feat(measure): make tool-call warning thresholds env-var configurable

[CodyLee117]

Summary

Adds TOKEN_OPTIMIZER_TOOL_CALL_WARN / TOKEN_OPTIMIZER_TOOL_CALL_CRITICAL env-var overrides for the cumulative-tool-call warning thresholds at skills/token-optimizer/scripts/measure.py (_TOOL_CALL_WARN_THRESHOLDS, around line 11419).

Defaults are unchanged from upstream (25 WARNING / 40 CRITICAL). No behavioral change for any user who doesn't set the env vars — this is opt-in only.

Rationale

The 25 / 40 thresholds (introduced in v5.6.7, commit 1185b3c6) derive from the COLM 2025 / codeongrass.com practitioner analysis on instruction-adherence degradation. That research predates Claude Opus 4.7's 1M-context tier — on those longer-context runs, sessions routinely cross 40 cumulative tool calls long before adherence actually starts degrading, so the CRITICAL warning fires too aggressively for that workload.

Concretely: a single Opus 4.7 1M session implementing a multi-slice feature (schema records → state integration → action handlers → tests → docs → commit + push) can land 80-120 tool calls while staying highly coherent. The current code wedges those sessions into a persistent CRITICAL state from call ~40 onward, with no in-product way to dismiss or tune.

Making the constants configurable lets:

• Long-context users tune higher (50/80, or 100/200 for deep research) without source edits or local patches.
• Shorter-context / non-Anthropic users keep the existing defaults, which still match the practitioner research for those workloads.
• Future research updates can ship as new default values without forcing every long-context user to re-patch.

Internal consistency

The pattern mirrors how the same file already exposes other knobs as env vars at lines 11425-11437:

```python
_CHECKPOINT_MAX_FILES = int(os.environ.get("TOKEN_OPTIMIZER_CHECKPOINT_FILES", "10"))
_CHECKPOINT_TTL_SECONDS = int(os.environ.get("TOKEN_OPTIMIZER_CHECKPOINT_TTL", "300"))
_CHECKPOINT_RETENTION_DAYS = int(os.environ.get("TOKEN_OPTIMIZER_CHECKPOINT_RETENTION_DAYS", "7"))
_RELEVANCE_THRESHOLD = float(os.environ.get("TOKEN_OPTIMIZER_RELEVANCE_THRESHOLD", "0.3"))
_EDIT_BATCH_WRITE_THRESHOLD = int(os.environ.get("TOKEN_OPTIMIZER_EDIT_BATCH_WRITE_THRESHOLD", "4"))
```

The override naming (TOKEN_OPTIMIZER_TOOL_CALL_WARN / _CRITICAL) matches the existing TOKEN_OPTIMIZER_* prefix convention.

Change

 # Tool call thresholds: instruction adherence degrades after ~15 tool calls
 # (COLM 2025, codeongrass.com practitioner analysis). Cumulative, not reset on compact.
+# Configurable via TOKEN_OPTIMIZER_TOOL_CALL_WARN / _CRITICAL env vars to suit
+# longer-context models (e.g. Claude Opus 4.7 1M). Defaults are unchanged from the
+# original literal — opt-in override only.
+_TOOL_CALL_WARN     = int(os.environ.get(\"TOKEN_OPTIMIZER_TOOL_CALL_WARN\", \"25\"))
+_TOOL_CALL_CRITICAL = int(os.environ.get(\"TOKEN_OPTIMIZER_TOOL_CALL_CRITICAL\", \"40\"))
 _TOOL_CALL_WARN_THRESHOLDS = [
-    (40, \"CRITICAL\", \"40+ tool calls, instruction adherence severely degraded\"),
-    (25, \"WARNING\", \"25+ tool calls, consider a fresh session\"),
+    (_TOOL_CALL_CRITICAL, \"CRITICAL\", f\"{_TOOL_CALL_CRITICAL}+ tool calls, instruction adherence severely degraded\"),
+    (_TOOL_CALL_WARN,     \"WARNING\",  f\"{_TOOL_CALL_WARN}+ tool calls, consider a fresh session\"),
 ]

The displayed count text is f-string-derived so it stays in sync with whatever threshold the user configures (e.g. "150+ tool calls" rather than the prior hardcoded "40+" when a user has set TOKEN_OPTIMIZER_TOOL_CALL_CRITICAL=150).

Test plan

• AST parse-check on patched measure.py (python3 -c \"import ast; ast.parse(...)\") — passes.
• Defaults preserved: with neither env var set, _TOOL_CALL_WARN_THRESHOLDS evaluates to [(40, \"CRITICAL\", \"40+ tool calls, …\"), (25, \"WARNING\", \"25+ tool calls, …\")] — identical to upstream.
• Overrides honored: with TOKEN_OPTIMIZER_TOOL_CALL_WARN=50 / TOKEN_OPTIMIZER_TOOL_CALL_CRITICAL=80 in the session env, the warning text correctly reports "50+" / "80+" and triggers at those counts.

Diff is +7/-2 lines, single file. Happy to add a README entry under a centralized "Environment variables" section if you'd like — let me know your preference and I'll fold it in.

🤖 Generated with Claude Code

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[CodyLee117]

I have read the CLA Document and I hereby sign the CLA

[alexgreensh]

Applied in v5.6.10 (commit db64beb) with credit in the release notes. Clean change, followed existing patterns perfectly. Thanks @CodyLee117 for the contribution! The safe env-var parsing helpers we added alongside your change also protect the existing TOKEN_OPTIMIZER_* configs from ValueError crashes on malformed input.