peteromallet Authority status (M11): Zero-authority history. All repair, audit, and deployment authority has been migrated to canonical delegation. This document is retained for reference only — it must not be used to materialize commands, grant authority, or drive automated actions.
name: planning description: "Planning Pipeline — Skill Reference"
Planning Pipeline — Skill Reference
Overview
The megaplan pipeline (packaged at arnold_pipelines/megaplan/) is the built-in megaplan plan-production substrate.
Its canonical registry identity is megaplan with the legacy alias planning → megaplan.
It orchestrates the full prepare → plan → critique/gate/revise loop →
finalize → execute → review lifecycle.
Gate verdict vocabulary
The gate stage emits a PipelineVerdict whose recommendation field is one
of the following GateRecommendation literals:
| Verdict | Meaning |
|---|---|
proceed |
Gate approved the plan; advance to finalize (and optionally execute). |
iterate |
Gate rejected the plan; re-enter the critique → revise loop. |
tiebreaker |
Evaluators are split; hand off to the tiebreaker stage for adjudication. |
escalate |
Quality ceiling reached at current tier; escalate to a higher-complexity model. |
Robustness levels
Robustness controls the depth of the critique/gate loop and the number of
evaluators engaged. Canonical names (accepted by --robustness / config):
| Level | Alias(es) | Behaviour |
|---|---|---|
bare |
tiny |
Single-pass, no gate loop. Fastest; for quick drafts. |
light |
— | One critique + revise round, minimal gate criteria. |
full |
standard |
Standard gate loop (default). Balanced quality/cost. |
thorough |
robust |
Extended gate loop, stricter criteria, more evaluators. |
extreme |
superrobust |
Maximum depth, all evaluators enabled. |
The default robustness when no --robustness flag is supplied is full.
Stage topology
prep → plan → critique ──→ gate ──proceed──→ finalize [→ execute → review]
↑ │
└──iterate────┘
│
└──tiebreaker──→ [adjudication] → finalize
Driver substrate: subprocess_isolated (execute/review) + graph+loop-node
(critique→gate→revise subloop).
arnold_api_version: 1.0
capabilities: plan, execute, review