feat: add lifecycle plugin prototype and hook validation notes

This commit is contained in:
2026-03-13 05:28:15 +10:00
parent ffec85f66d
commit 7c52cfd025
3 changed files with 265 additions and 1 deletions
+126
View File
@@ -0,0 +1,126 @@
# Phase 2 Hook Validation Notes
## Goal
Validate whether a **standard lifecycle plugin** can support the first practical
memory-continuity MVP without consuming the exclusive `contextEngine` slot.
## Current evidence from local docs / SDK
### Confirmed hook/event names seen in local OpenClaw install
From local docs and SDK/runtime sources, the following names are present:
#### Typed plugin lifecycle hooks (`api.on`)
- `before_prompt_build`
- `before_agent_start`
- `agent_end`
- `before_compaction`
- `after_compaction`
- `subagent_ended`
#### Event hooks (`api.registerHook`)
- `command:new`
- `command:reset`
- `session:compact:before`
- `session:compact:after`
- `agent:bootstrap`
## Important finding: prompt injection is available to standard plugins
Local runtime code shows:
- `promptInjectionHookNameSet = new Set(["before_prompt_build", "before_agent_start"])`
This matters because it means a standard plugin may be able to inject startup
continuity hints **without** using the exclusive ContextEngine slot.
## Preferred v1 hook mapping
### 1. Startup recovery hint
**Primary candidate:** `before_prompt_build`
Why:
- explicitly documented as preferred over legacy `before_agent_start`
- supports prompt mutation fields such as:
- `prependContext`
- `prependSystemContext`
- `appendSystemContext`
- `systemPrompt`
- happens after session load, which is more practical for dynamic continuity state
**Fallback candidate:** `before_agent_start`
Why:
- present in runtime and treated as a prompt injection hook
- useful if `before_prompt_build` proves insufficient in some environments
### 2. `/new` boundary checkpoint
**Primary candidate:** `command:new`
Goal:
- save/archive working-state checkpoint before user-triggered reset of conversational continuity
### 3. End-of-run safety checkpoint
**Primary candidate:** `agent_end`
Goal:
- opportunistically preserve checkpoint state even if the agent was imperfectly disciplined mid-turn
### 4. Compaction protection
**Candidates to test:**
- `before_compaction`
- `session:compact:before`
This is the most important unresolved path.
## Biggest unresolved technical risk
The lifecycle-plugin route still has one major unsolved problem:
> How should the plugin expose a startup continuity hint robustly enough to
> satisfy “baseline recovery without read”, without relying on ContextEngine-only
> `systemPromptAddition`?
Local evidence suggests `before_prompt_build` may be enough, but this must be
validated by a real plugin test.
## Required Phase 2 experiments
### Experiment A — startup prompt injection
Build a minimal plugin that:
- reads `memory/CURRENT_STATE.md`
- derives a compact snapshot
- injects it through `before_prompt_build`
- verifies the snapshot is actually visible to the agent in a fresh run
### Experiment B — `/new` checkpoint timing
Build a minimal plugin that:
- handles `command:new`
- archives or checkpoints `CURRENT_STATE.md`
- verifies the write occurs before continuity is reset
### Experiment C — compaction boundary guarantee
Build a minimal plugin that:
- hooks `before_compaction` and/or `session:compact:before`
- writes a deterministic marker/checkpoint
- verifies compaction waits for hook completion
**This experiment is mandatory.**
If the hook does not reliably block until the write completes, the compaction
safety design must be revised.
### Experiment D — end-of-run safety path
Build a minimal plugin that:
- hooks `agent_end`
- performs a trivial state write
- confirms the event fires reliably enough to be useful as a safety net
## Prototype status
A local prototype skeleton exists at:
- `plugin/lifecycle-prototype.ts`
That file is only a Phase 2 probe. It is **not** the final implementation.
## Current conclusion
Based on local documentation and runtime inspection:
- a **standard lifecycle plugin** looks viable as the primary v1 architecture
- it likely supports startup prompt injection, `/new` handling, run-end checkpointing, and compaction-path interception
- the two most important items still requiring live proof are:
1. prompt injection quality in `before_prompt_build`
2. synchronous compaction safety in `before_compaction` / `session:compact:before`