feat: add global visibility scope workflow for OpenClaw learnings

[ZaneChen76]

Summary

This PR improves self-improving-agent so cross-agent/global persistence is an explicit first-class workflow in OpenClaw.

What changed

• Added Visibility Scope (OpenClaw) section with 3 scopes:
  • session
  • agent (default)
  • global
• Added explicit trigger phrases for global persistence requests (English + Chinese).
• Added Quick Reference entry for cross-agent persistence requests.
• Added canonical reference file: references/global-learnings.md with a reusable global-entry template (GLRN-*).

Why

Previously, learnings were usually written in the current agent workspace (.learnings/*), which is often agent-local.
This caused confusion when users expected "all agents/sessions" visibility.

This PR keeps current behavior, but documents a clear and actionable path when users request global visibility.

Backward compatibility

• No breaking changes.
• Existing local .learnings/* workflow remains unchanged.

Notes

If accepted, I can follow up with a small automation helper script that performs multi-agent pointer sync when scope: global is selected.

[pskoett]

I will have a look at this seems like a good idea

[miskaone]

Code review

Found 5 issues:

1. Global learnings path is inside the skill install directory -- data loss on upgrade. The workflow instructs writing to ~/.openclaw/skills/self-improving-agent/references/global-learnings.md (line 96), which gets overwritten on clawdhub install self-improving-agent. Any persisted global learnings will be silently destroyed. The path should be under ~/.openclaw/workspace/ or another stable location outside the versioned skill package.

self-improving-agent/SKILL.md

Lines 95 to 97 in 9591800

	1. Write canonical entry to a shared file (example: `~/.openclaw/skills/self-improving-agent/references/global-learnings.md`).
	2. Add/refresh a short pointer block in each agent's `AGENTS.md` (or equivalent workspace prompt file) that references the canonical global entry.

2. Direct conflict with PR fix: add anti-loop guardrails to prevent infinite tool call loops #5. PR fix: add anti-loop guardrails to prevent infinite tool call loops #5 deletes the entire Quick Reference table and all OpenClaw-specific content that this PR extends. PR fix: add anti-loop guardrails to prevent infinite tool call loops #5 also adds anti-loop guardrails (max 3 tool calls, no chaining) that this PR's multi-step global broadcast workflow (write shared file, update N agent files, announce) would violate. These two PRs cannot both merge without one undoing the other.

3. GLRN ID prefix not added to the ID Generation section. The new global-learnings.md template uses GLRN-YYYYMMDD-XXX (line 8), but SKILL.md's ID Generation section (line 252) only defines LRN, ERR, and FEAT. Agents following the documented ID scheme have no guidance for the new type.

self-improving-agent/references/global-learnings.md

Lines 7 to 9 in 9591800

	```markdown
	## [GLRN-YYYYMMDD-XXX] title

self-improving-agent/SKILL.md

Lines 249 to 254 in 9591800

	## ID Generation
	Format: `TYPE-YYYYMMDD-XXX`
	- TYPE: `LRN` (learning), `ERR` (error), `FEAT` (feature)
	- YYYYMMDD: Current date
	- XXX: Sequential number or random 3 chars (e.g., `001`, `A7B`)

4. Unbounded AGENTS.md write scope with no enumeration or idempotency. Step 2 (line 97) says "Add/refresh a short pointer block in each agent's AGENTS.md" but provides no mechanism to discover which agents exist, no deduplication check, and no dry-run option. Repeated invocations will keep appending pointer blocks.

self-improving-agent/SKILL.md

Lines 96 to 98 in 9591800

	1. Write canonical entry to a shared file (example: `~/.openclaw/skills/self-improving-agent/references/global-learnings.md`).
	2. Add/refresh a short pointer block in each agent's `AGENTS.md` (or equivalent workspace prompt file) that references the canonical global entry.
	3. Announce completion and affected agents.

5. Status vocabulary mismatch. The new template uses active | superseded (line 13 of global-learnings.md) while all existing templates use pending | resolved | in_progress | wont_fix. Existing grep-based review commands (grep "Status**: pending") will silently miss global entries.

self-improving-agent/references/global-learnings.md

Lines 12 to 14 in 9591800

	**Priority**: low | medium | high | critical
	**Status**: active | superseded

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}