obra/superpowers

In scripts/run-pipeline-planner.ps1, inside the '## Step 2 -- Get item details' section, after the bullet '- comments -- array of objects; find the one with stage="assess" and decision="approved" to get the operator's L2 steering comment', insert a new bullet on its own line: '- item.enrichment_snapshot -- present in the response but DO NOT use it. The assess artifact is the complete authoritative context for this item.' This makes the stateless input contract explicit at the point where the model reads the API response.

In scripts/run-pipeline-planner.ps1, replace the existing 'The assessment artifact (artifacts[stage="assess"].artifact) contains:' bullet list (currently 4 bullets) with an expanded version listing all 9 actual artifact fields: '- value: one sentence on what the repo brings to the stack'; '- confidence: low | medium | high'; '- fit_to_stack: {score: string, reasoning: string} -- how well it fits the current tech stack'; '- fit_to_vision: {score: string, reasoning: string} -- how well it fits the product vision'; '- options: array of {name, description, risk, effort, tradeoffs: {pros: string[], cons: string[]}, target_project} -- each implementation approach'; '- recommendation: proceed | defer | skip'; '- recommendation_rationale: one sentence justifying the recommendation'; '- why_it_matters: prose context on the item significance'; '- what_implementation_looks_like: prose description of the minimal implementation path'. Documenting every field prevents the model from going to enrichment_snapshot when a field is not listed.

In scripts/run-pipeline-planner.ps1, add a new '## Stage contract' section between the end of the '## What the pipeline is' block and the start of the '## Pipeline API' block. Content (verbatim): 'The planner has exactly two inputs: (1) the L2 assess artifact - the structured JSON artifact written by the assessor, and (2) the operator approval comment (may be null). It must not read or use item.enrichment_snapshot. Output is the L3 plan artifact whose JSON schema is defined in Step 3 below.'

Create a new file context/pipeline-stage-contracts.md. Start with a one-line intro: 'Each pipeline stage reads ONLY the prior stage artifact -- not raw enrichment data or item.enrichment_snapshot.' Then add two sections. '## L2 Assess artifact (input to L3 Plan)' listing all assess fields with types: value (string), confidence (string: low|medium|high), fit_to_stack (object: {score: string, reasoning: string}), fit_to_vision (object: {score: string, reasoning: string}), options (array of {name: string, description: string, risk: string, effort: string, tradeoffs: {pros: string[], cons: string[]}, target_project: string}), recommendation (string: proceed|defer|skip), recommendation_rationale (string), why_it_matters (string), what_implementation_looks_like (string). Then '## L3 Plan artifact (input to L4 Execute)' listing all plan fields with types: objective (string), target_project (string), target_files (string[]), steps (array of {order: number, action: string, detail: string, file: string}), acceptance_criteria (string[]), estimated_effort (string), risks (string[]), operator_note_addressed (string). Keep it under 60 lines -- schema only, no prose beyond the intro.