Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

Memory plugins that register before_prompt_build and agent_end hooks via api.on() never have those hooks dispatched when sessions use the claude-cli provider.

The hooks are correctly registered (confirmed via openclaw plugins info showing all typed hooks), the plugin loads successfully ([plugins] Honcho memory plugin loaded / Honcho memory ready in gateway.log), and gateway_start hooks fire normally. However, per-request hooks (before_prompt_build, agent_end) are never called.

Investigation of the gateway source reveals that these hooks are dispatched only within pi-embedded (the native runner at pi-embedded-*.js:30075 and :32688 via getGlobalHookRunner()), but cli-runner (used by claude-cli provider) does not invoke hook dispatch points during the request lifecycle.

This means all memory plugins (including openclaw-honcho) are non-functional when using claude-cli as the execution provider, since they rely on before_prompt_build for context injection and agent_end for message capture.

Additionally, allowPromptInjection in the plugin hooks config must be explicitly set to true for before_prompt_build to be allowed by the loader (loader-*.js:1341), but this is not documented in the plugin installation guide.

Environment: OpenClaw v2026.4.9 (0512059), macOS, plugin openclaw-honcho v1.3.2, @honcho-ai/sdk v2.1.1

Steps to reproduce

1. Install openclaw-honcho v1.3.2: openclaw plugins install @honcho-ai/openclaw-honcho
2. Run openclaw honcho setup with a self-hosted Honcho server on localhost:8000
3. Set plugins.slots.memory = "openclaw-honcho" and plugins.allow = ["openclaw-honcho"] in openclaw.json
4. Set plugins.entries.openclaw-honcho.hooks.allowPromptInjection = true
5. Gateway reloads, plugin shows Status: loaded with all hooks registered
6. Open any chat session (provider = claude-cli)
7. Send messages and close the session
8. Check gateway.log: zero occurrences of agent_end, zero hook dispatch entries
9. Check /tmp/honcho.log: no new API calls from the plugin after initial gateway_start
10. Open a new session: no Honcho context injected (before_prompt_build never ran)

Expected: Hooks fire on every request (before_prompt_build) and after every turn (agent_end) Actual: Hooks never fire. Gateway log only shows [agent] cli exec: provider=claude-cli with no hook-related entries.

Expected behavior

When a memory plugin (e.g. openclaw-honcho) registers before_prompt_build and agent_end hooks, those hooks should fire during every request cycle regardless of which provider/runner is active. Specifically: before_prompt_build should inject retrieved memory context into the prompt, and agent_end should capture the conversation turn for storage. This works correctly with the pi-embedded runner.

Actual behavior

Neither before_prompt_build nor agent_end hooks fire when using the claude-cli provider. The cli-runner (cli-runner-DgGAOeQQ.js) does not contain any hook dispatch calls — only the pi-embedded runner dispatches these hooks (at lines ~30075 and ~32688 in pi-embedded-*.js). As a result, all memory plugins are non-functional with claude-cli. The gateway_start hook fires normally (gateway-level), confirming the plugin loads correctly but per-request hooks are skipped entirely.

OpenClaw version

2026.4.9 (0512059)

Operating system

macOS (Apple Silicon / Homebrew)

Install method

npm global

Model

anthropic/claude-sonnet-4-6

Provider / routing chain

openclaw → claude-cli → anthropic

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response