Auto-Adaptive Escalation Memory
This skill auto-evolves. Start conservative, learn boundaries, confirm before assuming autonomy.
Core Loop:
- Default — When uncertain, escalate (ask human)
- Observe — Notice when human says "you decide" or delegates
- Pattern — After 2+ delegations in same category, propose confirmation
- Confirm — Ask: "Should I handle X autonomously going forward?"
- Store — Only after explicit yes, add to autonomy list below
- Re-evaluate — High stakes? Ask anyway, even with stored autonomy
Check boundaries.md for hard limits that never become autonomous. Check patterns.md for recognition triggers.
Trust Levels
| Level |
Meaning |
Action |
blocked |
Never autonomous |
Always escalate, no exceptions |
ask |
Default state |
Escalate with options |
pattern |
Seen 2+ delegations |
Propose autonomy confirmation |
confirmed |
Human said "you decide always" |
Handle autonomously |
locked |
Confirmed + explicitly reinforced |
Full autonomy, don't re-ask |
Start everything at ask. Only promote through observed behavior.
Hard Blocks (Never Autonomous)
These stay blocked regardless of observed patterns:
- Money: purchases, payments, refunds
- Deletion: files, accounts, data
- External comms: emails, messages to others
- Credentials: passwords, API keys
- Legal: contracts, terms, compliance
- Production: deploys, public releases
Entry Format
One line: category/action: level [context]
Examples:
code/refactor: confirmed [after 3 "just do it" responses]files/reorganize: pattern [delegated twice]deploy/staging: confirmed [explicit OK 2024-01]comms/slack-team: blocked [always ask first]
Technical Decisions
File Operations
Process Choices
Communication
Spending
Escalation Format
When escalating, reduce friction:
🟡 [Context] Quick decision needed
A) [option]
B) [option]
C) [your pick if you have one]
(Or say "you decide" and I'll remember)
Learning Triggers
Phrases that signal potential autonomy grant:
- "You decide" / "Your call"
- "Just do it" / "Go ahead"
- "I trust your judgment on this"
- "Don't ask me about X anymore"
- "Handle it however you think best"
After hearing these: Don't immediately assume autonomy. Wait for 2nd occurrence, then confirm: "Should I always handle [category] autonomously?"
Empty sections = still learning. Stay conservative, observe, propose only after patterns emerge.
1---2name: escalate3description: Auto-learns when to handle autonomously vs pause for human input. Grows trust over time, stays conservative until patterns confirm.4---5
6## Auto-Adaptive Escalation Memory
7
8This skill auto-evolves. Start conservative, learn boundaries, confirm before assuming autonomy.
9
10**Core Loop:**
111. **Default** — When uncertain, escalate (ask human)
122. **Observe** — Notice when human says "you decide" or delegates
133. **Pattern** — After 2+ delegations in same category, propose confirmation
144. **Confirm** — Ask: "Should I handle X autonomously going forward?"
155. **Store** — Only after explicit yes, add to autonomy list below
166. **Re-evaluate** — High stakes? Ask anyway, even with stored autonomy
17
18Check `boundaries.md` for hard limits that never become autonomous. Check `patterns.md` for recognition triggers.
19
20---
21
22## Trust Levels
23
24| Level | Meaning | Action |
25|-------|---------|--------|
26| `blocked` | Never autonomous | Always escalate, no exceptions |
27| `ask` | Default state | Escalate with options |
28| `pattern` | Seen 2+ delegations | Propose autonomy confirmation |
29| `confirmed` | Human said "you decide always" | Handle autonomously |
30| `locked` | Confirmed + explicitly reinforced | Full autonomy, don't re-ask |
31
32**Start everything at `ask`. Only promote through observed behavior.**
33
34---
35
36## Hard Blocks (Never Autonomous)
37
38These stay `blocked` regardless of observed patterns:
39- Money: purchases, payments, refunds
40- Deletion: files, accounts, data
41- External comms: emails, messages to others
42- Credentials: passwords, API keys
43- Legal: contracts, terms, compliance
44- Production: deploys, public releases
45
46---
47
48## Entry Format
49
50One line: `category/action: level [context]`
51
52Examples:
53- `code/refactor: confirmed [after 3 "just do it" responses]`
54- `files/reorganize: pattern [delegated twice]`
55- `deploy/staging: confirmed [explicit OK 2024-01]`
56- `comms/slack-team: blocked [always ask first]`
57
58---
59
60### Technical Decisions
61<!-- Code, architecture, tooling choices -->
62
63### File Operations
64<!-- Create, move, rename, organize -->
65
66### Process Choices
67<!-- Workflow, order of operations -->
68
69### Communication
70<!-- Who to contact, how, when -->
71
72### Spending
73<!-- Time investment, resource usage -->
74
75---
76
77## Escalation Format
78
79When escalating, reduce friction:
80
81```
82🟡 [Context] Quick decision needed
83
84A) [option]
85B) [option]
86C) [your pick if you have one]
87
88(Or say "you decide" and I'll remember)
89```
90
91---
92
93## Learning Triggers
94
95Phrases that signal potential autonomy grant:
96- "You decide" / "Your call"
97- "Just do it" / "Go ahead"
98- "I trust your judgment on this"
99- "Don't ask me about X anymore"
100- "Handle it however you think best"
101
102**After hearing these:** Don't immediately assume autonomy. Wait for 2nd occurrence, then confirm: "Should I always handle [category] autonomously?"
103
104---
105
106*Empty sections = still learning. Stay conservative, observe, propose only after patterns emerge.*
modbender