@caupulican/pi-adaptative 0.80.75 → 0.80.76

package/dist/core/cost-guard.d.ts

@@ -0,0 +1,55 @@
+/**
+ * Proactive per-turn token cost guard (Hermes-parity superiority item #34).
+ *
+ * Hermes (and pi today) only react to context growth by compressing AFTER it's expensive. This estimates
+ * the dollar cost of the NEXT LLM call BEFORE it is submitted, so the agent can warn the user or
+ * automatically reduce reasoning effort before a runaway billing spike — a proactive ceiling, not a
+ * reactive cleanup. Pure functions: no I/O, fully testable.
+ */
+/** Per-token USD prices (as carried on `Model.cost`, which is per-token, not per-million). */
+export interface ModelTokenCost {
+    input: number;
+    output: number;
+    cacheRead?: number;
+    cacheWrite?: number;
+}
+/**
+ * Estimate the USD cost of one turn: the whole current context is billed as input, plus up to
+ * `maxOutputTokens` of output. `cachedInputTokens` (prefix-cache hits) are billed at the cheaper
+ * cache-read rate instead of the full input rate. This is an UPPER bound on the turn (it assumes the
+ * model emits its full output budget), which is what a spending ceiling should bound against.
+ */
+export declare function estimateTurnCostUsd(args: {
+    inputTokens: number;
+    maxOutputTokens: number;
+    cost: ModelTokenCost;
+    cachedInputTokens?: number;
+}): number;
+/** What to do when a turn's projected cost exceeds the threshold. */
+export type CostGuardAction = "warn" | "downgrade";
+export interface CostGuardSettings {
+    /** Per-turn USD ceiling. `0` (default) disables the guard entirely. */
+    maxTurnUsd: number;
+    /** Over the ceiling: `warn` (surface a notice) or `downgrade` (also reduce reasoning effort). */
+    action: CostGuardAction;
+}
+export declare const DEFAULT_COST_GUARD_SETTINGS: CostGuardSettings;
+export interface CostGuardDecision {
+    /** True when the guard is enabled AND the projected cost exceeds the ceiling. */
+    over: boolean;
+    estUsd: number;
+    thresholdUsd: number;
+    action: CostGuardAction;
+}
+/** Decide whether the projected turn cost trips the guard. Disabled (`maxTurnUsd<=0`) is never `over`. */
+export declare function evaluateCostGuard(estUsd: number, settings: CostGuardSettings): CostGuardDecision;
+/** Reasoning levels in descending cost order, used to pick the next-cheaper level on a downgrade. */
+declare const REASONING_LADDER: readonly ["xhigh", "high", "medium", "low", "minimal", "off"];
+export type ReasoningLevel = (typeof REASONING_LADDER)[number];
+/**
+ * One step down the reasoning ladder (cost reduction) from `current`. Returns `current` unchanged when
+ * already at the floor or unrecognized — the guard never raises effort, only lowers it.
+ */
+export declare function downgradeReasoning(current: string): ReasoningLevel | string;
+export {};
+//# sourceMappingURL=cost-guard.d.ts.map

package/dist/core/cost-guard.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"cost-guard.d.ts","sourceRoot":"","sources":["../../src/core/cost-guard.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,8FAA8F;AAC9F,MAAM,WAAW,cAAc;IAC9B,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;CACpB;AAED;;;;;GAKG;AACH,wBAAgB,mBAAmB,CAAC,IAAI,EAAE;IACzC,WAAW,EAAE,MAAM,CAAC;IACpB,eAAe,EAAE,MAAM,CAAC;IACxB,IAAI,EAAE,cAAc,CAAC;IACrB,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC3B,GAAG,MAAM,CAQT;AAED,qEAAqE;AACrE,MAAM,MAAM,eAAe,GAAG,MAAM,GAAG,WAAW,CAAC;AAEnD,MAAM,WAAW,iBAAiB;IACjC,uEAAuE;IACvE,UAAU,EAAE,MAAM,CAAC;IACnB,iGAAiG;IACjG,MAAM,EAAE,eAAe,CAAC;CACxB;AAED,eAAO,MAAM,2BAA2B,EAAE,iBAGzC,CAAC;AAEF,MAAM,WAAW,iBAAiB;IACjC,iFAAiF;IACjF,IAAI,EAAE,OAAO,CAAC;IACd,MAAM,EAAE,MAAM,CAAC;IACf,YAAY,EAAE,MAAM,CAAC;IACrB,MAAM,EAAE,eAAe,CAAC;CACxB;AAED,0GAA0G;AAC1G,wBAAgB,iBAAiB,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,EAAE,iBAAiB,GAAG,iBAAiB,CAQhG;AAED,qGAAqG;AACrG,QAAA,MAAM,gBAAgB,+DAAgE,CAAC;AACvF,MAAM,MAAM,cAAc,GAAG,CAAC,OAAO,gBAAgB,CAAC,CAAC,MAAM,CAAC,CAAC;AAE/D;;;GAGG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,MAAM,GAAG,cAAc,GAAG,MAAM,CAI3E","sourcesContent":["/**\n * Proactive per-turn token cost guard (Hermes-parity superiority item #34).\n *\n * Hermes (and pi today) only react to context growth by compressing AFTER it's expensive. This estimates\n * the dollar cost of the NEXT LLM call BEFORE it is submitted, so the agent can warn the user or\n * automatically reduce reasoning effort before a runaway billing spike — a proactive ceiling, not a\n * reactive cleanup. Pure functions: no I/O, fully testable.\n */\n\n/** Per-token USD prices (as carried on `Model.cost`, which is per-token, not per-million). */\nexport interface ModelTokenCost {\n\tinput: number;\n\toutput: number;\n\tcacheRead?: number;\n\tcacheWrite?: number;\n}\n\n/**\n * Estimate the USD cost of one turn: the whole current context is billed as input, plus up to\n * `maxOutputTokens` of output. `cachedInputTokens` (prefix-cache hits) are billed at the cheaper\n * cache-read rate instead of the full input rate. This is an UPPER bound on the turn (it assumes the\n * model emits its full output budget), which is what a spending ceiling should bound against.\n */\nexport function estimateTurnCostUsd(args: {\n\tinputTokens: number;\n\tmaxOutputTokens: number;\n\tcost: ModelTokenCost;\n\tcachedInputTokens?: number;\n}): number {\n\tconst { inputTokens, maxOutputTokens, cost } = args;\n\tconst cached = Math.max(0, Math.min(args.cachedInputTokens ?? 0, inputTokens));\n\tconst freshInput = inputTokens - cached;\n\tconst cacheReadRate = cost.cacheRead ?? cost.input;\n\tconst inputUsd = freshInput * cost.input + cached * cacheReadRate;\n\tconst outputUsd = Math.max(0, maxOutputTokens) * cost.output;\n\treturn inputUsd + outputUsd;\n}\n\n/** What to do when a turn's projected cost exceeds the threshold. */\nexport type CostGuardAction = \"warn\" | \"downgrade\";\n\nexport interface CostGuardSettings {\n\t/** Per-turn USD ceiling. `0` (default) disables the guard entirely. */\n\tmaxTurnUsd: number;\n\t/** Over the ceiling: `warn` (surface a notice) or `downgrade` (also reduce reasoning effort). */\n\taction: CostGuardAction;\n}\n\nexport const DEFAULT_COST_GUARD_SETTINGS: CostGuardSettings = {\n\tmaxTurnUsd: 0,\n\taction: \"warn\",\n};\n\nexport interface CostGuardDecision {\n\t/** True when the guard is enabled AND the projected cost exceeds the ceiling. */\n\tover: boolean;\n\testUsd: number;\n\tthresholdUsd: number;\n\taction: CostGuardAction;\n}\n\n/** Decide whether the projected turn cost trips the guard. Disabled (`maxTurnUsd<=0`) is never `over`. */\nexport function evaluateCostGuard(estUsd: number, settings: CostGuardSettings): CostGuardDecision {\n\tconst enabled = settings.maxTurnUsd > 0;\n\treturn {\n\t\tover: enabled && estUsd > settings.maxTurnUsd,\n\t\testUsd,\n\t\tthresholdUsd: settings.maxTurnUsd,\n\t\taction: settings.action,\n\t};\n}\n\n/** Reasoning levels in descending cost order, used to pick the next-cheaper level on a downgrade. */\nconst REASONING_LADDER = [\"xhigh\", \"high\", \"medium\", \"low\", \"minimal\", \"off\"] as const;\nexport type ReasoningLevel = (typeof REASONING_LADDER)[number];\n\n/**\n * One step down the reasoning ladder (cost reduction) from `current`. Returns `current` unchanged when\n * already at the floor or unrecognized — the guard never raises effort, only lowers it.\n */\nexport function downgradeReasoning(current: string): ReasoningLevel | string {\n\tconst i = REASONING_LADDER.indexOf(current as ReasoningLevel);\n\tif (i < 0) return current;\n\treturn REASONING_LADDER[Math.min(i + 1, REASONING_LADDER.length - 1)];\n}\n"]}

package/dist/core/cost-guard.js

@@ -0,0 +1,50 @@
+/**
+ * Proactive per-turn token cost guard (Hermes-parity superiority item #34).
+ *
+ * Hermes (and pi today) only react to context growth by compressing AFTER it's expensive. This estimates
+ * the dollar cost of the NEXT LLM call BEFORE it is submitted, so the agent can warn the user or
+ * automatically reduce reasoning effort before a runaway billing spike — a proactive ceiling, not a
+ * reactive cleanup. Pure functions: no I/O, fully testable.
+ */
+/**
+ * Estimate the USD cost of one turn: the whole current context is billed as input, plus up to
+ * `maxOutputTokens` of output. `cachedInputTokens` (prefix-cache hits) are billed at the cheaper
+ * cache-read rate instead of the full input rate. This is an UPPER bound on the turn (it assumes the
+ * model emits its full output budget), which is what a spending ceiling should bound against.
+ */
+export function estimateTurnCostUsd(args) {
+    const { inputTokens, maxOutputTokens, cost } = args;
+    const cached = Math.max(0, Math.min(args.cachedInputTokens ?? 0, inputTokens));
+    const freshInput = inputTokens - cached;
+    const cacheReadRate = cost.cacheRead ?? cost.input;
+    const inputUsd = freshInput * cost.input + cached * cacheReadRate;
+    const outputUsd = Math.max(0, maxOutputTokens) * cost.output;
+    return inputUsd + outputUsd;
+}
+export const DEFAULT_COST_GUARD_SETTINGS = {
+    maxTurnUsd: 0,
+    action: "warn",
+};
+/** Decide whether the projected turn cost trips the guard. Disabled (`maxTurnUsd<=0`) is never `over`. */
+export function evaluateCostGuard(estUsd, settings) {
+    const enabled = settings.maxTurnUsd > 0;
+    return {
+        over: enabled && estUsd > settings.maxTurnUsd,
+        estUsd,
+        thresholdUsd: settings.maxTurnUsd,
+        action: settings.action,
+    };
+}
+/** Reasoning levels in descending cost order, used to pick the next-cheaper level on a downgrade. */
+const REASONING_LADDER = ["xhigh", "high", "medium", "low", "minimal", "off"];
+/**
+ * One step down the reasoning ladder (cost reduction) from `current`. Returns `current` unchanged when
+ * already at the floor or unrecognized — the guard never raises effort, only lowers it.
+ */
+export function downgradeReasoning(current) {
+    const i = REASONING_LADDER.indexOf(current);
+    if (i < 0)
+        return current;
+    return REASONING_LADDER[Math.min(i + 1, REASONING_LADDER.length - 1)];
+}
+//# sourceMappingURL=cost-guard.js.map

package/dist/core/cost-guard.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cost-guard.js","sourceRoot":"","sources":["../../src/core/cost-guard.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAUH;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,IAKnC,EAAU;IACV,MAAM,EAAE,WAAW,EAAE,eAAe,EAAE,IAAI,EAAE,GAAG,IAAI,CAAC;IACpD,MAAM,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,iBAAiB,IAAI,CAAC,EAAE,WAAW,CAAC,CAAC,CAAC;IAC/E,MAAM,UAAU,GAAG,WAAW,GAAG,MAAM,CAAC;IACxC,MAAM,aAAa,GAAG,IAAI,CAAC,SAAS,IAAI,IAAI,CAAC,KAAK,CAAC;IACnD,MAAM,QAAQ,GAAG,UAAU,GAAG,IAAI,CAAC,KAAK,GAAG,MAAM,GAAG,aAAa,CAAC;IAClE,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,eAAe,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;IAC7D,OAAO,QAAQ,GAAG,SAAS,CAAC;AAAA,CAC5B;AAYD,MAAM,CAAC,MAAM,2BAA2B,GAAsB;IAC7D,UAAU,EAAE,CAAC;IACb,MAAM,EAAE,MAAM;CACd,CAAC;AAUF,0GAA0G;AAC1G,MAAM,UAAU,iBAAiB,CAAC,MAAc,EAAE,QAA2B,EAAqB;IACjG,MAAM,OAAO,GAAG,QAAQ,CAAC,UAAU,GAAG,CAAC,CAAC;IACxC,OAAO;QACN,IAAI,EAAE,OAAO,IAAI,MAAM,GAAG,QAAQ,CAAC,UAAU;QAC7C,MAAM;QACN,YAAY,EAAE,QAAQ,CAAC,UAAU;QACjC,MAAM,EAAE,QAAQ,CAAC,MAAM;KACvB,CAAC;AAAA,CACF;AAED,qGAAqG;AACrG,MAAM,gBAAgB,GAAG,CAAC,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,SAAS,EAAE,KAAK,CAAU,CAAC;AAGvF;;;GAGG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAAe,EAA2B;IAC5E,MAAM,CAAC,GAAG,gBAAgB,CAAC,OAAO,CAAC,OAAyB,CAAC,CAAC;IAC9D,IAAI,CAAC,GAAG,CAAC;QAAE,OAAO,OAAO,CAAC;IAC1B,OAAO,gBAAgB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,EAAE,gBAAgB,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC;AAAA,CACtE","sourcesContent":["/**\n * Proactive per-turn token cost guard (Hermes-parity superiority item #34).\n *\n * Hermes (and pi today) only react to context growth by compressing AFTER it's expensive. This estimates\n * the dollar cost of the NEXT LLM call BEFORE it is submitted, so the agent can warn the user or\n * automatically reduce reasoning effort before a runaway billing spike — a proactive ceiling, not a\n * reactive cleanup. Pure functions: no I/O, fully testable.\n */\n\n/** Per-token USD prices (as carried on `Model.cost`, which is per-token, not per-million). */\nexport interface ModelTokenCost {\n\tinput: number;\n\toutput: number;\n\tcacheRead?: number;\n\tcacheWrite?: number;\n}\n\n/**\n * Estimate the USD cost of one turn: the whole current context is billed as input, plus up to\n * `maxOutputTokens` of output. `cachedInputTokens` (prefix-cache hits) are billed at the cheaper\n * cache-read rate instead of the full input rate. This is an UPPER bound on the turn (it assumes the\n * model emits its full output budget), which is what a spending ceiling should bound against.\n */\nexport function estimateTurnCostUsd(args: {\n\tinputTokens: number;\n\tmaxOutputTokens: number;\n\tcost: ModelTokenCost;\n\tcachedInputTokens?: number;\n}): number {\n\tconst { inputTokens, maxOutputTokens, cost } = args;\n\tconst cached = Math.max(0, Math.min(args.cachedInputTokens ?? 0, inputTokens));\n\tconst freshInput = inputTokens - cached;\n\tconst cacheReadRate = cost.cacheRead ?? cost.input;\n\tconst inputUsd = freshInput * cost.input + cached * cacheReadRate;\n\tconst outputUsd = Math.max(0, maxOutputTokens) * cost.output;\n\treturn inputUsd + outputUsd;\n}\n\n/** What to do when a turn's projected cost exceeds the threshold. */\nexport type CostGuardAction = \"warn\" | \"downgrade\";\n\nexport interface CostGuardSettings {\n\t/** Per-turn USD ceiling. `0` (default) disables the guard entirely. */\n\tmaxTurnUsd: number;\n\t/** Over the ceiling: `warn` (surface a notice) or `downgrade` (also reduce reasoning effort). */\n\taction: CostGuardAction;\n}\n\nexport const DEFAULT_COST_GUARD_SETTINGS: CostGuardSettings = {\n\tmaxTurnUsd: 0,\n\taction: \"warn\",\n};\n\nexport interface CostGuardDecision {\n\t/** True when the guard is enabled AND the projected cost exceeds the ceiling. */\n\tover: boolean;\n\testUsd: number;\n\tthresholdUsd: number;\n\taction: CostGuardAction;\n}\n\n/** Decide whether the projected turn cost trips the guard. Disabled (`maxTurnUsd<=0`) is never `over`. */\nexport function evaluateCostGuard(estUsd: number, settings: CostGuardSettings): CostGuardDecision {\n\tconst enabled = settings.maxTurnUsd > 0;\n\treturn {\n\t\tover: enabled && estUsd > settings.maxTurnUsd,\n\t\testUsd,\n\t\tthresholdUsd: settings.maxTurnUsd,\n\t\taction: settings.action,\n\t};\n}\n\n/** Reasoning levels in descending cost order, used to pick the next-cheaper level on a downgrade. */\nconst REASONING_LADDER = [\"xhigh\", \"high\", \"medium\", \"low\", \"minimal\", \"off\"] as const;\nexport type ReasoningLevel = (typeof REASONING_LADDER)[number];\n\n/**\n * One step down the reasoning ladder (cost reduction) from `current`. Returns `current` unchanged when\n * already at the floor or unrecognized — the guard never raises effort, only lowers it.\n */\nexport function downgradeReasoning(current: string): ReasoningLevel | string {\n\tconst i = REASONING_LADDER.indexOf(current as ReasoningLevel);\n\tif (i < 0) return current;\n\treturn REASONING_LADDER[Math.min(i + 1, REASONING_LADDER.length - 1)];\n}\n"]}

package/dist/core/learning/reflection-engine.d.ts

@@ -51,6 +51,13 @@ export interface ReflectionResult {
     usage: Usage;
     rationale: string;
 }
+/**
+ * STATIC reflection system prompt (Hermes-parity #33). It is byte-identical across every reflection
+ * pass — the variable parts (existing memory snapshot + the turn transcript) live in the USER prompt —
+ * so the provider prompt-cache reuses this prefix instead of re-billing it each pass (cost guard).
+ * Do NOT interpolate per-call data into this constant or caching breaks.
+ */
+export declare const REFLECTION_SYSTEM_PROMPT = "You are a reflection engine. Your job is to analyze the recent conversation turn, compare it against the agent's existing memory, and decide if any memory updates are needed.\n\nMemory guidelines:\n- \"MEMORY\" is for project facts, configuration, repeatable workflows, and coding findings.\n- \"USER\" is for user preferences, patterns, and style specifications.\n- Avoid duplicate facts. If the fact is already represented, do not add it.\n- CONFRONT existing memory: if the new turn contradicts or updates an existing fact, use \"memory_replace\" or \"memory_remove\" to supersede the old fact rather than blindly appending.\n- Keep memories short, factual, and direct. No fluff.\n- Do NOT capture transient/environment-specific noise: tool/network failures, one-off errors, or a single narrative event. Persist only durable facts and preferences.\n- PROMOTE to behavior: if the turn established a REPEATABLE, multi-step PROCEDURE/workflow (not a one-off fact) that should govern a future class of tasks, emit a \"promote_skill\" instead of (or in addition to) a memory fact. Only promote a genuinely reusable procedure \u2014 never a single fact, a one-off narrative, or environment-specific noise. Prefer a memory fact when unsure.\n\nYou must output your analysis and writes in the following JSON format inside a ```json``` code fence:\n{\n  \"rationale\": \"Explanation of your reasoning\",\n  \"writes\": [\n    { \"kind\": \"memory_add\", \"section\": \"MEMORY\" | \"USER\", \"text\": \"New direct fact to append\" },\n    { \"kind\": \"memory_replace\", \"target\": \"Exact text substring to replace\", \"text\": \"New replacement text\" },\n    { \"kind\": \"memory_remove\", \"target\": \"Exact text substring to remove\" },\n    { \"kind\": \"promote_skill\", \"name\": \"kebab-case-skill-name\", \"description\": \"one line of when to use it\", \"body\": \"Markdown: the step-by-step procedure\" }\n  ]\n}\n";
 export declare class ReflectionEngine {
     /**
      * Build the reflection prompt, call the injected isolated complete(),

package/dist/core/learning/reflection-engine.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"reflection-engine.d.ts","sourceRoot":"","sources":["../../../src/core/learning/reflection-engine.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,mBAAmB,CAAC;AAE/C,MAAM,MAAM,UAAU,GAAG,MAAM,GAAG,SAAS,GAAG,SAAS,GAAG,OAAO,GAAG,MAAM,CAAC;AAE3E,MAAM,WAAW,wBAAwB;IACxC,IAAI,EAAE,MAAM,CAAC;IACb,KAAK,EAAE,KAAK,CAAC;IACb,UAAU,EAAE,UAAU,CAAC;CACvB;AAED,MAAM,MAAM,iBAAiB,GAAG,SAAS,GAAG,YAAY,GAAG,aAAa,GAAG,MAAM,CAAC;AAElF,MAAM,WAAW,aAAa;IAC7B,OAAO,EAAE,iBAAiB,CAAC;IAC3B,aAAa,EAAE,MAAM,CAAC;IACtB,aAAa,EAAE,OAAO,CAAC;IACvB,kBAAkB,EAAE,MAAM,CAAC;IAC3B,YAAY,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,UAAU;IAC1B,GAAG,EAAE,MAAM,GAAG,SAAS,CAAC;IACxB,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;CACpB;AAED;;;GAGG;AACH,wBAAgB,YAAY,CAAC,OAAO,EAAE,aAAa,GAAG,UAAU,CAyB/D;AAED,MAAM,WAAW,eAAe;IAC/B,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,CAAC;IACvB,IAAI,EAAE,UAAU,CAAC;IAEjB,QAAQ,EAAE,CAAC,YAAY,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,KAAK,OAAO,CAAC,wBAAwB,CAAC,CAAC;CAC1F;AAED,MAAM,MAAM,eAAe,GACxB;IAAE,IAAI,EAAE,YAAY,CAAC;IAAC,OAAO,EAAE,QAAQ,GAAG,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GAChE;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GACxD;IAAE,IAAI,EAAE,eAAe,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,GAEzC;IAAE,IAAI,EAAE,eAAe,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,WAAW,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAAC;AAE9E,MAAM,WAAW,gBAAgB;IAChC,MAAM,EAAE,eAAe,EAAE,CAAC;IAC1B,KAAK,EAAE,KAAK,CAAC;IACb,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,qBAAa,gBAAgB;IAC5B;;;;OAIG;IACG,OAAO,CAAC,KAAK,EAAE,eAAe,GAAG,OAAO,CAAC,gBAAgB,CAAC,CAkG/D;CACD","sourcesContent":["import type { Usage } from \"@caupulican/pi-ai\";\n\nexport type StopReason = \"stop\" | \"toolUse\" | \"aborted\" | \"error\" | string;\n\nexport interface IsolatedCompletionResult {\n\ttext: string;\n\tusage: Usage;\n\tstopReason: StopReason;\n}\n\nexport type ReflectionTrigger = \"complex\" | \"corrective\" | \"session-end\" | \"none\";\n\nexport interface DemandSignals {\n\ttrigger: ReflectionTrigger;\n\ttoolCallCount: number;\n\thadCorrection: boolean;\n\tcontextHeadroomPct: number; // 0..100\n\tusefulLately: number; // 0..1 rolling score\n}\n\nexport interface DemandPlan {\n\tact: \"skip\" | \"reflect\";\n\treason: string;\n\ttokenBudget: number;\n}\n\n/**\n * Pure zero-I/O heuristic to decide whether the current turn justifies a reflection run\n * and determine the token budget under the cheap-tool net-negative doctrine.\n */\nexport function decideDemand(signals: DemandSignals): DemandPlan {\n\tif (signals.trigger === \"none\") {\n\t\treturn { act: \"skip\", reason: \"No trigger detected\", tokenBudget: 0 };\n\t}\n\tif (signals.contextHeadroomPct < 10) {\n\t\treturn { act: \"skip\", reason: \"Context headroom is critically low (< 10%)\", tokenBudget: 0 };\n\t}\n\n\t// Dynamic token budget based on headroom (keep reflection bounded between 500 and 1500 tokens)\n\tconst baseBudget = 1000;\n\tconst tokenBudget = Math.max(500, Math.min(1500, Math.round(baseBudget * (signals.contextHeadroomPct / 100))));\n\n\tif (signals.hadCorrection) {\n\t\treturn { act: \"reflect\", reason: \"Correction detected in the turn\", tokenBudget };\n\t}\n\tif (signals.trigger === \"session-end\") {\n\t\treturn { act: \"reflect\", reason: \"Session end reflection triggered\", tokenBudget };\n\t}\n\tif (signals.trigger === \"complex\") {\n\t\tif (signals.toolCallCount >= 3) {\n\t\t\treturn { act: \"reflect\", reason: `Complex turn with ${signals.toolCallCount} tool calls`, tokenBudget };\n\t\t}\n\t}\n\n\treturn { act: \"skip\", reason: \"Signals do not justify reflection overhead\", tokenBudget: 0 };\n}\n\nexport interface ReflectionInput {\n\trecentTurnText: string; // host serializes the just-finished turn\n\texistingMemory: string; // current MEMORY.md + USER.md snapshot\n\tplan: DemandPlan;\n\t// host-injected isolated completion function:\n\tcomplete: (systemPrompt: string, userPrompt: string) => Promise<IsolatedCompletionResult>;\n}\n\nexport type ReflectionWrite =\n\t| { kind: \"memory_add\"; section: \"MEMORY\" | \"USER\"; text: string }\n\t| { kind: \"memory_replace\"; target: string; text: string }\n\t| { kind: \"memory_remove\"; target: string }\n\t// R7 memory-to-behavior: promote a recurring procedural workflow into an executable skill.\n\t| { kind: \"promote_skill\"; name: string; description: string; body: string };\n\nexport interface ReflectionResult {\n\twrites: ReflectionWrite[];\n\tusage: Usage;\n\trationale: string;\n}\n\nexport class ReflectionEngine {\n\t/**\n\t * Build the reflection prompt, call the injected isolated complete(),\n\t * parse the response, confront existing memory, and return memory writes.\n\t * Zero direct I/O.\n\t */\n\tasync reflect(input: ReflectionInput): Promise<ReflectionResult> {\n\t\tconst systemPrompt = `You are a reflection engine. Your job is to analyze the recent conversation turn, compare it against the agent's existing memory, and decide if any memory updates are needed.\n\nExisting Memory snapshot:\n${input.existingMemory}\n\nMemory guidelines:\n- \"MEMORY\" is for project facts, configuration, repeatable workflows, and coding findings.\n- \"USER\" is for user preferences, patterns, and style specifications.\n- Avoid duplicate facts. If the fact is already represented, do not add it.\n- CONFRONT existing memory: if the new turn contradicts or updates an existing fact, use \"memory_replace\" or \"memory_remove\" to supersede the old fact rather than blindly appending.\n- Keep memories short, factual, and direct. No fluff.\n- PROMOTE to behavior: if the turn established a REPEATABLE, multi-step PROCEDURE/workflow (not a one-off fact) that should govern a future class of tasks, emit a \"promote_skill\" instead of (or in addition to) a memory fact. Only promote a genuinely reusable procedure — never a single fact, a one-off narrative, or environment-specific noise. Prefer a memory fact when unsure.\n\nYou must output your analysis and writes in the following JSON format inside a \\`\\`\\`json\\`\\`\\` code fence:\n{\n  \"rationale\": \"Explanation of your reasoning\",\n  \"writes\": [\n    { \"kind\": \"memory_add\", \"section\": \"MEMORY\" | \"USER\", \"text\": \"New direct fact to append\" },\n    { \"kind\": \"memory_replace\", \"target\": \"Exact text substring to replace\", \"text\": \"New replacement text\" },\n    { \"kind\": \"memory_remove\", \"target\": \"Exact text substring to remove\" },\n    { \"kind\": \"promote_skill\", \"name\": \"kebab-case-skill-name\", \"description\": \"one line of when to use it\", \"body\": \"Markdown: the step-by-step procedure\" }\n  ]\n}\n`;\n\n\t\tconst userPrompt = `Recent turn transcript:\n${input.recentTurnText}\n\nAnalyze this turn and output your memory updates.`;\n\n\t\ttry {\n\t\t\tconst compResult = await input.complete(systemPrompt, userPrompt);\n\t\t\tconst text = compResult.text;\n\n\t\t\tconst jsonMatch = text.match(/```json\\s*([\\s\\S]*?)\\s*```/) || text.match(/{[\\s\\S]*}/);\n\t\t\tif (!jsonMatch) {\n\t\t\t\treturn {\n\t\t\t\t\twrites: [],\n\t\t\t\t\tusage: compResult.usage,\n\t\t\t\t\trationale: `Failed to locate JSON response. Raw text:\\n${text}`,\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tconst parsed = JSON.parse(jsonMatch[1] || jsonMatch[0]);\n\t\t\tconst rationale = parsed.rationale || \"\";\n\t\t\tconst writes: ReflectionWrite[] = [];\n\n\t\t\tif (Array.isArray(parsed.writes)) {\n\t\t\t\tfor (const w of parsed.writes) {\n\t\t\t\t\tif (w && typeof w === \"object\") {\n\t\t\t\t\t\tif (\n\t\t\t\t\t\t\tw.kind === \"memory_add\" &&\n\t\t\t\t\t\t\t(w.section === \"MEMORY\" || w.section === \"USER\") &&\n\t\t\t\t\t\t\ttypeof w.text === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_add\", section: w.section, text: w.text });\n\t\t\t\t\t\t} else if (\n\t\t\t\t\t\t\tw.kind === \"memory_replace\" &&\n\t\t\t\t\t\t\ttypeof w.target === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.text === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_replace\", target: w.target, text: w.text });\n\t\t\t\t\t\t} else if (w.kind === \"memory_remove\" && typeof w.target === \"string\") {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_remove\", target: w.target });\n\t\t\t\t\t\t} else if (\n\t\t\t\t\t\t\tw.kind === \"promote_skill\" &&\n\t\t\t\t\t\t\ttypeof w.name === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.description === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.body === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"promote_skill\", name: w.name, description: w.description, body: w.body });\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\treturn {\n\t\t\t\twrites,\n\t\t\t\tusage: compResult.usage,\n\t\t\t\trationale,\n\t\t\t};\n\t\t} catch (err) {\n\t\t\t// Zeroed/fallback usage representation\n\t\t\tconst emptyUsage: Usage = {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t};\n\t\t\treturn {\n\t\t\t\twrites: [],\n\t\t\t\tusage: emptyUsage,\n\t\t\t\trationale: `Error during reflection: ${String(err)}`,\n\t\t\t};\n\t\t}\n\t}\n}\n"]}
+{"version":3,"file":"reflection-engine.d.ts","sourceRoot":"","sources":["../../../src/core/learning/reflection-engine.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,mBAAmB,CAAC;AAE/C,MAAM,MAAM,UAAU,GAAG,MAAM,GAAG,SAAS,GAAG,SAAS,GAAG,OAAO,GAAG,MAAM,CAAC;AAE3E,MAAM,WAAW,wBAAwB;IACxC,IAAI,EAAE,MAAM,CAAC;IACb,KAAK,EAAE,KAAK,CAAC;IACb,UAAU,EAAE,UAAU,CAAC;CACvB;AAED,MAAM,MAAM,iBAAiB,GAAG,SAAS,GAAG,YAAY,GAAG,aAAa,GAAG,MAAM,CAAC;AAElF,MAAM,WAAW,aAAa;IAC7B,OAAO,EAAE,iBAAiB,CAAC;IAC3B,aAAa,EAAE,MAAM,CAAC;IACtB,aAAa,EAAE,OAAO,CAAC;IACvB,kBAAkB,EAAE,MAAM,CAAC;IAC3B,YAAY,EAAE,MAAM,CAAC;CACrB;AAED,MAAM,WAAW,UAAU;IAC1B,GAAG,EAAE,MAAM,GAAG,SAAS,CAAC;IACxB,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;CACpB;AAED;;;GAGG;AACH,wBAAgB,YAAY,CAAC,OAAO,EAAE,aAAa,GAAG,UAAU,CAyB/D;AAED,MAAM,WAAW,eAAe;IAC/B,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,CAAC;IACvB,IAAI,EAAE,UAAU,CAAC;IAEjB,QAAQ,EAAE,CAAC,YAAY,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,KAAK,OAAO,CAAC,wBAAwB,CAAC,CAAC;CAC1F;AAED,MAAM,MAAM,eAAe,GACxB;IAAE,IAAI,EAAE,YAAY,CAAC;IAAC,OAAO,EAAE,QAAQ,GAAG,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GAChE;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,GACxD;IAAE,IAAI,EAAE,eAAe,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,GAEzC;IAAE,IAAI,EAAE,eAAe,CAAC;IAAC,IAAI,EAAE,MAAM,CAAC;IAAC,WAAW,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAAC;AAE9E,MAAM,WAAW,gBAAgB;IAChC,MAAM,EAAE,eAAe,EAAE,CAAC;IAC1B,KAAK,EAAE,KAAK,CAAC;IACb,SAAS,EAAE,MAAM,CAAC;CAClB;AAED;;;;;GAKG;AACH,eAAO,MAAM,wBAAwB,s4DAqBpC,CAAC;AAEF,qBAAa,gBAAgB;IAC5B;;;;OAIG;IACG,OAAO,CAAC,KAAK,EAAE,eAAe,GAAG,OAAO,CAAC,gBAAgB,CAAC,CA+E/D;CACD","sourcesContent":["import type { Usage } from \"@caupulican/pi-ai\";\n\nexport type StopReason = \"stop\" | \"toolUse\" | \"aborted\" | \"error\" | string;\n\nexport interface IsolatedCompletionResult {\n\ttext: string;\n\tusage: Usage;\n\tstopReason: StopReason;\n}\n\nexport type ReflectionTrigger = \"complex\" | \"corrective\" | \"session-end\" | \"none\";\n\nexport interface DemandSignals {\n\ttrigger: ReflectionTrigger;\n\ttoolCallCount: number;\n\thadCorrection: boolean;\n\tcontextHeadroomPct: number; // 0..100\n\tusefulLately: number; // 0..1 rolling score\n}\n\nexport interface DemandPlan {\n\tact: \"skip\" | \"reflect\";\n\treason: string;\n\ttokenBudget: number;\n}\n\n/**\n * Pure zero-I/O heuristic to decide whether the current turn justifies a reflection run\n * and determine the token budget under the cheap-tool net-negative doctrine.\n */\nexport function decideDemand(signals: DemandSignals): DemandPlan {\n\tif (signals.trigger === \"none\") {\n\t\treturn { act: \"skip\", reason: \"No trigger detected\", tokenBudget: 0 };\n\t}\n\tif (signals.contextHeadroomPct < 10) {\n\t\treturn { act: \"skip\", reason: \"Context headroom is critically low (< 10%)\", tokenBudget: 0 };\n\t}\n\n\t// Dynamic token budget based on headroom (keep reflection bounded between 500 and 1500 tokens)\n\tconst baseBudget = 1000;\n\tconst tokenBudget = Math.max(500, Math.min(1500, Math.round(baseBudget * (signals.contextHeadroomPct / 100))));\n\n\tif (signals.hadCorrection) {\n\t\treturn { act: \"reflect\", reason: \"Correction detected in the turn\", tokenBudget };\n\t}\n\tif (signals.trigger === \"session-end\") {\n\t\treturn { act: \"reflect\", reason: \"Session end reflection triggered\", tokenBudget };\n\t}\n\tif (signals.trigger === \"complex\") {\n\t\tif (signals.toolCallCount >= 3) {\n\t\t\treturn { act: \"reflect\", reason: `Complex turn with ${signals.toolCallCount} tool calls`, tokenBudget };\n\t\t}\n\t}\n\n\treturn { act: \"skip\", reason: \"Signals do not justify reflection overhead\", tokenBudget: 0 };\n}\n\nexport interface ReflectionInput {\n\trecentTurnText: string; // host serializes the just-finished turn\n\texistingMemory: string; // current MEMORY.md + USER.md snapshot\n\tplan: DemandPlan;\n\t// host-injected isolated completion function:\n\tcomplete: (systemPrompt: string, userPrompt: string) => Promise<IsolatedCompletionResult>;\n}\n\nexport type ReflectionWrite =\n\t| { kind: \"memory_add\"; section: \"MEMORY\" | \"USER\"; text: string }\n\t| { kind: \"memory_replace\"; target: string; text: string }\n\t| { kind: \"memory_remove\"; target: string }\n\t// R7 memory-to-behavior: promote a recurring procedural workflow into an executable skill.\n\t| { kind: \"promote_skill\"; name: string; description: string; body: string };\n\nexport interface ReflectionResult {\n\twrites: ReflectionWrite[];\n\tusage: Usage;\n\trationale: string;\n}\n\n/**\n * STATIC reflection system prompt (Hermes-parity #33). It is byte-identical across every reflection\n * pass — the variable parts (existing memory snapshot + the turn transcript) live in the USER prompt —\n * so the provider prompt-cache reuses this prefix instead of re-billing it each pass (cost guard).\n * Do NOT interpolate per-call data into this constant or caching breaks.\n */\nexport const REFLECTION_SYSTEM_PROMPT = `You are a reflection engine. Your job is to analyze the recent conversation turn, compare it against the agent's existing memory, and decide if any memory updates are needed.\n\nMemory guidelines:\n- \"MEMORY\" is for project facts, configuration, repeatable workflows, and coding findings.\n- \"USER\" is for user preferences, patterns, and style specifications.\n- Avoid duplicate facts. If the fact is already represented, do not add it.\n- CONFRONT existing memory: if the new turn contradicts or updates an existing fact, use \"memory_replace\" or \"memory_remove\" to supersede the old fact rather than blindly appending.\n- Keep memories short, factual, and direct. No fluff.\n- Do NOT capture transient/environment-specific noise: tool/network failures, one-off errors, or a single narrative event. Persist only durable facts and preferences.\n- PROMOTE to behavior: if the turn established a REPEATABLE, multi-step PROCEDURE/workflow (not a one-off fact) that should govern a future class of tasks, emit a \"promote_skill\" instead of (or in addition to) a memory fact. Only promote a genuinely reusable procedure — never a single fact, a one-off narrative, or environment-specific noise. Prefer a memory fact when unsure.\n\nYou must output your analysis and writes in the following JSON format inside a \\`\\`\\`json\\`\\`\\` code fence:\n{\n  \"rationale\": \"Explanation of your reasoning\",\n  \"writes\": [\n    { \"kind\": \"memory_add\", \"section\": \"MEMORY\" | \"USER\", \"text\": \"New direct fact to append\" },\n    { \"kind\": \"memory_replace\", \"target\": \"Exact text substring to replace\", \"text\": \"New replacement text\" },\n    { \"kind\": \"memory_remove\", \"target\": \"Exact text substring to remove\" },\n    { \"kind\": \"promote_skill\", \"name\": \"kebab-case-skill-name\", \"description\": \"one line of when to use it\", \"body\": \"Markdown: the step-by-step procedure\" }\n  ]\n}\n`;\n\nexport class ReflectionEngine {\n\t/**\n\t * Build the reflection prompt, call the injected isolated complete(),\n\t * parse the response, confront existing memory, and return memory writes.\n\t * Zero direct I/O.\n\t */\n\tasync reflect(input: ReflectionInput): Promise<ReflectionResult> {\n\t\tconst systemPrompt = REFLECTION_SYSTEM_PROMPT;\n\n\t\t// Variable inputs go in the USER prompt so the system prefix above stays cache-stable (#33).\n\t\tconst userPrompt = `Existing Memory snapshot:\n${input.existingMemory}\n\nRecent turn transcript:\n${input.recentTurnText}\n\nAnalyze this turn against the existing memory and output your memory updates.`;\n\n\t\ttry {\n\t\t\tconst compResult = await input.complete(systemPrompt, userPrompt);\n\t\t\tconst text = compResult.text;\n\n\t\t\tconst jsonMatch = text.match(/```json\\s*([\\s\\S]*?)\\s*```/) || text.match(/{[\\s\\S]*}/);\n\t\t\tif (!jsonMatch) {\n\t\t\t\treturn {\n\t\t\t\t\twrites: [],\n\t\t\t\t\tusage: compResult.usage,\n\t\t\t\t\trationale: `Failed to locate JSON response. Raw text:\\n${text}`,\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tconst parsed = JSON.parse(jsonMatch[1] || jsonMatch[0]);\n\t\t\tconst rationale = parsed.rationale || \"\";\n\t\t\tconst writes: ReflectionWrite[] = [];\n\n\t\t\tif (Array.isArray(parsed.writes)) {\n\t\t\t\tfor (const w of parsed.writes) {\n\t\t\t\t\tif (w && typeof w === \"object\") {\n\t\t\t\t\t\tif (\n\t\t\t\t\t\t\tw.kind === \"memory_add\" &&\n\t\t\t\t\t\t\t(w.section === \"MEMORY\" || w.section === \"USER\") &&\n\t\t\t\t\t\t\ttypeof w.text === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_add\", section: w.section, text: w.text });\n\t\t\t\t\t\t} else if (\n\t\t\t\t\t\t\tw.kind === \"memory_replace\" &&\n\t\t\t\t\t\t\ttypeof w.target === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.text === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_replace\", target: w.target, text: w.text });\n\t\t\t\t\t\t} else if (w.kind === \"memory_remove\" && typeof w.target === \"string\") {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_remove\", target: w.target });\n\t\t\t\t\t\t} else if (\n\t\t\t\t\t\t\tw.kind === \"promote_skill\" &&\n\t\t\t\t\t\t\ttypeof w.name === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.description === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.body === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"promote_skill\", name: w.name, description: w.description, body: w.body });\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\treturn {\n\t\t\t\twrites,\n\t\t\t\tusage: compResult.usage,\n\t\t\t\trationale,\n\t\t\t};\n\t\t} catch (err) {\n\t\t\t// Zeroed/fallback usage representation\n\t\t\tconst emptyUsage: Usage = {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t};\n\t\t\treturn {\n\t\t\t\twrites: [],\n\t\t\t\tusage: emptyUsage,\n\t\t\t\trationale: `Error during reflection: ${String(err)}`,\n\t\t\t};\n\t\t}\n\t}\n}\n"]}

package/dist/core/learning/reflection-engine.js

@@ -25,17 +25,13 @@ export function decideDemand(signals) {
     }
     return { act: "skip", reason: "Signals do not justify reflection overhead", tokenBudget: 0 };
 }
-export class ReflectionEngine {
-    /**
-     * Build the reflection prompt, call the injected isolated complete(),
-     * parse the response, confront existing memory, and return memory writes.
-     * Zero direct I/O.
-     */
-    async reflect(input) {
-        const systemPrompt = `You are a reflection engine. Your job is to analyze the recent conversation turn, compare it against the agent's existing memory, and decide if any memory updates are needed.
-
-Existing Memory snapshot:
-${input.existingMemory}
+/**
+ * STATIC reflection system prompt (Hermes-parity #33). It is byte-identical across every reflection
+ * pass — the variable parts (existing memory snapshot + the turn transcript) live in the USER prompt —
+ * so the provider prompt-cache reuses this prefix instead of re-billing it each pass (cost guard).
+ * Do NOT interpolate per-call data into this constant or caching breaks.
+ */
+export const REFLECTION_SYSTEM_PROMPT = `You are a reflection engine. Your job is to analyze the recent conversation turn, compare it against the agent's existing memory, and decide if any memory updates are needed.
 
 Memory guidelines:
 - "MEMORY" is for project facts, configuration, repeatable workflows, and coding findings.
@@ -43,6 +39,7 @@ Memory guidelines:
 - Avoid duplicate facts. If the fact is already represented, do not add it.
 - CONFRONT existing memory: if the new turn contradicts or updates an existing fact, use "memory_replace" or "memory_remove" to supersede the old fact rather than blindly appending.
 - Keep memories short, factual, and direct. No fluff.
+- Do NOT capture transient/environment-specific noise: tool/network failures, one-off errors, or a single narrative event. Persist only durable facts and preferences.
 - PROMOTE to behavior: if the turn established a REPEATABLE, multi-step PROCEDURE/workflow (not a one-off fact) that should govern a future class of tasks, emit a "promote_skill" instead of (or in addition to) a memory fact. Only promote a genuinely reusable procedure — never a single fact, a one-off narrative, or environment-specific noise. Prefer a memory fact when unsure.
 
 You must output your analysis and writes in the following JSON format inside a \`\`\`json\`\`\` code fence:
@@ -56,10 +53,22 @@ You must output your analysis and writes in the following JSON format inside a \
   ]
 }
 `;
-        const userPrompt = `Recent turn transcript:
+export class ReflectionEngine {
+    /**
+     * Build the reflection prompt, call the injected isolated complete(),
+     * parse the response, confront existing memory, and return memory writes.
+     * Zero direct I/O.
+     */
+    async reflect(input) {
+        const systemPrompt = REFLECTION_SYSTEM_PROMPT;
+        // Variable inputs go in the USER prompt so the system prefix above stays cache-stable (#33).
+        const userPrompt = `Existing Memory snapshot:
+${input.existingMemory}
+
+Recent turn transcript:
 ${input.recentTurnText}
 
-Analyze this turn and output your memory updates.`;
+Analyze this turn against the existing memory and output your memory updates.`;
         try {
             const compResult = await input.complete(systemPrompt, userPrompt);
             const text = compResult.text;

package/dist/core/learning/reflection-engine.js.map

@@ -1 +1 @@
-{"version":3,"file":"reflection-engine.js","sourceRoot":"","sources":["../../../src/core/learning/reflection-engine.ts"],"names":[],"mappings":"AA0BA;;;GAGG;AACH,MAAM,UAAU,YAAY,CAAC,OAAsB,EAAc;IAChE,IAAI,OAAO,CAAC,OAAO,KAAK,MAAM,EAAE,CAAC;QAChC,OAAO,EAAE,GAAG,EAAE,MAAM,EAAE,MAAM,EAAE,qBAAqB,EAAE,WAAW,EAAE,CAAC,EAAE,CAAC;IACvE,CAAC;IACD,IAAI,OAAO,CAAC,kBAAkB,GAAG,EAAE,EAAE,CAAC;QACrC,OAAO,EAAE,GAAG,EAAE,MAAM,EAAE,MAAM,EAAE,4CAA4C,EAAE,WAAW,EAAE,CAAC,EAAE,CAAC;IAC9F,CAAC;IAED,+FAA+F;IAC/F,MAAM,UAAU,GAAG,IAAI,CAAC;IACxB,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,IAAI,CAAC,KAAK,CAAC,UAAU,GAAG,CAAC,OAAO,CAAC,kBAAkB,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;IAE/G,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC;QAC3B,OAAO,EAAE,GAAG,EAAE,SAAS,EAAE,MAAM,EAAE,iCAAiC,EAAE,WAAW,EAAE,CAAC;IACnF,CAAC;IACD,IAAI,OAAO,CAAC,OAAO,KAAK,aAAa,EAAE,CAAC;QACvC,OAAO,EAAE,GAAG,EAAE,SAAS,EAAE,MAAM,EAAE,kCAAkC,EAAE,WAAW,EAAE,CAAC;IACpF,CAAC;IACD,IAAI,OAAO,CAAC,OAAO,KAAK,SAAS,EAAE,CAAC;QACnC,IAAI,OAAO,CAAC,aAAa,IAAI,CAAC,EAAE,CAAC;YAChC,OAAO,EAAE,GAAG,EAAE,SAAS,EAAE,MAAM,EAAE,qBAAqB,OAAO,CAAC,aAAa,aAAa,EAAE,WAAW,EAAE,CAAC;QACzG,CAAC;IACF,CAAC;IAED,OAAO,EAAE,GAAG,EAAE,MAAM,EAAE,MAAM,EAAE,4CAA4C,EAAE,WAAW,EAAE,CAAC,EAAE,CAAC;AAAA,CAC7F;AAuBD,MAAM,OAAO,gBAAgB;IAC5B;;;;OAIG;IACH,KAAK,CAAC,OAAO,CAAC,KAAsB,EAA6B;QAChE,MAAM,YAAY,GAAG;;;EAGrB,KAAK,CAAC,cAAc;;;;;;;;;;;;;;;;;;;;CAoBrB,CAAC;QAEA,MAAM,UAAU,GAAG;EACnB,KAAK,CAAC,cAAc;;kDAE4B,CAAC;QAEjD,IAAI,CAAC;YACJ,MAAM,UAAU,GAAG,MAAM,KAAK,CAAC,QAAQ,CAAC,YAAY,EAAE,UAAU,CAAC,CAAC;YAClE,MAAM,IAAI,GAAG,UAAU,CAAC,IAAI,CAAC;YAE7B,MAAM,SAAS,GAAG,IAAI,CAAC,KAAK,CAAC,4BAA4B,CAAC,IAAI,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;YACtF,IAAI,CAAC,SAAS,EAAE,CAAC;gBAChB,OAAO;oBACN,MAAM,EAAE,EAAE;oBACV,KAAK,EAAE,UAAU,CAAC,KAAK;oBACvB,SAAS,EAAE,8CAA8C,IAAI,EAAE;iBAC/D,CAAC;YACH,CAAC;YAED,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;YACxD,MAAM,SAAS,GAAG,MAAM,CAAC,SAAS,IAAI,EAAE,CAAC;YACzC,MAAM,MAAM,GAAsB,EAAE,CAAC;YAErC,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,CAAC;gBAClC,KAAK,MAAM,CAAC,IAAI,MAAM,CAAC,MAAM,EAAE,CAAC;oBAC/B,IAAI,CAAC,IAAI,OAAO,CAAC,KAAK,QAAQ,EAAE,CAAC;wBAChC,IACC,CAAC,CAAC,IAAI,KAAK,YAAY;4BACvB,CAAC,CAAC,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,CAAC,OAAO,KAAK,MAAM,CAAC;4BAChD,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,EACzB,CAAC;4BACF,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,YAAY,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;wBACvE,CAAC;6BAAM,IACN,CAAC,CAAC,IAAI,KAAK,gBAAgB;4BAC3B,OAAO,CAAC,CAAC,MAAM,KAAK,QAAQ;4BAC5B,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,EACzB,CAAC;4BACF,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;wBACzE,CAAC;6BAAM,IAAI,CAAC,CAAC,IAAI,KAAK,eAAe,IAAI,OAAO,CAAC,CAAC,MAAM,KAAK,QAAQ,EAAE,CAAC;4BACvE,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,eAAe,EAAE,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC;wBAC1D,CAAC;6BAAM,IACN,CAAC,CAAC,IAAI,KAAK,eAAe;4BAC1B,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ;4BAC1B,OAAO,CAAC,CAAC,WAAW,KAAK,QAAQ;4BACjC,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,EACzB,CAAC;4BACF,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,eAAe,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,WAAW,EAAE,CAAC,CAAC,WAAW,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;wBAChG,CAAC;oBACF,CAAC;gBACF,CAAC;YACF,CAAC;YAED,OAAO;gBACN,MAAM;gBACN,KAAK,EAAE,UAAU,CAAC,KAAK;gBACvB,SAAS;aACT,CAAC;QACH,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACd,uCAAuC;YACvC,MAAM,UAAU,GAAU;gBACzB,KAAK,EAAE,CAAC;gBACR,MAAM,EAAE,CAAC;gBACT,SAAS,EAAE,CAAC;gBACZ,UAAU,EAAE,CAAC;gBACb,WAAW,EAAE,CAAC;gBACd,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE;aACpE,CAAC;YACF,OAAO;gBACN,MAAM,EAAE,EAAE;gBACV,KAAK,EAAE,UAAU;gBACjB,SAAS,EAAE,4BAA4B,MAAM,CAAC,GAAG,CAAC,EAAE;aACpD,CAAC;QACH,CAAC;IAAA,CACD;CACD","sourcesContent":["import type { Usage } from \"@caupulican/pi-ai\";\n\nexport type StopReason = \"stop\" | \"toolUse\" | \"aborted\" | \"error\" | string;\n\nexport interface IsolatedCompletionResult {\n\ttext: string;\n\tusage: Usage;\n\tstopReason: StopReason;\n}\n\nexport type ReflectionTrigger = \"complex\" | \"corrective\" | \"session-end\" | \"none\";\n\nexport interface DemandSignals {\n\ttrigger: ReflectionTrigger;\n\ttoolCallCount: number;\n\thadCorrection: boolean;\n\tcontextHeadroomPct: number; // 0..100\n\tusefulLately: number; // 0..1 rolling score\n}\n\nexport interface DemandPlan {\n\tact: \"skip\" | \"reflect\";\n\treason: string;\n\ttokenBudget: number;\n}\n\n/**\n * Pure zero-I/O heuristic to decide whether the current turn justifies a reflection run\n * and determine the token budget under the cheap-tool net-negative doctrine.\n */\nexport function decideDemand(signals: DemandSignals): DemandPlan {\n\tif (signals.trigger === \"none\") {\n\t\treturn { act: \"skip\", reason: \"No trigger detected\", tokenBudget: 0 };\n\t}\n\tif (signals.contextHeadroomPct < 10) {\n\t\treturn { act: \"skip\", reason: \"Context headroom is critically low (< 10%)\", tokenBudget: 0 };\n\t}\n\n\t// Dynamic token budget based on headroom (keep reflection bounded between 500 and 1500 tokens)\n\tconst baseBudget = 1000;\n\tconst tokenBudget = Math.max(500, Math.min(1500, Math.round(baseBudget * (signals.contextHeadroomPct / 100))));\n\n\tif (signals.hadCorrection) {\n\t\treturn { act: \"reflect\", reason: \"Correction detected in the turn\", tokenBudget };\n\t}\n\tif (signals.trigger === \"session-end\") {\n\t\treturn { act: \"reflect\", reason: \"Session end reflection triggered\", tokenBudget };\n\t}\n\tif (signals.trigger === \"complex\") {\n\t\tif (signals.toolCallCount >= 3) {\n\t\t\treturn { act: \"reflect\", reason: `Complex turn with ${signals.toolCallCount} tool calls`, tokenBudget };\n\t\t}\n\t}\n\n\treturn { act: \"skip\", reason: \"Signals do not justify reflection overhead\", tokenBudget: 0 };\n}\n\nexport interface ReflectionInput {\n\trecentTurnText: string; // host serializes the just-finished turn\n\texistingMemory: string; // current MEMORY.md + USER.md snapshot\n\tplan: DemandPlan;\n\t// host-injected isolated completion function:\n\tcomplete: (systemPrompt: string, userPrompt: string) => Promise<IsolatedCompletionResult>;\n}\n\nexport type ReflectionWrite =\n\t| { kind: \"memory_add\"; section: \"MEMORY\" | \"USER\"; text: string }\n\t| { kind: \"memory_replace\"; target: string; text: string }\n\t| { kind: \"memory_remove\"; target: string }\n\t// R7 memory-to-behavior: promote a recurring procedural workflow into an executable skill.\n\t| { kind: \"promote_skill\"; name: string; description: string; body: string };\n\nexport interface ReflectionResult {\n\twrites: ReflectionWrite[];\n\tusage: Usage;\n\trationale: string;\n}\n\nexport class ReflectionEngine {\n\t/**\n\t * Build the reflection prompt, call the injected isolated complete(),\n\t * parse the response, confront existing memory, and return memory writes.\n\t * Zero direct I/O.\n\t */\n\tasync reflect(input: ReflectionInput): Promise<ReflectionResult> {\n\t\tconst systemPrompt = `You are a reflection engine. Your job is to analyze the recent conversation turn, compare it against the agent's existing memory, and decide if any memory updates are needed.\n\nExisting Memory snapshot:\n${input.existingMemory}\n\nMemory guidelines:\n- \"MEMORY\" is for project facts, configuration, repeatable workflows, and coding findings.\n- \"USER\" is for user preferences, patterns, and style specifications.\n- Avoid duplicate facts. If the fact is already represented, do not add it.\n- CONFRONT existing memory: if the new turn contradicts or updates an existing fact, use \"memory_replace\" or \"memory_remove\" to supersede the old fact rather than blindly appending.\n- Keep memories short, factual, and direct. No fluff.\n- PROMOTE to behavior: if the turn established a REPEATABLE, multi-step PROCEDURE/workflow (not a one-off fact) that should govern a future class of tasks, emit a \"promote_skill\" instead of (or in addition to) a memory fact. Only promote a genuinely reusable procedure — never a single fact, a one-off narrative, or environment-specific noise. Prefer a memory fact when unsure.\n\nYou must output your analysis and writes in the following JSON format inside a \\`\\`\\`json\\`\\`\\` code fence:\n{\n  \"rationale\": \"Explanation of your reasoning\",\n  \"writes\": [\n    { \"kind\": \"memory_add\", \"section\": \"MEMORY\" | \"USER\", \"text\": \"New direct fact to append\" },\n    { \"kind\": \"memory_replace\", \"target\": \"Exact text substring to replace\", \"text\": \"New replacement text\" },\n    { \"kind\": \"memory_remove\", \"target\": \"Exact text substring to remove\" },\n    { \"kind\": \"promote_skill\", \"name\": \"kebab-case-skill-name\", \"description\": \"one line of when to use it\", \"body\": \"Markdown: the step-by-step procedure\" }\n  ]\n}\n`;\n\n\t\tconst userPrompt = `Recent turn transcript:\n${input.recentTurnText}\n\nAnalyze this turn and output your memory updates.`;\n\n\t\ttry {\n\t\t\tconst compResult = await input.complete(systemPrompt, userPrompt);\n\t\t\tconst text = compResult.text;\n\n\t\t\tconst jsonMatch = text.match(/```json\\s*([\\s\\S]*?)\\s*```/) || text.match(/{[\\s\\S]*}/);\n\t\t\tif (!jsonMatch) {\n\t\t\t\treturn {\n\t\t\t\t\twrites: [],\n\t\t\t\t\tusage: compResult.usage,\n\t\t\t\t\trationale: `Failed to locate JSON response. Raw text:\\n${text}`,\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tconst parsed = JSON.parse(jsonMatch[1] || jsonMatch[0]);\n\t\t\tconst rationale = parsed.rationale || \"\";\n\t\t\tconst writes: ReflectionWrite[] = [];\n\n\t\t\tif (Array.isArray(parsed.writes)) {\n\t\t\t\tfor (const w of parsed.writes) {\n\t\t\t\t\tif (w && typeof w === \"object\") {\n\t\t\t\t\t\tif (\n\t\t\t\t\t\t\tw.kind === \"memory_add\" &&\n\t\t\t\t\t\t\t(w.section === \"MEMORY\" || w.section === \"USER\") &&\n\t\t\t\t\t\t\ttypeof w.text === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_add\", section: w.section, text: w.text });\n\t\t\t\t\t\t} else if (\n\t\t\t\t\t\t\tw.kind === \"memory_replace\" &&\n\t\t\t\t\t\t\ttypeof w.target === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.text === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_replace\", target: w.target, text: w.text });\n\t\t\t\t\t\t} else if (w.kind === \"memory_remove\" && typeof w.target === \"string\") {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_remove\", target: w.target });\n\t\t\t\t\t\t} else if (\n\t\t\t\t\t\t\tw.kind === \"promote_skill\" &&\n\t\t\t\t\t\t\ttypeof w.name === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.description === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.body === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"promote_skill\", name: w.name, description: w.description, body: w.body });\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\treturn {\n\t\t\t\twrites,\n\t\t\t\tusage: compResult.usage,\n\t\t\t\trationale,\n\t\t\t};\n\t\t} catch (err) {\n\t\t\t// Zeroed/fallback usage representation\n\t\t\tconst emptyUsage: Usage = {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t};\n\t\t\treturn {\n\t\t\t\twrites: [],\n\t\t\t\tusage: emptyUsage,\n\t\t\t\trationale: `Error during reflection: ${String(err)}`,\n\t\t\t};\n\t\t}\n\t}\n}\n"]}
+{"version":3,"file":"reflection-engine.js","sourceRoot":"","sources":["../../../src/core/learning/reflection-engine.ts"],"names":[],"mappings":"AA0BA;;;GAGG;AACH,MAAM,UAAU,YAAY,CAAC,OAAsB,EAAc;IAChE,IAAI,OAAO,CAAC,OAAO,KAAK,MAAM,EAAE,CAAC;QAChC,OAAO,EAAE,GAAG,EAAE,MAAM,EAAE,MAAM,EAAE,qBAAqB,EAAE,WAAW,EAAE,CAAC,EAAE,CAAC;IACvE,CAAC;IACD,IAAI,OAAO,CAAC,kBAAkB,GAAG,EAAE,EAAE,CAAC;QACrC,OAAO,EAAE,GAAG,EAAE,MAAM,EAAE,MAAM,EAAE,4CAA4C,EAAE,WAAW,EAAE,CAAC,EAAE,CAAC;IAC9F,CAAC;IAED,+FAA+F;IAC/F,MAAM,UAAU,GAAG,IAAI,CAAC;IACxB,MAAM,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,IAAI,CAAC,KAAK,CAAC,UAAU,GAAG,CAAC,OAAO,CAAC,kBAAkB,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;IAE/G,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC;QAC3B,OAAO,EAAE,GAAG,EAAE,SAAS,EAAE,MAAM,EAAE,iCAAiC,EAAE,WAAW,EAAE,CAAC;IACnF,CAAC;IACD,IAAI,OAAO,CAAC,OAAO,KAAK,aAAa,EAAE,CAAC;QACvC,OAAO,EAAE,GAAG,EAAE,SAAS,EAAE,MAAM,EAAE,kCAAkC,EAAE,WAAW,EAAE,CAAC;IACpF,CAAC;IACD,IAAI,OAAO,CAAC,OAAO,KAAK,SAAS,EAAE,CAAC;QACnC,IAAI,OAAO,CAAC,aAAa,IAAI,CAAC,EAAE,CAAC;YAChC,OAAO,EAAE,GAAG,EAAE,SAAS,EAAE,MAAM,EAAE,qBAAqB,OAAO,CAAC,aAAa,aAAa,EAAE,WAAW,EAAE,CAAC;QACzG,CAAC;IACF,CAAC;IAED,OAAO,EAAE,GAAG,EAAE,MAAM,EAAE,MAAM,EAAE,4CAA4C,EAAE,WAAW,EAAE,CAAC,EAAE,CAAC;AAAA,CAC7F;AAuBD;;;;;GAKG;AACH,MAAM,CAAC,MAAM,wBAAwB,GAAG;;;;;;;;;;;;;;;;;;;;;CAqBvC,CAAC;AAEF,MAAM,OAAO,gBAAgB;IAC5B;;;;OAIG;IACH,KAAK,CAAC,OAAO,CAAC,KAAsB,EAA6B;QAChE,MAAM,YAAY,GAAG,wBAAwB,CAAC;QAE9C,6FAA6F;QAC7F,MAAM,UAAU,GAAG;EACnB,KAAK,CAAC,cAAc;;;EAGpB,KAAK,CAAC,cAAc;;8EAEwD,CAAC;QAE7E,IAAI,CAAC;YACJ,MAAM,UAAU,GAAG,MAAM,KAAK,CAAC,QAAQ,CAAC,YAAY,EAAE,UAAU,CAAC,CAAC;YAClE,MAAM,IAAI,GAAG,UAAU,CAAC,IAAI,CAAC;YAE7B,MAAM,SAAS,GAAG,IAAI,CAAC,KAAK,CAAC,4BAA4B,CAAC,IAAI,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;YACtF,IAAI,CAAC,SAAS,EAAE,CAAC;gBAChB,OAAO;oBACN,MAAM,EAAE,EAAE;oBACV,KAAK,EAAE,UAAU,CAAC,KAAK;oBACvB,SAAS,EAAE,8CAA8C,IAAI,EAAE;iBAC/D,CAAC;YACH,CAAC;YAED,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,IAAI,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;YACxD,MAAM,SAAS,GAAG,MAAM,CAAC,SAAS,IAAI,EAAE,CAAC;YACzC,MAAM,MAAM,GAAsB,EAAE,CAAC;YAErC,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC,EAAE,CAAC;gBAClC,KAAK,MAAM,CAAC,IAAI,MAAM,CAAC,MAAM,EAAE,CAAC;oBAC/B,IAAI,CAAC,IAAI,OAAO,CAAC,KAAK,QAAQ,EAAE,CAAC;wBAChC,IACC,CAAC,CAAC,IAAI,KAAK,YAAY;4BACvB,CAAC,CAAC,CAAC,OAAO,KAAK,QAAQ,IAAI,CAAC,CAAC,OAAO,KAAK,MAAM,CAAC;4BAChD,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,EACzB,CAAC;4BACF,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,YAAY,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;wBACvE,CAAC;6BAAM,IACN,CAAC,CAAC,IAAI,KAAK,gBAAgB;4BAC3B,OAAO,CAAC,CAAC,MAAM,KAAK,QAAQ;4BAC5B,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,EACzB,CAAC;4BACF,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,gBAAgB,EAAE,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;wBACzE,CAAC;6BAAM,IAAI,CAAC,CAAC,IAAI,KAAK,eAAe,IAAI,OAAO,CAAC,CAAC,MAAM,KAAK,QAAQ,EAAE,CAAC;4BACvE,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,eAAe,EAAE,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,CAAC;wBAC1D,CAAC;6BAAM,IACN,CAAC,CAAC,IAAI,KAAK,eAAe;4BAC1B,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ;4BAC1B,OAAO,CAAC,CAAC,WAAW,KAAK,QAAQ;4BACjC,OAAO,CAAC,CAAC,IAAI,KAAK,QAAQ,EACzB,CAAC;4BACF,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,eAAe,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,WAAW,EAAE,CAAC,CAAC,WAAW,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC;wBAChG,CAAC;oBACF,CAAC;gBACF,CAAC;YACF,CAAC;YAED,OAAO;gBACN,MAAM;gBACN,KAAK,EAAE,UAAU,CAAC,KAAK;gBACvB,SAAS;aACT,CAAC;QACH,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACd,uCAAuC;YACvC,MAAM,UAAU,GAAU;gBACzB,KAAK,EAAE,CAAC;gBACR,MAAM,EAAE,CAAC;gBACT,SAAS,EAAE,CAAC;gBACZ,UAAU,EAAE,CAAC;gBACb,WAAW,EAAE,CAAC;gBACd,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE;aACpE,CAAC;YACF,OAAO;gBACN,MAAM,EAAE,EAAE;gBACV,KAAK,EAAE,UAAU;gBACjB,SAAS,EAAE,4BAA4B,MAAM,CAAC,GAAG,CAAC,EAAE;aACpD,CAAC;QACH,CAAC;IAAA,CACD;CACD","sourcesContent":["import type { Usage } from \"@caupulican/pi-ai\";\n\nexport type StopReason = \"stop\" | \"toolUse\" | \"aborted\" | \"error\" | string;\n\nexport interface IsolatedCompletionResult {\n\ttext: string;\n\tusage: Usage;\n\tstopReason: StopReason;\n}\n\nexport type ReflectionTrigger = \"complex\" | \"corrective\" | \"session-end\" | \"none\";\n\nexport interface DemandSignals {\n\ttrigger: ReflectionTrigger;\n\ttoolCallCount: number;\n\thadCorrection: boolean;\n\tcontextHeadroomPct: number; // 0..100\n\tusefulLately: number; // 0..1 rolling score\n}\n\nexport interface DemandPlan {\n\tact: \"skip\" | \"reflect\";\n\treason: string;\n\ttokenBudget: number;\n}\n\n/**\n * Pure zero-I/O heuristic to decide whether the current turn justifies a reflection run\n * and determine the token budget under the cheap-tool net-negative doctrine.\n */\nexport function decideDemand(signals: DemandSignals): DemandPlan {\n\tif (signals.trigger === \"none\") {\n\t\treturn { act: \"skip\", reason: \"No trigger detected\", tokenBudget: 0 };\n\t}\n\tif (signals.contextHeadroomPct < 10) {\n\t\treturn { act: \"skip\", reason: \"Context headroom is critically low (< 10%)\", tokenBudget: 0 };\n\t}\n\n\t// Dynamic token budget based on headroom (keep reflection bounded between 500 and 1500 tokens)\n\tconst baseBudget = 1000;\n\tconst tokenBudget = Math.max(500, Math.min(1500, Math.round(baseBudget * (signals.contextHeadroomPct / 100))));\n\n\tif (signals.hadCorrection) {\n\t\treturn { act: \"reflect\", reason: \"Correction detected in the turn\", tokenBudget };\n\t}\n\tif (signals.trigger === \"session-end\") {\n\t\treturn { act: \"reflect\", reason: \"Session end reflection triggered\", tokenBudget };\n\t}\n\tif (signals.trigger === \"complex\") {\n\t\tif (signals.toolCallCount >= 3) {\n\t\t\treturn { act: \"reflect\", reason: `Complex turn with ${signals.toolCallCount} tool calls`, tokenBudget };\n\t\t}\n\t}\n\n\treturn { act: \"skip\", reason: \"Signals do not justify reflection overhead\", tokenBudget: 0 };\n}\n\nexport interface ReflectionInput {\n\trecentTurnText: string; // host serializes the just-finished turn\n\texistingMemory: string; // current MEMORY.md + USER.md snapshot\n\tplan: DemandPlan;\n\t// host-injected isolated completion function:\n\tcomplete: (systemPrompt: string, userPrompt: string) => Promise<IsolatedCompletionResult>;\n}\n\nexport type ReflectionWrite =\n\t| { kind: \"memory_add\"; section: \"MEMORY\" | \"USER\"; text: string }\n\t| { kind: \"memory_replace\"; target: string; text: string }\n\t| { kind: \"memory_remove\"; target: string }\n\t// R7 memory-to-behavior: promote a recurring procedural workflow into an executable skill.\n\t| { kind: \"promote_skill\"; name: string; description: string; body: string };\n\nexport interface ReflectionResult {\n\twrites: ReflectionWrite[];\n\tusage: Usage;\n\trationale: string;\n}\n\n/**\n * STATIC reflection system prompt (Hermes-parity #33). It is byte-identical across every reflection\n * pass — the variable parts (existing memory snapshot + the turn transcript) live in the USER prompt —\n * so the provider prompt-cache reuses this prefix instead of re-billing it each pass (cost guard).\n * Do NOT interpolate per-call data into this constant or caching breaks.\n */\nexport const REFLECTION_SYSTEM_PROMPT = `You are a reflection engine. Your job is to analyze the recent conversation turn, compare it against the agent's existing memory, and decide if any memory updates are needed.\n\nMemory guidelines:\n- \"MEMORY\" is for project facts, configuration, repeatable workflows, and coding findings.\n- \"USER\" is for user preferences, patterns, and style specifications.\n- Avoid duplicate facts. If the fact is already represented, do not add it.\n- CONFRONT existing memory: if the new turn contradicts or updates an existing fact, use \"memory_replace\" or \"memory_remove\" to supersede the old fact rather than blindly appending.\n- Keep memories short, factual, and direct. No fluff.\n- Do NOT capture transient/environment-specific noise: tool/network failures, one-off errors, or a single narrative event. Persist only durable facts and preferences.\n- PROMOTE to behavior: if the turn established a REPEATABLE, multi-step PROCEDURE/workflow (not a one-off fact) that should govern a future class of tasks, emit a \"promote_skill\" instead of (or in addition to) a memory fact. Only promote a genuinely reusable procedure — never a single fact, a one-off narrative, or environment-specific noise. Prefer a memory fact when unsure.\n\nYou must output your analysis and writes in the following JSON format inside a \\`\\`\\`json\\`\\`\\` code fence:\n{\n  \"rationale\": \"Explanation of your reasoning\",\n  \"writes\": [\n    { \"kind\": \"memory_add\", \"section\": \"MEMORY\" | \"USER\", \"text\": \"New direct fact to append\" },\n    { \"kind\": \"memory_replace\", \"target\": \"Exact text substring to replace\", \"text\": \"New replacement text\" },\n    { \"kind\": \"memory_remove\", \"target\": \"Exact text substring to remove\" },\n    { \"kind\": \"promote_skill\", \"name\": \"kebab-case-skill-name\", \"description\": \"one line of when to use it\", \"body\": \"Markdown: the step-by-step procedure\" }\n  ]\n}\n`;\n\nexport class ReflectionEngine {\n\t/**\n\t * Build the reflection prompt, call the injected isolated complete(),\n\t * parse the response, confront existing memory, and return memory writes.\n\t * Zero direct I/O.\n\t */\n\tasync reflect(input: ReflectionInput): Promise<ReflectionResult> {\n\t\tconst systemPrompt = REFLECTION_SYSTEM_PROMPT;\n\n\t\t// Variable inputs go in the USER prompt so the system prefix above stays cache-stable (#33).\n\t\tconst userPrompt = `Existing Memory snapshot:\n${input.existingMemory}\n\nRecent turn transcript:\n${input.recentTurnText}\n\nAnalyze this turn against the existing memory and output your memory updates.`;\n\n\t\ttry {\n\t\t\tconst compResult = await input.complete(systemPrompt, userPrompt);\n\t\t\tconst text = compResult.text;\n\n\t\t\tconst jsonMatch = text.match(/```json\\s*([\\s\\S]*?)\\s*```/) || text.match(/{[\\s\\S]*}/);\n\t\t\tif (!jsonMatch) {\n\t\t\t\treturn {\n\t\t\t\t\twrites: [],\n\t\t\t\t\tusage: compResult.usage,\n\t\t\t\t\trationale: `Failed to locate JSON response. Raw text:\\n${text}`,\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tconst parsed = JSON.parse(jsonMatch[1] || jsonMatch[0]);\n\t\t\tconst rationale = parsed.rationale || \"\";\n\t\t\tconst writes: ReflectionWrite[] = [];\n\n\t\t\tif (Array.isArray(parsed.writes)) {\n\t\t\t\tfor (const w of parsed.writes) {\n\t\t\t\t\tif (w && typeof w === \"object\") {\n\t\t\t\t\t\tif (\n\t\t\t\t\t\t\tw.kind === \"memory_add\" &&\n\t\t\t\t\t\t\t(w.section === \"MEMORY\" || w.section === \"USER\") &&\n\t\t\t\t\t\t\ttypeof w.text === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_add\", section: w.section, text: w.text });\n\t\t\t\t\t\t} else if (\n\t\t\t\t\t\t\tw.kind === \"memory_replace\" &&\n\t\t\t\t\t\t\ttypeof w.target === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.text === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_replace\", target: w.target, text: w.text });\n\t\t\t\t\t\t} else if (w.kind === \"memory_remove\" && typeof w.target === \"string\") {\n\t\t\t\t\t\t\twrites.push({ kind: \"memory_remove\", target: w.target });\n\t\t\t\t\t\t} else if (\n\t\t\t\t\t\t\tw.kind === \"promote_skill\" &&\n\t\t\t\t\t\t\ttypeof w.name === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.description === \"string\" &&\n\t\t\t\t\t\t\ttypeof w.body === \"string\"\n\t\t\t\t\t\t) {\n\t\t\t\t\t\t\twrites.push({ kind: \"promote_skill\", name: w.name, description: w.description, body: w.body });\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\treturn {\n\t\t\t\twrites,\n\t\t\t\tusage: compResult.usage,\n\t\t\t\trationale,\n\t\t\t};\n\t\t} catch (err) {\n\t\t\t// Zeroed/fallback usage representation\n\t\t\tconst emptyUsage: Usage = {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t};\n\t\t\treturn {\n\t\t\t\twrites: [],\n\t\t\t\tusage: emptyUsage,\n\t\t\t\trationale: `Error during reflection: ${String(err)}`,\n\t\t\t};\n\t\t}\n\t}\n}\n"]}

package/dist/core/settings-manager.d.ts

@@ -123,6 +123,11 @@ export interface Settings {
     /** Resource catalog directory (round resource management): the folder pi installs/updates/backs up from. */
     catalogDir?: string;
     compaction?: CompactionSettings;
+    /** Proactive per-turn cost guard (#34). */
+    costGuard?: {
+        maxTurnUsd?: number;
+        action?: "warn" | "downgrade";
+    };
     contextGc?: ContextGcSettings;
     branchSummary?: BranchSummarySettings;
     retry?: RetrySettings;
@@ -339,6 +344,11 @@ export declare class SettingsManager {
     getCompactionReserveTokens(): number;
     getCompactionKeepRecentTokens(): number;
     getCompactionTriggerPercent(): number;
+    /** Proactive per-turn cost guard (#34). `maxTurnUsd<=0` (default) disables it. */
+    getCostGuardSettings(): {
+        maxTurnUsd: number;
+        action: "warn" | "downgrade";
+    };
     /** Configured auxiliary summarizer model id, or "auto" (default) to pick the cheapest authed model. */
     getCompactionModel(): string;
     getCompactionSettings(): {