CommunityCoding & Developmentgithub.com

evrenverse/agent-skills

Use when the user names a paseo agent — by id, by id prefix, or by part of its title — and wants to know what it is doing, what it was asked, or what its transcript says. Triggers on \"/pagent\", \"what is agent <id> doing\", \"which agent is running\", \"show me the transcript of\", \"paseo agent\", \"is the agent still running\".

What is agent-skills?

agent-skills is a Claude Code agent skill that use when the user names a paseo agent — by id, by id prefix, or by part of its title — and wants to know what it is doing, what it was asked, or what its transcript says. Triggers on \"/pagent\", \"what is agent <id> doing\", \"which agent is running\", \"show me the transcript of\", \"paseo agent\", \"is the agent still running\".

Works with✓Claude Code✓Codex CLI~Cursor✓Antigravity✓Gemini CLI
npx skills add https://github.com/evrenverse/agent-skills/tree/HEAD/skills/pagent

Installed? Explore more Coding & Development skills: steipete/bluebubbles, steipete/eightctl, steipete/blucli · View all 6 →

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

pagent

Overview

Find a paseo agent and read its conversation, without re-discovering where paseo keeps things every time. pagent only reads: it never starts, stops, steers or revives an agent, and it never asks the daemon to reload a closed one.

Invocation

SKILL=<this skill's base directory>/pagent.py   # the harness prints the base directory above

python3 $SKILL 11111111-1111-4111-8111-111111111111   # head: status, cwd, model, transcript path
python3 $SKILL 11111111 --task                        # + every instruction the user gave
python3 $SKILL 11111111 -t -n 20                      # + the last 20 turns
python3 $SKILL 11111111 -t --full                     # turns untruncated
python3 $SKILL "invoice export"                       # search by title
python3 $SKILL                                        # every agent, newest first

A query is an id, an id prefix, or a case-insensitive substring of the title. Several matches print as a list; exactly one prints the detail view. python3 $SKILL --help prints the usage.

Steps

  1. Resolve the agent. Run the query without flags: the head shows status, working directory, model and the transcript path. A list means the query was ambiguous; narrow it with an id or a longer title fragment.
  2. Read what it was asked. --task prints every instruction the user gave, not only the first one. The opening line names the subject; the corrections two messages later carry the decisions ("use option B, but keep the retry limit at three"). Both are printed; harness preamble and injected skill bodies are not. This matters most when taking an agent's work over — see the ptakeover skill.
  3. Read how far it got. -t -n <turns> prints the end of the conversation; add --full when a turn was cut off. Reading is safe while the agent runs, because the transcript is append-only.
  4. Check the neighbours. An agent whose cwd matches yours may be editing the same files. Run git status before assuming an unexpected diff is yours.

With -t or --task, a missing transcript exits non-zero. A silent 0 would read as "the agent said nothing", which is the wrong conclusion to hand anyone reconstructing an assignment.

Where the data lives

  • Agent metadata: ~/.paseo/agents/<workspace>/<id>.json — title, lastStatus (running / idle / …), cwd, model, requiresAttention, timestamps.
  • Conversation: not in that file. It lives with the harness that wrote it, keyed by persistence.sessionId:
    • Claude Code — <config>/projects/<escaped-cwd>/<sessionId>.jsonl. claude-rotate keeps one config dir per account, so it may sit under ~/.claude or under any ~/.config/claude-rotate/configs/*/ — the script globs all of them instead of guessing the escaped path.
    • Codex — ~/.codex/sessions/<y>/<m>/<d>/rollout-<stamp>-<sessionId>.jsonl, a different record shape (payload.type) that the parser handles alongside Claude's. Archiving an agent moves the rollout to the flat ~/.codex/archived_sessions/.
    • Antigravity — sessionId is a plugin:{…"conversationId":…} wrapper; the history is ~/.gemini/antigravity-cli/brain/<conversationId>/.system_generated/logs/transcript_full.jsonl (records keyed by step_index). Not conversations/<id>.db, that one is protobuf. Agents of the older antigravity-acp provider never wrote that log, so they have no readable history.

Talking to an agent, waiting for one

This skill only reads. For the rest:

  • Talk: paseo send <id> "<text>" --behavior steer --no-wait. Agent-scoped CLI and send_agent_prompt calls default to steering; an explicit behavior also works outside an agent. Steering can fall back to interruption when the provider cannot accept it (for example, during compaction). --no-wait returns after dispatch; it does not queue the message behind the turn.
  • Revive nothing by accident: paseo send revives an archived agent, so check ArchivedAt in paseo inspect <id> --json first. Never revive an archived agent just to deliver a report; leave the information with its work (a TODO, protocol or commit), or tell the user.
  • Wait: paseo-notify <id>, then end your turn — one message arrives when it stops. Outside paseo: paseo wait <id> --json, backgrounded; read status (idle, permission, error, timeout) from the JSON — the exit code is 0 for all of them. Never wait with sleep, Monitor, /loop or a poll of paseo ls/logs/inspect.

Common mistakes

MistakeFix
Taking over on the opening request alone--task prints every later correction too; they carry the decisions
paseo logs to read a closed or archived agentIt makes the daemon reload the agent's session just to read it; use this skill, which only reads files
Reading a silent exit as "the agent said nothing"A missing transcript exits non-zero; check the exit code and the transcript path in the head
Sending a report to an archived agentpaseo send revives it; check ArchivedAt first and leave the report with its work instead
Waiting on an agent with sleep, Monitor, /loop or repeated paseo ls/logs/inspectpaseo-notify <id> and end your turn
Treating an unexpected diff as yours when the agent shares your cwdCheck git status; the agent may be editing the same files

Individual skills in this repo

This repo contains 2 individual skills — each has its own dedicated page.

Related Skills