memory-self-heal

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

```bash
# 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
- Ap...