memory-self-heal（记忆自愈）

Memory Self-Heal Skill

Use this skill when the agent starts failing repeatedly, stalls, or keeps asking the user for steps that could be inferred from prior evidence.

Goals

1. Recover execution without user micromanagement
2. Reuse previous fixes from memory/logs/tasks
3. Escalate only with minimal unblock input when truly blocked
4. Leave reusable evidence for future runs

When To Trigger

Trigger when any of these appear:

• Same or similar error occurs 2+ times in one task
• Tool call fails due to argument mismatch, missing config, auth wall, or context overflow
• Agent claims completion without verifiable artifact
• Task progress stalls (no new artifact across 2 cycles)

Inputs

• Current task objective
• Latest error/output
• Available evidence locations (memory, tasks, logs)

Portable Evidence Scan Order

Scan these in order; skip missing paths silently:

1. memory/ (or equivalent workspace memory path)
2. tasks/ or queue files
3. runtime logs / channel logs
4. skill docs (skills/*/SKILL.md) for known fallback recipes
5. core docs (TOOLS.md, CAPABILITIES.md, AGENTS.md)

Shell examples (use whichever shell is active):

# PowerShell
Get-ChildItem -Recurse memory, tasks -ErrorAction SilentlyContinue |
  Select-String -Pattern "error|blocked|retry|fallback|auth|token|proxy|timeout|context" -Context 2

# POSIX shell
rg -n "error|blocked|retry|fallback|auth|token|proxy|timeout|context" memory tasks 2>/dev/null

Failure Classification

Classify first, then act:

• syntax_or_args: command syntax/argument mismatch
• auth_or_config: key/token/env/config missing or invalid
• network_or_reachability: timeout, DNS, handshake, region restrictions
• ui_login_wall: page requires manual login/attach
• resource_limit: context window, rate limit, memory pressure
• false_done: no artifact/evidence but reported complete
• unknown: no confident class

Recovery Policy (3-Tier)

Attempt 1: Direct Fix

• Apply best-known fix from memory for same class/signature
• Re-run the smallest validating action
• Record result

Attempt 2: Safe Fallback

• Switch to alternate tool/path with lower fragility
• Narrow scope (smaller input, shorter query, one target)
• Re-run validation

Attempt 3: Controlled Escalation

• Mark blocked with minimum unblock input
• Provide exact next action user must do (one command or one UI step)
• Do not loop further until new input arrives

Safety Rules

• Never auto-run destructive operations without confirmation
• Never log secrets/tokens in memory files
• Max 3 retries per blocker signature per task
• Prefer deterministic steps over broad speculative retries

Completion Contract

Do not claim done unless all are true:

• At least one artifact exists and is readable (file/link/output)
• The original task objective is explicitly mapped to artifact(s)
• No unresolved blocker for current objective

Required output block:

DONE_CHECKLIST
- Objective met: yes/no
- Artifact: <path or URL or command output ref>
- Validation: <what was checked>
- Remaining blocker: <none or exact unblock input>

Memory Writeback Template

Append one concise entry after each self-heal cycle:

## Self-heal: <date-time> <short task>
- Signature: <normalized error signature>
- Class: <classification>
- Attempt1: <action> -> <result>
- Attempt2: <action> -> <result>
- Final: <success | blocked>
- Artifact/Evidence: <path|url|log ref>
- Reusable rule: <one-line rule>

Generic Known Fixes (Seed Set)

• Command mismatch on Windows: prefer native PowerShell cmdlets
• Token mismatch/auth failure: verify active config source and token scope
• WebSocket/timeouts: test reachability + proxy/no_proxy consistency
• Context overflow: split task into smaller units and reduce payload
• False completion: enforce artifact validation before final response

Integration Notes

• Works with autonomy/task-tracker skills but does not depend on them
• If a project has custom memory paths, adapt scan roots dynamically
• Keep entries short to avoid memory bloat