Problem
Every agent reads the ENTIRE decisions.md at spawn time (~86K tokens for this repo at 345KB). The lifecycle spawn API already accepts decisions?: string but nobody populates it. This is the single highest-ROI fix for the decisions bloat problem.
From
Flight, EECOM, Procedures reviews of #20. Identified as highest-value architectural change.
Proposed
- Add
filterDecisionsForAgent(decisions, agentDomain) function in SDK
- Coordinator reads decisions.md once at session start
- Before each spawn, coordinator filters to only relevant decisions (by domain tag or recency)
- Passes filtered string via existing
decisions param on lifecycle API
- Token consumption drops from ~86K to ~2K per spawn
Dependencies
Impact
Reduces per-agent spawn context by ~95% without deleting any data.
Owner
EECOM (Core Dev) + Procedures (Prompt Engineer)
Problem
Every agent reads the ENTIRE decisions.md at spawn time (~86K tokens for this repo at 345KB). The lifecycle spawn API already accepts
decisions?: stringbut nobody populates it. This is the single highest-ROI fix for the decisions bloat problem.From
Flight, EECOM, Procedures reviews of #20. Identified as highest-value architectural change.
Proposed
filterDecisionsForAgent(decisions, agentDomain)function in SDKdecisionsparam on lifecycle APIDependencies
[runtime],[testing], etc.) — see chore(scribe): Mandate ### YYYY-MM-DD: format in Scribe charter #24 (Scribe format)Impact
Reduces per-agent spawn context by ~95% without deleting any data.
Owner
EECOM (Core Dev) + Procedures (Prompt Engineer)