Audit of snapcompact settings against the accuracy question: "Am I using the most accurate snapcompaction settings?"
Moving parts (one config edit each, plus the docs page):
agent/config.yml: compaction.thresholdPercent 95 → 85 — restores the max_tokens gap-math written into agent/models.yml; 95 routes compaction through the overflow-refuse path.
agent/models.yml frank-archon row: imageCountBudget: 16 — custom providers inherit the unknown-provider floor of 5 images/request and silently halve the snapcompact archive.
Docs: omp-docs/core/snapcompact.md (measured recall/billing table from a live probe, the catalog-only frame-pricing law, the 85 % trigger law) + index + log entry.
Verdict on the knobs themselves: snapcompact.shape: auto, systemPrompt: none, toolResults: false, deterministic-first methodOrder — all verified accuracy-optimal by probe (auto shape 8/8 recall at 47 % of raw-text billed tokens; denser grids drop characters).
Audit of snapcompact settings against the accuracy question: "Am I using the most accurate snapcompaction settings?"
Moving parts (one config edit each, plus the docs page):
1. `agent/config.yml`: `compaction.thresholdPercent` 95 → 85 — restores the max_tokens gap-math written into `agent/models.yml`; 95 routes compaction through the overflow-refuse path.
2. `agent/models.yml` frank-archon row: `imageCountBudget: 16` — custom providers inherit the unknown-provider floor of 5 images/request and silently halve the snapcompact archive.
3. Docs: `omp-docs/core/snapcompact.md` (measured recall/billing table from a live probe, the catalog-only frame-pricing law, the 85 % trigger law) + index + log entry.
Verdict on the knobs themselves: `snapcompact.shape: auto`, `systemPrompt: none`, `toolResults: false`, deterministic-first `methodOrder` — all verified accuracy-optimal by probe (auto shape 8/8 recall at 47 % of raw-text billed tokens; denser grids drop characters).
---
cwd: /home/blank/.omp/wt/wt-20261027-181709-7c8a99d · session: 01a18015-1638-7446-b40e-b8d2016680d1
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Audit of snapcompact settings against the accuracy question: "Am I using the most accurate snapcompaction settings?"
Moving parts (one config edit each, plus the docs page):
agent/config.yml:compaction.thresholdPercent95 → 85 — restores the max_tokens gap-math written intoagent/models.yml; 95 routes compaction through the overflow-refuse path.agent/models.ymlfrank-archon row:imageCountBudget: 16— custom providers inherit the unknown-provider floor of 5 images/request and silently halve the snapcompact archive.omp-docs/core/snapcompact.md(measured recall/billing table from a live probe, the catalog-only frame-pricing law, the 85 % trigger law) + index + log entry.Verdict on the knobs themselves:
snapcompact.shape: auto,systemPrompt: none,toolResults: false, deterministic-firstmethodOrder— all verified accuracy-optimal by probe (auto shape 8/8 recall at 47 % of raw-text billed tokens; denser grids drop characters).cwd: /home/blank/.omp/wt/wt-20261027-181709-7c8a99d · session: 01a18015-1638-7446-b40e-b8d2016680d1