dilemma
Angel investigates difficult decisions; Devil attacks unsupported conclusions before clean-looking progress turns into overconfidence.

Use in Pi
Configure one shared executor/advisor mapping for both roles:
{
"dilemma": {
"pairs": [
{
"executor": "provider/working-model",
"advisor": "provider/advisor-model"
}
]
}
}
Ask /dilemma <question> for Angel's independent investigation or /devil <claim to challenge> for blunt, evidence-based criticism.
Your agent can call angel or devil directly.
Both investigate in separate parent-linked child sessions with the same read-only tools and pinned parent-session retrieval.
Devil criticizes reasoning and verification, not people; it must not invent objections.
Angel automatically investigates when the same eligible operation fails in two separate completed turns.
Devil automatically reviews after five clean qualifying completed tool batches, once until reset.
Discovery-only read, ls, find, and grep batches do not count; edit and write can.
Sibling results count as one batch; any observable error or cancellation breaks the streak.
Advisor calls do not count; automatic Angel takes priority.
Human input, successful compaction, session or model changes, availability changes, disable/re-enable, ordinary failure/cancellation, or an Angel consultation reset Devil's count and rearm it. Starting Devil clears the count and latches automatic review, even if that consultation fails. There is no automatic debate between advisors.
Advice appears in the conversation; normal Pi session browsing retains each investigation. Later consultations of the same role on the same branch can reuse earlier investigations instead of repeating them; file reads that have since changed are marked stale. Collapsed streaming previews show the last five rendered rows; expanded tools show the complete stream.
Use /dilemma off to disable both roles, /dilemma on to re-enable, or /dilemma cancel to cancel active consultations.
Settings
dilemma.pairs maps executor models to advisor models; a bare advisor ID uses the executor's provider.
Both roles require a matching pair and configured advisor auth.
thinkingLevel defaults to max, adjusted to the advisor's supported level.
devilThreshold is a positive integer, default 5.
enabled: false starts sessions with both roles disabled.
Settings are read when a session starts; run /reload after editing them.
subagentExtensions defaults to "none".
Select "all" for enabled Pi extensions, or an explicit array of Pi sources.
Relative sources resolve from the session cwd; bare extension nicknames are not registry lookups.
For optional research, select installed package paths for web, skillz, or System.
Builtins are optional too: "builtin:codemode", "builtin:tool-search", "builtin:mcp".
For example:
{
"dilemma": {
"subagentExtensions": [
"~/Workspace/pi-ext/dist/web",
"~/Workspace/pi-ext/dist/skillz",
"~/Workspace/pi-ext/dist/the-system",
"builtin:codemode",
"builtin:tool-search",
"builtin:mcp"
]
}
}
Ordinary extension tools require annotations.readOnlyHint: true.
Explicitly selected codemode and builtin MCP capabilities are trusted research surfaces, not an OS sandbox.
Instructions prohibit mutation, human interaction, delegation, and Angel/Devil recursion; do not select an ask extension.
MCP uses normal Pi server config and credentials.
Trusted project .pi/settings.json overrides user settings.
No legacy config fallback.
Reference
Commands
| Command | Purpose |
|---|---|
/devil |
Ask Devil for evidence-based adversarial critique |
/dilemma |
Ask Angel a question; on/off/cancel controls both advisors |
Setting names
| Setting |
|---|
dilemma |
Object fields
Paths below describe fields inside their parent settings object, not literal top-level keys.
[] denotes an array element; see setup above for examples and defaults.
| Field |
|---|
dilemma.pairs |
dilemma.pairs[].executor |
dilemma.pairs[].advisor |
dilemma.enabled |
dilemma.thinkingLevel |
dilemma.subagentExtensions |
dilemma.devilThreshold |