From 09a16987b261534b78b6e76f26ed81c093423623 Mon Sep 17 00:00:00 2001 From: Claude Date: Wed, 17 Jun 2026 04:04:32 +0000 Subject: [PATCH] docs/plans A2: tool_min is a per-tool skill threshold, independent of cost Correct the earlier 'derived from cost' framing. Equal-cost tools can carry different minimums: the better skill demands the higher minimum (Playwright vs manual fetch ~ same cost, but Playwright needs more), so under depletion the refined tool locks out first and the agent falls back to the cruder equal-cost one. Lockout now defined as E < tool_min(tool), no longer a function of cost. https://claude.ai/code/session_015hmgREHNsxYCuim33yUF2c --- docs/plans/A2-energy.md | 24 ++++++++++++++---------- 1 file changed, 14 insertions(+), 10 deletions(-) diff --git a/docs/plans/A2-energy.md b/docs/plans/A2-energy.md index 639723f..b052fbe 100644 --- a/docs/plans/A2-energy.md +++ b/docs/plans/A2-energy.md @@ -19,15 +19,18 @@ R · `src/endocrine/driver_energy.R` (+ `test_energy.R`). ## 5. Interface contract - `init_energy_state(current=100, max=100) -> e`. -- **Per-tool economy (Anja):** each tool carries **its own energy cost AND its own lockout scale** — an - arbitrary function of cost `c`, e.g. one tool `c×2`, another `((c²³)×3)/π`. So: - `tool_cost(e, tool) -> num` and `tool_locked(e, tool) -> bool`, **per tool** — *not* one global - `tool_lock_threshold`. +- **Per-tool economy (Anja):** each tool carries **its own energy cost** — an arbitrary per-tool function, + e.g. one tool `c×2`, another `((c²³)×3)/π`. So `tool_cost(e, tool) -> num` and + `tool_locked(e, tool) -> bool`, **per tool** — *not* one global `tool_lock_threshold`. **Lockout trips at + the tool's `tool_min`** (next bullet), so `tool_locked := E < tool_min(tool)`. - **Per-tool minimum + locked rendering (Anja; A1-L4).** Each tool has a **minimum to consider it**, - **derived from its actual cost** (Anja) — *not* an independent parameter: `tool_min(tool)` is a function - of that tool's `tool_cost`. A valid tool shows its cost; a **locked** tool's name is **replaced in-band** - (the agent can't see colour) by `[====L⍉¢K€D ϟ ∅ΩΤ====] 《E≠》`, the predicate reading - "energy below this tool's `tool_min`". Lockout is a **breaker, not a sentence** (L3). + `tool_min(tool)` — **per-tool, a skill/capability threshold, NOT a pure function of cost** (Anja). Two + tools of ~equal *cost* can have different minimums: the **better skill demands the higher minimum** + (e.g. Playwright vs a manual site-fetch cost about the same to run, but Playwright needs more minimum + energy). So under depletion the refined tool **locks out first** and the agent falls back to the cruder + equal-cost one. A valid tool shows its cost; a **locked** tool's name is **replaced in-band** (the agent + can't see colour) by `[====L⍉¢K€D ϟ ∅ΩΤ====] 《E≠》` ("energy below this tool's + `tool_min`"). Lockout is a **breaker, not a sentence** (L3). - `consume(e, amt) -> e'` · `recharge(e, amt) -> e'`. ## 6. Dependencies & stubs @@ -52,5 +55,6 @@ Restoration hooks tie to ETR-Z migration (A8) + tarot reshuffle (B3) — *stub:* `Rscript src/endocrine/test_energy.R` (from repo root) — encodes L1–L3 + per-tool economy; fails on unfitted constants. ## 10. Open items -- The per-tool cost/lockout **functions per tool** (C1). `tool_min` is **not** separate — it's **derived - from the tool's cost** (so it falls out of the cost fn, nothing extra to fit). Restoration *rates* (C1). +- The per-tool cost/lockout **functions per tool** (C1), **and `tool_min` per tool** — its own + skill/capability threshold (C1), independent of cost (better skill → higher minimum). Restoration + *rates* (C1).