apex-code 0.5.0 → 0.5.2

package/dist/core/cache-warmer.js

@@ -0,0 +1,356 @@
+import { calculateCost, } from "@earendil-works/pi-ai";
+import { getProviderEnvValue } from "@earendil-works/pi-ai/utils/provider-env";
+/** Streaming warming never continues past this long after the real request that started it. */
+const MAX_WARMING_AGE_MS = 60 * 60_000;
+/** Idle warming uses a shorter horizon because continuation estimates become less reliable with age. */
+const MAX_IDLE_WARMING_AGE_MS = 30 * 60_000;
+/** A refresh is sent only when it is expected to save at least this many dollars. */
+const CACHE_WARMING_MINIMUM_EXPECTED_SAVINGS = 0.05;
+/**
+ * Chance that a real request arrives before the cache entry expires while the
+ * agent sits idle. Measured from our own usage; per-session estimates were not
+ * better than this constant.
+ */
+const IDLE_CONTINUATION_PROBABILITY = 0.15;
+/** Refresh at 90% of the TTL while preserving at least ten seconds of margin. */
+export function getCacheWarmingDelayMs(ttlMs) {
+    if (ttlMs <= 10_000)
+        return undefined;
+    return Math.max(1, Math.floor(Math.min(ttlMs * 0.9, ttlMs - 10_000)));
+}
+/**
+ * Lifetime of the prompt cache entry a request writes, from the model's
+ * `promptCache` tier for the retention the request used. Undefined when the
+ * model has no lifetime for that tier or caching is off.
+ */
+export function getPromptCacheTtlMs(model, options) {
+    const retention = options?.cacheRetention ??
+        (getProviderEnvValue("PI_CACHE_RETENTION", options?.env) === "long" ? "long" : "short");
+    if (retention === "none")
+        return undefined;
+    const seconds = model.promptCache?.[retention];
+    return seconds === undefined ? undefined : seconds * 1000;
+}
+/**
+ * Whether replaying the request with a one-token output cap leaves its cache
+ * entry untouched. Anthropic's budget-based thinking (Claude models without
+ * adaptive thinking) derives `budget_tokens` from `max_tokens`; the replay
+ * would get a different budget, which Anthropic keys the message cache on,
+ * and the model could still think for thousands of tokens.
+ */
+export function isReplayable(model, options) {
+    if (!options?.reasoning || model.api !== "anthropic-messages")
+        return true;
+    return model.compat?.forceAdaptiveThinking === true;
+}
+/** Prompt size of the most recent real request on the branch, as reported by the provider. */
+function lastPromptTokens(entries) {
+    for (let index = entries.length - 1; index >= 0; index--) {
+        const entry = entries[index];
+        if (entry.type === "message" && entry.message.role === "assistant") {
+            const usage = entry.message.usage;
+            return usage.input + usage.cacheRead + usage.cacheWrite;
+        }
+    }
+    return 0;
+}
+function price(model, tokens) {
+    const usage = {
+        input: 0,
+        output: 0,
+        cacheRead: 0,
+        cacheWrite: 0,
+        totalTokens: 0,
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+        ...tokens,
+    };
+    return calculateCost(model, usage).total;
+}
+/**
+ * Keeps one prompt cache entry alive by re-sending its request with a
+ * one-token output cap before the entry expires. `start` replaces any
+ * previous run; warm requests never extend the fixed safety windows.
+ */
+export class CacheWarmer {
+    run;
+    inactive;
+    models;
+    sessionManager;
+    getMode;
+    /** Lets extensions override `event.action`; failures fall back to pi's decision. */
+    decide;
+    /** Called with the persisted usage entry after each successful refresh. */
+    onWarmed;
+    constructor(models, sessionManager, getMode, decide = async (event) => event.action) {
+        this.models = models;
+        this.sessionManager = sessionManager;
+        this.getMode = getMode;
+        this.decide = decide;
+        this.inactive = { state: "inactive", reason: "waiting for first request" };
+    }
+    get status() {
+        if (this.getMode() === "off")
+            return { state: "inactive", reason: "cache warming disabled" };
+        const run = this.run;
+        if (!run)
+            return this.inactive;
+        if (!run.isCurrent())
+            return { state: "inactive", reason: "conversation context changed" };
+        const decision = this.evaluate(run);
+        const refreshing = run.timer === undefined;
+        if (!decision.economicsAvailable && !refreshing) {
+            return { state: "inactive", reason: "cache economics unavailable" };
+        }
+        return {
+            state: refreshing ? "refreshing" : "scheduled",
+            nextWarmAt: run.nextWarmAt,
+            decision,
+            extensionOverride: run.extensionOverride,
+        };
+    }
+    /** Keep the prompt cache entry written by `request` warm while `isCurrent` holds. */
+    start(request, isCurrent) {
+        this.clearRun();
+        const mode = this.getMode();
+        if (mode === "off") {
+            this.stop("cache warming disabled");
+            return;
+        }
+        if (!isReplayable(request.model, request.options)) {
+            this.stop("request cannot be replayed safely");
+            return;
+        }
+        const ttlMs = getPromptCacheTtlMs(request.model, request.options);
+        if (ttlMs === undefined) {
+            this.stop(request.options.cacheRetention === "none"
+                ? "request disabled prompt caching"
+                : "cache lifetime unavailable");
+            return;
+        }
+        const delayMs = getCacheWarmingDelayMs(ttlMs);
+        if (delayMs === undefined) {
+            this.stop("cache lifetime unavailable");
+            return;
+        }
+        this.run = {
+            ...request,
+            isCurrent,
+            ttlMs,
+            delayMs,
+            refreshDeadlineAt: 0,
+            startedAt: Date.now(),
+            controller: new AbortController(),
+            phase: "streaming",
+            nextWarmAt: 0,
+            extensionOverride: false,
+        };
+        this.schedule(this.run);
+    }
+    onAgentSettled() {
+        const run = this.run;
+        if (!run)
+            return;
+        if (this.getMode() === "streaming") {
+            this.stop("agent run settled");
+            return;
+        }
+        run.phase = "idle";
+        const deadline = run.startedAt + MAX_IDLE_WARMING_AGE_MS;
+        if (run.nextWarmAt > deadline || Date.now() >= deadline) {
+            this.stop("30-minute idle safety limit reached");
+        }
+    }
+    /** Reconcile an active run after the persisted warming mode changes. */
+    onModeChanged() {
+        const run = this.run;
+        if (!run)
+            return;
+        const reason = this.getModeStopReason(run);
+        if (reason)
+            this.stop(reason);
+    }
+    cancel() {
+        this.stop("inactive");
+    }
+    clearRun() {
+        const run = this.run;
+        if (!run)
+            return;
+        this.run = undefined;
+        if (run.timer)
+            clearTimeout(run.timer);
+        run.controller.abort();
+    }
+    stop(reason, stopped) {
+        this.clearRun();
+        this.inactive = { state: "inactive", reason, ...stopped };
+    }
+    schedule(run) {
+        run.extensionOverride = false;
+        run.nextWarmAt = Date.now() + run.delayMs;
+        // A timer can run late after sleep or event-loop blockage. Keep half of
+        // the planned pre-expiry margin for that delay and request dispatch; a
+        // late refresh is likely a full-price cache write, not a cache warm.
+        run.refreshDeadlineAt = run.nextWarmAt + Math.floor((run.ttlMs - run.delayMs) / 2);
+        const deadline = run.startedAt + (run.phase === "idle" ? MAX_IDLE_WARMING_AGE_MS : MAX_WARMING_AGE_MS);
+        if (run.nextWarmAt > deadline || Date.now() >= deadline) {
+            this.stop(run.phase === "idle" ? "30-minute idle safety limit reached" : "one-hour safety limit reached");
+            return;
+        }
+        run.timer = setTimeout(() => void this.refresh(run), Math.max(0, run.nextWarmAt - Date.now()));
+        run.timer.unref?.();
+    }
+    async refresh(run) {
+        run.timer = undefined;
+        if (!this.validateRun(run))
+            return;
+        if (this.refreshDeadlineMissed(run))
+            return;
+        const decision = this.evaluate(run);
+        const { warmCost, missCost, continuationProbability } = decision;
+        let action = decision.action;
+        try {
+            action = await this.decide({
+                type: "cache_warming_decision",
+                warmCost,
+                missCost,
+                continuationProbability,
+                action,
+            });
+        }
+        catch {
+            // Extension failures fall back to pi's own decision.
+        }
+        if (!this.validateRun(run) || this.refreshDeadlineMissed(run))
+            return;
+        const extensionOverride = action !== decision.action;
+        if (action === "stop") {
+            const reason = extensionOverride
+                ? "stopped by extension"
+                : decision.economicsAvailable
+                    ? "expected savings below threshold"
+                    : "cache economics unavailable";
+            this.stop(reason, { decision, extensionOverride });
+            return;
+        }
+        run.extensionOverride = extensionOverride;
+        try {
+            const message = await this.models
+                .streamSimple(run.model, run.context, {
+                ...run.options,
+                maxTokens: 1,
+                maxRetries: 0,
+                signal: run.controller.signal,
+            })
+                .result();
+            if (!this.validateRun(run))
+                return;
+            if (message.stopReason !== "error" && message.stopReason !== "aborted") {
+                const entry = this.sessionManager.appendUsage("cache_warm", message.provider, message.responseModel ?? message.model, message.usage, extensionOverride ? "extension override" : undefined);
+                this.onWarmed?.(entry);
+            }
+        }
+        catch {
+            // Cache warming is best-effort and must not affect the active agent run.
+        }
+        if (this.run === run)
+            this.schedule(run);
+    }
+    refreshDeadlineMissed(run) {
+        if (Date.now() <= run.refreshDeadlineAt)
+            return false;
+        this.stop("cache refresh deadline missed");
+        return true;
+    }
+    validateRun(run) {
+        if (this.run !== run)
+            return false;
+        const reason = this.getModeStopReason(run) ?? (!run.isCurrent() ? "conversation context changed" : undefined);
+        if (!reason)
+            return true;
+        this.stop(reason);
+        return false;
+    }
+    getModeStopReason(run) {
+        const mode = this.getMode();
+        if (mode === "off")
+            return "cache warming disabled";
+        if (mode === "streaming" && run.phase === "idle")
+            return "agent run settled";
+        return undefined;
+    }
+    evaluate(run) {
+        const model = run.model;
+        const promptTokens = lastPromptTokens(this.sessionManager.getBranch());
+        const cacheHitCost = price(model, { cacheRead: promptTokens });
+        const cacheMissCost = price(model, model.cost.cacheWrite > 0 ? { cacheWrite: promptTokens } : { input: promptTokens });
+        const warmCost = price(model, { cacheRead: promptTokens, output: 1 });
+        const missCost = Math.max(0, cacheMissCost - cacheHitCost);
+        const continuationProbability = run.phase === "idle" ? IDLE_CONTINUATION_PROBABILITY : 1;
+        const economicsAvailable = promptTokens > 0 && (cacheHitCost > 0 || cacheMissCost > 0);
+        const expectedSavings = continuationProbability * missCost - warmCost;
+        return {
+            phase: run.phase,
+            warmCost,
+            missCost,
+            continuationProbability,
+            expectedSavings,
+            economicsAvailable,
+            action: expectedSavings >= CACHE_WARMING_MINIMUM_EXPECTED_SAVINGS ? "warm" : "stop",
+        };
+    }
+}
+function formatDollars(value) {
+    return value < 0 ? `-$${Math.abs(value).toFixed(3)}` : `$${value.toFixed(3)}`;
+}
+function formatCacheWarmingEconomics(decision) {
+    if (!decision.economicsAvailable)
+        return "cache economics unavailable";
+    const probability = Math.round(decision.continuationProbability * 100);
+    const probabilityText = decision.phase === "streaming"
+        ? `${probability}% continuation probability while agent is running`
+        : `${probability}% continuation probability`;
+    const comparison = decision.action === "warm" ? ">=" : "<";
+    return `${probabilityText}, expected savings ${formatDollars(decision.expectedSavings)} ${comparison} $${CACHE_WARMING_MINIMUM_EXPECTED_SAVINGS.toFixed(3)}`;
+}
+function formatCacheWarmingDecisionTime(nextWarmAt, now) {
+    if (nextWarmAt === undefined || nextWarmAt <= now)
+        return "Decision now";
+    let remainingSeconds = Math.ceil((nextWarmAt - now) / 1000);
+    const hours = Math.floor(remainingSeconds / 3600);
+    remainingSeconds %= 3600;
+    const minutes = Math.floor(remainingSeconds / 60);
+    const seconds = remainingSeconds % 60;
+    const parts = [];
+    if (hours > 0)
+        parts.push(`${hours}h`);
+    if (minutes > 0)
+        parts.push(`${minutes}m`);
+    if (seconds > 0 || parts.length === 0)
+        parts.push(`${seconds}s`);
+    return `Decision in ${parts.join(" ")}`;
+}
+/** One-line status for `/session`. */
+export function formatCacheWarmingStatus(status, now = Date.now()) {
+    const decision = status.decision;
+    // A decision is attached once pi (or an extension) acted on it; "inactive"
+    // without one never got that far.
+    if (!decision || (status.state === "inactive" && !decision.economicsAvailable && !status.extensionOverride)) {
+        return `Inactive (${status.reason ?? "unknown reason"})`;
+    }
+    const details = status.extensionOverride
+        ? `extension override, ${formatCacheWarmingEconomics(decision)}`
+        : `${formatCacheWarmingEconomics(decision)} -> ${decision.action}`;
+    if (status.state === "inactive")
+        return `Stopped (${details})`;
+    if (status.state === "refreshing")
+        return `Warming cache (${details})`;
+    return `${formatCacheWarmingDecisionTime(status.nextWarmAt, now)} (${details})`;
+}
+/** One-line transcript text for persisted cache-warming usage. */
+export function formatCacheWarmingUsage(entry) {
+    const note = entry.note ? ` (${entry.note})` : "";
+    const cost = entry.usage.cost.total.toFixed(6).replace(/(\.\d{3}\d*?)0+$/, "$1");
+    return `Cache warmed${note}: $${cost}`;
+}
+//# sourceMappingURL=cache-warmer.js.map

package/dist/core/cache-warmer.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cache-warmer.js","sourceRoot":"","sources":["../../src/core/cache-warmer.ts"],"names":[],"mappings":"AAAA,OAAO,EAGN,aAAa,GAKb,MAAM,uBAAuB,CAAC;AAC/B,OAAO,EAAE,mBAAmB,EAAE,MAAM,0CAA0C,CAAC;AAK/E,+FAA+F;AAC/F,MAAM,kBAAkB,GAAG,EAAE,GAAG,MAAM,CAAC;AACvC,wGAAwG;AACxG,MAAM,uBAAuB,GAAG,EAAE,GAAG,MAAM,CAAC;AAC5C,qFAAqF;AACrF,MAAM,sCAAsC,GAAG,IAAI,CAAC;AACpD;;;;GAIG;AACH,MAAM,6BAA6B,GAAG,IAAI,CAAC;AAE3C,iFAAiF;AACjF,MAAM,UAAU,sBAAsB,CAAC,KAAa,EAAsB;IACzE,IAAI,KAAK,IAAI,MAAM;QAAE,OAAO,SAAS,CAAC;IACtC,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,GAAG,GAAG,EAAE,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;AAAA,CACtE;AAED;;;;GAIG;AACH,MAAM,UAAU,mBAAmB,CAAC,KAAiB,EAAE,OAAwC,EAAsB;IACpH,MAAM,SAAS,GACd,OAAO,EAAE,cAAc;QACvB,CAAC,mBAAmB,CAAC,oBAAoB,EAAE,OAAO,EAAE,GAAG,CAAC,KAAK,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC;IACzF,IAAI,SAAS,KAAK,MAAM;QAAE,OAAO,SAAS,CAAC;IAC3C,MAAM,OAAO,GAAG,KAAK,CAAC,WAAW,EAAE,CAAC,SAAS,CAAC,CAAC;IAC/C,OAAO,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,OAAO,GAAG,IAAI,CAAC;AAAA,CAC1D;AAED;;;;;;GAMG;AACH,MAAM,UAAU,YAAY,CAAC,KAAiB,EAAE,OAAwC,EAAW;IAClG,IAAI,CAAC,OAAO,EAAE,SAAS,IAAI,KAAK,CAAC,GAAG,KAAK,oBAAoB;QAAE,OAAO,IAAI,CAAC;IAC3E,OAAQ,KAAqC,CAAC,MAAM,EAAE,qBAAqB,KAAK,IAAI,CAAC;AAAA,CACrF;AAED,8FAA8F;AAC9F,SAAS,gBAAgB,CAAC,OAAuB,EAAU;IAC1D,KAAK,IAAI,KAAK,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,KAAK,IAAI,CAAC,EAAE,KAAK,EAAE,EAAE,CAAC;QAC1D,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC;QAC7B,IAAI,KAAK,CAAC,IAAI,KAAK,SAAS,IAAI,KAAK,CAAC,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YACpE,MAAM,KAAK,GAAG,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;YAClC,OAAO,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC;QACzD,CAAC;IACF,CAAC;IACD,OAAO,CAAC,CAAC;AAAA,CACT;AAED,SAAS,KAAK,CACb,KAAiB,EACjB,MAA6E,EACpE;IACT,MAAM,KAAK,GAAU;QACpB,KAAK,EAAE,CAAC;QACR,MAAM,EAAE,CAAC;QACT,SAAS,EAAE,CAAC;QACZ,UAAU,EAAE,CAAC;QACb,WAAW,EAAE,CAAC;QACd,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE;QACpE,GAAG,MAAM;KACT,CAAC;IACF,OAAO,aAAa,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC,KAAK,CAAC;AAAA,CACzC;AAuED;;;;GAIG;AACH,MAAM,OAAO,WAAW;IACf,GAAG,CAAa;IAChB,QAAQ,CAAqB;IACpB,MAAM,CAAqC;IAC3C,cAAc,CAAoD;IAClE,OAAO,CAAyB;IACjD,oFAAoF;IACnE,MAAM,CAAoE;IAC3F,2EAA2E;IAC3E,QAAQ,CAA+B;IAEvC,YACC,MAA0C,EAC1C,cAAiE,EACjE,OAA+B,EAC/B,MAAM,GAAsE,KAAK,EAAE,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,MAAM,EACxG;QACD,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,cAAc,GAAG,cAAc,CAAC;QACrC,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC;QACvB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACrB,IAAI,CAAC,QAAQ,GAAG,EAAE,KAAK,EAAE,UAAU,EAAE,MAAM,EAAE,2BAA2B,EAAE,CAAC;IAAA,CAC3E;IAED,IAAI,MAAM,GAAuB;QAChC,IAAI,IAAI,CAAC,OAAO,EAAE,KAAK,KAAK;YAAE,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,MAAM,EAAE,wBAAwB,EAAE,CAAC;QAC7F,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC;QACrB,IAAI,CAAC,GAAG;YAAE,OAAO,IAAI,CAAC,QAAQ,CAAC;QAC/B,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE;YAAE,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,MAAM,EAAE,8BAA8B,EAAE,CAAC;QAC3F,MAAM,QAAQ,GAAG,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC;QACpC,MAAM,UAAU,GAAG,GAAG,CAAC,KAAK,KAAK,SAAS,CAAC;QAC3C,IAAI,CAAC,QAAQ,CAAC,kBAAkB,IAAI,CAAC,UAAU,EAAE,CAAC;YACjD,OAAO,EAAE,KAAK,EAAE,UAAU,EAAE,MAAM,EAAE,6BAA6B,EAAE,CAAC;QACrE,CAAC;QACD,OAAO;YACN,KAAK,EAAE,UAAU,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,WAAW;YAC9C,UAAU,EAAE,GAAG,CAAC,UAAU;YAC1B,QAAQ;YACR,iBAAiB,EAAE,GAAG,CAAC,iBAAiB;SACxC,CAAC;IAAA,CACF;IAED,qFAAqF;IACrF,KAAK,CAAC,OAAyB,EAAE,SAAwB,EAAQ;QAChE,IAAI,CAAC,QAAQ,EAAE,CAAC;QAChB,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,EAAE,CAAC;QAC5B,IAAI,IAAI,KAAK,KAAK,EAAE,CAAC;YACpB,IAAI,CAAC,IAAI,CAAC,wBAAwB,CAAC,CAAC;YACpC,OAAO;QACR,CAAC;QACD,IAAI,CAAC,YAAY,CAAC,OAAO,CAAC,KAAK,EAAE,OAAO,CAAC,OAAO,CAAC,EAAE,CAAC;YACnD,IAAI,CAAC,IAAI,CAAC,mCAAmC,CAAC,CAAC;YAC/C,OAAO;QACR,CAAC;QACD,MAAM,KAAK,GAAG,mBAAmB,CAAC,OAAO,CAAC,KAAK,EAAE,OAAO,CAAC,OAAO,CAAC,CAAC;QAClE,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;YACzB,IAAI,CAAC,IAAI,CACR,OAAO,CAAC,OAAO,CAAC,cAAc,KAAK,MAAM;gBACxC,CAAC,CAAC,iCAAiC;gBACnC,CAAC,CAAC,4BAA4B,CAC/B,CAAC;YACF,OAAO;QACR,CAAC;QACD,MAAM,OAAO,GAAG,sBAAsB,CAAC,KAAK,CAAC,CAAC;QAC9C,IAAI,OAAO,KAAK,SAAS,EAAE,CAAC;YAC3B,IAAI,CAAC,IAAI,CAAC,4BAA4B,CAAC,CAAC;YACxC,OAAO;QACR,CAAC;QACD,IAAI,CAAC,GAAG,GAAG;YACV,GAAG,OAAO;YACV,SAAS;YACT,KAAK;YACL,OAAO;YACP,iBAAiB,EAAE,CAAC;YACpB,SAAS,EAAE,IAAI,CAAC,GAAG,EAAE;YACrB,UAAU,EAAE,IAAI,eAAe,EAAE;YACjC,KAAK,EAAE,WAAW;YAClB,UAAU,EAAE,CAAC;YACb,iBAAiB,EAAE,KAAK;SACxB,CAAC;QACF,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IAAA,CACxB;IAED,cAAc,GAAS;QACtB,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC;QACrB,IAAI,CAAC,GAAG;YAAE,OAAO;QACjB,IAAI,IAAI,CAAC,OAAO,EAAE,KAAK,WAAW,EAAE,CAAC;YACpC,IAAI,CAAC,IAAI,CAAC,mBAAmB,CAAC,CAAC;YAC/B,OAAO;QACR,CAAC;QACD,GAAG,CAAC,KAAK,GAAG,MAAM,CAAC;QACnB,MAAM,QAAQ,GAAG,GAAG,CAAC,SAAS,GAAG,uBAAuB,CAAC;QACzD,IAAI,GAAG,CAAC,UAAU,GAAG,QAAQ,IAAI,IAAI,CAAC,GAAG,EAAE,IAAI,QAAQ,EAAE,CAAC;YACzD,IAAI,CAAC,IAAI,CAAC,qCAAqC,CAAC,CAAC;QAClD,CAAC;IAAA,CACD;IAED,wEAAwE;IACxE,aAAa,GAAS;QACrB,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC;QACrB,IAAI,CAAC,GAAG;YAAE,OAAO;QACjB,MAAM,MAAM,GAAG,IAAI,CAAC,iBAAiB,CAAC,GAAG,CAAC,CAAC;QAC3C,IAAI,MAAM;YAAE,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAAA,CAC9B;IAED,MAAM,GAAS;QACd,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC;IAAA,CACtB;IAEO,QAAQ,GAAS;QACxB,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC;QACrB,IAAI,CAAC,GAAG;YAAE,OAAO;QACjB,IAAI,CAAC,GAAG,GAAG,SAAS,CAAC;QACrB,IAAI,GAAG,CAAC,KAAK;YAAE,YAAY,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;QACvC,GAAG,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;IAAA,CACvB;IAEO,IAAI,CAAC,MAAc,EAAE,OAAoE,EAAQ;QACxG,IAAI,CAAC,QAAQ,EAAE,CAAC;QAChB,IAAI,CAAC,QAAQ,GAAG,EAAE,KAAK,EAAE,UAAU,EAAE,MAAM,EAAE,GAAG,OAAO,EAAE,CAAC;IAAA,CAC1D;IAEO,QAAQ,CAAC,GAAc,EAAQ;QACtC,GAAG,CAAC,iBAAiB,GAAG,KAAK,CAAC;QAC9B,GAAG,CAAC,UAAU,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,GAAG,CAAC,OAAO,CAAC;QAC1C,wEAAwE;QACxE,uEAAuE;QACvE,qEAAqE;QACrE,GAAG,CAAC,iBAAiB,GAAG,GAAG,CAAC,UAAU,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,KAAK,GAAG,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC;QACnF,MAAM,QAAQ,GAAG,GAAG,CAAC,SAAS,GAAG,CAAC,GAAG,CAAC,KAAK,KAAK,MAAM,CAAC,CAAC,CAAC,uBAAuB,CAAC,CAAC,CAAC,kBAAkB,CAAC,CAAC;QACvG,IAAI,GAAG,CAAC,UAAU,GAAG,QAAQ,IAAI,IAAI,CAAC,GAAG,EAAE,IAAI,QAAQ,EAAE,CAAC;YACzD,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,KAAK,KAAK,MAAM,CAAC,CAAC,CAAC,qCAAqC,CAAC,CAAC,CAAC,+BAA+B,CAAC,CAAC;YAC1G,OAAO;QACR,CAAC;QACD,GAAG,CAAC,KAAK,GAAG,UAAU,CAAC,GAAG,EAAE,CAAC,KAAK,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,GAAG,CAAC,UAAU,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;QAC/F,GAAG,CAAC,KAAK,CAAC,KAAK,EAAE,EAAE,CAAC;IAAA,CACpB;IAEO,KAAK,CAAC,OAAO,CAAC,GAAc,EAAiB;QACpD,GAAG,CAAC,KAAK,GAAG,SAAS,CAAC;QACtB,IAAI,CAAC,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC;YAAE,OAAO;QACnC,IAAI,IAAI,CAAC,qBAAqB,CAAC,GAAG,CAAC;YAAE,OAAO;QAC5C,MAAM,QAAQ,GAAG,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC;QACpC,MAAM,EAAE,QAAQ,EAAE,QAAQ,EAAE,uBAAuB,EAAE,GAAG,QAAQ,CAAC;QACjE,IAAI,MAAM,GAAG,QAAQ,CAAC,MAAM,CAAC;QAC7B,IAAI,CAAC;YACJ,MAAM,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC;gBAC1B,IAAI,EAAE,wBAAwB;gBAC9B,QAAQ;gBACR,QAAQ;gBACR,uBAAuB;gBACvB,MAAM;aACN,CAAC,CAAC;QACJ,CAAC;QAAC,MAAM,CAAC;YACR,qDAAqD;QACtD,CAAC;QACD,IAAI,CAAC,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,IAAI,IAAI,CAAC,qBAAqB,CAAC,GAAG,CAAC;YAAE,OAAO;QACtE,MAAM,iBAAiB,GAAG,MAAM,KAAK,QAAQ,CAAC,MAAM,CAAC;QACrD,IAAI,MAAM,KAAK,MAAM,EAAE,CAAC;YACvB,MAAM,MAAM,GAAG,iBAAiB;gBAC/B,CAAC,CAAC,sBAAsB;gBACxB,CAAC,CAAC,QAAQ,CAAC,kBAAkB;oBAC5B,CAAC,CAAC,kCAAkC;oBACpC,CAAC,CAAC,6BAA6B,CAAC;YAClC,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,EAAE,QAAQ,EAAE,iBAAiB,EAAE,CAAC,CAAC;YACnD,OAAO;QACR,CAAC;QAED,GAAG,CAAC,iBAAiB,GAAG,iBAAiB,CAAC;QAC1C,IAAI,CAAC;YACJ,MAAM,OAAO,GAAG,MAAM,IAAI,CAAC,MAAM;iBAC/B,YAAY,CAAC,GAAG,CAAC,KAAK,EAAE,GAAG,CAAC,OAAO,EAAE;gBACrC,GAAG,GAAG,CAAC,OAAO;gBACd,SAAS,EAAE,CAAC;gBACZ,UAAU,EAAE,CAAC;gBACb,MAAM,EAAE,GAAG,CAAC,UAAU,CAAC,MAAM;aAC7B,CAAC;iBACD,MAAM,EAAE,CAAC;YACX,IAAI,CAAC,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC;gBAAE,OAAO;YACnC,IAAI,OAAO,CAAC,UAAU,KAAK,OAAO,IAAI,OAAO,CAAC,UAAU,KAAK,SAAS,EAAE,CAAC;gBACxE,MAAM,KAAK,GAAG,IAAI,CAAC,cAAc,CAAC,WAAW,CAC5C,YAAY,EACZ,OAAO,CAAC,QAAQ,EAChB,OAAO,CAAC,aAAa,IAAI,OAAO,CAAC,KAAK,EACtC,OAAO,CAAC,KAAK,EACb,iBAAiB,CAAC,CAAC,CAAC,oBAAoB,CAAC,CAAC,CAAC,SAAS,CACpD,CAAC;gBACF,IAAI,CAAC,QAAQ,EAAE,CAAC,KAAK,CAAC,CAAC;YACxB,CAAC;QACF,CAAC;QAAC,MAAM,CAAC;YACR,yEAAyE;QAC1E,CAAC;QACD,IAAI,IAAI,CAAC,GAAG,KAAK,GAAG;YAAE,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC;IAAA,CACzC;IAEO,qBAAqB,CAAC,GAAc,EAAW;QACtD,IAAI,IAAI,CAAC,GAAG,EAAE,IAAI,GAAG,CAAC,iBAAiB;YAAE,OAAO,KAAK,CAAC;QACtD,IAAI,CAAC,IAAI,CAAC,+BAA+B,CAAC,CAAC;QAC3C,OAAO,IAAI,CAAC;IAAA,CACZ;IAEO,WAAW,CAAC,GAAc,EAAW;QAC5C,IAAI,IAAI,CAAC,GAAG,KAAK,GAAG;YAAE,OAAO,KAAK,CAAC;QACnC,MAAM,MAAM,GAAG,IAAI,CAAC,iBAAiB,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,8BAA8B,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC;QAC9G,IAAI,CAAC,MAAM;YAAE,OAAO,IAAI,CAAC;QACzB,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAClB,OAAO,KAAK,CAAC;IAAA,CACb;IAEO,iBAAiB,CAAC,GAAc,EAAsB;QAC7D,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,EAAE,CAAC;QAC5B,IAAI,IAAI,KAAK,KAAK;YAAE,OAAO,wBAAwB,CAAC;QACpD,IAAI,IAAI,KAAK,WAAW,IAAI,GAAG,CAAC,KAAK,KAAK,MAAM;YAAE,OAAO,mBAAmB,CAAC;QAC7E,OAAO,SAAS,CAAC;IAAA,CACjB;IAEO,QAAQ,CAAC,GAAc,EAAwB;QACtD,MAAM,KAAK,GAAG,GAAG,CAAC,KAAK,CAAC;QACxB,MAAM,YAAY,GAAG,gBAAgB,CAAC,IAAI,CAAC,cAAc,CAAC,SAAS,EAAE,CAAC,CAAC;QACvE,MAAM,YAAY,GAAG,KAAK,CAAC,KAAK,EAAE,EAAE,SAAS,EAAE,YAAY,EAAE,CAAC,CAAC;QAC/D,MAAM,aAAa,GAAG,KAAK,CAC1B,KAAK,EACL,KAAK,CAAC,IAAI,CAAC,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,UAAU,EAAE,YAAY,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,YAAY,EAAE,CAClF,CAAC;QACF,MAAM,QAAQ,GAAG,KAAK,CAAC,KAAK,EAAE,EAAE,SAAS,EAAE,YAAY,EAAE,MAAM,EAAE,CAAC,EAAE,CAAC,CAAC;QACtE,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,aAAa,GAAG,YAAY,CAAC,CAAC;QAC3D,MAAM,uBAAuB,GAAG,GAAG,CAAC,KAAK,KAAK,MAAM,CAAC,CAAC,CAAC,6BAA6B,CAAC,CAAC,CAAC,CAAC,CAAC;QACzF,MAAM,kBAAkB,GAAG,YAAY,GAAG,CAAC,IAAI,CAAC,YAAY,GAAG,CAAC,IAAI,aAAa,GAAG,CAAC,CAAC,CAAC;QACvF,MAAM,eAAe,GAAG,uBAAuB,GAAG,QAAQ,GAAG,QAAQ,CAAC;QACtE,OAAO;YACN,KAAK,EAAE,GAAG,CAAC,KAAK;YAChB,QAAQ;YACR,QAAQ;YACR,uBAAuB;YACvB,eAAe;YACf,kBAAkB;YAClB,MAAM,EAAE,eAAe,IAAI,sCAAsC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM;SACnF,CAAC;IAAA,CACF;CACD;AAED,SAAS,aAAa,CAAC,KAAa,EAAU;IAC7C,OAAO,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC;AAAA,CAC9E;AAED,SAAS,2BAA2B,CAAC,QAA8B,EAAU;IAC5E,IAAI,CAAC,QAAQ,CAAC,kBAAkB;QAAE,OAAO,6BAA6B,CAAC;IACvE,MAAM,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,uBAAuB,GAAG,GAAG,CAAC,CAAC;IACvE,MAAM,eAAe,GACpB,QAAQ,CAAC,KAAK,KAAK,WAAW;QAC7B,CAAC,CAAC,GAAG,WAAW,mDAAmD;QACnE,CAAC,CAAC,GAAG,WAAW,4BAA4B,CAAC;IAC/C,MAAM,UAAU,GAAG,QAAQ,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC;IAC3D,OAAO,GAAG,eAAe,sBAAsB,aAAa,CAAC,QAAQ,CAAC,eAAe,CAAC,IAAI,UAAU,KAAK,sCAAsC,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC;AAAA,CAC7J;AAED,SAAS,8BAA8B,CAAC,UAA8B,EAAE,GAAW,EAAU;IAC5F,IAAI,UAAU,KAAK,SAAS,IAAI,UAAU,IAAI,GAAG;QAAE,OAAO,cAAc,CAAC;IACzE,IAAI,gBAAgB,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,UAAU,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,CAAC;IAC5D,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,gBAAgB,GAAG,IAAI,CAAC,CAAC;IAClD,gBAAgB,IAAI,IAAI,CAAC;IACzB,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,gBAAgB,GAAG,EAAE,CAAC,CAAC;IAClD,MAAM,OAAO,GAAG,gBAAgB,GAAG,EAAE,CAAC;IACtC,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,IAAI,KAAK,GAAG,CAAC;QAAE,KAAK,CAAC,IAAI,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC;IACvC,IAAI,OAAO,GAAG,CAAC;QAAE,KAAK,CAAC,IAAI,CAAC,GAAG,OAAO,GAAG,CAAC,CAAC;IAC3C,IAAI,OAAO,GAAG,CAAC,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAAE,KAAK,CAAC,IAAI,CAAC,GAAG,OAAO,GAAG,CAAC,CAAC;IACjE,OAAO,eAAe,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;AAAA,CACxC;AAED,sCAAsC;AACtC,MAAM,UAAU,wBAAwB,CAAC,MAA0B,EAAE,GAAG,GAAG,IAAI,CAAC,GAAG,EAAE,EAAU;IAC9F,MAAM,QAAQ,GAAG,MAAM,CAAC,QAAQ,CAAC;IACjC,2EAA2E;IAC3E,kCAAkC;IAClC,IAAI,CAAC,QAAQ,IAAI,CAAC,MAAM,CAAC,KAAK,KAAK,UAAU,IAAI,CAAC,QAAQ,CAAC,kBAAkB,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC,EAAE,CAAC;QAC7G,OAAO,aAAa,MAAM,CAAC,MAAM,IAAI,gBAAgB,GAAG,CAAC;IAC1D,CAAC;IACD,MAAM,OAAO,GAAG,MAAM,CAAC,iBAAiB;QACvC,CAAC,CAAC,uBAAuB,2BAA2B,CAAC,QAAQ,CAAC,EAAE;QAChE,CAAC,CAAC,GAAG,2BAA2B,CAAC,QAAQ,CAAC,OAAO,QAAQ,CAAC,MAAM,EAAE,CAAC;IACpE,IAAI,MAAM,CAAC,KAAK,KAAK,UAAU;QAAE,OAAO,YAAY,OAAO,GAAG,CAAC;IAC/D,IAAI,MAAM,CAAC,KAAK,KAAK,YAAY;QAAE,OAAO,kBAAkB,OAAO,GAAG,CAAC;IACvE,OAAO,GAAG,8BAA8B,CAAC,MAAM,CAAC,UAAU,EAAE,GAAG,CAAC,KAAK,OAAO,GAAG,CAAC;AAAA,CAChF;AAED,kEAAkE;AAClE,MAAM,UAAU,uBAAuB,CAAC,KAAiB,EAAU;IAClE,MAAM,IAAI,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,KAAK,KAAK,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC;IAClD,MAAM,IAAI,GAAG,KAAK,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,kBAAkB,EAAE,IAAI,CAAC,CAAC;IACjF,OAAO,eAAe,IAAI,MAAM,IAAI,EAAE,CAAC;AAAA,CACvC","sourcesContent":["import {\n\ttype Api,\n\ttype Context,\n\tcalculateCost,\n\ttype Model,\n\ttype ModelsSimpleStreamOptions,\n\ttype SimpleStreamOptions,\n\ttype Usage,\n} from \"@earendil-works/pi-ai\";\nimport { getProviderEnvValue } from \"@earendil-works/pi-ai/utils/provider-env\";\nimport type { ModelRuntime } from \"./model-runtime.ts\";\nimport type { SessionEntry, SessionManager, UsageEntry } from \"./session-manager.ts\";\nimport type { CacheWarmingMode } from \"./settings-manager.ts\";\n\n/** Streaming warming never continues past this long after the real request that started it. */\nconst MAX_WARMING_AGE_MS = 60 * 60_000;\n/** Idle warming uses a shorter horizon because continuation estimates become less reliable with age. */\nconst MAX_IDLE_WARMING_AGE_MS = 30 * 60_000;\n/** A refresh is sent only when it is expected to save at least this many dollars. */\nconst CACHE_WARMING_MINIMUM_EXPECTED_SAVINGS = 0.05;\n/**\n * Chance that a real request arrives before the cache entry expires while the\n * agent sits idle. Measured from our own usage; per-session estimates were not\n * better than this constant.\n */\nconst IDLE_CONTINUATION_PROBABILITY = 0.15;\n\n/** Refresh at 90% of the TTL while preserving at least ten seconds of margin. */\nexport function getCacheWarmingDelayMs(ttlMs: number): number | undefined {\n\tif (ttlMs <= 10_000) return undefined;\n\treturn Math.max(1, Math.floor(Math.min(ttlMs * 0.9, ttlMs - 10_000)));\n}\n\n/**\n * Lifetime of the prompt cache entry a request writes, from the model's\n * `promptCache` tier for the retention the request used. Undefined when the\n * model has no lifetime for that tier or caching is off.\n */\nexport function getPromptCacheTtlMs(model: Model<Api>, options: SimpleStreamOptions | undefined): number | undefined {\n\tconst retention =\n\t\toptions?.cacheRetention ??\n\t\t(getProviderEnvValue(\"PI_CACHE_RETENTION\", options?.env) === \"long\" ? \"long\" : \"short\");\n\tif (retention === \"none\") return undefined;\n\tconst seconds = model.promptCache?.[retention];\n\treturn seconds === undefined ? undefined : seconds * 1000;\n}\n\n/**\n * Whether replaying the request with a one-token output cap leaves its cache\n * entry untouched. Anthropic's budget-based thinking (Claude models without\n * adaptive thinking) derives `budget_tokens` from `max_tokens`; the replay\n * would get a different budget, which Anthropic keys the message cache on,\n * and the model could still think for thousands of tokens.\n */\nexport function isReplayable(model: Model<Api>, options: SimpleStreamOptions | undefined): boolean {\n\tif (!options?.reasoning || model.api !== \"anthropic-messages\") return true;\n\treturn (model as Model<\"anthropic-messages\">).compat?.forceAdaptiveThinking === true;\n}\n\n/** Prompt size of the most recent real request on the branch, as reported by the provider. */\nfunction lastPromptTokens(entries: SessionEntry[]): number {\n\tfor (let index = entries.length - 1; index >= 0; index--) {\n\t\tconst entry = entries[index];\n\t\tif (entry.type === \"message\" && entry.message.role === \"assistant\") {\n\t\t\tconst usage = entry.message.usage;\n\t\t\treturn usage.input + usage.cacheRead + usage.cacheWrite;\n\t\t}\n\t}\n\treturn 0;\n}\n\nfunction price(\n\tmodel: Model<Api>,\n\ttokens: Partial<Pick<Usage, \"input\" | \"output\" | \"cacheRead\" | \"cacheWrite\">>,\n): number {\n\tconst usage: Usage = {\n\t\tinput: 0,\n\t\toutput: 0,\n\t\tcacheRead: 0,\n\t\tcacheWrite: 0,\n\t\ttotalTokens: 0,\n\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t...tokens,\n\t};\n\treturn calculateCost(model, usage).total;\n}\n\nexport type CacheWarmingAction = \"warm\" | \"stop\";\n\n/** Inputs and outcome of one warm-or-stop decision, as shown by `/session`. */\nexport interface CacheWarmingDecision {\n\t/** \"streaming\" while the agent run that sent the request is still active. */\n\tphase: \"streaming\" | \"idle\";\n\t/** Price of this refresh: a cache read of the prompt plus one output token. */\n\twarmCost: number;\n\t/** Extra price of the next real request if the cache entry is lost. */\n\tmissCost: number;\n\t/** Estimated chance that a real request arrives before the entry expires. */\n\tcontinuationProbability: number;\n\t/** `continuationProbability * missCost - warmCost`. */\n\texpectedSavings: number;\n\t/** False when the prompt size or the model's prices are unknown. */\n\teconomicsAvailable: boolean;\n\t/** Pi's decision: \"warm\" when `expectedSavings` is at least $0.05. */\n\taction: CacheWarmingAction;\n}\n\n/**\n * Fired before each refresh with pi's decision filled in. Everything else an\n * extension might want (model, idle state, context size) is on the context.\n */\nexport interface CacheWarmingDecisionEvent\n\textends Pick<CacheWarmingDecision, \"warmCost\" | \"missCost\" | \"continuationProbability\" | \"action\"> {\n\ttype: \"cache_warming_decision\";\n}\n\nexport interface CacheWarmingDecisionEventResult {\n\t/** Override whether this refresh is sent. \"stop\" ends warming until the next real request. */\n\taction?: CacheWarmingAction;\n}\n\nexport interface CacheWarmingStatus {\n\t/** \"scheduled\": a refresh timer is armed; \"refreshing\": a warm request is in flight. */\n\tstate: \"inactive\" | \"scheduled\" | \"refreshing\";\n\t/** Why nothing is scheduled. */\n\treason?: string;\n\tnextWarmAt?: number;\n\t/** The pending decision, or the decision that stopped warming. */\n\tdecision?: CacheWarmingDecision;\n\t/** True when an extension changed `decision.action`. */\n\textensionOverride?: boolean;\n}\n\n/** The request whose prompt cache entry should be kept warm, exactly as it was sent. */\nexport interface CacheWarmRequest {\n\tmodel: Model<Api>;\n\tcontext: Context;\n\toptions: ModelsSimpleStreamOptions;\n}\n\ninterface ActiveRun extends CacheWarmRequest {\n\t/** False once the session's model or messages no longer match the request. */\n\tisCurrent: () => boolean;\n\tttlMs: number;\n\tdelayMs: number;\n\t/** Latest safe time to send this refresh, leaving half the original expiry margin. */\n\trefreshDeadlineAt: number;\n\tstartedAt: number;\n\tcontroller: AbortController;\n\tphase: \"streaming\" | \"idle\";\n\tnextWarmAt: number;\n\t/** Set while a refresh that an extension forced is in flight. */\n\textensionOverride: boolean;\n\ttimer?: ReturnType<typeof setTimeout>;\n}\n\n/**\n * Keeps one prompt cache entry alive by re-sending its request with a\n * one-token output cap before the entry expires. `start` replaces any\n * previous run; warm requests never extend the fixed safety windows.\n */\nexport class CacheWarmer {\n\tprivate run?: ActiveRun;\n\tprivate inactive: CacheWarmingStatus;\n\tprivate readonly models: Pick<ModelRuntime, \"streamSimple\">;\n\tprivate readonly sessionManager: Pick<SessionManager, \"appendUsage\" | \"getBranch\">;\n\tprivate readonly getMode: () => CacheWarmingMode;\n\t/** Lets extensions override `event.action`; failures fall back to pi's decision. */\n\tprivate readonly decide: (event: CacheWarmingDecisionEvent) => Promise<CacheWarmingAction>;\n\t/** Called with the persisted usage entry after each successful refresh. */\n\tonWarmed?: (entry: UsageEntry) => void;\n\n\tconstructor(\n\t\tmodels: Pick<ModelRuntime, \"streamSimple\">,\n\t\tsessionManager: Pick<SessionManager, \"appendUsage\" | \"getBranch\">,\n\t\tgetMode: () => CacheWarmingMode,\n\t\tdecide: (event: CacheWarmingDecisionEvent) => Promise<CacheWarmingAction> = async (event) => event.action,\n\t) {\n\t\tthis.models = models;\n\t\tthis.sessionManager = sessionManager;\n\t\tthis.getMode = getMode;\n\t\tthis.decide = decide;\n\t\tthis.inactive = { state: \"inactive\", reason: \"waiting for first request\" };\n\t}\n\n\tget status(): CacheWarmingStatus {\n\t\tif (this.getMode() === \"off\") return { state: \"inactive\", reason: \"cache warming disabled\" };\n\t\tconst run = this.run;\n\t\tif (!run) return this.inactive;\n\t\tif (!run.isCurrent()) return { state: \"inactive\", reason: \"conversation context changed\" };\n\t\tconst decision = this.evaluate(run);\n\t\tconst refreshing = run.timer === undefined;\n\t\tif (!decision.economicsAvailable && !refreshing) {\n\t\t\treturn { state: \"inactive\", reason: \"cache economics unavailable\" };\n\t\t}\n\t\treturn {\n\t\t\tstate: refreshing ? \"refreshing\" : \"scheduled\",\n\t\t\tnextWarmAt: run.nextWarmAt,\n\t\t\tdecision,\n\t\t\textensionOverride: run.extensionOverride,\n\t\t};\n\t}\n\n\t/** Keep the prompt cache entry written by `request` warm while `isCurrent` holds. */\n\tstart(request: CacheWarmRequest, isCurrent: () => boolean): void {\n\t\tthis.clearRun();\n\t\tconst mode = this.getMode();\n\t\tif (mode === \"off\") {\n\t\t\tthis.stop(\"cache warming disabled\");\n\t\t\treturn;\n\t\t}\n\t\tif (!isReplayable(request.model, request.options)) {\n\t\t\tthis.stop(\"request cannot be replayed safely\");\n\t\t\treturn;\n\t\t}\n\t\tconst ttlMs = getPromptCacheTtlMs(request.model, request.options);\n\t\tif (ttlMs === undefined) {\n\t\t\tthis.stop(\n\t\t\t\trequest.options.cacheRetention === \"none\"\n\t\t\t\t\t? \"request disabled prompt caching\"\n\t\t\t\t\t: \"cache lifetime unavailable\",\n\t\t\t);\n\t\t\treturn;\n\t\t}\n\t\tconst delayMs = getCacheWarmingDelayMs(ttlMs);\n\t\tif (delayMs === undefined) {\n\t\t\tthis.stop(\"cache lifetime unavailable\");\n\t\t\treturn;\n\t\t}\n\t\tthis.run = {\n\t\t\t...request,\n\t\t\tisCurrent,\n\t\t\tttlMs,\n\t\t\tdelayMs,\n\t\t\trefreshDeadlineAt: 0,\n\t\t\tstartedAt: Date.now(),\n\t\t\tcontroller: new AbortController(),\n\t\t\tphase: \"streaming\",\n\t\t\tnextWarmAt: 0,\n\t\t\textensionOverride: false,\n\t\t};\n\t\tthis.schedule(this.run);\n\t}\n\n\tonAgentSettled(): void {\n\t\tconst run = this.run;\n\t\tif (!run) return;\n\t\tif (this.getMode() === \"streaming\") {\n\t\t\tthis.stop(\"agent run settled\");\n\t\t\treturn;\n\t\t}\n\t\trun.phase = \"idle\";\n\t\tconst deadline = run.startedAt + MAX_IDLE_WARMING_AGE_MS;\n\t\tif (run.nextWarmAt > deadline || Date.now() >= deadline) {\n\t\t\tthis.stop(\"30-minute idle safety limit reached\");\n\t\t}\n\t}\n\n\t/** Reconcile an active run after the persisted warming mode changes. */\n\tonModeChanged(): void {\n\t\tconst run = this.run;\n\t\tif (!run) return;\n\t\tconst reason = this.getModeStopReason(run);\n\t\tif (reason) this.stop(reason);\n\t}\n\n\tcancel(): void {\n\t\tthis.stop(\"inactive\");\n\t}\n\n\tprivate clearRun(): void {\n\t\tconst run = this.run;\n\t\tif (!run) return;\n\t\tthis.run = undefined;\n\t\tif (run.timer) clearTimeout(run.timer);\n\t\trun.controller.abort();\n\t}\n\n\tprivate stop(reason: string, stopped?: Pick<CacheWarmingStatus, \"decision\" | \"extensionOverride\">): void {\n\t\tthis.clearRun();\n\t\tthis.inactive = { state: \"inactive\", reason, ...stopped };\n\t}\n\n\tprivate schedule(run: ActiveRun): void {\n\t\trun.extensionOverride = false;\n\t\trun.nextWarmAt = Date.now() + run.delayMs;\n\t\t// A timer can run late after sleep or event-loop blockage. Keep half of\n\t\t// the planned pre-expiry margin for that delay and request dispatch; a\n\t\t// late refresh is likely a full-price cache write, not a cache warm.\n\t\trun.refreshDeadlineAt = run.nextWarmAt + Math.floor((run.ttlMs - run.delayMs) / 2);\n\t\tconst deadline = run.startedAt + (run.phase === \"idle\" ? MAX_IDLE_WARMING_AGE_MS : MAX_WARMING_AGE_MS);\n\t\tif (run.nextWarmAt > deadline || Date.now() >= deadline) {\n\t\t\tthis.stop(run.phase === \"idle\" ? \"30-minute idle safety limit reached\" : \"one-hour safety limit reached\");\n\t\t\treturn;\n\t\t}\n\t\trun.timer = setTimeout(() => void this.refresh(run), Math.max(0, run.nextWarmAt - Date.now()));\n\t\trun.timer.unref?.();\n\t}\n\n\tprivate async refresh(run: ActiveRun): Promise<void> {\n\t\trun.timer = undefined;\n\t\tif (!this.validateRun(run)) return;\n\t\tif (this.refreshDeadlineMissed(run)) return;\n\t\tconst decision = this.evaluate(run);\n\t\tconst { warmCost, missCost, continuationProbability } = decision;\n\t\tlet action = decision.action;\n\t\ttry {\n\t\t\taction = await this.decide({\n\t\t\t\ttype: \"cache_warming_decision\",\n\t\t\t\twarmCost,\n\t\t\t\tmissCost,\n\t\t\t\tcontinuationProbability,\n\t\t\t\taction,\n\t\t\t});\n\t\t} catch {\n\t\t\t// Extension failures fall back to pi's own decision.\n\t\t}\n\t\tif (!this.validateRun(run) || this.refreshDeadlineMissed(run)) return;\n\t\tconst extensionOverride = action !== decision.action;\n\t\tif (action === \"stop\") {\n\t\t\tconst reason = extensionOverride\n\t\t\t\t? \"stopped by extension\"\n\t\t\t\t: decision.economicsAvailable\n\t\t\t\t\t? \"expected savings below threshold\"\n\t\t\t\t\t: \"cache economics unavailable\";\n\t\t\tthis.stop(reason, { decision, extensionOverride });\n\t\t\treturn;\n\t\t}\n\n\t\trun.extensionOverride = extensionOverride;\n\t\ttry {\n\t\t\tconst message = await this.models\n\t\t\t\t.streamSimple(run.model, run.context, {\n\t\t\t\t\t...run.options,\n\t\t\t\t\tmaxTokens: 1,\n\t\t\t\t\tmaxRetries: 0,\n\t\t\t\t\tsignal: run.controller.signal,\n\t\t\t\t})\n\t\t\t\t.result();\n\t\t\tif (!this.validateRun(run)) return;\n\t\t\tif (message.stopReason !== \"error\" && message.stopReason !== \"aborted\") {\n\t\t\t\tconst entry = this.sessionManager.appendUsage(\n\t\t\t\t\t\"cache_warm\",\n\t\t\t\t\tmessage.provider,\n\t\t\t\t\tmessage.responseModel ?? message.model,\n\t\t\t\t\tmessage.usage,\n\t\t\t\t\textensionOverride ? \"extension override\" : undefined,\n\t\t\t\t);\n\t\t\t\tthis.onWarmed?.(entry);\n\t\t\t}\n\t\t} catch {\n\t\t\t// Cache warming is best-effort and must not affect the active agent run.\n\t\t}\n\t\tif (this.run === run) this.schedule(run);\n\t}\n\n\tprivate refreshDeadlineMissed(run: ActiveRun): boolean {\n\t\tif (Date.now() <= run.refreshDeadlineAt) return false;\n\t\tthis.stop(\"cache refresh deadline missed\");\n\t\treturn true;\n\t}\n\n\tprivate validateRun(run: ActiveRun): boolean {\n\t\tif (this.run !== run) return false;\n\t\tconst reason = this.getModeStopReason(run) ?? (!run.isCurrent() ? \"conversation context changed\" : undefined);\n\t\tif (!reason) return true;\n\t\tthis.stop(reason);\n\t\treturn false;\n\t}\n\n\tprivate getModeStopReason(run: ActiveRun): string | undefined {\n\t\tconst mode = this.getMode();\n\t\tif (mode === \"off\") return \"cache warming disabled\";\n\t\tif (mode === \"streaming\" && run.phase === \"idle\") return \"agent run settled\";\n\t\treturn undefined;\n\t}\n\n\tprivate evaluate(run: ActiveRun): CacheWarmingDecision {\n\t\tconst model = run.model;\n\t\tconst promptTokens = lastPromptTokens(this.sessionManager.getBranch());\n\t\tconst cacheHitCost = price(model, { cacheRead: promptTokens });\n\t\tconst cacheMissCost = price(\n\t\t\tmodel,\n\t\t\tmodel.cost.cacheWrite > 0 ? { cacheWrite: promptTokens } : { input: promptTokens },\n\t\t);\n\t\tconst warmCost = price(model, { cacheRead: promptTokens, output: 1 });\n\t\tconst missCost = Math.max(0, cacheMissCost - cacheHitCost);\n\t\tconst continuationProbability = run.phase === \"idle\" ? IDLE_CONTINUATION_PROBABILITY : 1;\n\t\tconst economicsAvailable = promptTokens > 0 && (cacheHitCost > 0 || cacheMissCost > 0);\n\t\tconst expectedSavings = continuationProbability * missCost - warmCost;\n\t\treturn {\n\t\t\tphase: run.phase,\n\t\t\twarmCost,\n\t\t\tmissCost,\n\t\t\tcontinuationProbability,\n\t\t\texpectedSavings,\n\t\t\teconomicsAvailable,\n\t\t\taction: expectedSavings >= CACHE_WARMING_MINIMUM_EXPECTED_SAVINGS ? \"warm\" : \"stop\",\n\t\t};\n\t}\n}\n\nfunction formatDollars(value: number): string {\n\treturn value < 0 ? `-$${Math.abs(value).toFixed(3)}` : `$${value.toFixed(3)}`;\n}\n\nfunction formatCacheWarmingEconomics(decision: CacheWarmingDecision): string {\n\tif (!decision.economicsAvailable) return \"cache economics unavailable\";\n\tconst probability = Math.round(decision.continuationProbability * 100);\n\tconst probabilityText =\n\t\tdecision.phase === \"streaming\"\n\t\t\t? `${probability}% continuation probability while agent is running`\n\t\t\t: `${probability}% continuation probability`;\n\tconst comparison = decision.action === \"warm\" ? \">=\" : \"<\";\n\treturn `${probabilityText}, expected savings ${formatDollars(decision.expectedSavings)} ${comparison} $${CACHE_WARMING_MINIMUM_EXPECTED_SAVINGS.toFixed(3)}`;\n}\n\nfunction formatCacheWarmingDecisionTime(nextWarmAt: number | undefined, now: number): string {\n\tif (nextWarmAt === undefined || nextWarmAt <= now) return \"Decision now\";\n\tlet remainingSeconds = Math.ceil((nextWarmAt - now) / 1000);\n\tconst hours = Math.floor(remainingSeconds / 3600);\n\tremainingSeconds %= 3600;\n\tconst minutes = Math.floor(remainingSeconds / 60);\n\tconst seconds = remainingSeconds % 60;\n\tconst parts: string[] = [];\n\tif (hours > 0) parts.push(`${hours}h`);\n\tif (minutes > 0) parts.push(`${minutes}m`);\n\tif (seconds > 0 || parts.length === 0) parts.push(`${seconds}s`);\n\treturn `Decision in ${parts.join(\" \")}`;\n}\n\n/** One-line status for `/session`. */\nexport function formatCacheWarmingStatus(status: CacheWarmingStatus, now = Date.now()): string {\n\tconst decision = status.decision;\n\t// A decision is attached once pi (or an extension) acted on it; \"inactive\"\n\t// without one never got that far.\n\tif (!decision || (status.state === \"inactive\" && !decision.economicsAvailable && !status.extensionOverride)) {\n\t\treturn `Inactive (${status.reason ?? \"unknown reason\"})`;\n\t}\n\tconst details = status.extensionOverride\n\t\t? `extension override, ${formatCacheWarmingEconomics(decision)}`\n\t\t: `${formatCacheWarmingEconomics(decision)} -> ${decision.action}`;\n\tif (status.state === \"inactive\") return `Stopped (${details})`;\n\tif (status.state === \"refreshing\") return `Warming cache (${details})`;\n\treturn `${formatCacheWarmingDecisionTime(status.nextWarmAt, now)} (${details})`;\n}\n\n/** One-line transcript text for persisted cache-warming usage. */\nexport function formatCacheWarmingUsage(entry: UsageEntry): string {\n\tconst note = entry.note ? ` (${entry.note})` : \"\";\n\tconst cost = entry.usage.cost.total.toFixed(6).replace(/(\\.\\d{3}\\d*?)0+$/, \"$1\");\n\treturn `Cache warmed${note}: $${cost}`;\n}\n"]}

package/dist/core/compaction/branch-summarization.d.ts

@@ -52,7 +52,7 @@ export interface GenerateBranchSummaryOptions {
     customInstructions?: string;
     /** If true, customInstructions replaces the default prompt instead of being appended */
     replaceInstructions?: boolean;
-    /** Tokens reserved for prompt + LLM response (default 16384) */
+    /** Tokens reserved when selecting branch history (default 16384) */
     reserveTokens?: number;
     /** Optional session stream function. Used to preserve SDK request behavior without mutating agent state. */
     streamFn?: StreamFn;

package/dist/core/compaction/branch-summarization.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"branch-summarization.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/branch-summarization.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,cAAc,EAAE,WAAW,EAAE,MAAM,uBAAuB,CAAC;AAEzE,OAAO,KAAK,EAAE,KAAK,EAAuB,KAAK,EAAE,MAAM,8BAA8B,CAAC;AACtF,OAAO,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAOnE,OAAO,KAAK,EAAE,sBAAsB,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAC;AAElF,OAAO,EAIN,KAAK,cAAc,EAInB,MAAM,YAAY,CAAC;AAMpB,MAAM,WAAW,mBAAmB;IACnC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,SAAS,CAAC,EAAE,MAAM,EAAE,CAAC;IACrB,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;IACzB,OAAO,CAAC,EAAE,OAAO,CAAC;IAClB,KAAK,CAAC,EAAE,MAAM,CAAC;CACf;AAED,qEAAqE;AACrE,MAAM,WAAW,oBAAoB;IACpC,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AAED,YAAY,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAEjD,MAAM,WAAW,iBAAiB;IACjC,mEAAmE;IACnE,QAAQ,EAAE,YAAY,EAAE,CAAC;IACzB,gDAAgD;IAChD,OAAO,EAAE,cAAc,CAAC;IACxB,yCAAyC;IACzC,WAAW,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,oBAAoB;IACpC,mDAAmD;IACnD,OAAO,EAAE,YAAY,EAAE,CAAC;IACxB,2DAA2D;IAC3D,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAC;CAChC;AAED,MAAM,WAAW,4BAA4B;IAC5C,qCAAqC;IACrC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IAClB,4BAA4B;IAC5B,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,oCAAoC;IACpC,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACjC,uDAAuD;IACvD,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC7B,oCAAoC;IACpC,MAAM,EAAE,WAAW,CAAC;IACpB,qDAAqD;IACrD,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,wFAAwF;IACxF,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,gEAAgE;IAChE,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,4GAA4G;IAC5G,QAAQ,CAAC,EAAE,QAAQ,CAAC;IACpB,+FAA+F;IAC/F,KAAK,CAAC,EAAE,WAAW,CAAC;IACpB,0EAA0E;IAC1E,SAAS,CAAC,EAAE,cAAc,CAAC;CAC3B;AAMD;;;;;;;;;;;GAWG;AACH,wBAAgB,8BAA8B,CAC7C,OAAO,EAAE,sBAAsB,EAC/B,SAAS,EAAE,MAAM,GAAG,IAAI,EACxB,QAAQ,EAAE,MAAM,GACd,oBAAoB,CAkCtB;AAoCD;;;;;;;;;;;;GAYG;AACH,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,WAAW,GAAE,MAAU,GAAG,iBAAiB,CAoDxG;AAwCD;;;;;GAKG;AACH,wBAAsB,qBAAqB,CAC1C,OAAO,EAAE,YAAY,EAAE,EACvB,OAAO,EAAE,4BAA4B,GACnC,OAAO,CAAC,mBAAmB,CAAC,CAoF9B","sourcesContent":["/**\n * Branch summarization for tree navigation.\n *\n * When navigating to a different point in the session tree, this generates\n * a summary of the branch being left so context isn't lost.\n */\n\nimport type { RetryCallbacks, RetryPolicy } from \"@earendil-works/pi-ai\";\nimport { contentText } from \"@earendil-works/pi-ai\";\nimport type { Model, SimpleStreamOptions, Usage } from \"@earendil-works/pi-ai/compat\";\nimport type { AgentMessage, StreamFn } from \"apex-code-agent-core\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport type { ReadonlySessionManager, SessionEntry } from \"../session-manager.ts\";\nimport { completeSummarization, estimateTokens, getSummarizationFailure } from \"./compaction.ts\";\nimport {\n\tcomputeFileLists,\n\tcreateFileOps,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tSUMMARIZATION_SYSTEM_PROMPT,\n\tserializeConversation,\n} from \"./utils.ts\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface BranchSummaryResult {\n\tsummary?: string;\n\tusage?: Usage;\n\treadFiles?: string[];\n\tmodifiedFiles?: string[];\n\taborted?: boolean;\n\terror?: string;\n}\n\n/** Details stored in BranchSummaryEntry.details for file tracking */\nexport interface BranchSummaryDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n}\n\nexport type { FileOperations } from \"./utils.ts\";\n\nexport interface BranchPreparation {\n\t/** Messages extracted for summarization, in chronological order */\n\tmessages: AgentMessage[];\n\t/** File operations extracted from tool calls */\n\tfileOps: FileOperations;\n\t/** Total estimated tokens in messages */\n\ttotalTokens: number;\n}\n\nexport interface CollectEntriesResult {\n\t/** Entries to summarize, in chronological order */\n\tentries: SessionEntry[];\n\t/** Common ancestor between old and new position, if any */\n\tcommonAncestorId: string | null;\n}\n\nexport interface GenerateBranchSummaryOptions {\n\t/** Model to use for summarization */\n\tmodel: Model<any>;\n\t/** API key for the model */\n\tapiKey?: string;\n\t/** Request headers for the model */\n\theaders?: Record<string, string>;\n\t/** Provider-scoped environment values for the model */\n\tenv?: Record<string, string>;\n\t/** Abort signal for cancellation */\n\tsignal: AbortSignal;\n\t/** Optional custom instructions for summarization */\n\tcustomInstructions?: string;\n\t/** If true, customInstructions replaces the default prompt instead of being appended */\n\treplaceInstructions?: boolean;\n\t/** Tokens reserved for prompt + LLM response (default 16384) */\n\treserveTokens?: number;\n\t/** Optional session stream function. Used to preserve SDK request behavior without mutating agent state. */\n\tstreamFn?: StreamFn;\n\t/** Retry policy for transient summarization errors. Reuses coding-agent's `settings.retry`. */\n\tretry?: RetryPolicy;\n\t/** Optional callbacks for retry reporting (e.g. TUI retry indicators). */\n\tcallbacks?: RetryCallbacks;\n}\n\n// ============================================================================\n// Entry Collection\n// ============================================================================\n\n/**\n * Collect entries that should be summarized when navigating from one position to another.\n *\n * Walks from oldLeafId back to the common ancestor with targetId, collecting entries\n * along the way. Does NOT stop at compaction boundaries - those are included and their\n * summaries become context.\n *\n * @param session - Session manager (read-only access)\n * @param oldLeafId - Current position (where we're navigating from)\n * @param targetId - Target position (where we're navigating to)\n * @returns Entries to summarize and the common ancestor\n */\nexport function collectEntriesForBranchSummary(\n\tsession: ReadonlySessionManager,\n\toldLeafId: string | null,\n\ttargetId: string,\n): CollectEntriesResult {\n\t// If no old position, nothing to summarize\n\tif (!oldLeafId) {\n\t\treturn { entries: [], commonAncestorId: null };\n\t}\n\n\t// Find common ancestor (deepest node that's on both paths)\n\tconst oldPath = new Set(session.getBranch(oldLeafId).map((e) => e.id));\n\tconst targetPath = session.getBranch(targetId);\n\n\t// targetPath is root-first, so iterate backwards to find deepest common ancestor\n\tlet commonAncestorId: string | null = null;\n\tfor (let i = targetPath.length - 1; i >= 0; i--) {\n\t\tif (oldPath.has(targetPath[i].id)) {\n\t\t\tcommonAncestorId = targetPath[i].id;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Collect entries from old leaf back to common ancestor\n\tconst entries: SessionEntry[] = [];\n\tlet current: string | null = oldLeafId;\n\n\twhile (current && current !== commonAncestorId) {\n\t\tconst entry = session.getEntry(current);\n\t\tif (!entry) break;\n\t\tentries.push(entry);\n\t\tcurrent = entry.parentId;\n\t}\n\n\t// Reverse to get chronological order\n\tentries.reverse();\n\n\treturn { entries, commonAncestorId };\n}\n\n// ============================================================================\n// Entry to Message Conversion\n// ============================================================================\n\n/**\n * Extract AgentMessage from a session entry.\n * Similar to getMessageFromEntry in compaction.ts but also handles compaction entries.\n */\nfunction getMessageFromEntry(entry: SessionEntry): AgentMessage | undefined {\n\tswitch (entry.type) {\n\t\tcase \"message\":\n\t\t\t// Skip tool results - context is in assistant's tool call\n\t\t\tif (entry.message.role === \"toolResult\") return undefined;\n\t\t\treturn entry.message;\n\n\t\tcase \"custom_message\":\n\t\t\treturn createCustomMessage(entry.customType, entry.content, entry.display, entry.details, entry.timestamp);\n\n\t\tcase \"branch_summary\":\n\t\t\treturn createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);\n\n\t\tcase \"compaction\":\n\t\t\treturn createCompactionSummaryMessage(entry.summary, entry.tokensBefore, entry.timestamp);\n\n\t\t// These don't contribute to conversation content\n\t\tcase \"thinking_level_change\":\n\t\tcase \"model_change\":\n\t\tcase \"custom\":\n\t\tcase \"label\":\n\t\tcase \"session_info\":\n\t\t\treturn undefined;\n\t}\n}\n\n/**\n * Prepare entries for summarization with token budget.\n *\n * Walks entries from NEWEST to OLDEST, adding messages until we hit the token budget.\n * This ensures we keep the most recent context when the branch is too long.\n *\n * Also collects file operations from:\n * - Tool calls in assistant messages\n * - Existing branch_summary entries' details (for cumulative tracking)\n *\n * @param entries - Entries in chronological order\n * @param tokenBudget - Maximum tokens to include (0 = no limit)\n */\nexport function prepareBranchEntries(entries: SessionEntry[], tokenBudget: number = 0): BranchPreparation {\n\tconst messages: AgentMessage[] = [];\n\tconst fileOps = createFileOps();\n\tlet totalTokens = 0;\n\n\t// First pass: collect file ops from ALL entries (even if they don't fit in token budget)\n\t// This ensures we capture cumulative file tracking from nested branch summaries\n\t// Only extract from pi-generated summaries (fromHook !== true), not extension-generated ones\n\tfor (const entry of entries) {\n\t\tif (entry.type === \"branch_summary\" && !entry.fromHook && entry.details) {\n\t\t\tconst details = entry.details as BranchSummaryDetails;\n\t\t\tif (Array.isArray(details.readFiles)) {\n\t\t\t\tfor (const f of details.readFiles) fileOps.read.add(f);\n\t\t\t}\n\t\t\tif (Array.isArray(details.modifiedFiles)) {\n\t\t\t\t// Modified files go into both edited and written for proper deduplication\n\t\t\t\tfor (const f of details.modifiedFiles) {\n\t\t\t\t\tfileOps.edited.add(f);\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t}\n\n\t// Second pass: walk from newest to oldest, adding messages until token budget\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tconst message = getMessageFromEntry(entry);\n\t\tif (!message) continue;\n\n\t\t// Extract file ops from assistant messages (tool calls)\n\t\textractFileOpsFromMessage(message, fileOps);\n\n\t\tconst tokens = estimateTokens(message);\n\n\t\t// Check budget before adding\n\t\tif (tokenBudget > 0 && totalTokens + tokens > tokenBudget) {\n\t\t\t// If this is a summary entry, try to fit it anyway as it's important context\n\t\t\tif (entry.type === \"compaction\" || entry.type === \"branch_summary\") {\n\t\t\t\tif (totalTokens < tokenBudget * 0.9) {\n\t\t\t\t\tmessages.unshift(message);\n\t\t\t\t\ttotalTokens += tokens;\n\t\t\t\t}\n\t\t\t}\n\t\t\t// Stop - we've hit the budget\n\t\t\tbreak;\n\t\t}\n\n\t\tmessages.unshift(message);\n\t\ttotalTokens += tokens;\n\t}\n\n\treturn { messages, fileOps, totalTokens };\n}\n\n// ============================================================================\n// Summary Generation\n// ============================================================================\n\nconst BRANCH_SUMMARY_PREAMBLE = `The user explored a different conversation branch before returning here.\nSummary of that exploration:\n\n`;\n\nconst BRANCH_SUMMARY_PROMPT = `Create a structured summary of this conversation branch for context when returning later.\n\nUse this EXACT format:\n\n## Goal\n[What was the user trying to accomplish in this branch?]\n\n## Constraints & Preferences\n- [Any constraints, preferences, or requirements mentioned]\n- [Or \"(none)\" if none were mentioned]\n\n## Progress\n### Done\n- [x] [Completed tasks/changes]\n\n### In Progress\n- [ ] [Work that was started but not finished]\n\n### Blocked\n- [Issues preventing progress, if any]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale]\n\n## Next Steps\n1. [What should happen next to continue this work]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages.`;\n\n/**\n * Generate a summary of abandoned branch entries.\n *\n * @param entries - Session entries to summarize (chronological order)\n * @param options - Generation options\n */\nexport async function generateBranchSummary(\n\tentries: SessionEntry[],\n\toptions: GenerateBranchSummaryOptions,\n): Promise<BranchSummaryResult> {\n\tconst {\n\t\tmodel,\n\t\tapiKey,\n\t\theaders,\n\t\tenv,\n\t\tsignal,\n\t\tcustomInstructions,\n\t\treplaceInstructions,\n\t\treserveTokens = 16384,\n\t\tstreamFn,\n\t\tretry,\n\t\tcallbacks,\n\t} = options;\n\n\t// Token budget = context window minus reserved space for prompt + response\n\tconst contextWindow = model.contextWindow || 128000;\n\tconst tokenBudget = contextWindow - reserveTokens;\n\n\tconst { messages, fileOps } = prepareBranchEntries(entries, tokenBudget);\n\n\tif (messages.length === 0) {\n\t\treturn { summary: \"No content to summarize\" };\n\t}\n\n\t// Transform to LLM-compatible messages, then serialize to text\n\t// Serialization prevents the model from treating it as a conversation to continue\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversation(llmMessages);\n\n\t// Build prompt\n\tlet instructions: string;\n\tif (replaceInstructions && customInstructions) {\n\t\tinstructions = customInstructions;\n\t} else if (customInstructions) {\n\t\tinstructions = `${BRANCH_SUMMARY_PROMPT}\\n\\nAdditional focus: ${customInstructions}`;\n\t} else {\n\t\tinstructions = BRANCH_SUMMARY_PROMPT;\n\t}\n\tconst promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n${instructions}`;\n\n\tconst summarizationMessages = [\n\t\t{\n\t\t\trole: \"user\" as const,\n\t\t\tcontent: [{ type: \"text\" as const, text: promptText }],\n\t\t\ttimestamp: Date.now(),\n\t\t},\n\t];\n\n\t// Call LLM for summarization. Prefer the session stream function so SDK\n\t// request behavior (timeouts, retries, attribution headers) stays consistent\n\t// without running through agent state/events. Retried via completeSummarization\n\t// so transient stream drops reuse the configured retry policy.\n\tconst context = { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages };\n\tconst requestOptions: SimpleStreamOptions = { apiKey, headers, env, signal, maxTokens: 2048 };\n\tconst response = await completeSummarization(model, context, requestOptions, streamFn, retry, callbacks);\n\n\t// Check if aborted or errored\n\tif (response.stopReason === \"aborted\") {\n\t\treturn { aborted: true };\n\t}\n\tconst failure = getSummarizationFailure(response, \"Branch summarization\");\n\tif (failure) {\n\t\treturn { error: failure };\n\t}\n\tif (response.content.some((block) => block.type === \"toolCall\")) {\n\t\treturn { error: \"Branch summarization attempted to call a tool\" };\n\t}\n\n\tlet summary = contentText(response.content);\n\n\t// Prepend preamble to provide context about the branch summary\n\tsummary = BRANCH_SUMMARY_PREAMBLE + summary;\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\treturn {\n\t\tsummary: summary || \"No summary generated\",\n\t\tusage: response.usage,\n\t\treadFiles,\n\t\tmodifiedFiles,\n\t};\n}\n"]}
+{"version":3,"file":"branch-summarization.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/branch-summarization.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,cAAc,EAAE,WAAW,EAAE,MAAM,uBAAuB,CAAC;AAEzE,OAAO,KAAK,EAAE,KAAK,EAAuB,KAAK,EAAE,MAAM,8BAA8B,CAAC;AACtF,OAAO,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAOnE,OAAO,KAAK,EAAE,sBAAsB,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAC;AAElF,OAAO,EAIN,KAAK,cAAc,EAInB,MAAM,YAAY,CAAC;AAMpB,MAAM,WAAW,mBAAmB;IACnC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,SAAS,CAAC,EAAE,MAAM,EAAE,CAAC;IACrB,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;IACzB,OAAO,CAAC,EAAE,OAAO,CAAC;IAClB,KAAK,CAAC,EAAE,MAAM,CAAC;CACf;AAED,qEAAqE;AACrE,MAAM,WAAW,oBAAoB;IACpC,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AAED,YAAY,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAEjD,MAAM,WAAW,iBAAiB;IACjC,mEAAmE;IACnE,QAAQ,EAAE,YAAY,EAAE,CAAC;IACzB,gDAAgD;IAChD,OAAO,EAAE,cAAc,CAAC;IACxB,yCAAyC;IACzC,WAAW,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,oBAAoB;IACpC,mDAAmD;IACnD,OAAO,EAAE,YAAY,EAAE,CAAC;IACxB,2DAA2D;IAC3D,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAC;CAChC;AAED,MAAM,WAAW,4BAA4B;IAC5C,qCAAqC;IACrC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IAClB,4BAA4B;IAC5B,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,oCAAoC;IACpC,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACjC,uDAAuD;IACvD,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC7B,oCAAoC;IACpC,MAAM,EAAE,WAAW,CAAC;IACpB,qDAAqD;IACrD,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,wFAAwF;IACxF,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,oEAAoE;IACpE,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,4GAA4G;IAC5G,QAAQ,CAAC,EAAE,QAAQ,CAAC;IACpB,+FAA+F;IAC/F,KAAK,CAAC,EAAE,WAAW,CAAC;IACpB,0EAA0E;IAC1E,SAAS,CAAC,EAAE,cAAc,CAAC;CAC3B;AAMD;;;;;;;;;;;GAWG;AACH,wBAAgB,8BAA8B,CAC7C,OAAO,EAAE,sBAAsB,EAC/B,SAAS,EAAE,MAAM,GAAG,IAAI,EACxB,QAAQ,EAAE,MAAM,GACd,oBAAoB,CAkCtB;AAoCD;;;;;;;;;;;;GAYG;AACH,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,WAAW,GAAE,MAAU,GAAG,iBAAiB,CAoDxG;AAwCD;;;;;GAKG;AACH,wBAAsB,qBAAqB,CAC1C,OAAO,EAAE,YAAY,EAAE,EACvB,OAAO,EAAE,4BAA4B,GACnC,OAAO,CAAC,mBAAmB,CAAC,CAsF9B","sourcesContent":["/**\n * Branch summarization for tree navigation.\n *\n * When navigating to a different point in the session tree, this generates\n * a summary of the branch being left so context isn't lost.\n */\n\nimport type { RetryCallbacks, RetryPolicy } from \"@earendil-works/pi-ai\";\nimport { contentText, normalizeContext } from \"@earendil-works/pi-ai\";\nimport type { Model, SimpleStreamOptions, Usage } from \"@earendil-works/pi-ai/compat\";\nimport type { AgentMessage, StreamFn } from \"apex-code-agent-core\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport type { ReadonlySessionManager, SessionEntry } from \"../session-manager.ts\";\nimport { completeSummarization, estimateTokens, getSummarizationFailure } from \"./compaction.ts\";\nimport {\n\tcomputeFileLists,\n\tcreateFileOps,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tSUMMARIZATION_SYSTEM_PROMPT,\n\tserializeConversation,\n} from \"./utils.ts\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface BranchSummaryResult {\n\tsummary?: string;\n\tusage?: Usage;\n\treadFiles?: string[];\n\tmodifiedFiles?: string[];\n\taborted?: boolean;\n\terror?: string;\n}\n\n/** Details stored in BranchSummaryEntry.details for file tracking */\nexport interface BranchSummaryDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n}\n\nexport type { FileOperations } from \"./utils.ts\";\n\nexport interface BranchPreparation {\n\t/** Messages extracted for summarization, in chronological order */\n\tmessages: AgentMessage[];\n\t/** File operations extracted from tool calls */\n\tfileOps: FileOperations;\n\t/** Total estimated tokens in messages */\n\ttotalTokens: number;\n}\n\nexport interface CollectEntriesResult {\n\t/** Entries to summarize, in chronological order */\n\tentries: SessionEntry[];\n\t/** Common ancestor between old and new position, if any */\n\tcommonAncestorId: string | null;\n}\n\nexport interface GenerateBranchSummaryOptions {\n\t/** Model to use for summarization */\n\tmodel: Model<any>;\n\t/** API key for the model */\n\tapiKey?: string;\n\t/** Request headers for the model */\n\theaders?: Record<string, string>;\n\t/** Provider-scoped environment values for the model */\n\tenv?: Record<string, string>;\n\t/** Abort signal for cancellation */\n\tsignal: AbortSignal;\n\t/** Optional custom instructions for summarization */\n\tcustomInstructions?: string;\n\t/** If true, customInstructions replaces the default prompt instead of being appended */\n\treplaceInstructions?: boolean;\n\t/** Tokens reserved when selecting branch history (default 16384) */\n\treserveTokens?: number;\n\t/** Optional session stream function. Used to preserve SDK request behavior without mutating agent state. */\n\tstreamFn?: StreamFn;\n\t/** Retry policy for transient summarization errors. Reuses coding-agent's `settings.retry`. */\n\tretry?: RetryPolicy;\n\t/** Optional callbacks for retry reporting (e.g. TUI retry indicators). */\n\tcallbacks?: RetryCallbacks;\n}\n\n// ============================================================================\n// Entry Collection\n// ============================================================================\n\n/**\n * Collect entries that should be summarized when navigating from one position to another.\n *\n * Walks from oldLeafId back to the common ancestor with targetId, collecting entries\n * along the way. Does NOT stop at compaction boundaries - those are included and their\n * summaries become context.\n *\n * @param session - Session manager (read-only access)\n * @param oldLeafId - Current position (where we're navigating from)\n * @param targetId - Target position (where we're navigating to)\n * @returns Entries to summarize and the common ancestor\n */\nexport function collectEntriesForBranchSummary(\n\tsession: ReadonlySessionManager,\n\toldLeafId: string | null,\n\ttargetId: string,\n): CollectEntriesResult {\n\t// If no old position, nothing to summarize\n\tif (!oldLeafId) {\n\t\treturn { entries: [], commonAncestorId: null };\n\t}\n\n\t// Find common ancestor (deepest node that's on both paths)\n\tconst oldPath = new Set(session.getBranch(oldLeafId).map((e) => e.id));\n\tconst targetPath = session.getBranch(targetId);\n\n\t// targetPath is root-first, so iterate backwards to find deepest common ancestor\n\tlet commonAncestorId: string | null = null;\n\tfor (let i = targetPath.length - 1; i >= 0; i--) {\n\t\tif (oldPath.has(targetPath[i].id)) {\n\t\t\tcommonAncestorId = targetPath[i].id;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Collect entries from old leaf back to common ancestor\n\tconst entries: SessionEntry[] = [];\n\tlet current: string | null = oldLeafId;\n\n\twhile (current && current !== commonAncestorId) {\n\t\tconst entry = session.getEntry(current);\n\t\tif (!entry) break;\n\t\tentries.push(entry);\n\t\tcurrent = entry.parentId;\n\t}\n\n\t// Reverse to get chronological order\n\tentries.reverse();\n\n\treturn { entries, commonAncestorId };\n}\n\n// ============================================================================\n// Entry to Message Conversion\n// ============================================================================\n\n/**\n * Extract AgentMessage from a session entry.\n * Similar to getMessageFromEntry in compaction.ts but also handles compaction entries.\n */\nfunction getMessageFromEntry(entry: SessionEntry): AgentMessage | undefined {\n\tswitch (entry.type) {\n\t\tcase \"message\":\n\t\t\t// Skip tool results - context is in assistant's tool call\n\t\t\tif (entry.message.role === \"toolResult\") return undefined;\n\t\t\treturn entry.message;\n\n\t\tcase \"custom_message\":\n\t\t\treturn createCustomMessage(entry.customType, entry.content, entry.display, entry.details, entry.timestamp);\n\n\t\tcase \"branch_summary\":\n\t\t\treturn createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);\n\n\t\tcase \"compaction\":\n\t\t\treturn createCompactionSummaryMessage(entry.summary, entry.tokensBefore, entry.timestamp);\n\n\t\t// These don't contribute to conversation content\n\t\tcase \"thinking_level_change\":\n\t\tcase \"model_change\":\n\t\tcase \"custom\":\n\t\tcase \"label\":\n\t\tcase \"session_info\":\n\t\t\treturn undefined;\n\t}\n}\n\n/**\n * Prepare entries for summarization with token budget.\n *\n * Walks entries from NEWEST to OLDEST, adding messages until we hit the token budget.\n * This ensures we keep the most recent context when the branch is too long.\n *\n * Also collects file operations from:\n * - Tool calls in assistant messages\n * - Existing branch_summary entries' details (for cumulative tracking)\n *\n * @param entries - Entries in chronological order\n * @param tokenBudget - Maximum tokens to include (0 = no limit)\n */\nexport function prepareBranchEntries(entries: SessionEntry[], tokenBudget: number = 0): BranchPreparation {\n\tconst messages: AgentMessage[] = [];\n\tconst fileOps = createFileOps();\n\tlet totalTokens = 0;\n\n\t// First pass: collect file ops from ALL entries (even if they don't fit in token budget)\n\t// This ensures we capture cumulative file tracking from nested branch summaries\n\t// Only extract from pi-generated summaries (fromHook !== true), not extension-generated ones\n\tfor (const entry of entries) {\n\t\tif (entry.type === \"branch_summary\" && !entry.fromHook && entry.details) {\n\t\t\tconst details = entry.details as BranchSummaryDetails;\n\t\t\tif (Array.isArray(details.readFiles)) {\n\t\t\t\tfor (const f of details.readFiles) fileOps.read.add(f);\n\t\t\t}\n\t\t\tif (Array.isArray(details.modifiedFiles)) {\n\t\t\t\t// Modified files go into both edited and written for proper deduplication\n\t\t\t\tfor (const f of details.modifiedFiles) {\n\t\t\t\t\tfileOps.edited.add(f);\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t}\n\n\t// Second pass: walk from newest to oldest, adding messages until token budget\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tconst message = getMessageFromEntry(entry);\n\t\tif (!message) continue;\n\n\t\t// Extract file ops from assistant messages (tool calls)\n\t\textractFileOpsFromMessage(message, fileOps);\n\n\t\tconst tokens = estimateTokens(message);\n\n\t\t// Check budget before adding\n\t\tif (tokenBudget > 0 && totalTokens + tokens > tokenBudget) {\n\t\t\t// If this is a summary entry, try to fit it anyway as it's important context\n\t\t\tif (entry.type === \"compaction\" || entry.type === \"branch_summary\") {\n\t\t\t\tif (totalTokens < tokenBudget * 0.9) {\n\t\t\t\t\tmessages.unshift(message);\n\t\t\t\t\ttotalTokens += tokens;\n\t\t\t\t}\n\t\t\t}\n\t\t\t// Stop - we've hit the budget\n\t\t\tbreak;\n\t\t}\n\n\t\tmessages.unshift(message);\n\t\ttotalTokens += tokens;\n\t}\n\n\treturn { messages, fileOps, totalTokens };\n}\n\n// ============================================================================\n// Summary Generation\n// ============================================================================\n\nconst BRANCH_SUMMARY_PREAMBLE = `The user explored a different conversation branch before returning here.\nSummary of that exploration:\n\n`;\n\nconst BRANCH_SUMMARY_PROMPT = `Create a structured summary of this conversation branch for context when returning later.\n\nUse this EXACT format:\n\n## Goal\n[What was the user trying to accomplish in this branch?]\n\n## Constraints & Preferences\n- [Any constraints, preferences, or requirements mentioned]\n- [Or \"(none)\" if none were mentioned]\n\n## Progress\n### Done\n- [x] [Completed tasks/changes]\n\n### In Progress\n- [ ] [Work that was started but not finished]\n\n### Blocked\n- [Issues preventing progress, if any]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale]\n\n## Next Steps\n1. [What should happen next to continue this work]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages.`;\n\n/**\n * Generate a summary of abandoned branch entries.\n *\n * @param entries - Session entries to summarize (chronological order)\n * @param options - Generation options\n */\nexport async function generateBranchSummary(\n\tentries: SessionEntry[],\n\toptions: GenerateBranchSummaryOptions,\n): Promise<BranchSummaryResult> {\n\tconst {\n\t\tmodel,\n\t\tapiKey,\n\t\theaders,\n\t\tenv,\n\t\tsignal,\n\t\tcustomInstructions,\n\t\treplaceInstructions,\n\t\treserveTokens = 16384,\n\t\tstreamFn,\n\t\tretry,\n\t\tcallbacks,\n\t} = options;\n\n\t// Token budget = context window minus reserved space for prompt + response\n\tconst contextWindow = model.contextWindow || 128000;\n\tconst tokenBudget = contextWindow - reserveTokens;\n\n\tconst { messages, fileOps } = prepareBranchEntries(entries, tokenBudget);\n\n\tif (messages.length === 0) {\n\t\treturn { summary: \"No content to summarize\" };\n\t}\n\n\t// Transform to LLM-compatible messages, then serialize to text\n\t// Serialization prevents the model from treating it as a conversation to continue\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversation(llmMessages);\n\n\t// Build prompt\n\tlet instructions: string;\n\tif (replaceInstructions && customInstructions) {\n\t\tinstructions = customInstructions;\n\t} else if (customInstructions) {\n\t\tinstructions = `${BRANCH_SUMMARY_PROMPT}\\n\\nAdditional focus: ${customInstructions}`;\n\t} else {\n\t\tinstructions = BRANCH_SUMMARY_PROMPT;\n\t}\n\tconst promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n${instructions}`;\n\n\tconst summarizationMessages = [\n\t\t{\n\t\t\trole: \"user\" as const,\n\t\t\tcontent: [{ type: \"text\" as const, text: promptText }],\n\t\t\ttimestamp: Date.now(),\n\t\t},\n\t];\n\n\tconst maxTokens = Math.min(4096, model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY);\n\n\t// Call LLM for summarization. Prefer the session stream function so SDK\n\t// request behavior (timeouts, retries, attribution headers) stays consistent\n\t// without running through agent state/events. Retried via completeSummarization\n\t// so transient stream drops reuse the configured retry policy.\n\tconst context = normalizeContext({ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages });\n\tconst requestOptions: SimpleStreamOptions = { apiKey, headers, env, signal, maxTokens };\n\tconst response = await completeSummarization(model, context, requestOptions, streamFn, retry, callbacks);\n\n\t// Check if aborted or errored\n\tif (response.stopReason === \"aborted\") {\n\t\treturn { aborted: true };\n\t}\n\tconst failure = getSummarizationFailure(response, \"Branch summarization\");\n\tif (failure) {\n\t\treturn { error: failure };\n\t}\n\tif (response.content.some((block) => block.type === \"toolCall\")) {\n\t\treturn { error: \"Branch summarization attempted to call a tool\" };\n\t}\n\n\tlet summary = contentText(response.content);\n\n\t// Prepend preamble to provide context about the branch summary\n\tsummary = BRANCH_SUMMARY_PREAMBLE + summary;\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\treturn {\n\t\tsummary: summary || \"No summary generated\",\n\t\tusage: response.usage,\n\t\treadFiles,\n\t\tmodifiedFiles,\n\t};\n}\n"]}

package/dist/core/compaction/branch-summarization.js

@@ -4,7 +4,7 @@
  * When navigating to a different point in the session tree, this generates
  * a summary of the branch being left so context isn't lost.
  */
-import { contentText } from "@earendil-works/pi-ai";
+import { contentText, normalizeContext } from "@earendil-works/pi-ai";
 import { convertToLlm, createBranchSummaryMessage, createCompactionSummaryMessage, createCustomMessage, } from "../messages.js";
 import { completeSummarization, estimateTokens, getSummarizationFailure } from "./compaction.js";
 import { computeFileLists, createFileOps, extractFileOpsFromMessage, formatFileOperations, SUMMARIZATION_SYSTEM_PROMPT, serializeConversation, } from "./utils.js";
@@ -216,12 +216,13 @@ export async function generateBranchSummary(entries, options) {
             timestamp: Date.now(),
         },
     ];
+    const maxTokens = Math.min(4096, model.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY);
     // Call LLM for summarization. Prefer the session stream function so SDK
     // request behavior (timeouts, retries, attribution headers) stays consistent
     // without running through agent state/events. Retried via completeSummarization
     // so transient stream drops reuse the configured retry policy.
-    const context = { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages };
-    const requestOptions = { apiKey, headers, env, signal, maxTokens: 2048 };
+    const context = normalizeContext({ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages });
+    const requestOptions = { apiKey, headers, env, signal, maxTokens };
     const response = await completeSummarization(model, context, requestOptions, streamFn, retry, callbacks);
     // Check if aborted or errored
     if (response.stopReason === "aborted") {