sase-chats

name	sase_chats
description	Inspect prior sase agent chat transcripts. Use when the user asks about previous chats, chat transcripts, prior agent conversations, "what did agent X say?", "summarize the previous agent", or any question about an earlier sase agent's prompt or response.

Before doing anything else, run this command to record that you are using this skill:

sase skill use sase_chats --reason "<one-line reason for using this skill>"

Quick reference for answering "what did the previous agent say?" questions about sase chat transcripts.

Primary command

sase chat list -j

Prints a stable-shape JSON array of recent transcripts, newest first. Each row has: path, basename, mtime, size_bytes, workflow, agent, timestamp, prompt_snippet, response_snippet, followed by the sync-provenance fields provenance, source_machine, source_username, project_key, agent_artifact_dir, agent_local_name, agent_global_name, sidecar_repo, sidecar_relpath, publication_pending, publication_last_error, publication_quarantined, publication_attempts, publication_disposition. Prefer JSON over the pretty table when summarizing.

provenance is one of local (written here, not published), shared (written here and published to an agents sidecar), remote (imported from another machine — source_machine names the origin), or unknown (provenance could not be determined). Never report unknown as local.

Publication is independent of provenance: a shared chat can still have outstanding work for a committed-but-not-pushed publication or a later revision. publication_disposition is queued when every matching request is active, quarantined when every request has stopped retrying, mixed when active and quarantined requests coexist, or null when no matching request exists. publication_attempts is the maximum matching attempt count, and publication_last_error comes from the deterministically most recently updated request.

The compatibility boolean publication_pending is true for queued and mixed, because retryable work exists. publication_quarantined is true only for a fully quarantined disposition. Preserve publication_last_error in diagnostics and tell the user to run sase agent sync --retry-quarantined when they want to release quarantined work for retry. Remote provenance is authoritative and carries no local publication state.

Useful list options:

• sase chat list -j -l 20 — cap how many rows are returned (default is conservative).
• sase chat list -j -m '<machine>' — only transcripts whose source machine matches.
• sase chat list -j -P remote — only transcripts with that provenance.
• sase chat list -j -q '<text>' — case-insensitive content/path/basename filter.

Looking up a specific transcript

When the user names an agent, resolve by name first:

sase chat show --agent <name>

This follows the same fallback as #fork: completed agent's done.json["response_path"] first, then agent_meta.json["chat_path"].

Step-suffixed child workflows (e.g. <name>.plan, <name>.commit) are recorded as separate transcript entries — --agent <name> does NOT include them. A submitted plan in particular appears as a <name>.plan chat whose response_snippet begins with Plan submitted for review. and names the ~/.sase/plans/<YYMM>/<file>.md path.

If --agent <name> exits non-zero, walk this fallback chain before giving up:

1. sase chat list -j -q '<name>' — catches step-suffixed siblings (<name>.plan, <name>.commit, ...) that the named lookup missed.
2. sase agent list -a -j filtered to <name> — includes recently DONE/FAILED agents, not just RUNNING.
3. If the agent is still RUNNING, hand off to the /sase_agents_status skill for the artifacts-dir workflow (live reply, workflow checkpoints, mid-run plan drafts).

Only after that chain is exhausted should you tell the user the agent has no recoverable transcript or artifact. When that chain reaches live artifacts, name the artifact paths you read and label the evidence source as draft/live (live_reply.md from a running agent) or stable/completed (checkpoints, submitted plans, done.json, or saved transcripts).

When the user gives a path or basename:

sase chat show --path <path>
sase chat show --basename <basename>

Exactly one selector (--agent, --path, or --basename) is required.

Picking the right --format

sase chat show defaults to -f raw (full transcript markdown). Switch when the question is narrower:

• -f response — just the agent's latest response. Use this for "what did agent X conclude?" or "what was the recommendation?". Exits non-zero if no response heading can be parsed; fall back to -f raw and say so rather than guessing.
• -f resume — flattened User/Assistant turns with nested resume references expanded. Use this when the user wants the full chronological context (e.g. "summarize the back-and-forth").

How to summarize

• Always cite the agent name and path/basename you read so the user can open the source.
• Distinguish prompt content from response content — they live under separate headings in the transcript.
• Prefer short excerpts over long quotes; only include a longer quote when the exact wording matters.
• If sase chat list -j returns an empty array, say "no transcripts found" plainly — do not fabricate rows.
• If a named-agent lookup fails AND the fallback chain above (content-filtered list, -a status, /sase_agents_status artifacts) turns up nothing, say so plainly instead of falling back to an unrelated transcript.

Not for continuing a conversation

This skill is read-only. To continue a prior conversation in a new agent, use #fork:<name> or #fork_by_chat:<basename> xprompts — they remain the right tool for that.