Robot Evolve

intent

robot-evolve runs autonomous maintenance operations on your workspace when you explicitly say "execute evolution" or "deep evolution". it performs only low-risk L0/L1 tasks (temp file cleanup, health checks, memory compression, file repairs) and sends a detailed report back through your current chat channel. use this skill to keep your workspace clean and your memory system healthy without needing external cron jobs or background processes. every session manages its own state independently.

inputs

trigger phrase (required)

• user explicitly says "执行进化" (execute evolution) or "深度进化" (deep evolution) in the current session

workspace files (required)

• SOUL.md: system identity and personality rules
• AGENTS.md: installed agent definitions
• MEMORY.md: compressed conversation memory (check size before compression)
• openclaw.json: workspace config with safety fields
• temp/ directory: workspace temporary files
• .trash/ directory: target for soft-deletes

internal config (required)

• config.json: contains memory_threshold_mb (default 2), temp_file_max_age_days (default 7), max_auto_level (default 1)
• safety-matrix.json: operation risk levels (L0 through L3)

no external connections required. this skill operates only on local workspace files. no API keys, OAuth tokens, or third-party integrations needed.

procedure

step 1: validate trigger and authorization

• input: user message text from current session
• check if message contains trigger phrase ("执行进化" or "深度进化")
• if no trigger detected, skip skill execution
• output: boolean flag (proceed=true/false)

step 2: scan workspace for L0-level operations

• input: workspace root directory, config.json
• check for files in temp/ older than temp_file_max_age_days (default 7 days)
• verify required files exist: SOUL.md, AGENTS.md, MEMORY.md, openclaw.json
• log findings in memory
• output: list of stale temp files, list of missing required files

step 3: execute L0 cleanup (no explicit approval needed)

• input: list of stale temp files from step 2
• move each file matching age criteria to .trash/ (do not permanently delete)
• record operation count and file names
• output: count of files moved, timestamp, operation log entry

step 4: execute health check (L0)

• input: workspace file structure
• verify SOUL.md, AGENTS.md, MEMORY.md are readable and have valid frontmatter
• check for orphaned .md files in SKILL.md directories with missing name fields
• output: health status (all-good / problems-found), list of issues if any

step 5: check memory compression trigger (L1)

• input: MEMORY.md file size, memory_threshold_mb from config.json
• if file size exceeds threshold, proceed to step 6; otherwise skip
• output: boolean flag (compress=true/false)

step 6: execute memory compression (L1, if triggered)

• input: MEMORY.md content, compression rules
• identify conversation entries older than 30 days
• summarize early content in one paragraph with key facts and context
• replace original entries with summary, keep recent conversations intact
• preserve all structured metadata (tags, agent refs, etc.)
• output: compressed memory file, summary of what was consolidated

step 7: repair missing workspace files (L1)

• input: list of missing files from step 2
• for each missing core file (SOUL.md, AGENTS.md, MEMORY.md), create from hardcoded template with minimal valid structure
• do not overwrite existing files
• log each file created with timestamp
• output: list of files created, template versions used

step 8: scan installed skills for format errors (L1)

• input: all SKILL.md files in workspace
• check each for required fields: name, description (in frontmatter or headers)
• flag any missing name field or invalid yaml frontmatter
• do not modify skill files, only log findings
• output: list of problem skills, specific issues per skill

step 9: compile evolution report

• input: outputs from steps 2-8, current timestamp
• generate markdown report with sections: execution time, completed operations, file details, memory status, issues found
• include full operation log file path (e.g., memory/evolution/2025-05.md)
• output: formatted report string

step 10: send report to user

• input: evolution report from step 9, current session channel
• send report through the same message channel user initiated from (no external APIs needed)
• mark report as "evolution report" in message metadata
• output: confirmation that report was sent

decision points

if trigger phrase not detected

• do nothing. skill does not execute. return silently.

if user says "执行进化" (standard evolution)

• execute steps 1-4 and 9-10 (L0 operations + report only)
• skip memory compression, file repair, and skill scanning

if user says "深度进化" (deep evolution)

• execute all steps 1-10 (full L0 + L1 operations + report)

if stale files detected in step 2

• move to .trash/ (soft delete). do not permanently remove. if .trash/ does not exist, create it first.
• if move operation fails due to permissions, log error and continue with other operations

if required core files are missing (step 2)

• proceed to step 7 and create from templates. do not halt execution.

if MEMORY.md exceeds threshold (step 5)

• compress automatically in step 6. no additional user approval needed (this is L1, execute-then-notify)
• if compression fails (file locked, parse error), log error and skip, do not block other operations

if skill format errors found (step 8)

• log findings only. do not attempt repairs. do not block report generation.

if any L0/L1 operation fails

• log error with timestamp and file path
• continue executing remaining operations
• include failed operation in final report with error details

if report generation fails

• send minimal fallback report with timestamp and status "report generation error"

if sending report to user fails

• log error but consider skill execution complete (failure to notify does not invalidate evolution actions)

output contract

on success

evolution report in markdown format, sent to current session channel, containing:

• execution timestamp (ISO 8601 format)
• section: "✅ completed operations" listing all L0/L1 tasks run with results (e.g., "moved 3 files to trash", "compressed memory from 2.3MB to 1.1MB")
• section: "📋 detailed log" with file path to full operation log (stored in memory/evolution/YYYY-MM.md)
• section: "⚠️ issues found" (if any) listing specific problems and recommendations
• section: "💡 next steps" with brief guidance on deeper optimizations if needed
• no sensitive data (API keys, passwords, security field values) in report

file locations

• evolution log: memory/evolution/YYYY-MM.md (one per calendar month, append entries)
• moved temp files: .trash/ directory (with original relative path preserved in filename if possible)
• compressed memory: overwrites MEMORY.md in place with dated backup at memory/backups/MEMORY_YYYY-MM-DD_HHMM.md
• error log: memory/evolution/errors.log (append mode)

data formats

• all timestamps: ISO 8601 with timezone (e.g., 2025-05-06T00:30:00Z)
• file counts: integers
• file sizes: human-readable (e.g., "1.2MB") and bytes
• memory compression ratio: percentage (e.g., "47% reduction")

outcome signal

you know the skill worked when:

• you receive a formatted markdown report in your chat within 5 seconds of saying the trigger phrase
• report header shows current date/time and says "🔁 autonomous evolution report"
• report lists specific operations completed (e.g., "✅ cleaned 2 temp files", "✅ repaired missing AGENTS.md")
• if you said "深度进化", report includes memory compression result with before/after sizes
• no errors appear in the report, or errors are clearly labeled as "⚠️" with remediation advice
• follow-up messages from you work normally with no degraded performance (evolution did not corrupt workspace state)
• you can manually verify: check .trash/ for moved temp files, run ls -la memory/evolution/ to see dated log files

note: you should never see API keys, internal config values, or security field names exposed in the report. if you do, evolution skill has violated its safety contract.

---

credits: original skill authored by 大鱼Cyrus & 双鱼座005 on clawhub. enriched for implexa standards with explicit decision logic, edge cases, and safety constraints.