v5.6.13

Bug Fixes

• Quality score stability: Enforce ResourceHealth monotonicity within a session. Score can only decrease (worsen) unless a compaction resets context. Prevents fill_pct measurement fluctuations from causing 10-20 point score swings in the statusline.

• False tool call warnings suppressed: Tool call threshold scaling changed from linear (5x) to superlinear (x^1.3) for 1M context windows, raising thresholds from warn=125/crit=200 to warn=202/crit=324. Additionally, tool call warnings are now gated on context fill >= 50%, since tool calls at low fill don't degrade instruction adherence.

• Two-row statusline: Status bar split into two rows for readability. Row 1: model, effort, project, context bar, ContextQ. Row 2: Eff, warnings, compactions, session duration, active agents. Session duration is now always visible (previously gated on quality < 75).

• Edge case hardening: Duration display guards against invalid timestamps (epoch 0, negative values, >7 day sessions). Monotonicity clamp handles corrupted cache where compaction count regresses.

v5.6.12

Self-healing for malformed hook commands in settings.json.

Hooks written by Claude during interactive sessions using $(ls -d $HOME/...) subshell patterns or double-$HOME paths now get automatically detected and removed on next session start. Correct hooks are immediately re-provisioned.

• Identity matcher handles subshell-wrapped commands via regex fallback
• Word-boundary matching prevents false positives on similar path names
• macOS-only guard for double-home detection
• Immediate hook re-provisioning for script-install users (no 24h gap)
• Codex runtime unaffected (guarded)

v5.6.11

Context-aware tool call thresholds

The tool-call warning thresholds now auto-scale with your context window size:

Context	Warn	Critical
200K	25	40
500K	62	100
1M	125	200

Previously, the hardcoded 25/40 thresholds (from COLM 2025 research on 200K models) fired false CRITICAL warnings on 1M context sessions that routinely hit 80-120 tool calls while staying coherent at 94+ quality scores.

Env var overrides (TOKEN_OPTIMIZER_TOOL_CALL_WARN / TOKEN_OPTIMIZER_TOOL_CALL_CRITICAL) still take precedence for manual tuning.

Thanks to @CodyLee117 whose PR #33 added the env var configurability that made this scaling possible.

v5.6.10

Codex: Global hooks - codex-install now writes to ~/.codex/hooks.json (global) by default. One install covers all projects.

Codex: Runtime isolation - Surgical guards prevent Claude Code settings writes under Codex while preserving dashboard/daemon/checkpoint operations.

Configurable tool-call thresholds - TOKEN_OPTIMIZER_TOOL_CALL_WARN and TOKEN_OPTIMIZER_TOOL_CALL_CRITICAL env vars. Defaults unchanged (25/40). Thanks to @CodyLee117 (#33).

Safe env-var parsing - All TOKEN_OPTIMIZER_* int/float configs warn on invalid values instead of crashing.

Fixes: Version alignment (was 5.6.7/5.6.8/5.6.9), install.sh sparse checkout, symlink guard on ~/.codex, dry-run no longer creates ~/.codex as a side effect.

v5.6.9

Fix stale model pricing in reference docs. Opus 4.5+ is $5/$25 per MTok, not $15/$75 (that was Opus 4.0/4.1). Haiku 4.5 is $1/$5, not $0.25/$1.25 (that was Haiku 3). Core pricing engine was already correct; only reference docs were stale. Recalculated all worked examples and cost ratios.

v5.6.8 - Security Hardening

Security Fixes

• Add post-clone integrity verification and commit SHA logging (High)
• Remove install/build commands from bash compression whitelist (High)
• Add log rotation (1MB cap) and remove sensitive command text from rewrite log (Medium)

v5.6.7

Quality score architecture split: ResourceHealth + SessionEfficiency

The rising-score bug is dead. The composite quality score has been split into two independent metrics:

ResourceHealth (displayed as ContextQ:) is now architecturally monotonic. It uses only fill percentage, compaction depth, and waste signals that cannot improve without compaction. This is the backward-compatible score value for all existing consumers.

SessionEfficiency (displayed as Eff:) tracks behavioral signals (stale reads, bloated results, decision density, tool call patterns) and can rise/fall freely, which is correct for those metrics.

New real-time signals:

• 40+ tool calls warning (CRITICAL threshold)
• 50% fill regime change detection
• Re-reading loop detection

Rolling window scoring from v5.6.6 is preserved. No changes needed for Codex or OpenClaw.

v5.6.6

Fix: Quality score monotonicity bug

The composite quality score could climb during sessions without compaction. Three ratio-based signals (stale_reads, bloated_results, decision_density = 48% combined weight) were computed over full session history, so denominators grew and scores improved even as context degraded.

Rolling window scoring: Ratio-based signals now use a sliding window of the last 20 operations instead of full session history. Context fill, compaction depth, and duplicates remain cumulative (correct by nature). Configurable via TOKEN_OPTIMIZER_QUALITY_WINDOW env var.

Independent fill warnings: New fill-percentage warnings at 75% (WARNING) and 85% (CRITICAL) fire in the statusline and as system messages regardless of composite score value. These cannot be masked by improving ratio signals.

No changes needed for Codex or OpenClaw -- Codex uses the same compute_quality_score function (fix applies automatically), and OpenClaw scoring uses only point-in-time signals (no denominator expansion possible).

v5.6.5 — read cache escape hatch, edit-blocking fix

Fixes a reported issue where the read cache could block file edits on Linux/Cursor environments.

What changed:

• Escape hatch: After 3 consecutive denials of the same file, the read cache allows a full re-read. Prevents models from getting stuck in deny loops.
• Belt-and-suspenders guard: main() validates tool_name before processing, catching environments where hook matchers are not enforced.
• Diagnostic telemetry: Unexpected tool invocations logged for debugging.
• Clearer denial messages: All deny responses tell the model it can still Edit or Read a specific range.

No configuration changes needed. Existing users on auto-update will get this automatically.

v5.6.4 — loop detection false-positive fix

Fixes

• Tightened live loop detection so short successful tool results no longer trigger Possible retry loop warnings.
• Retry-churn warnings now require concrete failure signals and repeated failures from the same tool family.
• Added conservative guards for benign outputs such as 0 failed and normal search-style exit code 1 results.

Version sync

• Bumped Claude Code plugin metadata from 5.6.3 to 5.6.4.
• Confirmed no id-manifest.json is present in this marketplace tree.

OpenClaw check

• Checked the OpenClaw implementation: this exact live short-result detector is deferred/not present there, and the existing retry-churn detector is session-level and failure-outcome based.
• Synced OpenClaw patch metadata to 2.4.1 and fixed stale README/CLI version strings.

Verification

• python3 -m py_compile skills/token-optimizer/scripts/measure.py
• Focused loop-detection regression checks
• npm run build in openclaw/
• python3 -m pytest tests (141 passed)