sub-agents. Neither is a prompt to substitute one of them; report what the tool said.

For a single session, keep calling session_transcript with its returned full session_id and next_after_event_line as the next after_event_line while has_more is true. For multiple sessions, retrieve and page one session at a time. For a batched follow-up, replace the original prefixes in session_ids with returned full IDs and use those same full IDs as after_event_lines keys. A page boundary is normal; preserve the tool's [MORE MESSAGES AVAILABLE ...] marker between pages. Do not hide capture issues, infer missing messages, use graph reconstruction, or fall back to raw provider events.