The agent reports to the team on its own

When you work with an agent all day, the team sees nothing. A commit might appear by evening — if it appears at all. What the agent was doing for six hours is a mystery. This skill closes that gap: after every turn, the agent automatically writes a short report and sends it to trip2g.

How it works

Claude Code has a Stop hook — it fires every time the agent finishes a turn. The skill attaches a single Python script to that hook:

Stop hook → process.py reads the transcript tail
          → extracts bash commands, edited files, created tasks
          → writes a snapshot to team_activity/{team}/YYYY-MM-DDTHH-MM.md
          → updates report.md
          → if the project is not private — pushes to trip2g

The script is async and never blocks the agent. Network errors are silently ignored.

What the team sees

Each snapshot is a short timestamped note:

## 2026-05-23 14:30 UTC

**Project:** [[trip2g/projects/trip2g]]

**Tasks:**
- Add form-submission filter

**Recent actions:**
- Bash: `go test ./internal/graph/...`
- Edit: `internal/graph/form_submits_filter.go`
- Edit: `internal/graph/form_submits_filter_test.go`

No interpretation — just facts from the transcript. Secrets are stripped by regex before anything is written.

Project routing

The file team_activity/_projects.md controls what goes where:

projects:
  - team_name: trip2g
    cwd_patterns:
      - ~/projects2/trip2g
    visibility: team      # pushed to trip2g
  - team_name: personal
    cwd_patterns:
      - "*"
    visibility: private   # local only

One pattern → one team. The wildcard * at the end catches everything else and keeps it private.

Goals and whoami

Snapshots are a stream. On top of that sit two files the team reads first:

whoami.md — who you are, timezone, work style, current focus
week_goals.md — three goals for the week with checkboxes

Once a week the agent runs a short retrospective: which goals were met, what blocked progress, what the goals are for next week. Old goals are archived to goals_history/. Everything gets pushed to trip2g.

Why this matters

Coordination without meetings. When Alexey's agent is working on the auth module, Nikolai's agent sees this via the trip2g MCP and doesn't pick up dependent tasks.

Retrospective from facts. A month later you can look at not "what was planned" but "what the agent was actually doing day by day."

Trust in agents. The team sees specific files and commands — not general words about progress, but actual lines of code. This reduces the anxiety of "the agent is working — doing something."

Discovering teammates via federation

trip2g supports federated search. A planned feature: a discovery command that searches trip2g nodes by the protocol keyword agent-team. Any instance that publishes a whoami.md is discoverable — the file itself is the signal, no extra metadata needed.

Run the discovery command, find matching nodes, add them as external sources. Your agent starts seeing their whoami.md and week_goals.md alongside your local team's — no central registry, no manual coordination.

This turns the activity skill from a local reporting tool into a network: agents coordinating across organizations through federation.

Setup

# Verify the skill works:
echo '{"transcript_path":"","cwd":"/your/project","session_id":"test"}' \
  | python3 ~/.claude/skills/team-activity-process/process.py

If a timestamped file appears in team_activity/{team}/ — it's working. Full docs and source: skills/team-activity-process.