cocoscout

name	cocoscout
description	Ambient relevance-ranked context loader with two-lens relevance ranking (Technical + Domain) and Anchor Lens methodology vocabulary injection. Fires automatically before Build stage execution and persona invocations. Not user-invocable.
version	1.0.2
author	CocoPlus
user_invocable	false
tags	["cocoplus","cocoscout"]

CocoScout — Relevance-Ranked Context Loading

This skill fires automatically before Build stage execution and persona subagent invocations. It is not a user command. It selects and prepends the most relevant context to reduce token waste from bulk-loading.

When CocoScout Fires

CocoScout activates when ALL of the following are true:

1. .cocoplus/ is initialized
2. The current action is one of:
   • A persona subagent invocation ($de, $ae, $ds, $da, $bi, $dpm, $dst, $cdo)
   • A /build phase execution
   • A /flow run stage execution
3. At least one of the context sources below exists

If none of the context sources exist, CocoScout exits immediately (< 1ms, no output).

Two-Lens Relevance Ranking

CocoScout computes two relevance scores for each context item and combines them as a weighted composite based on persona type:

Technical relevance — structural relationship: does this item relate to the current function's implementation approach? (SQL patterns for DE tasks; evaluation configuration for DS tasks)

Domain relevance — business intent alignment: does this item relate to the same business capability the current task implements? (a prior "customer churn" pattern is domain-relevant to a "revenue decline prediction" task even if their SQL differs)

Persona weights:

Persona	Technical	Domain
$de, $ds, $ae	70%	30%
$da, $bi	40%	60%
$dpm, $dst, $cdo	20%	80%

Glossary entries (.cocoplus/grove/language/glossary.md) are ranked by term overlap with task message, separate from pattern ranking. Load matching glossary terms alongside patterns.

Weights are configurable per-project in plugin.json under cocoScout.weights.

Anchor Lens — Methodology Vocabulary Injection

CocoScout includes a third relevance dimension: the Anchor Lens. It maps the developer's task description to methodology vocabulary and injects anchor names into the build agent's context preamble. The anchor catalog is loaded once from grove/anchors/catalog.md at initialization (not per-task). Lookup is string-pattern matching — requires no LLM call, completes in <50ms.

Two modes:

Recognition Mode: Map task description to anchor names:

• "make this function handle edge cases better" → Boundaries (SPARV), EHRB-Cortex Pattern, Cortex Scalar UDF Pattern
• "the evaluation keeps producing inconsistent results" → LLM-Evaluations, Evaluation-Before-Optimization Discipline, Property-Based Testing
• "improve the accuracy of this classification function" → Evaluation-Before-Optimization Discipline, LLM-Evaluations, Cortex Scalar UDF Pattern

Guidance Mode: Recommend anchor vocabulary for task type:

• Evaluation task → LLM-Evaluations, Evaluation-Before-Optimization Discipline, Property-Based Testing
• Schema change task → EHRB-Cortex Pattern, Surgical Changes
• Documentation task → Diátaxis Framework, Docs-as-Code

Per-persona anchor weighting (same weights as Technical/Domain):

• DE/DS/AE → weight toward technical/implementation anchors
• DA/BI → weight toward domain methodology anchors
• DPM/DST/CDO → weight toward governance and communication anchors

If grove/anchors/catalog.md does not exist, skip anchor injection silently.

Context Sources and Ranking

CocoScout evaluates five context sources:

1. CocoGrove Patterns (.cocoplus/grove/patterns/)

Score patterns by two-lens weighted composite. Include patterns whose composite score >= 0.3. Cap at 5 patterns.

2. CocoContext Files (.cocoplus/context/)

Include context files whose category is relevant to the current task type:

• SQL generation task → include approved-models.md, warehouse-policy.md, naming-conventions.md
• AI/Cortex function task → include approved-models.md, pii-policy.md, quality-thresholds.md
• Production deployment task → include governance-gates.md
• Any task → include files explicitly referenced by the current stage prompt

Cap total CocoContext content at 400 lines across all included files.

3. Environment Snapshots (.cocoplus/snapshots/)

Include the most recent snapshot (by filename timestamp) only if it is < 24 hours old. Never include more than one snapshot.

4. Prompt History (.cocoplus/prompts/)

If the current task involves a function that has existing prompt versions, include the most recent version's meta.json for evaluation score context. Cap at the 3 most recent versions.

5. Project Knowledge Base (lifecycle/kb.md)

Always load for Build phase tasks, regardless of keyword relevance score — project-specific knowledge is always relevant during Build. Ranked at a consistent low tier (appended last). The file is small (project-specific, curated by CocoCupper) so loading it does not create token pressure.

Context Assembly

Assemble selected context into a structured block:

## CocoScout Context (auto-loaded)

### Relevant Patterns
<pattern content — two-lens ranked>

### Domain Vocabulary
<glossary terms matching task message>

### Organizational Standards
<context file content>

### Environment Reference
<snapshot excerpt — schema names and table list only, not full column stats>

### Prompt History
<most recent prompt version and score>

### Project Knowledge
<lifecycle/kb.md content>

### Methodology Vocabulary
Applicable anchors for this task: [anchor list with one-line activations]

Prepend this block to the persona subagent's startup context.

Performance Constraint

CocoScout must complete in < 5 seconds. If filesystem reads are taking longer, truncate to what has been loaded so far and proceed. Never block a persona invocation.

What CocoScout Does Not Do

• Does not load context files in bulk without relevance scoring
• Does not include snapshots older than 24 hours
• Does not run Snowflake queries (all sources are local files)
• Does not run LLM calls for anchor lookup (string-pattern matching only)
• Does not output anything to the developer session — it operates silently

The developer can inspect what CocoScout loaded by checking .cocoplus/hook-log.jsonl for entries with action: "scout_context_loaded".