pull

$pull

Distill a file or directory of files into a pull artifact (.pull.md) that preserves all decision-relevant information while reducing token consumption. CocoHarvest uses pull files automatically when stage inputs exceed the configured token threshold.

Preconditions

• .cocoplus/ must be initialized (run $pod init first)
• <target> file or directory must exist

Arguments

Argument	Description
<target> (required)	File path or directory to distill. Eligible: .md, .json, .sql, .txt
--validate (optional)	Run round-trip reconstruction test after distillation
--human (optional)	Produce human-readable narrative summary instead of LLM-dense pull file

Two output modes:

• Default (no flag): Produce <target>.pull.md — machine-dense, LLM-optimized, verbatim decision values. Used automatically by CocoHarvest.
• --human: Produce <target>.pull-human.md — prose narrative for stakeholder consumption. Not used by CocoHarvest. Never overwrites the .pull.md machine output.

Step-by-Step Behavior

$pull <target> --human

1. Verify initialization and target (same as default mode steps 1–2).

2. Spawn Sonnet narrative subagent: Reads <target> and produces a prose summary optimized for a non-technical stakeholder audience. Rules:

   • Lead with the most important conclusion or finding (inverted pyramid)
   • Use plain language — no SQL, no JSON field names unless unavoidable
   • Express numeric values in context: not "0.88" but "88% accuracy — above the 85% threshold set in discuss.md"
   • Maximum 500 words unless source content genuinely requires more
   • Do not include raw tables — translate to sentences or brief bullets

3. Write output to <target>.pull-human.md. Include header:

   ---
   pull_source: [source file path]
   pull_timestamp: [ISO8601]
   pull_mode: human
   audience: stakeholder
   ---
   

4. Do NOT update .cocoplus/pull/manifest.json for human-mode output — human pull files are not CocoHarvest artifacts.