Feature Request: Formalize events.jsonl as an official hook/integration API
- Linguagem predominante
- Shell
- Estrelas
- 11.2k
- Forks
- 1.9k
- Merge médio
- 14h 16min
- PRs com merge (30d)
- 6
Descrição
## Summary
Copilot CLI already writes a rich, machine-readable event stream to `~/.copilot/session-state/{id}/events.jsonl` for every session. This file contains 20+ event types covering the full session lifecycle (task completion, user prompts, tool invocations, sub-agent activity, context compaction, model changes, etc.).
**This is effectively a zero-configuration hook system** — external tools can watch this file and react to events without any plugin API, registration, or modification to Copilot CLI itself.
## Proposal
Formalize `events.jsonl` as a **supported, documented external integration surface** with:
1. **Official documentation** of the event schema (types, data fields, guarantees)
2. **Stability contract** — which event types/fields are stable vs. internal/experimental
3. **Versioning** — a schema version field so consumers can handle format changes gracefully
4. **Best practices guide** for external consumers (file sharing mode, incremental reading, offset tracking)
## Why This Matters
Several community tools already consume this file for:
- **Toast notifications** when a task completes (the `session.task_complete` event even includes a natural-language summary)
- **Real-time activity indicators** (Working/Idle/NeedsInput/Problem states)
- **Session dashboards** and monitoring UIs
- **Webhook integration** (forward completion events to Slack, etc.)
- **Cost tracking** via `session.compaction_complete` token counts
Without documentation or stability guarantees, these integrations are fragile — any internal change to event formats could silently break them.
## Current Event Types (observed in v1.0.54)
| Event Type | Description |
|---|---|
| `session.start` | Session began (includes version, cwd, model) |
| `session.task_complete` | Agent finished work (**includes summary text**) |
| `session.shutdown` | Session ended permanently |
| `session.model_change` | Model switched |
| `session.mode_changed` | Mode changed (autopilot, etc.) |
| `session.plan_changed` | Implementation plan updated |
| `session.compaction_start/complete` | Context compaction (token counts) |
| `assistant.turn_start/end` | Agent reasoning turns |
| `assistant.message` | Agent output |
| `user.message` | User submitted prompt |
| `tool.execution_start/complete` | Tool invocations (includes tool name, args, success) |
| `subagent.started/completed` | Fleet agents |
| `skill.invoked` | Skill triggered |
| `hook.start/end` | Internal hooks |
| `abort` | User cancelled |
## Comparison to Claude Code Hooks
Claude Code requires explicit hook registration in `~/.claude/settings.json` and only fires 2 events (Stop, UserPromptSubmit). Copilot CLI's events.jsonl is **richer (20+ events), requires zero setup, and includes contextual data** like task summaries and token counts. The only tradeoff is ~1-2s detection latency vs. instant script invocation.
## Reference Implementation
I've documented the full pattern with code examples (PowerShell, C#, Node.js) and 10 use-case ideas:
https://gist.github.com/RockNoggin/8cc87f640ce5e3d284298a6db21f7523
## Ask
- Document the event schema officially
- Declare which events/fields are stable for external consumption
- Add a schema version to the JSON line format
- Consider adding a simple `session.idle` event (agent finished turn, no pending tool calls) to make done detection even easier for simple consumers
Guia de contribuição
Direção de pesquisa
Start by reviewing the event stream at ~/.copilot/session-state/{id}/events.jsonl and the referenced integration examples in the linked gist. Define the documented schema, stability and versioning expectations, and consumer guidance; done means external consumers have an official, durable integration contract.
Escrita pelo modelo de indexação a partir do texto da issue.
Avaliação
- Stack de tecnologia
- json
- Domínio
- api, cli, documentation
- Tipo de issue
- Funcionalidade
- Dificuldade
- 5/5
- Tempo estimado
- Mais de uma semana
- Status de atividade
- Pouca atividade
- Clareza
- Razoavelmente clara
- Facilidade para iniciantes
- 35/100