token-optimizer-opencode 1.0.8 → 1.0.14

package/dist/nudges/fresh-session-nudge.d.ts

@@ -0,0 +1,72 @@
+/**
+ * Fresh-session nudge: fires once per session when context is BOTH long
+ * (fill >= FRESH_NUDGE_MIN_FILL_PCT) AND degraded (quality < FRESH_NUDGE_QUALITY_THRESHOLD).
+ *
+ * Confidently reassures the user that Token Optimizer has checkpointed their
+ * active task so a fresh session resumes exactly where they stopped, and shows
+ * the concrete tokens they would reclaim by starting fresh now.
+ *
+ * Takes PRECEDENCE over the ordinary quality/compact nudge (the caller skips
+ * that when this fires — both messages would be noise).
+ *
+ * Ported from Python _maybe_fresh_session_nudge / _fresh_session_savings_estimate
+ * in skills/token-optimizer/scripts/measure.py.
+ */
+/**
+ * Look up the API input rate ($/M tokens) for the given model.
+ * Substring-matches the lowercase model id against the table — same
+ * strategy as contextWindowForModel. Falls back to Sonnet ($3.00/M).
+ */
+export declare function modelInputRatePer1M(model?: string): number;
+/**
+ * API-equivalent dollar value of the reclaimed tokens, priced at the session's
+ * own model input rate. Returns 0 on any error (best-effort).
+ * Mirrors Python _fresh_session_savings_usd.
+ */
+export declare function freshSessionSavingsUsd(savedTokens: number, model?: string): number;
+export declare const FRESH_NUDGE_QUALITY_THRESHOLD: number;
+export declare const FRESH_NUDGE_MIN_FILL_PCT: number;
+export interface FreshNudgeResult {
+    shouldNudge: boolean;
+    message: string | null;
+}
+/**
+ * Estimate tokens reclaimed by starting a fresh session now.
+ * current context size = (fillPct / 100) * contextWindow
+ * savings = current context - lean block re-injection overhead
+ *
+ * @param fillPct        0-100 (percentage, not fraction)
+ * @param model          optional model id — used only as a last-resort fallback
+ *                       when sessionWindow is unavailable
+ * @param sessionWindow  the EXACT context-window value the fill% was measured
+ *                       against (pass the same value used in computeQualityScore).
+ *                       When provided this takes priority over re-deriving from the
+ *                       model, which guarantees token count == fill% of that window
+ *                       (e.g. 54% of 1_000_000 ≈ 540K, never ~107K on a 200K fallback).
+ * @returns [savedTokens, contextWindow]
+ */
+export declare function freshSessionSavingsEstimate(fillPct: number, model?: string, sessionWindow?: number): [number, number];
+/**
+ * Check whether the fresh-session nudge should fire for this turn.
+ *
+ * @param currentScore        current quality/resource-health score (0-100)
+ * @param fillPct             current context fill as 0-100 (percentage, not fraction)
+ * @param previousScore       score from the previous turn (null = no prior score yet)
+ * @param freshNudgeFired     whether the nudge already fired this session
+ * @param nudgesEnabled       whether quality nudges are enabled in config
+ * @param continuityEnabled   whether checkpoint continuity is enabled. The nudge's
+ *                            whole pitch ("start fresh, your place is saved") only
+ *                            holds when continuity actually restores the checkpoint
+ *                            in the new session. With continuity off, suppress the
+ *                            nudge so the ordinary quality nudge (/compact) takes
+ *                            over instead of promising a restore that never happens.
+ * @param model               optional model id — fallback for context-window lookup
+ * @param sessionWindow       the EXACT context-window value the fill% was measured
+ *                            against; threads through to freshSessionSavingsEstimate
+ *                            so the token count is consistent with the fill% display
+ * @param qualityThreshold    score below which (with fill) the nudge may fire; defaults
+ *                            to the env-tunable module constant, overridable via config
+ * @param minFillPct          fill% at/above which the nudge may fire; same default rule
+ */
+export declare function checkFreshSessionNudge(currentScore: number, fillPct: number, previousScore: number | null, freshNudgeFired: boolean, nudgesEnabled: boolean, continuityEnabled: boolean, model?: string, sessionWindow?: number, qualityThreshold?: number, minFillPct?: number): FreshNudgeResult;
+//# sourceMappingURL=fresh-session-nudge.d.ts.map

package/dist/nudges/fresh-session-nudge.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"fresh-session-nudge.d.ts","sourceRoot":"","sources":["../../src/nudges/fresh-session-nudge.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAoDH;;;;GAIG;AACH,wBAAgB,mBAAmB,CAAC,KAAK,CAAC,EAAE,MAAM,GAAG,MAAM,CAc1D;AAED;;;;GAIG;AACH,wBAAgB,sBAAsB,CAAC,WAAW,EAAE,MAAM,EAAE,KAAK,CAAC,EAAE,MAAM,GAAG,MAAM,CAOlF;AAUD,eAAO,MAAM,6BAA6B,QAAoD,CAAC;AAC/F,eAAO,MAAM,wBAAwB,QAAqD,CAAC;AAK3F,MAAM,WAAW,gBAAgB;IAC/B,WAAW,EAAE,OAAO,CAAC;IACrB,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;CACxB;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,2BAA2B,CAAC,OAAO,EAAE,MAAM,EAAE,KAAK,CAAC,EAAE,MAAM,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAYrH;AAED;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,wBAAgB,sBAAsB,CACpC,YAAY,EAAE,MAAM,EACpB,OAAO,EAAE,MAAM,EACf,aAAa,EAAE,MAAM,GAAG,IAAI,EAC5B,eAAe,EAAE,OAAO,EACxB,aAAa,EAAE,OAAO,EACtB,iBAAiB,EAAE,OAAO,EAC1B,KAAK,CAAC,EAAE,MAAM,EACd,aAAa,CAAC,EAAE,MAAM,EACtB,gBAAgB,GAAE,MAAsC,EACxD,UAAU,GAAE,MAAiC,GAC5C,gBAAgB,CAoClB"}

package/dist/nudges/fresh-session-nudge.js

@@ -0,0 +1,190 @@
+/**
+ * Fresh-session nudge: fires once per session when context is BOTH long
+ * (fill >= FRESH_NUDGE_MIN_FILL_PCT) AND degraded (quality < FRESH_NUDGE_QUALITY_THRESHOLD).
+ *
+ * Confidently reassures the user that Token Optimizer has checkpointed their
+ * active task so a fresh session resumes exactly where they stopped, and shows
+ * the concrete tokens they would reclaim by starting fresh now.
+ *
+ * Takes PRECEDENCE over the ordinary quality/compact nudge (the caller skips
+ * that when this fires — both messages would be noise).
+ *
+ * Ported from Python _maybe_fresh_session_nudge / _fresh_session_savings_estimate
+ * in skills/token-optimizer/scripts/measure.py.
+ */
+import { contextWindowForModel } from "../util/context-window.js";
+// ---------------------------------------------------------------------------
+// Per-model input rates ($/M tokens). Mirrors Python PRICING_TIERS["anthropic"]
+// and the non-Claude model table. Fallback is Sonnet at $3.00/M.
+// ---------------------------------------------------------------------------
+const MODEL_INPUT_RATES = {
+    // Anthropic Claude
+    fable: 10.0,
+    opus: 5.0,
+    sonnet: 3.0,
+    haiku: 1.0,
+    // GPT-5 family
+    "gpt-5.5-pro": 30.0,
+    "gpt-5.5": 5.0,
+    "gpt-5.4": 2.5,
+    "gpt-5.4-mini": 0.75,
+    "gpt-5.4-nano": 0.2,
+    "gpt-5.3-codex": 1.75,
+    "gpt-5.2-codex": 1.75,
+    "gpt-5.2": 1.75,
+    "gpt-5.1-codex-mini": 0.25,
+    "gpt-5.1-codex": 1.25,
+    "gpt-5.1": 1.25,
+    "gpt-5-codex": 1.25,
+    "gpt-5": 1.25,
+    "gpt-5-mini": 0.25,
+    "gpt-5-nano": 0.05,
+    // GPT-4 family
+    "gpt-4.1": 2.0,
+    "gpt-4.1-mini": 0.4,
+    "gpt-4.1-nano": 0.1,
+    "gpt-4o": 2.5,
+    "gpt-4o-mini": 0.15,
+    // OpenAI reasoning
+    "o3-pro": 20.0,
+    o3: 2.0,
+    "o3-mini": 1.1,
+    "o4-mini": 1.1,
+    // Google Gemini
+    "gemini-2.5-pro": 1.25,
+    "gemini-2.5-flash": 0.3,
+    "gemini-2.5-flash-lite": 0.1,
+    "gemini-2.0-flash": 0.1,
+    "gemini-2.0-flash-lite": 0.075,
+};
+/** Sonnet input rate used as the default fallback ($/M tokens). */
+const FALLBACK_INPUT_RATE_PER_MTOK = 3.0;
+/**
+ * Look up the API input rate ($/M tokens) for the given model.
+ * Substring-matches the lowercase model id against the table — same
+ * strategy as contextWindowForModel. Falls back to Sonnet ($3.00/M).
+ */
+export function modelInputRatePer1M(model) {
+    if (!model)
+        return FALLBACK_INPUT_RATE_PER_MTOK;
+    const lower = model.toLowerCase();
+    const direct = MODEL_INPUT_RATES[lower];
+    if (direct !== undefined)
+        return direct;
+    // Substring scan (longest key wins via insertion order — most specific first
+    // in the table above). Mirrors Python's model-tier resolution.
+    for (const [key, rate] of Object.entries(MODEL_INPUT_RATES)) {
+        if (lower.includes(key))
+            return rate;
+    }
+    return FALLBACK_INPUT_RATE_PER_MTOK;
+}
+/**
+ * API-equivalent dollar value of the reclaimed tokens, priced at the session's
+ * own model input rate. Returns 0 on any error (best-effort).
+ * Mirrors Python _fresh_session_savings_usd.
+ */
+export function freshSessionSavingsUsd(savedTokens, model) {
+    try {
+        const rate = modelInputRatePer1M(model);
+        return Math.max(0, savedTokens * rate / 1_000_000);
+    }
+    catch {
+        return 0;
+    }
+}
+function intEnv(key, fallback) {
+    const raw = process.env[key]?.trim();
+    if (!raw)
+        return fallback;
+    const parsed = parseInt(raw, 10);
+    return isNaN(parsed) ? fallback : parsed;
+}
+// Env-tunable thresholds, matching Python constants.
+export const FRESH_NUDGE_QUALITY_THRESHOLD = intEnv("TOKEN_OPTIMIZER_FRESH_NUDGE_QUALITY", 70);
+export const FRESH_NUDGE_MIN_FILL_PCT = intEnv("TOKEN_OPTIMIZER_FRESH_NUDGE_MIN_FILL", 50);
+/** Tokens re-injected by a fresh lean-resume (the small overhead the new session pays). */
+const FRESH_NUDGE_LEAN_BLOCK_TOKENS = 1000;
+/**
+ * Estimate tokens reclaimed by starting a fresh session now.
+ * current context size = (fillPct / 100) * contextWindow
+ * savings = current context - lean block re-injection overhead
+ *
+ * @param fillPct        0-100 (percentage, not fraction)
+ * @param model          optional model id — used only as a last-resort fallback
+ *                       when sessionWindow is unavailable
+ * @param sessionWindow  the EXACT context-window value the fill% was measured
+ *                       against (pass the same value used in computeQualityScore).
+ *                       When provided this takes priority over re-deriving from the
+ *                       model, which guarantees token count == fill% of that window
+ *                       (e.g. 54% of 1_000_000 ≈ 540K, never ~107K on a 200K fallback).
+ * @returns [savedTokens, contextWindow]
+ */
+export function freshSessionSavingsEstimate(fillPct, model, sessionWindow) {
+    // Priority: explicit session window > model-based lookup > default fallback.
+    // Re-deriving from the model is a last resort: it can silently disagree with
+    // the window the fill % was actually computed against (e.g. 200K fallback on a
+    // 1M session makes the token estimate 5x too low).
+    const contextWindow = (sessionWindow && sessionWindow > 0)
+        ? sessionWindow
+        : contextWindowForModel(model ?? "");
+    const clampedFill = Math.max(0, Math.min(100, fillPct));
+    const currentCtx = Math.round((clampedFill / 100) * contextWindow);
+    const saved = Math.max(0, currentCtx - FRESH_NUDGE_LEAN_BLOCK_TOKENS);
+    return [saved, contextWindow];
+}
+/**
+ * Check whether the fresh-session nudge should fire for this turn.
+ *
+ * @param currentScore        current quality/resource-health score (0-100)
+ * @param fillPct             current context fill as 0-100 (percentage, not fraction)
+ * @param previousScore       score from the previous turn (null = no prior score yet)
+ * @param freshNudgeFired     whether the nudge already fired this session
+ * @param nudgesEnabled       whether quality nudges are enabled in config
+ * @param continuityEnabled   whether checkpoint continuity is enabled. The nudge's
+ *                            whole pitch ("start fresh, your place is saved") only
+ *                            holds when continuity actually restores the checkpoint
+ *                            in the new session. With continuity off, suppress the
+ *                            nudge so the ordinary quality nudge (/compact) takes
+ *                            over instead of promising a restore that never happens.
+ * @param model               optional model id — fallback for context-window lookup
+ * @param sessionWindow       the EXACT context-window value the fill% was measured
+ *                            against; threads through to freshSessionSavingsEstimate
+ *                            so the token count is consistent with the fill% display
+ * @param qualityThreshold    score below which (with fill) the nudge may fire; defaults
+ *                            to the env-tunable module constant, overridable via config
+ * @param minFillPct          fill% at/above which the nudge may fire; same default rule
+ */
+export function checkFreshSessionNudge(currentScore, fillPct, previousScore, freshNudgeFired, nudgesEnabled, continuityEnabled, model, sessionWindow, qualityThreshold = FRESH_NUDGE_QUALITY_THRESHOLD, minFillPct = FRESH_NUDGE_MIN_FILL_PCT) {
+    if (!nudgesEnabled)
+        return { shouldNudge: false, message: null };
+    // The nudge promises "Token Optimizer has checkpointed your task, so a new
+    // session picks up where you stopped." That is only true when continuity is on.
+    // If the user disabled it, do not inject that promise into the system prompt --
+    // bail so the ordinary quality nudge handles the long+degraded session instead.
+    if (!continuityEnabled)
+        return { shouldNudge: false, message: null };
+    // Post-compaction suppression: no prior score means this is a fresh/just-compacted
+    // session. Let the ordinary nudge seed the baseline first.
+    if (previousScore === null)
+        return { shouldNudge: false, message: null };
+    // Once per session.
+    if (freshNudgeFired)
+        return { shouldNudge: false, message: null };
+    // Both conditions must hold: long session AND degraded quality.
+    if (!(currentScore < qualityThreshold && fillPct >= minFillPct)) {
+        return { shouldNudge: false, message: null };
+    }
+    const [saved] = freshSessionSavingsEstimate(fillPct, model, sessionWindow);
+    const savedStr = saved >= 1000 ? `~${Math.floor(saved / 1000)}K` : `~${saved}`;
+    const fillRounded = Math.round(fillPct);
+    const scoreRounded = Math.round(currentScore);
+    const usd = freshSessionSavingsUsd(saved, model);
+    const costStr = usd >= 0.01 ? `, about $${usd.toFixed(2)} in API-equivalent cost` : "";
+    const message = `[Token Optimizer] This session is long (${fillRounded}% full) and context quality has fallen to ${scoreRounded}. ` +
+        `Starting a fresh session now would reclaim ${savedStr} tokens (~${fillRounded}% of your window)${costStr}. ` +
+        `You won't lose your place: Token Optimizer has checkpointed your active task, key decisions, files, and tool results, ` +
+        `so a new session picks up exactly where you stopped. Just open one and say "continue this" — the context is rebuilt for free.`;
+    return { shouldNudge: true, message };
+}
+//# sourceMappingURL=fresh-session-nudge.js.map

package/dist/nudges/fresh-session-nudge.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"fresh-session-nudge.js","sourceRoot":"","sources":["../../src/nudges/fresh-session-nudge.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,EAAE,qBAAqB,EAAE,MAAM,2BAA2B,CAAC;AAElE,8EAA8E;AAC9E,gFAAgF;AAChF,iEAAiE;AACjE,8EAA8E;AAC9E,MAAM,iBAAiB,GAA2B;IAChD,mBAAmB;IACnB,KAAK,EAAE,IAAI;IACX,IAAI,EAAE,GAAG;IACT,MAAM,EAAE,GAAG;IACX,KAAK,EAAE,GAAG;IACV,eAAe;IACf,aAAa,EAAE,IAAI;IACnB,SAAS,EAAE,GAAG;IACd,SAAS,EAAE,GAAG;IACd,cAAc,EAAE,IAAI;IACpB,cAAc,EAAE,GAAG;IACnB,eAAe,EAAE,IAAI;IACrB,eAAe,EAAE,IAAI;IACrB,SAAS,EAAE,IAAI;IACf,oBAAoB,EAAE,IAAI;IAC1B,eAAe,EAAE,IAAI;IACrB,SAAS,EAAE,IAAI;IACf,aAAa,EAAE,IAAI;IACnB,OAAO,EAAE,IAAI;IACb,YAAY,EAAE,IAAI;IAClB,YAAY,EAAE,IAAI;IAClB,eAAe;IACf,SAAS,EAAE,GAAG;IACd,cAAc,EAAE,GAAG;IACnB,cAAc,EAAE,GAAG;IACnB,QAAQ,EAAE,GAAG;IACb,aAAa,EAAE,IAAI;IACnB,mBAAmB;IACnB,QAAQ,EAAE,IAAI;IACd,EAAE,EAAE,GAAG;IACP,SAAS,EAAE,GAAG;IACd,SAAS,EAAE,GAAG;IACd,gBAAgB;IAChB,gBAAgB,EAAE,IAAI;IACtB,kBAAkB,EAAE,GAAG;IACvB,uBAAuB,EAAE,GAAG;IAC5B,kBAAkB,EAAE,GAAG;IACvB,uBAAuB,EAAE,KAAK;CAC/B,CAAC;AAEF,mEAAmE;AACnE,MAAM,4BAA4B,GAAG,GAAG,CAAC;AAEzC;;;;GAIG;AACH,MAAM,UAAU,mBAAmB,CAAC,KAAc;IAChD,IAAI,CAAC,KAAK;QAAE,OAAO,4BAA4B,CAAC;IAChD,MAAM,KAAK,GAAG,KAAK,CAAC,WAAW,EAAE,CAAC;IAElC,MAAM,MAAM,GAAG,iBAAiB,CAAC,KAAK,CAAC,CAAC;IACxC,IAAI,MAAM,KAAK,SAAS;QAAE,OAAO,MAAM,CAAC;IAExC,6EAA6E;IAC7E,+DAA+D;IAC/D,KAAK,MAAM,CAAC,GAAG,EAAE,IAAI,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,iBAAiB,CAAC,EAAE,CAAC;QAC5D,IAAI,KAAK,CAAC,QAAQ,CAAC,GAAG,CAAC;YAAE,OAAO,IAAI,CAAC;IACvC,CAAC;IAED,OAAO,4BAA4B,CAAC;AACtC,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,sBAAsB,CAAC,WAAmB,EAAE,KAAc;IACxE,IAAI,CAAC;QACH,MAAM,IAAI,GAAG,mBAAmB,CAAC,KAAK,CAAC,CAAC;QACxC,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,WAAW,GAAG,IAAI,GAAG,SAAS,CAAC,CAAC;IACrD,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,CAAC,CAAC;IACX,CAAC;AACH,CAAC;AAED,SAAS,MAAM,CAAC,GAAW,EAAE,QAAgB;IAC3C,MAAM,GAAG,GAAG,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC,EAAE,IAAI,EAAE,CAAC;IACrC,IAAI,CAAC,GAAG;QAAE,OAAO,QAAQ,CAAC;IAC1B,MAAM,MAAM,GAAG,QAAQ,CAAC,GAAG,EAAE,EAAE,CAAC,CAAC;IACjC,OAAO,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC;AAC3C,CAAC;AAED,qDAAqD;AACrD,MAAM,CAAC,MAAM,6BAA6B,GAAG,MAAM,CAAC,qCAAqC,EAAE,EAAE,CAAC,CAAC;AAC/F,MAAM,CAAC,MAAM,wBAAwB,GAAG,MAAM,CAAC,sCAAsC,EAAE,EAAE,CAAC,CAAC;AAE3F,2FAA2F;AAC3F,MAAM,6BAA6B,GAAG,IAAI,CAAC;AAO3C;;;;;;;;;;;;;;GAcG;AACH,MAAM,UAAU,2BAA2B,CAAC,OAAe,EAAE,KAAc,EAAE,aAAsB;IACjG,6EAA6E;IAC7E,6EAA6E;IAC7E,+EAA+E;IAC/E,mDAAmD;IACnD,MAAM,aAAa,GAAG,CAAC,aAAa,IAAI,aAAa,GAAG,CAAC,CAAC;QACxD,CAAC,CAAC,aAAa;QACf,CAAC,CAAC,qBAAqB,CAAC,KAAK,IAAI,EAAE,CAAC,CAAC;IACvC,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC,CAAC;IACxD,MAAM,UAAU,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,WAAW,GAAG,GAAG,CAAC,GAAG,aAAa,CAAC,CAAC;IACnE,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,UAAU,GAAG,6BAA6B,CAAC,CAAC;IACtE,OAAO,CAAC,KAAK,EAAE,aAAa,CAAC,CAAC;AAChC,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;GAqBG;AACH,MAAM,UAAU,sBAAsB,CACpC,YAAoB,EACpB,OAAe,EACf,aAA4B,EAC5B,eAAwB,EACxB,aAAsB,EACtB,iBAA0B,EAC1B,KAAc,EACd,aAAsB,EACtB,mBAA2B,6BAA6B,EACxD,aAAqB,wBAAwB;IAE7C,IAAI,CAAC,aAAa;QAAE,OAAO,EAAE,WAAW,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IAEjE,2EAA2E;IAC3E,gFAAgF;IAChF,gFAAgF;IAChF,gFAAgF;IAChF,IAAI,CAAC,iBAAiB;QAAE,OAAO,EAAE,WAAW,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IAErE,mFAAmF;IACnF,2DAA2D;IAC3D,IAAI,aAAa,KAAK,IAAI;QAAE,OAAO,EAAE,WAAW,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IAEzE,oBAAoB;IACpB,IAAI,eAAe;QAAE,OAAO,EAAE,WAAW,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IAElE,gEAAgE;IAChE,IAAI,CAAC,CAAC,YAAY,GAAG,gBAAgB,IAAI,OAAO,IAAI,UAAU,CAAC,EAAE,CAAC;QAChE,OAAO,EAAE,WAAW,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IAC/C,CAAC;IAED,MAAM,CAAC,KAAK,CAAC,GAAG,2BAA2B,CAAC,OAAO,EAAE,KAAK,EAAE,aAAa,CAAC,CAAC;IAC3E,MAAM,QAAQ,GAAG,KAAK,IAAI,IAAI,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC,KAAK,CAAC,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,IAAI,KAAK,EAAE,CAAC;IAC/E,MAAM,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;IACxC,MAAM,YAAY,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,CAAC;IAE9C,MAAM,GAAG,GAAG,sBAAsB,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;IACjD,MAAM,OAAO,GAAG,GAAG,IAAI,IAAI,CAAC,CAAC,CAAC,YAAY,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,yBAAyB,CAAC,CAAC,CAAC,EAAE,CAAC;IAEvF,MAAM,OAAO,GACX,2CAA2C,WAAW,6CAA6C,YAAY,IAAI;QACnH,8CAA8C,QAAQ,aAAa,WAAW,oBAAoB,OAAO,IAAI;QAC7G,wHAAwH;QACxH,+HAA+H,CAAC;IAElI,OAAO,EAAE,WAAW,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC;AACxC,CAAC"}

package/dist/nudges/verbosity-steer.d.ts

@@ -0,0 +1,28 @@
+/**
+ * Verbosity-steer nudge: inject a tiered conciseness nudge when context is
+ * under pressure. Mirrors the Python `run_verbosity_steer` function in
+ * measure.py.
+ *
+ * Tiered messaging:
+ *   25-74% fill + degraded quality (<75)  → gentle nudge
+ *   75-89% fill                           → strong nudge with specific directives
+ *   90%+ fill                             → suppressed (adding tokens makes it worse)
+ *
+ * Cooldown: max 3 nudges per session, 5 min between nudges.
+ * Shares the same nudge_count / last_nudge_time fields as the quality nudge
+ * so the two features share a single cooldown counter.
+ */
+import type { SessionStore } from "../storage/session-store.js";
+export interface VerbositySteerResult {
+    shouldNudge: boolean;
+    message: string | null;
+    tier: "gentle" | "strong" | "suppressed" | "none";
+}
+export declare function checkVerbositySteer(store: SessionStore, fillPct: number, qualityScore: number): VerbositySteerResult;
+/**
+ * Estimate output token savings from a verbosity-steer nudge.
+ * The nudge causes the model to produce ~10-15% fewer output tokens
+ * on affected responses. Returns [tokensSaved, tier] for logging.
+ */
+export declare function verbositySteerSavingsEstimate(fillPct: number): [number, string];
+//# sourceMappingURL=verbosity-steer.d.ts.map

package/dist/nudges/verbosity-steer.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"verbosity-steer.d.ts","sourceRoot":"","sources":["../../src/nudges/verbosity-steer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,6BAA6B,CAAC;AAShE,MAAM,WAAW,oBAAoB;IACnC,WAAW,EAAE,OAAO,CAAC;IACrB,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IACvB,IAAI,EAAE,QAAQ,GAAG,QAAQ,GAAG,YAAY,GAAG,MAAM,CAAC;CACnD;AAED,wBAAgB,mBAAmB,CACjC,KAAK,EAAE,YAAY,EACnB,OAAO,EAAE,MAAM,EACf,YAAY,EAAE,MAAM,GACnB,oBAAoB,CAkCtB;AAED;;;;GAIG;AACH,wBAAgB,6BAA6B,CAC3C,OAAO,EAAE,MAAM,GACd,CAAC,MAAM,EAAE,MAAM,CAAC,CAIlB"}

package/dist/nudges/verbosity-steer.js

@@ -0,0 +1,61 @@
+/**
+ * Verbosity-steer nudge: inject a tiered conciseness nudge when context is
+ * under pressure. Mirrors the Python `run_verbosity_steer` function in
+ * measure.py.
+ *
+ * Tiered messaging:
+ *   25-74% fill + degraded quality (<75)  → gentle nudge
+ *   75-89% fill                           → strong nudge with specific directives
+ *   90%+ fill                             → suppressed (adding tokens makes it worse)
+ *
+ * Cooldown: max 3 nudges per session, 5 min between nudges.
+ * Shares the same nudge_count / last_nudge_time fields as the quality nudge
+ * so the two features share a single cooldown counter.
+ */
+const COOLDOWN_SEC = 300; // 5 minutes
+const SESSION_CAP = 3;
+const GENTLE_FILL_THRESHOLD = 25;
+const STRONG_FILL_THRESHOLD = 75;
+const CRITICAL_FILL_THRESHOLD = 90;
+const QUALITY_THRESHOLD = 75;
+export function checkVerbositySteer(store, fillPct, qualityScore) {
+    const cache = store.getQualityCache();
+    const nudgeCount = cache?.nudge_count ?? 0;
+    const lastNudgeTime = cache?.last_nudge_time ?? 0;
+    const now = Date.now() / 1000;
+    // Cooldown and session cap (shared with quality nudge)
+    if (nudgeCount >= SESSION_CAP)
+        return { shouldNudge: false, message: null, tier: "none" };
+    if (now - lastNudgeTime < COOLDOWN_SEC)
+        return { shouldNudge: false, message: null, tier: "none" };
+    // At 90%+ fill, don't add more tokens — suppress entirely
+    if (fillPct >= CRITICAL_FILL_THRESHOLD) {
+        return { shouldNudge: false, message: null, tier: "suppressed" };
+    }
+    // Strong tier: 75%+ fill regardless of quality
+    if (fillPct >= STRONG_FILL_THRESHOLD) {
+        const message = `[Token Optimizer] Context at ${Math.round(fillPct)}% capacity, quality ${Math.round(qualityScore)}/100. ` +
+            "Reason as deeply as you need — but keep your visible output lean: no preamble, " +
+            "no restating the request, no explanations unless asked. Every token saved extends the session.";
+        return { shouldNudge: true, message, tier: "strong" };
+    }
+    // Gentle tier: 25%+ fill with degraded quality
+    if (fillPct >= GENTLE_FILL_THRESHOLD && qualityScore < QUALITY_THRESHOLD) {
+        const message = `[Token Optimizer] Context at ${Math.round(fillPct)}% capacity, quality ${Math.round(qualityScore)}/100. ` +
+            "Reason fully, then keep your output lean — skip restating the request and " +
+            "omit unnecessary preamble. Every token saved extends the session.";
+        return { shouldNudge: true, message, tier: "gentle" };
+    }
+    return { shouldNudge: false, message: null, tier: "none" };
+}
+/**
+ * Estimate output token savings from a verbosity-steer nudge.
+ * The nudge causes the model to produce ~10-15% fewer output tokens
+ * on affected responses. Returns [tokensSaved, tier] for logging.
+ */
+export function verbositySteerSavingsEstimate(fillPct) {
+    const avgResponseTokens = 800;
+    const reduction = fillPct >= STRONG_FILL_THRESHOLD ? 0.15 : 0.10;
+    return [Math.round(avgResponseTokens * reduction), fillPct >= STRONG_FILL_THRESHOLD ? "strong" : "gentle"];
+}
+//# sourceMappingURL=verbosity-steer.js.map

package/dist/nudges/verbosity-steer.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"verbosity-steer.js","sourceRoot":"","sources":["../../src/nudges/verbosity-steer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAIH,MAAM,YAAY,GAAG,GAAG,CAAC,CAAC,YAAY;AACtC,MAAM,WAAW,GAAG,CAAC,CAAC;AACtB,MAAM,qBAAqB,GAAG,EAAE,CAAC;AACjC,MAAM,qBAAqB,GAAG,EAAE,CAAC;AACjC,MAAM,uBAAuB,GAAG,EAAE,CAAC;AACnC,MAAM,iBAAiB,GAAG,EAAE,CAAC;AAQ7B,MAAM,UAAU,mBAAmB,CACjC,KAAmB,EACnB,OAAe,EACf,YAAoB;IAEpB,MAAM,KAAK,GAAG,KAAK,CAAC,eAAe,EAAE,CAAC;IACtC,MAAM,UAAU,GAAG,KAAK,EAAE,WAAW,IAAI,CAAC,CAAC;IAC3C,MAAM,aAAa,GAAG,KAAK,EAAE,eAAe,IAAI,CAAC,CAAC;IAClD,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,IAAI,CAAC;IAE9B,uDAAuD;IACvD,IAAI,UAAU,IAAI,WAAW;QAAE,OAAO,EAAE,WAAW,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;IAC1F,IAAI,GAAG,GAAG,aAAa,GAAG,YAAY;QAAE,OAAO,EAAE,WAAW,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;IAEnG,0DAA0D;IAC1D,IAAI,OAAO,IAAI,uBAAuB,EAAE,CAAC;QACvC,OAAO,EAAE,WAAW,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,YAAY,EAAE,CAAC;IACnE,CAAC;IAED,+CAA+C;IAC/C,IAAI,OAAO,IAAI,qBAAqB,EAAE,CAAC;QACrC,MAAM,OAAO,GACX,gCAAgC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,uBAAuB,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,QAAQ;YAC1G,iFAAiF;YACjF,gGAAgG,CAAC;QACnG,OAAO,EAAE,WAAW,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC;IACxD,CAAC;IAED,+CAA+C;IAC/C,IAAI,OAAO,IAAI,qBAAqB,IAAI,YAAY,GAAG,iBAAiB,EAAE,CAAC;QACzE,MAAM,OAAO,GACX,gCAAgC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,uBAAuB,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,QAAQ;YAC1G,4EAA4E;YAC5E,mEAAmE,CAAC;QACtE,OAAO,EAAE,WAAW,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC;IACxD,CAAC;IAED,OAAO,EAAE,WAAW,EAAE,KAAK,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC;AAC7D,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,6BAA6B,CAC3C,OAAe;IAEf,MAAM,iBAAiB,GAAG,GAAG,CAAC;IAC9B,MAAM,SAAS,GAAG,OAAO,IAAI,qBAAqB,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC;IACjE,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,iBAAiB,GAAG,SAAS,CAAC,EAAE,OAAO,IAAI,qBAAqB,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC;AAC7G,CAAC"}

package/dist/pricing.d.ts

@@ -0,0 +1,58 @@
+/**
+ * Minimal per-MTok rate card for OpenCode's realized-savings counterfactual.
+ *
+ * OpenCode persists a pre-computed `cost_usd` per session (priced at its own
+ * model when recorded), so the ACTUAL arm needs no rate card. The
+ * COUNTERFACTUAL arm, however, reprices the same token VOLUME at a DIFFERENT
+ * (baseline) model mix, which requires a per-class rate card the stored cost
+ * cannot supply. This module is that card.
+ *
+ * Rates mirror openclaw/src/pricing.ts exactly (USD per token; verified
+ * May 30, 2026). cacheWrite = 5-minute-TTL rate (1.25x input); cacheWrite1h =
+ * 1-hour-TTL rate (2x input), only set for Claude models that support it.
+ *
+ * A "mix" is a {modelKey -> share} map (shares sum to ~1). price() blends the
+ * per-token cost across the mix; unpriced models fall back to a proxy rate.
+ */
+export interface ModelPricing {
+    input: number;
+    output: number;
+    cacheRead: number;
+    /** 5-minute cache-write rate (1.25x input). Used when TTL is unknown. */
+    cacheWrite: number;
+    /** 1-hour cache-write rate (2x input). Only set for Claude models. */
+    cacheWrite1h?: number;
+}
+/** Default pricing (USD per token). Mirrors openclaw/src/pricing.ts. */
+export declare const DEFAULT_PRICING: Record<string, ModelPricing>;
+/** Price the proxy rate card uses when a model is unpriced. */
+export declare const PROXY_MODEL = "sonnet";
+/**
+ * Normalize a model ID into a pricing key. Mirrors openclaw/src/pricing.ts.
+ * Handles provider prefixes (anthropic/claude-sonnet-4-6 -> sonnet) and version
+ * suffixes (gpt-5.2-2026-03 -> gpt-5.2). Returns lowercased raw on no match.
+ */
+export declare function normalizeModelName(modelId: string): string;
+/** A model mix: modelKey (or display name) -> token share. Shares sum to ~1. */
+export type ModelMix = Record<string, number>;
+/**
+ * Price the fresh+cache_read POOL and OUTPUT at a model mix (NO cache-write).
+ * Linear in tokens, so aggregate window totals price the whole window directly.
+ * Mirrors measure.py's `price(fi, cr, out, shares)`.
+ */
+export declare function price(F: number, CR: number, O: number, mix: ModelMix): number;
+/**
+ * Price cache-write at a model mix, TTL-aware: 1h writes bill at 2x input, 5m
+ * at 1.25x. Cache-write IS a routing lever (billed at the writing model's
+ * rate), so each arm prices CW at its OWN mix. Mirrors measure.py's `price_cw`.
+ * OpenCode's session_log has no 5m/1h split, so all writes are treated as 5m
+ * (conservative) unless cw1h is supplied.
+ */
+export declare function price_cw(CW: number, mix: ModelMix, CW_5m?: number, CW_1h?: number): number;
+/**
+ * Cost of 1M fresh-input tokens at a mix. Used to reprice the compression
+ * add-back: baseline_input_rate / current_input_rate. Mirrors measure.py's
+ * `price(1_000_000, 0, 0, shares)`.
+ */
+export declare function inputRatePerMTok(mix: ModelMix): number;
+//# sourceMappingURL=pricing.d.ts.map

package/dist/pricing.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"pricing.d.ts","sourceRoot":"","sources":["../src/pricing.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,MAAM,WAAW,YAAY;IAC3B,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,yEAAyE;IACzE,UAAU,EAAE,MAAM,CAAC;IACnB,sEAAsE;IACtE,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,wEAAwE;AACxE,eAAO,MAAM,eAAe,EAAE,MAAM,CAAC,MAAM,EAAE,YAAY,CAqExD,CAAC;AAEF,+DAA+D;AAC/D,eAAO,MAAM,WAAW,WAAW,CAAC;AA2BpC;;;;GAIG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAuE1D;AAQD,gFAAgF;AAChF,MAAM,MAAM,QAAQ,GAAG,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;AA8B9C;;;;GAIG;AACH,wBAAgB,KAAK,CAAC,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,GAAG,EAAE,QAAQ,GAAG,MAAM,CAM7E;AAED;;;;;;GAMG;AACH,wBAAgB,QAAQ,CAAC,EAAE,EAAE,MAAM,EAAE,GAAG,EAAE,QAAQ,EAAE,KAAK,CAAC,EAAE,MAAM,EAAE,KAAK,CAAC,EAAE,MAAM,GAAG,MAAM,CAO1F;AAED;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,GAAG,EAAE,QAAQ,GAAG,MAAM,CAEtD"}