zelari-code 1.41.0 → 1.42.0

package/dist/cli/budget/llmCompact.js

@@ -1,23 +1,48 @@
 /**
- * Optional LLM pass for history compaction summaries.
- * Falls back to null on any failure so callers use extractive summary.
+ * LLM compaction via CACHE-AWARE PREFIX REPLAY (v1.36.0, P9).
  *
- * Disable with ZELARI_LLM_COMPACT=0.
+ * Pre-1.36 this module built a COLD request: its own COMPACT_SYSTEM, a
+ * flattened transcript, and a raw fetch to the chat-completions endpoint.
+ * That request diverged from the live conversation from token 0 — zero
+ * prompt-cache reuse — and the rewritten history invalidated the warm
+ * prefix for the NEXT conversation turn too.
  *
- * @since v1.21.0
+ * The replay approach sends:
+ *
+ *   SYSTEM(original) + TOOLS(original) + DROPPED PREFIX + COMPACTION_INSTRUCTION
+ *
+ * so everything up to the trailing instruction is byte-identical to the
+ * previous routed request and hits the provider prefix cache (DeepSeek/
+ * OpenAI/GLM bill cached tokens at ~1/10). Tools stay advertised even
+ * though the summarizer must not call them: removing them would change
+ * the prefix token sequence and kill cache reuse.
+ *
+ * Fallback: any failure → null → caller uses the extractive summary.
+ * Disable entirely with ZELARI_LLM_COMPACT=0.
+ *
+ * @since v1.21.0 (cold-request version)
+ * @updated v1.36.0 — replay-based, providerStream-routed
  */
-import { getModelForProvider, getProviderConfig, getCustomEndpoint, } from '../providerConfig.js';
-import { resolveApiKeyWithMeta } from '../keyStore.js';
-import { PROVIDER_ENDPOINTS, } from '../provider/openai-compatible.js';
-const COMPACT_SYSTEM = `You compress earlier turns of a coding-agent session into a dense continuity brief.
-Output plain text (no markdown fences) with these sections:
-1) Goal — what the user wants
-2) Decisions — choices already made
-3) Done — completed work / files changed
-4) Open — remaining tasks / blockers
-5) Constraints — important rules the agent must keep
+export const COMPACTION_INSTRUCTION = `
+You are now acting as a compaction engine for this coding-agent session.
 
-Be factual and concise. Max ~400 words. Do not invent work that was not present.`;
+Condense the conversation ABOVE into a compact checkpoint sufficient to continue the task.
+
+Preserve:
+- user's goal and evolving intent
+- decisions already made
+- exact file paths and identifiers
+- code changes already completed
+- commands/errors that still matter
+- constraints
+- unfinished work
+- the single most likely next action
+
+Do not call tools.
+Do not mention this summarization request.
+Output only the checkpoint.
+Be concise.
+`.trim();
 export function isLlmCompactEnabled() {
     const v = process.env.ZELARI_LLM_COMPACT?.trim().toLowerCase();
     if (v === '0' || v === 'false' || v === 'off' || v === 'no')
@@ -25,90 +50,79 @@ export function isLlmCompactEnabled() {
     // default on when env unset
     return true;
 }
+/** Explicit model override for the summarizer (ZELARI_COMPACT_MODEL). */
+export function compactModelOverride() {
+    const v = process.env.ZELARI_COMPACT_MODEL?.trim();
+    return v ? v : undefined;
+}
+/** Hard ceiling so a stuck summarizer can't hang the dispatch loop. */
+const REPLAY_TIMEOUT_MS = 60_000;
 /**
- * Ask the active provider for a continuity summary.
- * Returns null if disabled, no key, or request fails.
+ * Summarize `droppedMessages` by replaying the ORIGINAL request prefix and
+ * appending the compaction instruction as the final user message.
+ *
+ * Returns `summary: null` when disabled, empty, failed, or when the model
+ * emitted a tool call (the summarizer must never act — only condense).
  */
-export async function llmSummarizeHistory(input) {
+export async function llmSummarizeHistoryReplay(input) {
+    const override = input.overrideModel ?? compactModelOverride();
+    const model = override ?? input.model;
+    const cacheReuseExpected = !override;
     if (!isLlmCompactEnabled())
-        return null;
-    if (!input.droppedTranscript.trim())
-        return null;
-    let config = null;
-    try {
-        config = await resolveCompactProviderConfig();
-    }
-    catch {
-        return null;
+        return { summary: null, model, cacheReuseExpected };
+    if (input.droppedMessages.length === 0) {
+        return { summary: null, model, cacheReuseExpected };
     }
-    if (!config)
-        return null;
-    const model = process.env.ZELARI_COMPACT_MODEL?.trim() || config.model;
+    // Replay = original prefix + instruction tail. The prefix must be
+    // byte-identical to the previous routed request for cache reuse.
+    const messages = [
+        ...input.systemMessages,
+        ...input.droppedMessages,
+        {
+            role: 'user',
+            content: COMPACTION_INSTRUCTION,
+        },
+    ];
     const controller = new AbortController();
-    const timeout = setTimeout(() => controller.abort(), 45_000);
+    const timeout = setTimeout(() => controller.abort(), REPLAY_TIMEOUT_MS);
     const onOuterAbort = () => controller.abort();
     input.signal?.addEventListener('abort', onOuterAbort, { once: true });
     try {
-        const url = `${config.baseUrl.replace(/\/$/, '')}/chat/completions`;
-        const res = await fetch(url, {
-            method: 'POST',
+        let text = '';
+        let emittedToolCall = false;
+        for await (const delta of input.providerStream({
+            provider: input.provider,
+            model,
+            messages,
+            // Tools stay advertised: dropping them would change the prefix token
+            // sequence and destroy cache reuse (explicit DSH decision). They are
+            // sorted canonically (same discipline as the live routed request and
+            // the snapshot fingerprints) so the replay prefix is byte-identical.
+            tools: [...input.tools].sort((a, b) => (a.name < b.name ? -1 : a.name > b.name ? 1 : 0)),
             signal: controller.signal,
-            headers: {
-                'content-type': 'application/json',
-                authorization: `Bearer ${config.apiKey}`,
-            },
-            body: JSON.stringify({
-                model,
+            generation: {
+                purpose: 'compaction',
                 temperature: 0.1,
-                max_tokens: 900,
-                stream: false,
-                messages: [
-                    { role: 'system', content: COMPACT_SYSTEM },
-                    {
-                        role: 'user',
-                        content: `Extractive sketch (may be incomplete):\n${input.extractive.slice(0, 2_500)}\n\n` +
-                            `Transcript of dropped turns:\n${input.droppedTranscript}`,
-                    },
-                ],
-            }),
-        });
-        if (!res.ok)
-            return null;
-        const json = (await res.json());
-        const text = json.choices?.[0]?.message?.content?.trim();
-        if (!text)
-            return null;
-        return ('[history-summary · llm]\n' +
-            text +
-            '\n\nContinue from the recent messages below; honor decisions already made above.');
+                maxTokens: 900,
+            },
+        })) {
+            if (delta.kind === 'text')
+                text += delta.delta;
+            if (delta.kind === 'tool_call')
+                emittedToolCall = true;
+        }
+        if (emittedToolCall)
+            return { summary: null, model, cacheReuseExpected };
+        if (!text.trim())
+            return { summary: null, model, cacheReuseExpected };
+        return { summary: text.trim(), model, cacheReuseExpected };
     }
     catch {
-        return null;
+        return { summary: null, model, cacheReuseExpected };
     }
     finally {
         clearTimeout(timeout);
         input.signal?.removeEventListener('abort', onOuterAbort);
     }
 }
-async function resolveCompactProviderConfig() {
-    const active = getProviderConfig().activeProviderId;
-    const meta = await resolveApiKeyWithMeta(active);
-    const apiKey = meta?.apiKey;
-    if (!apiKey)
-        return null;
-    const custom = getCustomEndpoint(active);
-    let baseUrl = custom ||
-        (active === 'openai-compatible' || active === 'custom'
-            ? process.env.OPENAI_BASE_URL ?? PROVIDER_ENDPOINTS[active]
-            : PROVIDER_ENDPOINTS[active]);
-    if (!baseUrl)
-        return null;
-    const model = getModelForProvider(active);
-    return {
-        apiKey,
-        baseUrl,
-        model,
-        providerId: active,
-    };
-}
 //# sourceMappingURL=llmCompact.js.map

package/dist/cli/budget/llmCompact.js.map

@@ -1 +1 @@
-{"version":3,"file":"llmCompact.js","sourceRoot":"","sources":["../../../src/cli/budget/llmCompact.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,OAAO,EACL,mBAAmB,EACnB,iBAAiB,EACjB,iBAAiB,GAClB,MAAM,sBAAsB,CAAC;AAC9B,OAAO,EAAE,qBAAqB,EAAE,MAAM,gBAAgB,CAAC;AACvD,OAAO,EACL,kBAAkB,GAEnB,MAAM,kCAAkC,CAAC;AAG1C,MAAM,cAAc,GAAG;;;;;;;;iFAQ0D,CAAC;AAElF,MAAM,UAAU,mBAAmB;IACjC,MAAM,CAAC,GAAG,OAAO,CAAC,GAAG,CAAC,kBAAkB,EAAE,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IAC/D,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK,OAAO,IAAI,CAAC,KAAK,KAAK,IAAI,CAAC,KAAK,IAAI;QAAE,OAAO,KAAK,CAAC;IAC1E,4BAA4B;IAC5B,OAAO,IAAI,CAAC;AACd,CAAC;AAED;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,mBAAmB,CAAC,KAIzC;IACC,IAAI,CAAC,mBAAmB,EAAE;QAAE,OAAO,IAAI,CAAC;IACxC,IAAI,CAAC,KAAK,CAAC,iBAAiB,CAAC,IAAI,EAAE;QAAE,OAAO,IAAI,CAAC;IAEjD,IAAI,MAAM,GAAkC,IAAI,CAAC;IACjD,IAAI,CAAC;QACH,MAAM,GAAG,MAAM,4BAA4B,EAAE,CAAC;IAChD,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,IAAI,CAAC;IACd,CAAC;IACD,IAAI,CAAC,MAAM;QAAE,OAAO,IAAI,CAAC;IAEzB,MAAM,KAAK,GACT,OAAO,CAAC,GAAG,CAAC,oBAAoB,EAAE,IAAI,EAAE,IAAI,MAAM,CAAC,KAAK,CAAC;IAE3D,MAAM,UAAU,GAAG,IAAI,eAAe,EAAE,CAAC;IACzC,MAAM,OAAO,GAAG,UAAU,CAAC,GAAG,EAAE,CAAC,UAAU,CAAC,KAAK,EAAE,EAAE,MAAM,CAAC,CAAC;IAC7D,MAAM,YAAY,GAAG,GAAG,EAAE,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;IAC9C,KAAK,CAAC,MAAM,EAAE,gBAAgB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;IAEtE,IAAI,CAAC;QACH,MAAM,GAAG,GAAG,GAAG,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,mBAAmB,CAAC;QACpE,MAAM,GAAG,GAAG,MAAM,KAAK,CAAC,GAAG,EAAE;YAC3B,MAAM,EAAE,MAAM;YACd,MAAM,EAAE,UAAU,CAAC,MAAM;YACzB,OAAO,EAAE;gBACP,cAAc,EAAE,kBAAkB;gBAClC,aAAa,EAAE,UAAU,MAAM,CAAC,MAAM,EAAE;aACzC;YACD,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC;gBACnB,KAAK;gBACL,WAAW,EAAE,GAAG;gBAChB,UAAU,EAAE,GAAG;gBACf,MAAM,EAAE,KAAK;gBACb,QAAQ,EAAE;oBACR,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,cAAc,EAAE;oBAC3C;wBACE,IAAI,EAAE,MAAM;wBACZ,OAAO,EACL,2CAA2C,KAAK,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,MAAM;4BACjF,iCAAiC,KAAK,CAAC,iBAAiB,EAAE;qBAC7D;iBACF;aACF,CAAC;SACH,CAAC,CAAC;QACH,IAAI,CAAC,GAAG,CAAC,EAAE;YAAE,OAAO,IAAI,CAAC;QACzB,MAAM,IAAI,GAAG,CAAC,MAAM,GAAG,CAAC,IAAI,EAAE,CAE7B,CAAC;QACF,MAAM,IAAI,GAAG,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;QACzD,IAAI,CAAC,IAAI;YAAE,OAAO,IAAI,CAAC;QACvB,OAAO,CACL,2BAA2B;YAC3B,IAAI;YACJ,kFAAkF,CACnF,CAAC;IACJ,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,IAAI,CAAC;IACd,CAAC;YAAS,CAAC;QACT,YAAY,CAAC,OAAO,CAAC,CAAC;QACtB,KAAK,CAAC,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,YAAY,CAAC,CAAC;IAC3D,CAAC;AACH,CAAC;AAED,KAAK,UAAU,4BAA4B;IACzC,MAAM,MAAM,GAAG,iBAAiB,EAAE,CAAC,gBAAgC,CAAC;IACpE,MAAM,IAAI,GAAG,MAAM,qBAAqB,CAAC,MAAM,CAAC,CAAC;IACjD,MAAM,MAAM,GAAG,IAAI,EAAE,MAAM,CAAC;IAC5B,IAAI,CAAC,MAAM;QAAE,OAAO,IAAI,CAAC;IAEzB,MAAM,MAAM,GAAG,iBAAiB,CAAC,MAAM,CAAC,CAAC;IACzC,IAAI,OAAO,GACT,MAAM;QACN,CAAC,MAAM,KAAK,mBAAmB,IAAI,MAAM,KAAK,QAAQ;YACpD,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,eAAe,IAAI,kBAAkB,CAAC,MAAM,CAAC;YAC3D,CAAC,CAAC,kBAAkB,CAAC,MAAM,CAAC,CAAC,CAAC;IAClC,IAAI,CAAC,OAAO;QAAE,OAAO,IAAI,CAAC;IAE1B,MAAM,KAAK,GAAG,mBAAmB,CAAC,MAAM,CAAC,CAAC;IAC1C,OAAO;QACL,MAAM;QACN,OAAO;QACP,KAAK;QACL,UAAU,EAAE,MAAM;KACnB,CAAC;AACJ,CAAC"}
+{"version":3,"file":"llmCompact.js","sourceRoot":"","sources":["../../../src/cli/budget/llmCompact.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAQH,MAAM,CAAC,MAAM,sBAAsB,GAAG;;;;;;;;;;;;;;;;;;;CAmBrC,CAAC,IAAI,EAAE,CAAC;AAET,MAAM,UAAU,mBAAmB;IACjC,MAAM,CAAC,GAAG,OAAO,CAAC,GAAG,CAAC,kBAAkB,EAAE,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;IAC/D,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,KAAK,OAAO,IAAI,CAAC,KAAK,KAAK,IAAI,CAAC,KAAK,IAAI;QAAE,OAAO,KAAK,CAAC;IAC1E,4BAA4B;IAC5B,OAAO,IAAI,CAAC;AACd,CAAC;AAED,yEAAyE;AACzE,MAAM,UAAU,oBAAoB;IAClC,MAAM,CAAC,GAAG,OAAO,CAAC,GAAG,CAAC,oBAAoB,EAAE,IAAI,EAAE,CAAC;IACnD,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;AAC3B,CAAC;AA8BD,uEAAuE;AACvE,MAAM,iBAAiB,GAAG,MAAM,CAAC;AAEjC;;;;;;GAMG;AACH,MAAM,CAAC,KAAK,UAAU,yBAAyB,CAC7C,KAA8B;IAE9B,MAAM,QAAQ,GAAG,KAAK,CAAC,aAAa,IAAI,oBAAoB,EAAE,CAAC;IAC/D,MAAM,KAAK,GAAG,QAAQ,IAAI,KAAK,CAAC,KAAK,CAAC;IACtC,MAAM,kBAAkB,GAAG,CAAC,QAAQ,CAAC;IAErC,IAAI,CAAC,mBAAmB,EAAE;QAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,kBAAkB,EAAE,CAAC;IAChF,IAAI,KAAK,CAAC,eAAe,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACvC,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,kBAAkB,EAAE,CAAC;IACtD,CAAC;IAED,kEAAkE;IAClE,iEAAiE;IACjE,MAAM,QAAQ,GAAmB;QAC/B,GAAG,KAAK,CAAC,cAAc;QACvB,GAAG,KAAK,CAAC,eAAe;QACxB;YACE,IAAI,EAAE,MAAM;YACZ,OAAO,EAAE,sBAAsB;SAChC;KACF,CAAC;IAEF,MAAM,UAAU,GAAG,IAAI,eAAe,EAAE,CAAC;IACzC,MAAM,OAAO,GAAG,UAAU,CAAC,GAAG,EAAE,CAAC,UAAU,CAAC,KAAK,EAAE,EAAE,iBAAiB,CAAC,CAAC;IACxE,MAAM,YAAY,GAAG,GAAG,EAAE,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;IAC9C,KAAK,CAAC,MAAM,EAAE,gBAAgB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;IAEtE,IAAI,CAAC;QACH,IAAI,IAAI,GAAG,EAAE,CAAC;QACd,IAAI,eAAe,GAAG,KAAK,CAAC;QAE5B,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,KAAK,CAAC,cAAc,CAAC;YAC7C,QAAQ,EAAE,KAAK,CAAC,QAAQ;YACxB,KAAK;YACL,QAAQ;YACR,qEAAqE;YACrE,qEAAqE;YACrE,qEAAqE;YACrE,qEAAqE;YACrE,KAAK,EAAE,CAAC,GAAG,KAAK,CAAC,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACxF,MAAM,EAAE,UAAU,CAAC,MAAM;YACzB,UAAU,EAAE;gBACV,OAAO,EAAE,YAAY;gBACrB,WAAW,EAAE,GAAG;gBAChB,SAAS,EAAE,GAAG;aACf;SACF,CAAC,EAAE,CAAC;YACH,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM;gBAAE,IAAI,IAAI,KAAK,CAAC,KAAK,CAAC;YAC/C,IAAI,KAAK,CAAC,IAAI,KAAK,WAAW;gBAAE,eAAe,GAAG,IAAI,CAAC;QACzD,CAAC;QAED,IAAI,eAAe;YAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,kBAAkB,EAAE,CAAC;QACzE,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE;YAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,kBAAkB,EAAE,CAAC;QAEtE,OAAO,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE,EAAE,KAAK,EAAE,kBAAkB,EAAE,CAAC;IAC7D,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,kBAAkB,EAAE,CAAC;IACtD,CAAC;YAAS,CAAC;QACT,YAAY,CAAC,OAAO,CAAC,CAAC;QACtB,KAAK,CAAC,MAAM,EAAE,mBAAmB,CAAC,OAAO,EAAE,YAAY,CAAC,CAAC;IAC3D,CAAC;AACH,CAAC"}

package/dist/cli/budget/requestMeter.js

@@ -0,0 +1,138 @@
+/**
+ * requestMeter — full-request occupancy estimation with provider-usage
+ * anchoring (v1.36.0 context/cache upgrade; P4/P5).
+ *
+ * The legacy `estimateHistoryTokens` measured ONLY the rolling history
+ * (`content` + `toolCalls` args). It ignored the system prompt, tool
+ * schemas, `reasoningContent`, and per-message role overhead — tens of
+ * thousands of tokens in tool-heavy sessions — so occupancy read low and
+ * compaction fired too late (or never).
+ *
+ * This meter measures the WHOLE request surface:
+ *
+ *   estimatedPromptTokens ≈ system + tools-schema + conversation
+ *
+ * and anchors to provider-reported usage when the header fingerprint
+ * matches the last routed request: provider usage is ground truth for
+ * the stable header, so we only re-estimate the conversation DELTA.
+ *
+ * `contextPressureTokens` NEVER subtracts cached tokens: the provider
+ * must still HOLD the whole prefix (cached or not) in the context window.
+ *
+ * @since v1.36.0
+ */
+import { stableStringify, sha256Hex, canonicalTools, } from '@zelari/core/harness';
+/** Local chars→tokens estimate (same heuristic as tokenBudget; local copy to
+ *  avoid a tokenBudget ⇄ requestMeter import cycle). */
+function estimateTokensLocal(text) {
+    if (!text)
+        return 0;
+    return Math.max(1, Math.ceil(text.length / 4));
+}
+/** ~4 tokens of wire overhead per message (role/id framing). */
+const MESSAGE_OVERHEAD_TOKENS = 4;
+/** Estimate one message across ALL its token-bearing fields. */
+export function estimateMessageTokens(m) {
+    let n = MESSAGE_OVERHEAD_TOKENS + estimateTokensLocal(m.content ?? '');
+    if (m.toolCalls) {
+        for (const tc of m.toolCalls) {
+            n += estimateTokensLocal(tc.name) + estimateTokensLocal(tc.id);
+            n += estimateTokensLocal(JSON.stringify(tc.args ?? {}));
+        }
+    }
+    if (m.reasoningContent)
+        n += estimateTokensLocal(m.reasoningContent);
+    if (m.toolCallId)
+        n += estimateTokensLocal(m.toolCallId);
+    return n;
+}
+/** Estimate the tool-schema surface (name + description + parameters). */
+export function estimateToolSchemaTokens(tools) {
+    let n = 0;
+    for (const t of tools) {
+        n += estimateTokensLocal(t.name) + estimateTokensLocal(t.description ?? '');
+        n += estimateTokensLocal(JSON.stringify(t.parameters ?? {}));
+    }
+    // Wire framing per advertised function.
+    return n + tools.length * MESSAGE_OVERHEAD_TOKENS;
+}
+/** Estimate the system-prompt surface. */
+export function estimateSystemTokens(systemMessages) {
+    let n = 0;
+    for (const m of systemMessages)
+        n += estimateMessageTokens(m);
+    return n;
+}
+/** Estimated tokens of the conversation tail. */
+export function estimateConversationTokens(conversation) {
+    let n = 0;
+    for (const m of conversation)
+        n += estimateMessageTokens(m);
+    return n;
+}
+/**
+ * Same fingerprint discipline as requestSnapshot: provider + model + system
+ * + canonical tools. Cheap enough to recompute per turn.
+ */
+export function headerFingerprintOf(input) {
+    return createFingerprintOnly(input);
+}
+function createFingerprintOnly(input) {
+    return sha256Hex(stableStringify({
+        provider: input.provider,
+        model: input.model,
+        systemMessages: input.systemMessages,
+        tools: canonicalTools(input.tools),
+    }));
+}
+/**
+ * Measure the full request surface.
+ *
+ * Anchoring: when `anchor.snapshot`'s headerFingerprint equals the current
+ * header fingerprint AND its usage arrived, the stable header is measured
+ * as `usage.promptTokens - estimatedConversationTokens(anchor.conversation)`
+ * (clamped to ≥0) — provider ground truth beats chars/4. The current
+ * conversation is always freshly estimated.
+ */
+export function measureRequest(input) {
+    const estimatedHeaderTokens = estimateSystemTokens(input.systemMessages) + estimateToolSchemaTokens(input.tools);
+    const currentConversationTokens = estimateConversationTokens(input.conversation);
+    let headerAnchored = false;
+    let headerTokens = estimatedHeaderTokens;
+    const anchorUsage = input.anchor?.usage;
+    const anchorSnapshot = input.anchor?.snapshot;
+    if (anchorUsage && anchorSnapshot) {
+        const currentHeaderFp = createFingerprintOnly({
+            provider: anchorSnapshot.provider,
+            model: anchorSnapshot.model,
+            systemMessages: input.systemMessages,
+            tools: input.tools,
+        });
+        if (currentHeaderFp === anchorSnapshot.headerFingerprint) {
+            // Provider saw header + its conversation; subtract the conversation
+            // estimate to isolate the header ground truth.
+            const anchorConv = estimateConversationTokens(anchorSnapshot.conversation);
+            const headerFromUsage = Math.max(0, anchorUsage.promptTokens - anchorConv);
+            // v1.36.0 (case 9): the provider truth is EXPECTED to dwarf the
+            // chars/4 estimate (chat templating, tool marshalling, hidden
+            // framing). Anchoring exists precisely for that gap — so only a
+            // non-positive (garbage/empty) value is rejected, never a large one.
+            if (headerFromUsage > 0) {
+                headerTokens = headerFromUsage;
+                headerAnchored = true;
+            }
+        }
+    }
+    const estimatedPromptTokens = headerTokens + currentConversationTokens;
+    const reservedOutput = input.reservedOutputTokens ?? 0;
+    const contextPressureTokens = estimatedPromptTokens + reservedOutput;
+    return {
+        estimatedPromptTokens,
+        estimatedHeaderTokens,
+        headerAnchored,
+        contextPressureTokens,
+        occupancy: Math.min(1, contextPressureTokens / input.contextLimit),
+        purpose: input.purpose ?? 'conversation',
+    };
+}
+//# sourceMappingURL=requestMeter.js.map

package/dist/cli/budget/requestMeter.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"requestMeter.js","sourceRoot":"","sources":["../../../src/cli/budget/requestMeter.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAGH,OAAO,EACL,eAAe,EACf,SAAS,EACT,cAAc,GACf,MAAM,sBAAsB,CAAC;AAG9B;wDACwD;AACxD,SAAS,mBAAmB,CAAC,IAAY;IACvC,IAAI,CAAC,IAAI;QAAE,OAAO,CAAC,CAAC;IACpB,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC;AACjD,CAAC;AAED,gEAAgE;AAChE,MAAM,uBAAuB,GAAG,CAAC,CAAC;AAiClC,gEAAgE;AAChE,MAAM,UAAU,qBAAqB,CAAC,CAAe;IACnD,IAAI,CAAC,GAAG,uBAAuB,GAAG,mBAAmB,CAAC,CAAC,CAAC,OAAO,IAAI,EAAE,CAAC,CAAC;IACvE,IAAI,CAAC,CAAC,SAAS,EAAE,CAAC;QAChB,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,SAAS,EAAE,CAAC;YAC7B,CAAC,IAAI,mBAAmB,CAAC,EAAE,CAAC,IAAI,CAAC,GAAG,mBAAmB,CAAC,EAAE,CAAC,EAAE,CAAC,CAAC;YAC/D,CAAC,IAAI,mBAAmB,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,CAAC,IAAI,IAAI,EAAE,CAAC,CAAC,CAAC;QAC1D,CAAC;IACH,CAAC;IACD,IAAI,CAAC,CAAC,gBAAgB;QAAE,CAAC,IAAI,mBAAmB,CAAC,CAAC,CAAC,gBAAgB,CAAC,CAAC;IACrE,IAAI,CAAC,CAAC,UAAU;QAAE,CAAC,IAAI,mBAAmB,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC;IACzD,OAAO,CAAC,CAAC;AACX,CAAC;AAED,0EAA0E;AAC1E,MAAM,UAAU,wBAAwB,CAAC,KAA+B;IACtE,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,KAAK,MAAM,CAAC,IAAI,KAAK,EAAE,CAAC;QACtB,CAAC,IAAI,mBAAmB,CAAC,CAAC,CAAC,IAAI,CAAC,GAAG,mBAAmB,CAAC,CAAC,CAAC,WAAW,IAAI,EAAE,CAAC,CAAC;QAC5E,CAAC,IAAI,mBAAmB,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,UAAU,IAAI,EAAE,CAAC,CAAC,CAAC;IAC/D,CAAC;IACD,wCAAwC;IACxC,OAAO,CAAC,GAAG,KAAK,CAAC,MAAM,GAAG,uBAAuB,CAAC;AACpD,CAAC;AAED,0CAA0C;AAC1C,MAAM,UAAU,oBAAoB,CAClC,cAAuC;IAEvC,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,KAAK,MAAM,CAAC,IAAI,cAAc;QAAE,CAAC,IAAI,qBAAqB,CAAC,CAAC,CAAC,CAAC;IAC9D,OAAO,CAAC,CAAC;AACX,CAAC;AAED,iDAAiD;AACjD,MAAM,UAAU,0BAA0B,CACxC,YAAqC;IAErC,IAAI,CAAC,GAAG,CAAC,CAAC;IACV,KAAK,MAAM,CAAC,IAAI,YAAY;QAAE,CAAC,IAAI,qBAAqB,CAAC,CAAC,CAAC,CAAC;IAC5D,OAAO,CAAC,CAAC;AACX,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,mBAAmB,CAAC,KAKnC;IACC,OAAO,qBAAqB,CAAC,KAAK,CAAC,CAAC;AACtC,CAAC;AAED,SAAS,qBAAqB,CAAC,KAK9B;IACC,OAAO,SAAS,CACd,eAAe,CAAC;QACd,QAAQ,EAAE,KAAK,CAAC,QAAQ;QACxB,KAAK,EAAE,KAAK,CAAC,KAAK;QAClB,cAAc,EAAE,KAAK,CAAC,cAAc;QACpC,KAAK,EAAE,cAAc,CAAC,KAAK,CAAC,KAAK,CAAC;KACnC,CAAC,CACH,CAAC;AACJ,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,cAAc,CAC5B,KAIC;IAED,MAAM,qBAAqB,GACzB,oBAAoB,CAAC,KAAK,CAAC,cAAc,CAAC,GAAG,wBAAwB,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;IACrF,MAAM,yBAAyB,GAAG,0BAA0B,CAAC,KAAK,CAAC,YAAY,CAAC,CAAC;IAEjF,IAAI,cAAc,GAAG,KAAK,CAAC;IAC3B,IAAI,YAAY,GAAG,qBAAqB,CAAC;IAEzC,MAAM,WAAW,GAAG,KAAK,CAAC,MAAM,EAAE,KAAK,CAAC;IACxC,MAAM,cAAc,GAAG,KAAK,CAAC,MAAM,EAAE,QAAQ,CAAC;IAC9C,IAAI,WAAW,IAAI,cAAc,EAAE,CAAC;QAClC,MAAM,eAAe,GAAG,qBAAqB,CAAC;YAC5C,QAAQ,EAAE,cAAc,CAAC,QAAQ;YACjC,KAAK,EAAE,cAAc,CAAC,KAAK;YAC3B,cAAc,EAAE,KAAK,CAAC,cAAc;YACpC,KAAK,EAAE,KAAK,CAAC,KAAK;SACnB,CAAC,CAAC;QACH,IAAI,eAAe,KAAK,cAAc,CAAC,iBAAiB,EAAE,CAAC;YACzD,oEAAoE;YACpE,+CAA+C;YAC/C,MAAM,UAAU,GAAG,0BAA0B,CAAC,cAAc,CAAC,YAAY,CAAC,CAAC;YAC3E,MAAM,eAAe,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,WAAW,CAAC,YAAY,GAAG,UAAU,CAAC,CAAC;YAC3E,gEAAgE;YAChE,8DAA8D;YAC9D,gEAAgE;YAChE,qEAAqE;YACrE,IAAI,eAAe,GAAG,CAAC,EAAE,CAAC;gBACxB,YAAY,GAAG,eAAe,CAAC;gBAC/B,cAAc,GAAG,IAAI,CAAC;YACxB,CAAC;QACH,CAAC;IACH,CAAC;IAED,MAAM,qBAAqB,GAAG,YAAY,GAAG,yBAAyB,CAAC;IACvE,MAAM,cAAc,GAAG,KAAK,CAAC,oBAAoB,IAAI,CAAC,CAAC;IACvD,MAAM,qBAAqB,GAAG,qBAAqB,GAAG,cAAc,CAAC;IAErE,OAAO;QACL,qBAAqB;QACrB,qBAAqB;QACrB,cAAc;QACd,qBAAqB;QACrB,SAAS,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,qBAAqB,GAAG,KAAK,CAAC,YAAY,CAAC;QAClE,OAAO,EAAE,KAAK,CAAC,OAAO,IAAI,cAAc;KACzC,CAAC;AACJ,CAAC"}

package/dist/cli/budget/requestSnapshotStore.js

@@ -0,0 +1,55 @@
+/**
+ * requestSnapshotStore — in-memory per-session store of routed request
+ * snapshots (v1.36.0 context/cache upgrade).
+ *
+ * Keeps the LAST snapshot per session (that's all cache-aware compaction
+ * needs: the most recent routed request = the warmest provider prefix)
+ * plus the provider-reported usage that answered it.
+ *
+ * Usage flow:
+ *   - AgentHarness `onRequestSnapshot` → `recordRequestSnapshot(sessionId, s)`
+ *   - `message_end` usage in useChatTurn → `recordRequestUsage(sessionId, usage)`
+ *     (associated to the snapshot that is currently pending for that session)
+ *   - `applyBudgetPolicyAsync` → `getRequestSnapshot(sessionId)` → replay base
+ *   - `/clear` | `/new` → `clearRequestSnapshots(sessionId)` /
+ *     `clearAllRequestSnapshots()` (CLI is single-session per process).
+ *
+ * @since v1.36.0
+ */
+const store = new Map();
+/** Track the most recently routed request for a session. */
+export function recordRequestSnapshot(sessionId, snapshot) {
+    // Keep the pending slot: usage recorded next will bind to this snapshot.
+    store.set(sessionId, { snapshot });
+}
+/**
+ * Attach provider-reported usage to the session's latest snapshot.
+ * Called when the `message_end` usage delta lands — after the snapshot.
+ */
+export function recordRequestUsage(sessionId, usage) {
+    const entry = store.get(sessionId);
+    if (!entry)
+        return;
+    entry.usage = usage;
+}
+/** Last routed snapshot (+ usage, when reported). Null when none yet. */
+export function getRequestSnapshot(sessionId) {
+    return store.get(sessionId)?.snapshot ?? null;
+}
+/** Last snapshot together with its provider-reported usage. */
+export function getRequestSnapshotWithUsage(sessionId) {
+    return store.get(sessionId) ?? null;
+}
+/** Drop one session's snapshot (session switch). */
+export function clearRequestSnapshots(sessionId) {
+    store.delete(sessionId);
+}
+/** Drop everything (/clear, /new — the CLI is mono-session per process). */
+export function clearAllRequestSnapshots() {
+    store.clear();
+}
+/** Test-only reset. */
+export function _resetRequestSnapshotStoreForTests() {
+    store.clear();
+}
+//# sourceMappingURL=requestSnapshotStore.js.map

package/dist/cli/budget/requestSnapshotStore.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"requestSnapshotStore.js","sourceRoot":"","sources":["../../../src/cli/budget/requestSnapshotStore.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAkBH,MAAM,KAAK,GAAG,IAAI,GAAG,EAAwB,CAAC;AAE9C,4DAA4D;AAC5D,MAAM,UAAU,qBAAqB,CACnC,SAAiB,EACjB,QAA+B;IAE/B,yEAAyE;IACzE,KAAK,CAAC,GAAG,CAAC,SAAS,EAAE,EAAE,QAAQ,EAAE,CAAC,CAAC;AACrC,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,kBAAkB,CAChC,SAAiB,EACjB,KAAyB;IAEzB,MAAM,KAAK,GAAG,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;IACnC,IAAI,CAAC,KAAK;QAAE,OAAO;IACnB,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC;AACtB,CAAC;AAED,yEAAyE;AACzE,MAAM,UAAU,kBAAkB,CAAC,SAAiB;IAClD,OAAO,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,EAAE,QAAQ,IAAI,IAAI,CAAC;AAChD,CAAC;AAED,+DAA+D;AAC/D,MAAM,UAAU,2BAA2B,CACzC,SAAiB;IAEjB,OAAO,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,IAAI,IAAI,CAAC;AACtC,CAAC;AAED,oDAAoD;AACpD,MAAM,UAAU,qBAAqB,CAAC,SAAiB;IACrD,KAAK,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC;AAC1B,CAAC;AAED,4EAA4E;AAC5E,MAAM,UAAU,wBAAwB;IACtC,KAAK,CAAC,KAAK,EAAE,CAAC;AAChB,CAAC;AAED,uBAAuB;AACvB,MAAM,UAAU,kCAAkC;IAChD,KAAK,CAAC,KAAK,EAAE,CAAC;AAChB,CAAC"}