v2.2.0: PRD Diet + Per-Harness Tuning

ATHENA v2.2.0 — PRD Diet

Release date: 2026-07-06
Repo: taylorparsons/athena-skill
Full change list: CHANGELOG.md (v2.2.0)

Follow-up to the v2.1.0 token diet. The cross-harness A/B experiments showed two remaining costs: full PRD reads (both Claude and Codex agents read all ~6.4K tokens of a mature PRD) and, on Codex, reasoning-effort spend that dwarfs context savings. v2.2.0 removes the first inside the framework and documents the fix for the second.

What changed

• New owl prune-prd: archives the PRD Metadata Inputs:/Decisions: link lists and Shipped/Implemented/Planned backlog lines (except those citing active features) to docs/audit/prd-archive.md, leaving pointer lines. Idempotent; Goals and FRs are never touched. Applied to this repository: PRD 26.9KB → 20.9KB.
• SKILL.md hardened: the PRD is never read in full — grep for the section being changed, read only that range; single requirements via owl show <FR-ID>.
• 3 new unit tests (23 total).

Value per harness

• Claude Code: removes the last recurring full-file read from the loop (~1.5–1.6K tokens per session on this repo, compounding across turns and growing with PRD age).

• Codex: the same read savings, plus the recommended pairing measured to matter more — a lower reasoning tier for routine loop work. Add to ~/.codex/config.toml:

  [profiles.athena]
  model_reasoning_effort = "medium"
  

  and run loop sessions with codex --profile athena. Keep the default high tier for design work. Also keep ~/.codex/skills free of non-skill files — Codex truncates skill descriptions when the listing exceeds its context budget.

Verification

• python3 -m unittest discover -s scripts/tests — 23/23 pass
• owl prune-prd run twice on this repository (second run: no-op); traceability lint unchanged