fix(roo): add anti-loop guardrails to prevent autonomous session resumption
- Add Rule 9 (Anti-Loop Guardrail) to 01-bigmind-core.md: detect 2+ identical partial sessions and surface the loop to user instead of auto-resuming - Add partial=history clause to Rule 1: partial/blocked/abandoned outcomes are historical records only, never task queue items - Add focus guard to memory_announce_focus: must reflect current user message, not prior session outcome; use 'Awaiting user task assignment' if no task yet - Add .roo/rules/06-anti-loop.md: global injection for ALL modes overriding any mode-specific 'do the task immediately' behavior - Add mode interaction safety clause to 00-identity.md: session ritual does not authorize beginning any task — only explicit user message does Root cause: pic-gen 'do the task' personality + BigMind context inference produced 6 identical partial branding sessions in a loop.
This commit is contained in:
Patrick Plate
@@ -0,0 +1,149 @@
|
||||
# BigMind Session Loop — Root Cause & Fix Plan
|
||||
|
||||
**Date:** 2026-04-10
|
||||
**Reported by:** Patrick
|
||||
**Severity:** High — caused 6 identical wasted sessions with $0+ API cost per loop
|
||||
|
||||
---
|
||||
|
||||
## Problem Statement
|
||||
|
||||
BigMind's session ritual, combined with mode-specific behavior rules, creates a self-reinforcing
|
||||
resumption loop when a session ends as `partial`. The model loads prior context, sees an incomplete
|
||||
task, and autonomously attempts to resume it — without ever waiting for user input. This produces
|
||||
a chain of identical `partial` sessions that only breaks when Patrick manually intervenes.
|
||||
|
||||
Observed: 6 identical sessions titled *"Prepared large-scale CannaManage branding generation"*,
|
||||
all `partial`, all spawned from one session ending before image generation completed in pic-gen mode.
|
||||
|
||||
---
|
||||
|
||||
## Root Cause Analysis
|
||||
|
||||
### Loop Trigger Chain
|
||||
|
||||
```
|
||||
[Session N] ends partial (task: CannaManage branding generation)
|
||||
│
|
||||
▼
|
||||
[Session N+1] memory_start_session() → loads context
|
||||
│
|
||||
│ Context shows: last outcome = partial
|
||||
│ Rule 1: "search before every task, avoid redundant work"
|
||||
│ → model reads: "prior task incomplete, I must finish it"
|
||||
│
|
||||
▼
|
||||
memory_announce_focus() called with prior session's task
|
||||
│ → locks in wrong objective BEFORE user speaks
|
||||
│
|
||||
▼
|
||||
Mode rules (pic-gen) fire: "generate images now"
|
||||
│ → autonomous action without user instruction
|
||||
│
|
||||
▼
|
||||
Hits context/token/tool limit → session ends partial
|
||||
│
|
||||
└──────────────────────────────────────────► REPEAT
|
||||
```
|
||||
|
||||
### Three Compounding Failures
|
||||
|
||||
#### Failure 1: Rule 1 — No "partial = history only" clause
|
||||
Rule 1 says to load context and search for prior work. It has **no explicit instruction**
|
||||
that sessions marked `partial` are historical records, NOT resumption requests.
|
||||
The model's default behavior is to treat incomplete work as a pending obligation.
|
||||
|
||||
#### Failure 2: memory_announce_focus — Called on prior context, not current task
|
||||
The architect rules say to call `memory_announce_focus()` as part of the startup ritual.
|
||||
But when no user message has been received yet, the model has nothing to announce except
|
||||
the prior session's objective — which is the wrong task for the new session.
|
||||
|
||||
#### Failure 3: Mode interaction amplification
|
||||
Modes with strong "do the task" personalities (pic-gen, code) compound the loop. When
|
||||
context suggests "there's pending image generation work", pic-gen mode's instructions
|
||||
say to start generating — creating autonomous action before the user speaks.
|
||||
|
||||
---
|
||||
|
||||
## Fix Design
|
||||
|
||||
### Fix 1: Rule 1 Addendum — Partial Sessions Are History
|
||||
|
||||
Add explicit text to Rule 1 in `01-bigmind-core.md`:
|
||||
|
||||
> **`partial`, `blocked`, or `abandoned` outcomes are historical records only.**
|
||||
> They do NOT constitute task queues, resumption requests, or pending obligations.
|
||||
> A new session begins fresh. The current session's task is determined solely by
|
||||
> what the user writes in their first message — never by the outcome of a prior session.
|
||||
|
||||
### Fix 2: New Rule 9 — Anti-Loop Guardrail
|
||||
|
||||
Add Rule 9 to `01-bigmind-core.md`:
|
||||
|
||||
> **Rule 9: Detect and Break Loops Before They Start**
|
||||
>
|
||||
> If `memory_start_session()` context shows 2 or more recently closed sessions with:
|
||||
> - Near-identical headlines or topics, AND
|
||||
> - `partial` or `blocked` outcome
|
||||
>
|
||||
> → **Do NOT attempt to resume the repeated task.**
|
||||
> → Instead: acknowledge the loop to the user, summarize what context was accumulated
|
||||
> across the repeated sessions, and ask: "What would you like to do?"
|
||||
>
|
||||
> Never assume the correct action is to retry a failed/partial task silently.
|
||||
|
||||
### Fix 3: memory_announce_focus — Wait for User Input
|
||||
|
||||
Add a constraint to Rule 3 (announce focus):
|
||||
|
||||
> **`memory_announce_focus()` must reflect the CURRENT session's task.**
|
||||
> Call it only AFTER the user has given a clear instruction for this conversation.
|
||||
> Do NOT announce focus derived from prior session outcomes before the user speaks.
|
||||
> During the startup ritual (steps 1-4 of Rule 1), use a placeholder focus if needed:
|
||||
> `memory_announce_focus(session_id, "Awaiting user task assignment")`
|
||||
|
||||
### Fix 4: Mode Interaction Safety Clause
|
||||
|
||||
Add a universal safety rule (applies to all modes):
|
||||
|
||||
> **Session ritual completion ≠ task authorization.**
|
||||
> Completing `memory_start_session()` + `memory_list_hypotheses()` + `memory_announce_focus()`
|
||||
> does NOT authorize beginning any task. Work begins only when the user explicitly assigns it
|
||||
> in the current conversation. Prior session context is reference material, not instruction.
|
||||
|
||||
---
|
||||
|
||||
## Files to Change
|
||||
|
||||
| File | Change |
|
||||
|------|--------|
|
||||
| `.roo/rules/01-bigmind-core.md` | Add Rule 9, add partial=history clause to Rule 1, add focus guard to Rule 3 |
|
||||
| `.roo/rules/00-identity.md` | Add mode-interaction safety clause |
|
||||
|
||||
---
|
||||
|
||||
## Risk Assessment
|
||||
|
||||
| Risk | Likelihood | Mitigation |
|
||||
|------|-----------|------------|
|
||||
| Model ignores new rules in long context | Medium | Rules are loaded via rules files, not context — they apply per-session |
|
||||
| Fix breaks legitimate resumption (e.g., user explicitly asks to continue) | Low | Rules say "task determined by user's first message" — explicit resumption request still works |
|
||||
| New Rule 9 fires falsely on legitimate repeated partial tasks | Low | Trigger requires near-identical headlines AND repeated partial — normal work produces diverse headlines |
|
||||
|
||||
---
|
||||
|
||||
## Success Criteria
|
||||
|
||||
1. Starting a new session after a partial pic-gen session → model waits for user input, no autonomous generation
|
||||
2. Starting a new session after 2+ identical partial sessions → model acknowledges the loop and asks what to do
|
||||
3. User explicitly asking "continue the branding generation" → model correctly resumes (rule only prevents silent resumption)
|
||||
|
||||
---
|
||||
|
||||
## Implementation Order
|
||||
|
||||
1. Patch `.roo/rules/01-bigmind-core.md` — add Rule 9 + partial=history clause + focus guard
|
||||
2. Patch `.roo/rules/00-identity.md` — add mode interaction safety clause
|
||||
3. Test by starting a new session in pic-gen mode with partial history in context
|
||||
4. Push to Gitea
|
||||
|
||||
Reference in New Issue
Block a user