fix(openclaw-plugin): default recall to assemble and bound afterTurn

[0xble]

Description

Default the OpenClaw plugin to an assemble-first recall path when OpenViking
owns the contextEngine slot, and fail open if the plugin's hot-path capture
work gets slow.

This avoids running hook-based recall and context-engine recall on the same
reply path, updates the setup helper to resolve the active OpenClaw config
file so existing installs upgrade cleanly, and bounds afterTurn auto-capture
so slow OV writes do not keep an OpenClaw run open after the model already
replied. It also keeps default plugin logs content-safe: session summaries,
message previews, and captured turn text are only emitted when debug routing
logs are enabled.

Related Issue

Related to #1283

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)
• Performance improvement
• Test update

Changes Made

• Moved recall assembly into assemble() by default and added a recallPath
  config so before_prompt_build only runs in explicit compatibility mode.
• Updated the default OpenClaw plugin install path to set
  recallPath=assemble, disable prompt injection in context-engine mode, and
  resolve the active OpenClaw config file instead of assuming
  openclaw.json.
• Bounded afterTurn auto-capture with a fail-open timeout so slow OpenViking
  session writes cannot block OpenClaw run completion after the assistant has
  already produced a reply.
• Gated content-bearing diagnostic logs behind logFindRequests and kept
  default info logs to non-content metadata such as counts, status, task, and
  trace IDs.
• Kept ingest-reply assist available when hook-mode recall client
  initialization fails, so a recall outage does not suppress the non-client
  prompt assist path.
• Added regression coverage for the assemble-first flow, compatibility mode,
  setup-helper config handling, content-safe default logging, recall init
  fallback, and afterTurn timeout fallback behavior.

Testing

• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have tested this on the following platforms:
  • Linux
  • macOS
  • Windows

Commands run:

• npm exec --package typescript -- tsc --noEmit
• npm test -- tests/ut/local-startup-failure.test.ts tests/ut/plugin-bypass-session-patterns.test.ts tests/ut/context-engine-assemble.test.ts
• npm test
• git diff --check
• cr review --type committed --base upstream/main --agent

Checklist

• My code follows the project's coding style
• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Screenshots (if applicable)

N/A

Additional Notes

• I reproduced the original failure locally on OpenClaw 2026.4.12 with the
  OpenViking plugin at v0.3.5.
• This PR stays plugin-scoped. It does not change OpenClaw's separate
  gateway-timeout fallback behavior.
• Merged current upstream main into the PR branch to resolve the reported
  conflict.

[CLAassistant]

All committers have signed the CLA.

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 1283 - Partially compliant Compliant requirements: Auto-recall logic moved to assemble() via recallPath="assemble" default before_prompt_build hook skipped when recallPath !== "hook" Dual conflict avoided by separating recall paths Non-compliant requirements: (empty) Requires further human verification: (empty)

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🏅 Score: 88

🧪 PR contains tests

🔒 No security concerns identified

✅ No TODO sections

🔀 Multiple PR themes Sub-PR theme: Improve OpenClaw config path detection Relevant files: examples/openclaw-plugin/setup-helper/install.js Sub-PR theme: Move recall to assemble() for OpenClaw 4.5 compatibility Relevant files: examples/openclaw-plugin/config.ts examples/openclaw-plugin/context-engine.ts examples/openclaw-plugin/index.ts examples/openclaw-plugin/recall-context.ts examples/openclaw-plugin/openclaw.plugin.json examples/openclaw-plugin/tests/ut/config.test.ts examples/openclaw-plugin/tests/ut/context-engine-assemble.test.ts examples/openclaw-plugin/tests/ut/local-startup-bad-config-real.test.ts examples/openclaw-plugin/tests/ut/local-startup-failure.test.ts examples/openclaw-plugin/tests/ut/plugin-bypass-session-patterns.test.ts examples/openclaw-plugin/tests/ut/plugin-normal-flow-real-server.test.ts Sub-PR theme: Update plugin documentation for new recall path Relevant files: examples/openclaw-plugin/INSTALL.md examples/openclaw-plugin/README.md

⚡ Recommended focus areas for review Bare Error Swallowing The resolveMemoryContent function uses a bare catch block that swallows exceptions without logging. This could hide real errors when reading memory content, making debugging harder. try { const fullContent = await readFn(item.uri); content = fullContent && typeof fullContent === "string" && fullContent.trim() ? fullContent.trim() : (item.abstract?.trim() || item.uri); } catch { content = item.abstract?.trim() || item.uri; }

[github-actions]

PR Code Suggestions ✨

No code suggestions found for the PR.

[Mijamind719]

“这些 logger.info 现在会在非 debug 模式下输出 session context / summary / message preview，默认日志会带上用户内容。建议收敛到 logFindRequests/debug 开关后，或只保留计数与 trace_id 等非内容信息。”

代码有confilict

[0xble]

Kept the ingest-reply-assist fallback even when recall init fails, pushed the conflict resolution, and tightened the noisy session-content logs so full summaries/message previews stay behind debug logging.

[0xble]

Did one simplify pass on the current branch head. This only removes a duplicated roughEstimate(messages) call in the assemble path and reuses the same token estimate for passthrough decisions; behavior stays the same.

[Mijamind719]

I found one functional regression in the new afterTurn / assemble flow.

1. Tool call history no longer round-trips through OpenViking session storage.
   • extractNewTurnMessages() now keeps only text blocks for normal user / assistant messages, so assistant-side toolUse blocks are dropped during capture (text-utils.ts, around extractPartText() and extractNewTurnMessages()).
   • toolResult messages are then persisted as role="user" tool parts (text-utils.ts, context-engine.ts).
   • On the read path, convertToAgentMessages() only reconstructs tool parts when msg.role === "assistant", so those stored tool results are ignored and come back as an empty user message (context-engine.ts).

I reproduced this locally with a focused afterTurn -> assemble round-trip using:

• one assistant message containing text + a toolUse
• one matching toolResult

The stored OV session contained:

• an assistant message with only the text part
• a user message with the tool part

Then assemble() returned:

• the assistant text
• an empty user message

instead of reconstructing the original tool-use / tool-result pair.

Impact: once a session is restored from OV, agents can lose the actual tool invocation/result history they may need for continuity, debugging, or follow-up decisions.

I’d recommend either:

• preserving tool-use / tool-result under a shape that can round-trip cleanly through convertToAgentMessages(), or
• teaching convertToAgentMessages() to reconstruct tool parts stored on non-assistant roles,

and adding a dedicated round-trip test that exercises afterTurn() followed by assemble().

For context, I also ran the current plugin test suite locally (vitest run), and it passed; this regression seems to be a missing round-trip coverage gap rather than an already-failing path.