@cubicecho/agent-core 2.18.1 → 2.20.0

package/dist/agent-loop.js

@@ -1,16 +1,17 @@
 import { charsPerTokenFor } from "./calibration.js";
-import { capabilitiesFor, effortFor, } from "./capabilities.js";
+import { capabilitiesFor, effortFor, modelCapabilitiesFor, } from "./capabilities.js";
 import { firstTokenMs, getClient, NO_KEY, timeoutMs } from "./client.js";
 import { continueTurn } from "./continuation.js";
 import { errorMessage } from "./errors.js";
 import { runMetrics } from "./events.js";
-import { gather, notify, turnIndex, turnMessages, withContext, } from "./hooks.js";
-import { toolsChars } from "./retry.js";
+import { configureHooks, gather, notify, turnIndex, turnMessages, withContext, } from "./hooks.js";
+import { rebaseLedger, recordRequest } from "./ledger.js";
+import { contextTokens, toolsChars } from "./retry.js";
 import { runTurn } from "./run-turn.js";
 import { relaxTools, sanitizeTools } from "./schema-compat.js";
 import { askJson, tryAsk } from "./side-task.js";
 import { parseToolArguments, recoverToolCalls } from "./tool-calls.js";
-import { catalogPrompt, expandNames, inCatalog, LOAD_TOOLS, LOAD_TOOLS_DEFINITION, loadedTools, loadResult, MAX_PER_LOAD, orderTools, PRESELECT_SCHEMA, preselectByKeywords, preselectInput, preselection, preselectSystem, requestedNames, } from "./tool-loading.js";
+import { CALL_TOOL, catalogPrompt, expandNames, inCatalog, LOAD_TOOLS, LOAD_TOOLS_DEFINITION, loadedTools, loadResult, MAX_PER_LOAD, orderTools, PRESELECT_SCHEMA, PROXY_TOOLS, preselectByKeywords, preselectInput, preselection, preselectSystem, proxiedCall, proxyCatalogPrompt, proxyLoadResult, requestedNames, shownCall, } from "./tool-loading.js";
 /**
  * The loop above a turn: send, run the tools the model asked for, send again, until it stops
  * asking.
@@ -206,17 +207,32 @@ function cacheDiagnosis(previous, messages, tools, usage) {
  * prompt unchanged from step to step, a load adds to the tool array — which every request sends
  * in the stable order `toolOrder` asks for — and
  * a catalogued tool called without being loaded is loaded and run rather than refused, and a
- * preselection shapes the first step. A turn cut off at `maxTokens` is said so as a notice,
- * because it otherwise reads exactly like a finished one — or, given `maxContinuations`, is
+ * preselection shapes the first step unless `preselectRouting` says to append it. A turn cut off
+ * at `maxTokens` is said so as a notice, because it otherwise reads exactly like a finished one — or, given `maxContinuations`, is
  * continued first. Every turn ends in a `usage` event carrying the turn's own report, with the
- * cache compared against the request before it; see `TurnUsage` and `runMetrics`.
+ * cache compared against the request before it and the request broken down by what filled it; see
+ * `TurnUsage` and `runMetrics`.
+ *
+ * Each step's reported prompt goes on a token ledger, in the same epoch as the step before where
+ * the request only appended to it and in a new one where it did not — a `beforeStep` that rewrote
+ * the transcript, a tool array a load grew — so that what the messages between two steps cost can
+ * be read back as a subtraction. The prompt recorded is the first request's, before any
+ * continuation, and a step that reported none, or reported a cache count above its prompt, is not
+ * recorded. The ledger comes back in the result and on every `usage` event; see `tokensBetween`.
+ *
+ * Proxied discovery is the same catalogue behind a tool array that never moves: every step
+ * declares `PROXY_TOOLS` and nothing else, a load answers with the definitions themselves, and a
+ * `call_tool` is dispatched, counted and reported as the tool it names. A preselection has no
+ * tool array to go into, so it is written into the transcript as a `load_tools` exchange after
+ * the question. `call_tool` is honoured on demand too, unless the host has a tool of that name —
+ * a session switched out of proxied mode still has it in its history, and the model copies it.
  *
  * @param options The config, transcript, tools and dispatcher, plus the optional hooks, events
  * and cancellation. See `AgentLoopOptions`.
  */
 export async function runAgentLoop(options) {
     const { config, system = "", tools = [], catalog = [], dispatch, hooks, signal } = options;
-    const { onTurn, beforeStep, parallel = false, recoverToolCalls: recover = true, maxContinuations = 0, toolOrder = true, dedupeToolCalls = true, } = options;
+    const { onRequest, onTurn, onToolCall, onToolResult, beforeStep, parallel = false, recoverToolCalls: recover = true, maxContinuations = 0, toolOrder = true, dedupeToolCalls = true, preselectRouting = "exclusive", } = options;
     const dedupable = typeof dedupeToolCalls === "function" ? dedupeToolCalls : () => dedupeToolCalls;
     const started = Date.now();
     // What the loop emitted, less the token deltas, for `runMetrics` at the end. Stamped here rather
@@ -245,11 +261,16 @@ export async function runAgentLoop(options) {
     const supports = capabilitiesFor(config.baseUrl, config.apiKey);
     const maxRetries = Math.max(0, Number(config.maxRetries) || 0);
     const notice = (text) => onEvent({ kind: "notice", text });
-    const onDemand = config.toolDiscovery === "ondemand" && catalog.length > 0;
-    const loaded = new Set(onDemand ? (options.loaded ?? []) : []);
+    // On demand, but with a tool array that never changes: definitions come back as `load_tools`
+    // results and run through `call_tool`. `onDemand` is true of both.
+    const proxied = config.toolDiscovery === "proxy" && catalog.length > 0;
+    const onDemand = proxied || (config.toolDiscovery === "ondemand" && catalog.length > 0);
+    // Proxied, `loaded` is what this run has put a definition in the history for, and starts empty.
+    const loaded = new Set(onDemand && !proxied ? (options.loaded ?? []) : []);
     const preselected = onDemand ? [...(options.preselected ?? [])] : [];
-    for (const name of preselected)
-        loaded.add(name);
+    if (!proxied)
+        for (const name of preselected)
+            loaded.add(name);
     const used = new Set();
     const definitions = new Map();
     for (const tool of tools) {
@@ -257,6 +278,8 @@ export async function runAgentLoop(options) {
             definitions.set(tool.function.name, tool);
         }
     }
+    // Looked through unless the host has a tool by that name, which is then the host's to answer.
+    const proxies = onDemand && !definitions.has(CALL_TOOL);
     // In the order the names are given, not the order of `tools`: `loaded` is a set, which iterates
     // in the order things were added, so a load appends and never reshuffles what went before.
     const byName = (names) => [...names].flatMap((name) => definitions.get(name) ?? []);
@@ -272,32 +295,84 @@ export async function runAgentLoop(options) {
             maxTokens: hooks.maxTokens,
         })
         : { context: "", notes: [] };
+    // Read once, so what the result hands back is what every step of this run said.
+    const preface = hooks?.preface ?? configureHooks().preface;
     const usage = { prompt: 0, completion: 0, total: 0, cached: 0 };
     const toolCalls = [];
     const loads = { toolsLoaded: 0, redundantLoads: 0, unknownToolNames: 0 };
     let previous;
+    let ledger = options.ledger ?? [];
+    // The request the ledger's last entry came from, which is what the next one is measured against.
+    // Nothing before this run's first: an earlier run's head is not something the loop can see.
+    let measured;
+    // Proxied, a shortlist has nowhere to go but the history: the tool array is fixed, so it is
+    // answered as though the model had loaded it, and the definitions sit after the question.
+    const shortlist = proxied
+        ? byName(new Set(preselected)).filter((tool) => tool.type === "function" && inCatalog(catalog, tool.function.name))
+        : [];
+    if (shortlist.length) {
+        const names = shortlist.map((tool) => (tool.type === "function" ? tool.function.name : ""));
+        // Numbered by where the call lands, so two questions in one transcript do not share an id.
+        const id = `preselect-${messages.length}`;
+        const args = JSON.stringify({ names });
+        // Held to its own length rather than `MAX_PER_LOAD`: that cap is for a model choosing, and a
+        // host that shortlisted more has already chosen.
+        const content = proxyLoadResult(expandNames(names, catalog, names.length), catalog, shortlist);
+        // Announced as a call the model made would be, so a host pairing calls with results by id
+        // shows this one like the rest.
+        onEvent({ kind: "tool-call", id, name: LOAD_TOOLS, text: preview(args) });
+        onToolCall?.({ id, name: LOAD_TOOLS, args: { names }, raw: args });
+        onEvent({ kind: "tool-result", id, name: LOAD_TOOLS, ok: true, text: preview(content) });
+        onToolResult?.({ id, name: LOAD_TOOLS, ok: true, content });
+        toolCalls.push({ id, name: LOAD_TOOLS, ok: true });
+        loads.toolsLoaded += names.length;
+        for (const name of names)
+            loaded.add(name);
+        messages.push({
+            role: "assistant",
+            content: null,
+            tool_calls: [{ id, type: "function", function: { name: LOAD_TOOLS, arguments: args } }],
+        }, { role: "tool", tool_call_id: id, content });
+    }
     for (let step = 0; step < config.maxToolIterations; step++) {
         // A stop aborts the request in flight, but a tool call already handed off runs to its own
         // end — so the signal is read between steps as well.
         signal?.throwIfAborted();
-        messages = (await beforeStep?.(messages, step)) ?? messages;
+        const before = messages;
+        const rewritten = await beforeStep?.(messages, step, {
+            ...(previous && previous.prompt > 0 ? { used: previous.prompt } : {}),
+            limit: config.contextLength ?? 0,
+            ledger,
+        });
+        // A rewrite may have folded a definition away, and a load answered "already loaded" would
+        // then point at nothing. Forgetting costs a definition sent twice at worst.
+        if (proxied && rewritten && rewritten !== messages)
+            loaded.clear();
+        messages = rewritten ?? messages;
+        ledger = rebaseLedger(ledger, before, messages);
         onEvent({ kind: "turn", text: `turn ${step + 1}` });
-        const routed = preselected.length > 0 && step === 0;
+        // Appended, the shortlist is already in `loaded` — after what was carried, and once — so the
+        // first step needs nothing of its own. Proxied, it is in the history.
+        const routed = !proxied && preselectRouting !== "append" && preselected.length > 0 && step === 0;
         // Ordered here rather than left to `buildBody`, so `names` below is what the request actually
         // declared — a diagnosis reading an order the server never saw calls an untouched tool array
         // `tools-changed`.
         const declared = orderTools(routed
             ? byName(new Set(preselected))
-            : onDemand
-                ? loadedTools([LOAD_TOOLS_DEFINITION], byName(loaded))
-                : tools, toolOrder);
+            : proxied
+                ? [...PROXY_TOOLS]
+                : onDemand
+                    ? loadedTools([LOAD_TOOLS_DEFINITION], byName(loaded))
+                    : tools, toolOrder);
         // Unmarked, so the system prompt is the same text on every step and a load does not throw
         // away the cache for the whole transcript. What is loaded is said in `declared` and in the
-        // `load_tools` result instead. The preselected first step is the one exception, by design.
-        const prompt = onDemand && !routed ? `${system}\n\n${catalogPrompt(catalog)}`.trim() : system;
+        // `load_tools` result instead. The preselected first step is the one exception, by design,
+        // and `preselectRouting: "append"` is the way out of it.
+        const catalogue = proxied ? proxyCatalogPrompt(catalog) : catalogPrompt(catalog);
+        const prompt = onDemand && !routed ? `${system}\n\n${catalogue}`.trim() : system;
         const request = [
             ...(prompt ? [{ role: "system", content: prompt }] : []),
-            ...withContext(messages, question ? messages.indexOf(question) : -1, gathered.context, hooks?.preface),
+            ...withContext(messages, question ? messages.indexOf(question) : -1, gathered.context, preface),
         ];
         const build = (supported, refused) => buildBody(config, supported, refused, request, declared, toolOrder);
         const turnOptions = {
@@ -315,15 +390,37 @@ export async function runAgentLoop(options) {
             onThinking: (text) => onEvent({ kind: "thinking", text }),
             onOutput: (text) => onEvent({ kind: "output", text }),
         };
+        // Built the way `negotiate` is about to build the first attempt — the same flags, the same
+        // model's refusals — so what the host is told is what is sent unless something is refused.
+        if (onRequest) {
+            const opening = build(supports, modelCapabilitiesFor(supports, config.model));
+            onRequest({ messages: opening.messages, tools: opening.tools ?? [], step });
+        }
         const first = await runTurn(client, supports, build, turnOptions);
         const names = declared.map((tool) => (tool.type === "function" ? tool.function.name : ""));
         // Compared before any continuation is joined on: the cache a request meets is the one its own
         // prompt found, and a continuation's prompt is this request's plus the reply so far.
+        // The breakdown measures the array `toolSchemaTokens` does, so where no prompt was reported
+        // the two give one number for the tool block rather than two.
+        const charsPerToken = charsPerTokenFor(supports, config.model);
         Object.assign(first.usage, cacheDiagnosis(previous, request, names, first.usage), {
             toolsDeclared: declared.length,
-            toolSchemaTokens: Math.ceil(toolsChars(declared) / charsPerTokenFor(supports, config.model)),
+            toolSchemaTokens: Math.ceil(toolsChars(declared) / charsPerToken),
+            context: contextTokens({ model: config.model, stream: true, messages: request, tools: declared }, { charsPerToken, promptTokens: first.usage.prompt }),
         });
         const firstPrompt = first.usage.prompt;
+        // Recorded from the first request for the same reason. A cache count above the prompt is a
+        // server reporting its prompt net of the cache, whose differences would measure nothing.
+        const sent = {
+            messages: request,
+            tools: names,
+            prompt: first.usage.cached > firstPrompt ? 0 : firstPrompt,
+            through: messages.length - 1,
+        };
+        const entered = recordRequest(ledger, measured, sent);
+        if (entered !== ledger)
+            measured = sent;
+        ledger = entered;
         const turn = maxContinuations > 0
             ? await continueTurn(client, supports, build, first, { ...turnOptions, maxContinuations })
             : first;
@@ -342,6 +439,7 @@ export async function runAgentLoop(options) {
                 totalTokens: usage.total,
                 cachedTokens: usage.cached,
                 turn: { ...turn.usage, finishReason: turn.finishReason },
+                ledger,
             },
         });
         if (turn.finishReason === "length") {
@@ -354,7 +452,7 @@ export async function runAgentLoop(options) {
         if (recover && !calls.length && content && (tools.length > 0 || onDemand)) {
             const names = tools.flatMap((tool) => (tool.type === "function" ? [tool.function.name] : []));
             const recovered = recoverToolCalls(content, {
-                names: onDemand ? [...names, LOAD_TOOLS] : names,
+                names: onDemand ? [...names, LOAD_TOOLS, ...(proxies ? [CALL_TOOL] : [])] : names,
             });
             if (recovered.toolCalls.length) {
                 calls = recovered.toolCalls;
@@ -391,11 +489,12 @@ export async function runAgentLoop(options) {
                 : {}),
         });
         if (!calls.length) {
+            let afterTurn = Promise.resolve([]);
             if (hooks) {
                 const at = question ? messages.indexOf(question) : -1;
                 // Not awaited: the answer is ready, and remembering it is not something to hold it for.
-                // `notify` never rejects.
-                void notify(hooks.run, "afterTurn", {
+                // Handed back instead, for a host that wants its notes.
+                afterTurn = notify(hooks.run, "afterTurn", {
                     ...hooks.context,
                     reply: turn.content,
                     ...(at >= 0
@@ -414,47 +513,87 @@ export async function runAgentLoop(options) {
                 messages,
                 usage,
                 toolCalls,
-                loaded: [...loaded],
+                loaded: proxied ? [] : [...loaded],
                 used: [...used],
                 notes: gathered.notes,
+                context: gathered.context,
+                preface,
+                afterTurn,
                 metrics: {
                     ...metrics,
                     ...(onDemand ? loads : {}),
                     wallMs: Date.now() - started,
                     outcome: turn.finishReason === "length" ? "truncated" : "answered",
                 },
+                ledger,
             };
         }
         // Per step, not per run: the answer to a call made two steps ago was true before the tools in
         // between ran, and the file the model read may be the file it has since written.
         const answered = new Map();
         const run = async ({ call, args, error: unreadable, normal }) => {
-            const { name, arguments: raw } = call.function;
-            onEvent({ kind: "tool-call", name, text: preview(raw) });
+            const proxy = proxies && call.function.name === CALL_TOOL;
+            // A `call_tool` is shown, counted and dispatched as the tool it names. Read from the repaired
+            // arguments where there are any, since the model's own may be almost-JSON.
+            const { name, input: raw } = proxy
+                ? shownCall(CALL_TOOL, args ? normal : call.function.arguments)
+                : { name: call.function.name, input: call.function.arguments };
+            // Looked through before the call is announced, so a host is told of the tool it names with
+            // that tool's arguments. A name outside the catalogue is refused here, so `call_tool` reaches
+            // nothing a load could not, and the refusal is the call's answer below.
+            let inner = args ?? {};
+            let refused;
+            if (proxy && args) {
+                try {
+                    inner = proxiedCall(args, catalog).input;
+                }
+                catch (error) {
+                    inner = {};
+                    refused = { error };
+                }
+            }
+            // Built for every call rather than only the dispatched ones, so a host hears of the calls
+            // the loop answers itself in the same shape. Arguments that could not be read are none.
+            const request = { id: call.id, name, args: inner, raw };
+            onEvent({ kind: "tool-call", id: call.id, name, text: preview(raw) });
+            onToolCall?.(request);
             let content;
             let ok = true;
             try {
                 if (!args)
                     throw unreadable;
-                if (onDemand && name === LOAD_TOOLS) {
+                // By the name the model called, not the one a `call_tool` wraps: `load_tools` is not in
+                // the catalogue, and one reached through `call_tool` is refused below like any other.
+                if (onDemand && call.function.name === LOAD_TOOLS) {
                     const resolved = expandNames(requestedNames(args), catalog);
-                    content = loadResult(resolved, catalog, loaded);
-                    for (const hit of resolved.matched)
+                    content = proxied
+                        ? proxyLoadResult(resolved, catalog, byName(resolved.matched), loaded)
+                        : loadResult(resolved, catalog, loaded);
+                    // Proxied, a load is only of what the result could define: a catalogued name the host
+                    // gave no definition for was not loaded, and is not counted or remembered as though it was.
+                    const hits = proxied
+                        ? resolved.matched.filter((hit) => definitions.has(hit))
+                        : resolved.matched;
+                    for (const hit of hits)
                         loads[loaded.has(hit) ? "redundantLoads" : "toolsLoaded"]++;
                     loads.unknownToolNames += resolved.unknown.length;
-                    for (const hit of resolved.matched)
+                    for (const hit of hits)
                         loaded.add(hit);
-                    ok = resolved.matched.length > 0;
+                    ok = hits.length > 0;
                 }
                 else {
+                    if (refused)
+                        throw refused.error;
                     // A model that skips `load_tools` and calls a catalogued tool by name is right about
-                    // what it wants; load it and run it rather than refusing.
-                    if (onDemand && inCatalog(catalog, name))
+                    // what it wants; load it and run it rather than refusing. Not proxied, where loaded
+                    // means its definition is in the history, and this call put none there.
+                    if (onDemand && !proxied && inCatalog(catalog, name))
                         loaded.add(name);
                     used.add(name);
-                    const request = { id: call.id, name, args, raw };
+                    // Keyed on the inner call, so one made through `call_tool` and one made natively share.
+                    const key = `${name}\0${proxy ? JSON.stringify(inner) : normal}`;
                     content = dedupable(request)
-                        ? await once(answered, `${name}\0${normal}`, () => dispatch(request, signal))
+                        ? await once(answered, key, () => dispatch(request, signal))
                         : await dispatch(request, signal);
                 }
             }
@@ -464,8 +603,10 @@ export async function runAgentLoop(options) {
                 content = errorMessage(error);
                 ok = false;
             }
-            onEvent({ kind: "tool-result", name, ok, text: preview(content) });
-            return { id: call.id, name, ok, content };
+            onEvent({ kind: "tool-result", id: call.id, name, ok, text: preview(content) });
+            const result = { id: call.id, name, ok, content };
+            onToolResult?.(result);
+            return result;
         };
         const outcomes = [];
         if (parallel) {
@@ -479,7 +620,7 @@ export async function runAgentLoop(options) {
             }
         }
         for (const { id, name, ok, content } of outcomes) {
-            toolCalls.push({ name, ok });
+            toolCalls.push({ id, name, ok });
             messages.push({ role: "tool", tool_call_id: id, content });
         }
     }

package/dist/compaction.d.ts

@@ -42,7 +42,9 @@ export interface PruneOptions {
  * after it, and by then the model has usually taken what it wanted from it; the stub keeps the
  * call answered — a call with no result is a malformed transcript — and says how much was there,
  * so a model that does need it again knows to ask. Short results are kept, since a stub saves
- * nothing on them. Returns the same array when there was nothing to clear. Rewrites the prefix;
+ * nothing on them, and so is a proxied `load_tools` result carrying definitions, which is the
+ * only copy of them the model has (`holdsDefinitions`); either still counts as one of the latest
+ * few. Returns the same array when there was nothing to clear. Rewrites the prefix;
  * see the module comment on when to run it.
  *
  * @param messages The transcript. Not written to.
@@ -62,6 +64,17 @@ export interface CompactionOptions {
     compactAt?: number;
     /** The fraction of `limit` the kept tail may fill. `KEEP_RATIO` by default. */
     keepRatio?: number;
+    /**
+     * The fraction of `limit` the whole request should come down to, in place of `keepRatio`: the
+     * cut is the earliest that leaves `used`, less what is folded, plus `summaryTokens`, at or under
+     * it. Absent, or not above zero, plans by `keepRatio` as before.
+     */
+    target?: number;
+    /**
+     * What the summary is assumed to cost when planning to a `target` — the ceiling its writer is
+     * held to, since the plan is made before it is written. 1024 by default, which is `summariser`'s.
+     */
+    summaryTokens?: number;
     /** One message's tokens. `messageTokens` by default, divided by `charsPerToken`. */
     estimate?: (message: Message) => number;
     /**
@@ -91,6 +104,12 @@ export interface CompactionPlan {
     toSummarise: Message[];
     /** The summary an earlier compaction left, which this one continues. */
     previous?: string;
+    /**
+     * What the request is expected to cost once folded, in tokens: what was in use, less what is
+     * summarised, plus the summary's ceiling. Only on a plan made to a `target`, and above the
+     * target where no legal cut reaches it.
+     */
+    after?: number;
 }
 /**
  * Where to fold a transcript that has grown into its window, or `undefined` when it should not be.
@@ -107,11 +126,20 @@ export interface CompactionPlan {
  * transcript — its system prompt a separate argument, no summary message in the array at all —
  * knows them exactly, and passes `from` and `previous` instead of hoping the scan agrees.
  *
+ * `keepRatio` bounds the kept tail alone, and the system prompt, the tool schemas and the summary
+ * sit in the window on top of it uncounted. Given a `target` the plan is made against what the
+ * request will be afterwards instead: the cut is the first user message at which `used`, less the
+ * folded messages, plus the summary's ceiling, is at or under that share of the window. Where no
+ * cut gets there the plan folds up to the last user message, which is as near as a legal one
+ * comes, and `after` says how far short it fell. An earlier summary the new one replaces is not
+ * credited, so `after` errs high by at most that much. It is only as good as `estimate` — see
+ * `estimateFrom` for one that is measured.
+ *
  * @param messages The transcript, system prompts included if the caller keeps them in it.
  * @param options The window, what is in use, the ratios, and where the last fold ended. See
  * `CompactionOptions`.
  */
-export declare function planCompaction(messages: Message[], { limit, used, compactAt, keepRatio, charsPerToken, estimate, from: givenFrom, previous: givenPrevious, }: CompactionOptions): CompactionPlan | undefined;
+export declare function planCompaction(messages: Message[], { limit, used, compactAt, keepRatio, target, summaryTokens, charsPerToken, estimate, from: givenFrom, previous: givenPrevious, }: CompactionOptions): CompactionPlan | undefined;
 /**
  * What the summariser is handed for a plan: the earlier summary if there was one, then each
  * message as its role and at most 4000 characters of its text.

package/dist/compaction.js

@@ -1,10 +1,13 @@
 import { consult, notify, turnMessages, } from "./hooks.js";
 import { messageTokens } from "./retry.js";
 import { ask } from "./side-task.js";
+import { holdsDefinitions } from "./tool-loading.js";
 /** The fraction of the window in use before a summary is worth its own round trip. */
 export const COMPACT_AT = 0.75;
 /** The fraction of the window the kept tail may fill, leaving room for the run to grow again. */
 export const KEEP_RATIO = 0.35;
+/** The ceiling `summariser` gives a summary, and what `planCompaction` assumes one will cost. */
+const SUMMARY_TOKENS = 1024;
 /** How much of any one message the summariser is shown. A pasted file is not worth it whole. */
 const SUMMARY_SLICE = 4000;
 /**
@@ -63,7 +66,9 @@ const systemHead = (messages) => {
  * after it, and by then the model has usually taken what it wanted from it; the stub keeps the
  * call answered — a call with no result is a malformed transcript — and says how much was there,
  * so a model that does need it again knows to ask. Short results are kept, since a stub saves
- * nothing on them. Returns the same array when there was nothing to clear. Rewrites the prefix;
+ * nothing on them, and so is a proxied `load_tools` result carrying definitions, which is the
+ * only copy of them the model has (`holdsDefinitions`); either still counts as one of the latest
+ * few. Returns the same array when there was nothing to clear. Rewrites the prefix;
  * see the module comment on when to run it.
  *
  * @param messages The transcript. Not written to.
@@ -81,6 +86,10 @@ export function pruneToolResults(messages, { keepLast = 5, maxChars = 256 } = {}
         const text = textOf(message.content);
         if (text.length <= maxChars || text.startsWith("[result cleared"))
             continue;
+        // A proxied load's result is the schema itself; stubbed, the tool is one the model can still
+        // name and no longer call correctly.
+        if (holdsDefinitions(text))
+            continue;
         out ??= [...messages];
         out[at] = {
             ...message,
@@ -104,11 +113,20 @@ export function pruneToolResults(messages, { keepLast = 5, maxChars = 256 } = {}
  * transcript — its system prompt a separate argument, no summary message in the array at all —
  * knows them exactly, and passes `from` and `previous` instead of hoping the scan agrees.
  *
+ * `keepRatio` bounds the kept tail alone, and the system prompt, the tool schemas and the summary
+ * sit in the window on top of it uncounted. Given a `target` the plan is made against what the
+ * request will be afterwards instead: the cut is the first user message at which `used`, less the
+ * folded messages, plus the summary's ceiling, is at or under that share of the window. Where no
+ * cut gets there the plan folds up to the last user message, which is as near as a legal one
+ * comes, and `after` says how far short it fell. An earlier summary the new one replaces is not
+ * credited, so `after` errs high by at most that much. It is only as good as `estimate` — see
+ * `estimateFrom` for one that is measured.
+ *
  * @param messages The transcript, system prompts included if the caller keeps them in it.
  * @param options The window, what is in use, the ratios, and where the last fold ended. See
  * `CompactionOptions`.
  */
-export function planCompaction(messages, { limit, used, compactAt = COMPACT_AT, keepRatio = KEEP_RATIO, charsPerToken, estimate = (message) => messageTokens(message, { charsPerToken }), from: givenFrom, previous: givenPrevious, }) {
+export function planCompaction(messages, { limit, used, compactAt = COMPACT_AT, keepRatio = KEEP_RATIO, target, summaryTokens = SUMMARY_TOKENS, charsPerToken, estimate = (message) => messageTokens(message, { charsPerToken }), from: givenFrom, previous: givenPrevious, }) {
     if (!(limit > 0))
         return undefined;
     const cost = used ?? messages.reduce((total, message) => total + estimate(message), 0);
@@ -117,20 +135,42 @@ export function planCompaction(messages, { limit, used, compactAt = COMPACT_AT,
     const head = systemHead(messages);
     const from = Math.min(Math.max(givenFrom ?? head.from, 0), messages.length);
     const previous = givenPrevious ?? head.previous;
-    const budget = limit * keepRatio;
-    let kept = 0;
     let cut = messages.length;
-    for (let at = messages.length - 1; at > from; at--) {
-        kept += estimate(messages[at]);
-        if (kept > budget)
-            break;
-        cut = at;
+    let after;
+    if (target !== undefined && target > 0) {
+        // Forward from the head, so the first cut that reaches the target is the one that folds least.
+        let folded = 0;
+        for (let at = from; at < messages.length; at++) {
+            if (messages[at].role === "user") {
+                cut = at;
+                after = cost - folded + summaryTokens;
+                if (after <= limit * target)
+                    break;
+            }
+            folded += estimate(messages[at]);
+        }
+    }
+    else {
+        const budget = limit * keepRatio;
+        let kept = 0;
+        for (let at = messages.length - 1; at > from; at--) {
+            kept += estimate(messages[at]);
+            if (kept > budget)
+                break;
+            cut = at;
+        }
+        while (cut < messages.length && messages[cut].role !== "user")
+            cut++;
     }
-    while (cut < messages.length && messages[cut].role !== "user")
-        cut++;
     if (cut >= messages.length || cut - from < 2)
         return undefined;
-    return { from, cut, toSummarise: messages.slice(from, cut), ...(previous ? { previous } : {}) };
+    return {
+        from,
+        cut,
+        toSummarise: messages.slice(from, cut),
+        ...(previous ? { previous } : {}),
+        ...(after === undefined ? {} : { after: Math.round(after) }),
+    };
 }
 /**
  * What the summariser is handed for a plan: the earlier summary if there was one, then each
@@ -158,7 +198,7 @@ export function summaryInput(plan) {
  * @param options Cancellation and notices; the ceiling is 1024 and the instruction
  * `SUMMARY_PROMPT` unless given.
  */
-export const summariser = (config, model, { system = SUMMARY_PROMPT, maxTokens = 1024, ...options } = {}) => (text) => ask(config, model, system, text, { maxTokens, ...options });
+export const summariser = (config, model, { system = SUMMARY_PROMPT, maxTokens = SUMMARY_TOKENS, ...options } = {}) => (text) => ask(config, model, system, text, { maxTokens, ...options });
 /**
  * The hooks and the summariser for a plan, as a record to store rather than a transcript to send.
  *

package/dist/config.d.ts

@@ -75,9 +75,11 @@ export interface ModelParams {
 export interface ToolPolicy {
     /**
      * "eager" sends every tool definition on every request. "ondemand" sends a name-only
-     * catalogue and lets the model pull in the schemas it needs. See `tool-loading.ts`.
+     * catalogue and lets the model pull in the schemas it needs. "proxy" is the same catalogue
+     * behind a tool array that never changes — `load_tools` and `call_tool` — for a server whose
+     * prompt cache a growing tool array throws away. See `tool-loading.ts`.
      */
-    toolDiscovery: "eager" | "ondemand";
+    toolDiscovery: "eager" | "ondemand" | "proxy";
     /** The model that does the preselection pass. Empty means don't preselect. */
     toolSelectModel: string;
     /** Hard stop on runaway tool loops. */

package/dist/events.d.ts

@@ -1,3 +1,4 @@
+import type { TokenLedger } from "./ledger.ts";
 import type { TurnUsage } from "./stream.ts";
 /**
  * What a run is doing, while it is doing it.
@@ -76,9 +77,9 @@ export type RunEventKind =
  | "thinking"
 /** Reply tokens, as they arrive. */
  | "output"
-/** The model asked for a tool, with the arguments it chose. */
+/** The model asked for a tool, with the arguments it chose. `id` is the call's. */
  | "tool-call"
-/** A tool came back, with what it said. */
+/** A tool came back, with what it said. `id` is the id of the call it answers. */
  | "tool-result"
 /** Something the runner did that is not the model's doing — a preselection, a retry. */
  | "notice"
@@ -110,6 +111,12 @@ export interface RunUsage {
      * from a caller emitting usage of its own.
      */
     turn?: TurnReport;
+    /**
+     * The run's token ledger as of this turn, whole rather than the entry it added — a host whose
+     * run then fails still holds the latest one, and a rewrite in `beforeStep` moves earlier entries.
+     * Absent from a caller emitting usage of its own.
+     */
+    ledger?: TokenLedger;
 }
 /** One turn's usage and measurements as a `usage` event carries them. */
 export interface TurnReport extends TurnUsage {
@@ -119,8 +126,8 @@ export interface TurnReport extends TurnUsage {
 /**
  * One thing that happened in a run, as a watcher receives it.
  *
- * Every field is always present — the empty ones are `""` or `null` rather than missing — so a
- * client reads it without guarding each key.
+ * Every field but `id` is always present — the empty ones are `""` or `null` rather than
+ * missing — so a client reads it without guarding each key.
  */
 export interface RunEvent {
     /** The run it belongs to. `emit` fills this in; a caller does not pass it. */
@@ -141,6 +148,12 @@ export interface RunEvent {
     text: string;
     /** Tool name on the tool kinds, otherwise empty. */
     name: string;
+    /**
+     * The call's id on the tool kinds, which pairs a `tool-result` with its `tool-call` when one
+     * tool is called twice in a step. The one field `emit` does not fill in: it is there only where
+     * the emitter gave one, and optional so an event built before it existed still compiles.
+     */
+    id?: string;
     /**
      * The caller's own flow step this happened inside, so a watcher can group a run the way the
      * work is written. Empty for events that belong to the run rather than to any one step, and