@agent-finops/core 0.6.0 → 0.7.0

package/dist/localAgentLogs.js

@@ -1,6 +1,8 @@
-import { readdir, readFile, stat } from "node:fs/promises";
+import { createReadStream } from "node:fs";
+import { lstat, open, readdir, readFile, stat } from "node:fs/promises";
 import { basename, isAbsolute, join, resolve, sep } from "node:path";
 import { homedir } from "node:os";
+import { createInterface } from "node:readline";
 import { estimateTokenCostUsd } from "./modelPricing.js";
 import { redactSecrets } from "./discovery.js";
 import { createCodexInvocationCollector } from "./toolInvocations.js";
@@ -52,6 +54,141 @@ function totalUsageTokens(usage) {
         (usage.cacheWrite5mTokens ?? 0) +
         (usage.cacheWrite1hTokens ?? 0);
 }
+function parseClaudeFinancialUsage(value, onDiagnostic) {
+    const inputTokens = tokenComponentOf(value.input_tokens);
+    const outputTokens = tokenComponentOf(value.output_tokens);
+    const cacheReadField = optionalTokenComponent(value, "cache_read_input_tokens");
+    const writeTotalField = optionalTokenComponent(value, "cache_creation_input_tokens");
+    const reportedTotalField = optionalTokenComponent(value, "total_tokens");
+    const cacheCreationPresent = Object.prototype.hasOwnProperty.call(value, "cache_creation");
+    const cacheCreation = isRecord(value.cache_creation) ? value.cache_creation : undefined;
+    const write5mField = cacheCreation
+        ? optionalTokenComponent(cacheCreation, "ephemeral_5m_input_tokens")
+        : { present: false };
+    const write1hField = cacheCreation
+        ? optionalTokenComponent(cacheCreation, "ephemeral_1h_input_tokens")
+        : { present: false };
+    const componentsSupported = inputTokens !== undefined &&
+        outputTokens !== undefined &&
+        (!cacheReadField.present || cacheReadField.value !== undefined) &&
+        (!writeTotalField.present || writeTotalField.value !== undefined) &&
+        (!reportedTotalField.present || reportedTotalField.value !== undefined) &&
+        (!cacheCreationPresent || Boolean(cacheCreation)) &&
+        (!write5mField.present || write5mField.value !== undefined) &&
+        (!write1hField.present || write1hField.value !== undefined);
+    const usage = {
+        // Retain every valid component for partial evidence, but never let a
+        // missing/invalid required field become a priceable zero-dollar call.
+        inputTokens: inputTokens ?? 0,
+        outputTokens: outputTokens ?? 0,
+        cacheReadTokens: cacheReadField.value ?? 0,
+        // Prefer the 5m/1h breakdown; fall back to the total as 5m (cheaper bound).
+        cacheWrite5mTokens: write5mField.value ?? writeTotalField.value ?? 0,
+        cacheWrite1hTokens: write1hField.value ?? 0
+    };
+    if (componentsSupported) {
+        return {
+            usage,
+            latestTurnUsage: toTurnUsage(usage, "assistant_message_usage")
+        };
+    }
+    onDiagnostic?.({ code: "unsupported_token_shape", count: 1 });
+    const reportedTotalTokens = reportedTotalField.value;
+    return {
+        usage,
+        usageSupport: "unsupported_token_shape",
+        ...(reportedTotalTokens !== undefined ? { reportedTotalTokens } : {})
+    };
+}
+function optionalTokenComponent(record, key) {
+    if (!Object.prototype.hasOwnProperty.call(record, key))
+        return { present: false };
+    const value = tokenComponentOf(record[key]);
+    return value === undefined ? { present: true } : { present: true, value };
+}
+function parseCodexCumulativeUsage(current, baseline) {
+    const rawInput = tokenComponentOf(current.input_tokens);
+    const rawOutput = tokenComponentOf(current.output_tokens);
+    const currentCachedField = optionalTokenComponent(current, "cached_input_tokens");
+    const currentTotalField = optionalTokenComponent(current, "total_tokens");
+    const rawCached = currentCachedField.value ?? 0;
+    const rawReportedTotal = currentTotalField.value;
+    const baselineInput = baseline ? tokenComponentOf(baseline.input_tokens) : undefined;
+    const baselineOutput = baseline ? tokenComponentOf(baseline.output_tokens) : undefined;
+    const baselineCachedField = baseline
+        ? optionalTokenComponent(baseline, "cached_input_tokens")
+        : { present: false };
+    const baselineTotalField = baseline
+        ? optionalTokenComponent(baseline, "total_tokens")
+        : { present: false };
+    const baselineCached = baselineCachedField.value ?? 0;
+    const baselineReportedTotal = baselineTotalField.value;
+    const currentSupported = rawInput !== undefined &&
+        rawOutput !== undefined &&
+        (!currentCachedField.present || currentCachedField.value !== undefined) &&
+        (!currentTotalField.present || currentTotalField.value !== undefined) &&
+        rawCached <= rawInput &&
+        (rawReportedTotal === undefined || rawReportedTotal >= rawInput + rawOutput) &&
+        !((rawReportedTotal ?? 0) > 0 && rawInput === 0 && rawOutput === 0);
+    const baselineSupported = !baseline || (baselineInput !== undefined &&
+        baselineOutput !== undefined &&
+        (!baselineCachedField.present || baselineCachedField.value !== undefined) &&
+        (!baselineTotalField.present || baselineTotalField.value !== undefined) &&
+        baselineCached <= baselineInput &&
+        (baselineReportedTotal === undefined ||
+            baselineReportedTotal >= baselineInput + baselineOutput) &&
+        !((baselineReportedTotal ?? 0) > 0 && baselineInput === 0 && baselineOutput === 0));
+    const monotonic = !baseline || (rawInput !== undefined && baselineInput !== undefined && rawInput >= baselineInput &&
+        rawOutput !== undefined && baselineOutput !== undefined && rawOutput >= baselineOutput &&
+        rawCached >= baselineCached &&
+        rawInput - rawCached >= baselineInput - baselineCached &&
+        (rawReportedTotal === undefined ||
+            baselineReportedTotal === undefined ||
+            rawReportedTotal >= baselineReportedTotal));
+    const input = Math.max(0, (rawInput ?? 0) - (baselineInput ?? 0));
+    const cached = Math.max(0, rawCached - baselineCached);
+    const output = Math.max(0, (rawOutput ?? 0) - (baselineOutput ?? 0));
+    const reportedTotalTokens = rawReportedTotal === undefined
+        ? undefined
+        : Math.max(0, rawReportedTotal - (baselineReportedTotal ?? 0));
+    return {
+        usage: {
+            // Codex input_tokens includes cached input; expose the non-cached split.
+            inputTokens: Math.max(0, input - cached),
+            outputTokens: output,
+            cacheReadTokens: cached
+        },
+        supported: currentSupported && baselineSupported && monotonic,
+        ...(reportedTotalTokens !== undefined ? { reportedTotalTokens } : {})
+    };
+}
+function parseCodexTurnUsage(value) {
+    const rawInput = tokenComponentOf(value.input_tokens);
+    const rawOutput = tokenComponentOf(value.output_tokens);
+    const cachedField = optionalTokenComponent(value, "cached_input_tokens");
+    const totalField = optionalTokenComponent(value, "total_tokens");
+    const cached = cachedField.value ?? 0;
+    const total = totalField.value;
+    const supported = rawInput !== undefined &&
+        rawOutput !== undefined &&
+        (!cachedField.present || cachedField.value !== undefined) &&
+        (!totalField.present || totalField.value !== undefined) &&
+        cached <= rawInput &&
+        (total === undefined || total >= rawInput + rawOutput);
+    if (!supported)
+        return { supported: false };
+    return {
+        supported: true,
+        usage: {
+            inputTokens: rawInput - cached,
+            outputTokens: rawOutput,
+            cacheReadTokens: cached,
+            contextTokens: rawInput,
+            totalTokens: total ?? rawInput + rawOutput,
+            source: "transcript_last_token_usage"
+        }
+    };
+}
 /** Parse one Claude Code transcript (JSONL). Exported for tests. */
 export function parseClaudeCodeTranscript(content, filePath = "", sinceMs, onDiagnostic) {
     const calls = [];
@@ -113,18 +250,7 @@ export function parseClaudeCodeTranscript(content, filePath = "", sinceMs, onDia
             pendingPrompts.length = 0;
             continue;
         }
-        const cacheCreation = isRecord(usage.cache_creation) ? usage.cache_creation : undefined;
-        const write5m = numberOf(cacheCreation?.ephemeral_5m_input_tokens);
-        const write1h = numberOf(cacheCreation?.ephemeral_1h_input_tokens);
-        const writeTotal = numberOf(usage.cache_creation_input_tokens) ?? 0;
-        const parsedUsage = {
-            inputTokens: numberOf(usage.input_tokens) ?? 0,
-            outputTokens: numberOf(usage.output_tokens) ?? 0,
-            cacheReadTokens: numberOf(usage.cache_read_input_tokens) ?? 0,
-            // Prefer the 5m/1h breakdown; fall back to the total as 5m (cheaper bound).
-            cacheWrite5mTokens: write5m ?? writeTotal,
-            cacheWrite1hTokens: write1h ?? 0
-        };
+        const parsedUsage = parseClaudeFinancialUsage(usage, onDiagnostic);
         const workingDirectory = absoluteWorkingDirectory(stringOf(entry.cwd));
         const project = projectFromCwd(workingDirectory) ?? projectFromTranscriptPath(filePath);
         const sessionId = stringOf(entry.sessionId);
@@ -135,9 +261,15 @@ export function parseClaudeCodeTranscript(content, filePath = "", sinceMs, onDia
             project,
             workingDirectory,
             sessionId,
-            latestTurnUsage: toTurnUsage(parsedUsage, "assistant_message_usage"),
+            ...(parsedUsage.latestTurnUsage
+                ? { latestTurnUsage: parsedUsage.latestTurnUsage }
+                : {}),
             usageScope: "turn",
-            usage: parsedUsage
+            ...(parsedUsage.usageSupport ? { usageSupport: parsedUsage.usageSupport } : {}),
+            ...(parsedUsage.reportedTotalTokens !== undefined
+                ? { reportedTotalTokens: parsedUsage.reportedTotalTokens }
+                : {}),
+            usage: parsedUsage.usage
         };
         calls.push(call);
         const activityKey = localActivityScopeKey(sessionId, workingDirectory, project);
@@ -312,31 +444,16 @@ export function parseCodexRollout(content, onEntry, onDiagnostic) {
     }
     if (!lastTotal || isSubagent && !rootTaskStarted)
         return [];
-    const rawInput = numberOf(lastTotal.input_tokens);
-    const rawOutput = numberOf(lastTotal.output_tokens);
-    const rawCached = numberOf(lastTotal.cached_input_tokens);
-    const rawReportedTotal = numberOf(lastTotal.total_tokens);
-    const baselineInput = numberOf(inheritedUsageBaseline?.input_tokens);
-    const baselineOutput = numberOf(inheritedUsageBaseline?.output_tokens);
-    const baselineReportedTotal = numberOf(inheritedUsageBaseline?.total_tokens);
-    const currentComponentsComplete = rawInput !== undefined && rawOutput !== undefined && !((rawReportedTotal ?? 0) > 0 && rawInput === 0 && rawOutput === 0);
-    const baselineComponentsComplete = !inheritedUsageBaseline || (baselineInput !== undefined && baselineOutput !== undefined && !((baselineReportedTotal ?? 0) > 0 && baselineInput === 0 && baselineOutput === 0));
-    const usageSupport = currentComponentsComplete && baselineComponentsComplete
+    const parsedUsage = parseCodexCumulativeUsage(lastTotal, inheritedUsageBaseline);
+    const parsedTurn = lastTurn
+        ? parseCodexTurnUsage(lastTurn)
+        : { supported: true };
+    const usageSupport = parsedUsage.supported && parsedTurn.supported
         ? "complete"
         : "unsupported_token_shape";
     if (usageSupport === "unsupported_token_shape") {
         onDiagnostic?.({ code: "unsupported_token_shape", count: 1 });
     }
-    const input = Math.max(0, (rawInput ?? 0) - (baselineInput ?? 0));
-    const cached = Math.max(0, (rawCached ?? 0) -
-        (numberOf(inheritedUsageBaseline?.cached_input_tokens) ?? 0));
-    const output = Math.max(0, (rawOutput ?? 0) - (baselineOutput ?? 0));
-    const reportedTotalTokens = rawReportedTotal === undefined
-        ? undefined
-        : Math.max(0, rawReportedTotal - (baselineReportedTotal ?? 0));
-    const latestTurnUsage = lastTurn
-        ? codexTurnUsage(lastTurn)
-        : undefined;
     const workingDirectory = absoluteWorkingDirectory(dominantCodexCwd(rootCwd, toolWorkdirs));
     const project = projectFromCwd(workingDirectory);
     const activity = buildLocalAgentActivity({
@@ -357,16 +474,15 @@ export function parseCodexRollout(content, onEntry, onDiagnostic) {
             sessionId,
             rateLimits: lastRateLimits,
             activity,
-            latestTurnUsage,
+            ...(usageSupport === "complete" && parsedTurn.usage
+                ? { latestTurnUsage: parsedTurn.usage }
+                : {}),
             usageScope: "session_cumulative",
             usageSupport,
-            ...(reportedTotalTokens !== undefined ? { reportedTotalTokens } : {}),
-            usage: {
-                // Codex input_tokens INCLUDES cached tokens; split them out.
-                inputTokens: Math.max(0, input - cached),
-                outputTokens: output,
-                cacheReadTokens: cached
-            }
+            ...(parsedUsage.reportedTotalTokens !== undefined
+                ? { reportedTotalTokens: parsedUsage.reportedTotalTokens }
+                : {}),
+            usage: parsedUsage.usage
         }];
 }
 function collectEmbeddedToolWorkdirs(input, toolWorkdirs) {
@@ -507,6 +623,648 @@ export async function loadLocalAgentUsage(options = {}) {
         ...(codexInvocationFiles ? { codexInvocationFiles } : {})
     };
 }
+/**
+ * Stream only the financial evidence required by init/status snapshots.
+ *
+ * Unlike `loadLocalAgentUsage`, this path never derives prompts, file focus,
+ * tool activity, or invocation evidence. It deliberately returns the same
+ * result contract so callers can preserve the existing evidence and
+ * diagnostics vocabulary without maintaining a second loader schema. Codex
+ * project attribution intentionally uses only root metadata; home-launched
+ * tool-workdir inference remains exclusive to the full qualitative loader.
+ */
+export async function loadLocalAgentFinancialUsage(options = {}) {
+    const home = homedir();
+    const claudeDir = options.claudeProjectsDir ?? join(home, ".claude", "projects");
+    const codexDir = options.codexSessionsDir ?? join(home, ".codex", "sessions");
+    const claudeCalls = [];
+    const codexCalls = [];
+    const since = options.sinceIso ? Date.parse(options.sinceIso) : undefined;
+    const sinceMs = typeof since === "number" && Number.isFinite(since) ? since : undefined;
+    const claudeDiagnostics = [];
+    const codexDiagnostics = [];
+    const sourceScans = [
+        {
+            ...emptySourceScan("claude-code"),
+            filesSkippedBeforeWindow: 0,
+            jsonlValidationCoverage: "complete"
+        },
+        {
+            ...emptySourceScan("codex"),
+            filesSkippedBeforeWindow: 0,
+            filesReadFinancially: 0,
+            bytesSkippedAsNonFinancialHistory: 0,
+            nonFinancialLinesPrefiltered: 0,
+            nonFinancialBytesPrefiltered: 0,
+            jsonlValidationCoverage: "complete"
+        }
+    ];
+    const claudeScan = sourceScans[0];
+    const codexScan = sourceScans[1];
+    const scanClaude = async () => {
+        for (const file of await listJsonlFiles(claudeDir, claudeScan, claudeDiagnostics)) {
+            if (!await shouldStreamFile(file, sinceMs, "claude-code", claudeScan, claudeDiagnostics)) {
+                continue;
+            }
+            const fileCalls = [];
+            const fileDiagnostics = [];
+            const seen = new Set();
+            let streamed;
+            try {
+                streamed = await streamJsonlRecords(file, (entry) => {
+                    const call = parseClaudeFinancialEntry(entry, file, sinceMs, seen, (diagnostic) => fileDiagnostics.push(diagnostic));
+                    if (call)
+                        fileCalls.push(call);
+                });
+            }
+            catch (error) {
+                recordUnreadableFile("claude-code", claudeScan, claudeDiagnostics, error);
+                continue;
+            }
+            if (!streamed.hadContent)
+                continue;
+            claudeScan.filesParsed += 1;
+            claudeCalls.push(...fileCalls);
+            for (const diagnostic of fileDiagnostics) {
+                recordParseDiagnostic("claude-code", claudeScan, claudeDiagnostics, diagnostic);
+            }
+            if (streamed.malformedLines > 0) {
+                recordParseDiagnostic("claude-code", claudeScan, claudeDiagnostics, {
+                    code: "malformed_jsonl",
+                    count: streamed.malformedLines
+                });
+            }
+        }
+    };
+    const scanCodex = async () => {
+        for (const file of await listJsonlFiles(codexDir, codexScan, codexDiagnostics)) {
+            if (!basename(file).startsWith("rollout-"))
+                continue;
+            if (!await shouldStreamFile(file, sinceMs, "codex", codexScan, codexDiagnostics)) {
+                continue;
+            }
+            let financialFile;
+            try {
+                financialFile = await readCodexFinancialFile(file);
+            }
+            catch (error) {
+                recordUnreadableFile("codex", codexScan, codexDiagnostics, error);
+                continue;
+            }
+            if (!financialFile.hadContent)
+                continue;
+            codexScan.filesParsed += 1;
+            codexScan.filesReadFinancially = (codexScan.filesReadFinancially ?? 0) + 1;
+            codexScan.bytesSkippedAsNonFinancialHistory =
+                (codexScan.bytesSkippedAsNonFinancialHistory ?? 0) + financialFile.bytesSkipped;
+            codexScan.nonFinancialLinesPrefiltered =
+                (codexScan.nonFinancialLinesPrefiltered ?? 0) + financialFile.prefilteredLines;
+            codexScan.nonFinancialBytesPrefiltered =
+                (codexScan.nonFinancialBytesPrefiltered ?? 0) + financialFile.prefilteredBytes;
+            if (financialFile.bytesSkipped > 0 || financialFile.prefilteredLines > 0) {
+                codexScan.jsonlValidationCoverage = "financial_events_only";
+            }
+            if (financialFile.malformedLines > 0) {
+                recordParseDiagnostic("codex", codexScan, codexDiagnostics, {
+                    code: "malformed_jsonl",
+                    count: financialFile.malformedLines
+                });
+            }
+            const state = createCodexFinancialStreamState();
+            for (const entry of financialFile.entries) {
+                consumeCodexFinancialEntry(state, entry);
+            }
+            const call = finishCodexFinancialStream(state, (diagnostic) => {
+                recordParseDiagnostic("codex", codexScan, codexDiagnostics, diagnostic);
+            });
+            if (call)
+                codexCalls.push(call);
+        }
+    };
+    // The sources are independent and can be traversed concurrently without
+    // changing per-file ordering or cumulative-session deduplication.
+    await Promise.all([scanClaude(), scanCodex()]);
+    const calls = [...claudeCalls, ...codexCalls];
+    const normalizedCalls = dedupeCumulativeSessionCalls(calls);
+    const filtered = typeof sinceMs === "number"
+        ? normalizedCalls.filter((call) => Date.parse(call.timestamp) >= sinceMs)
+        : normalizedCalls;
+    return {
+        records: aggregateCalls(filtered),
+        calls: filtered,
+        filesParsed: claudeScan.filesParsed + codexScan.filesParsed,
+        agentsDetected: [...new Set(filtered.map((call) => call.agent))],
+        sourceScans,
+        diagnostics: [...claudeDiagnostics, ...codexDiagnostics]
+    };
+}
+async function streamJsonlRecords(file, onRecord) {
+    const input = createReadStream(file, { encoding: "utf8" });
+    const lines = createInterface({ input, crlfDelay: Infinity });
+    let hadContent = false;
+    let malformedLines = 0;
+    try {
+        for await (const line of lines) {
+            hadContent = true;
+            if (!line.trim())
+                continue;
+            let entry;
+            try {
+                entry = JSON.parse(line);
+            }
+            catch {
+                malformedLines += 1;
+                continue;
+            }
+            // Parsed event type is checked before any deeper traversal. This keeps
+            // qualitative payloads out of the fast path while still validating
+            // every JSONL line and reporting malformed coverage honestly.
+            if (isRecord(entry))
+                onRecord(entry);
+        }
+    }
+    finally {
+        lines.close();
+        input.destroy();
+    }
+    return { hadContent, malformedLines };
+}
+const FINANCIAL_REVERSE_CHUNK_BYTES = 16 * 1024 * 1024;
+/**
+ * Codex token counters are cumulative. Reading the root metadata plus the
+ * newest complete financial tail is therefore equivalent to replaying copied
+ * prompt/tool history, while avoiding multi-gigabyte inherited histories in
+ * compaction and subagent rollouts.
+ *
+ * The reverse scan has proof-based stopping conditions. A normal rollout must
+ * expose the newest total, last-turn usage, model, and rate-limit snapshot.
+ * Subagent rollouts always continue to byte zero because the full parser's
+ * first qualifying task boundary cannot be proved from timestamps alone. If
+ * any normal-session proof is missing, it also continues to byte zero; no
+ * byte/line cap silently truncates evidence.
+ */
+async function readCodexFinancialFile(file) {
+    const handle = await open(file, "r");
+    try {
+        const fileStat = await handle.stat();
+        if (fileStat.size === 0) {
+            return {
+                hadContent: false,
+                malformedLines: 0,
+                entries: [],
+                bytesSkipped: 0,
+                prefilteredLines: 0,
+                prefilteredBytes: 0
+            };
+        }
+        const rootEntry = await readFirstJsonlRecord(handle, fileStat.size);
+        const rootPayload = rootEntry?.type === "session_meta" && isRecord(rootEntry.payload)
+            ? rootEntry.payload
+            : undefined;
+        const isSubagent = Boolean(rootPayload) && (stringOf(rootPayload?.thread_source) === "subagent" ||
+            isRecord(rootPayload?.source) && "subagent" in rootPayload.source);
+        const entriesReverse = [];
+        let malformedLines = 0;
+        let prefilteredLines = 0;
+        let prefilteredBytes = 0;
+        let position = fileStat.size;
+        // Chunks for one cross-boundary line are retained newest-first. Appending
+        // is O(1); they are ordered only once when the line's start is found.
+        let suffixPartsReverse = [];
+        let stoppedEarly = false;
+        let bytesSkipped = 0;
+        const proof = createCodexReverseProof(isSubagent, !rootPayload);
+        while (position > 0 && !stoppedEarly) {
+            const chunkStart = Math.max(0, position - FINANCIAL_REVERSE_CHUNK_BYTES);
+            const length = position - chunkStart;
+            const chunk = Buffer.allocUnsafe(length);
+            const { bytesRead } = await handle.read(chunk, 0, length, chunkStart);
+            if (bytesRead !== length) {
+                throw newErrorWithCode("EIO");
+            }
+            const current = chunk.subarray(0, bytesRead);
+            let lineEnd = current.length;
+            let newline = current.lastIndexOf(0x0a, lineEnd - 1);
+            while (newline >= 0) {
+                const leadingPart = current.subarray(newline + 1, lineEnd);
+                const parts = suffixPartsReverse.length > 0
+                    ? [leadingPart, ...suffixPartsReverse.slice().reverse()]
+                    : [leadingPart];
+                const observed = parseCodexFinancialLine(parts);
+                if (observed.malformed)
+                    malformedLines += 1;
+                if (observed.prefilteredBytes > 0) {
+                    prefilteredLines += 1;
+                    prefilteredBytes += observed.prefilteredBytes;
+                }
+                if (observed.entry) {
+                    entriesReverse.push(observed.entry);
+                    if (observeCodexReverseProof(proof, observed.entry)) {
+                        stoppedEarly = true;
+                        bytesSkipped = chunkStart + newline;
+                        break;
+                    }
+                }
+                suffixPartsReverse = [];
+                lineEnd = newline;
+                newline = current.lastIndexOf(0x0a, lineEnd - 1);
+            }
+            if (!stoppedEarly) {
+                const prefix = current.subarray(0, lineEnd);
+                if (prefix.length > 0)
+                    suffixPartsReverse.push(prefix);
+                position = chunkStart;
+            }
+        }
+        if (!stoppedEarly && suffixPartsReverse.length > 0) {
+            const observed = parseCodexFinancialLine(suffixPartsReverse.slice().reverse());
+            if (observed.malformed)
+                malformedLines += 1;
+            if (observed.prefilteredBytes > 0) {
+                prefilteredLines += 1;
+                prefilteredBytes += observed.prefilteredBytes;
+            }
+            if (observed.entry)
+                entriesReverse.push(observed.entry);
+        }
+        const entries = entriesReverse.reverse();
+        // A proof-complete tail does not contain byte-zero root metadata. Inject
+        // the separately parsed first record so session identity/start/cwd remain
+        // identical to the full parser. A full scan already contains that record.
+        if (stoppedEarly && rootEntry)
+            entries.unshift(rootEntry);
+        return {
+            hadContent: true,
+            malformedLines,
+            entries,
+            bytesSkipped: Math.max(0, bytesSkipped),
+            prefilteredLines,
+            prefilteredBytes
+        };
+    }
+    finally {
+        await handle.close();
+    }
+}
+async function readFirstJsonlRecord(handle, fileSize) {
+    const chunks = [];
+    let position = 0;
+    while (position < fileSize) {
+        const length = Math.min(FINANCIAL_REVERSE_CHUNK_BYTES, fileSize - position);
+        const chunk = Buffer.allocUnsafe(length);
+        const { bytesRead } = await handle.read(chunk, 0, length, position);
+        if (bytesRead <= 0)
+            break;
+        const value = chunk.subarray(0, bytesRead);
+        const newline = value.indexOf(0x0a);
+        if (newline >= 0) {
+            chunks.push(value.subarray(0, newline));
+            break;
+        }
+        chunks.push(value);
+        position += bytesRead;
+    }
+    if (chunks.length === 0)
+        return undefined;
+    try {
+        const parsed = JSON.parse(Buffer.concat(chunks).toString("utf8"));
+        return isRecord(parsed) ? parsed : undefined;
+    }
+    catch {
+        return undefined;
+    }
+}
+const CODEX_ENVELOPE_PREFIX_BYTES = 16 * 1024;
+function parseCodexFinancialLine(parts) {
+    const byteLength = parts.reduce((total, part) => total + part.length, 0);
+    if (byteLength === 0)
+        return { malformed: false, prefilteredBytes: 0 };
+    const prefix = bufferPartsPrefix(parts, CODEX_ENVELOPE_PREFIX_BYTES).toString("utf8");
+    const classification = classifyCodexFinancialEnvelope(prefix);
+    if (classification === "nonfinancial") {
+        return { malformed: false, prefilteredBytes: byteLength };
+    }
+    const text = Buffer.concat(parts, byteLength).toString("utf8").trim();
+    if (!text)
+        return { malformed: false, prefilteredBytes: 0 };
+    let parsed;
+    try {
+        parsed = JSON.parse(text);
+    }
+    catch {
+        return { malformed: true, prefilteredBytes: 0 };
+    }
+    if (!isRecord(parsed))
+        return { malformed: false, prefilteredBytes: 0 };
+    const payload = isRecord(parsed.payload) ? parsed.payload : undefined;
+    const financial = parsed.type === "session_meta" ||
+        parsed.type === "turn_context" ||
+        parsed.type === "event_msg" && (payload?.type === "token_count" || payload?.type === "task_started");
+    return financial
+        ? { entry: parsed, malformed: false, prefilteredBytes: 0 }
+        : { malformed: false, prefilteredBytes: byteLength };
+}
+function bufferPartsPrefix(parts, limit) {
+    const selected = [];
+    let remaining = limit;
+    for (const part of parts) {
+        if (remaining <= 0)
+            break;
+        const value = part.subarray(0, Math.min(part.length, remaining));
+        if (value.length > 0)
+            selected.push(value);
+        remaining -= value.length;
+    }
+    return selected.length === 1
+        ? selected[0]
+        : Buffer.concat(selected, limit - remaining);
+}
+function classifyCodexFinancialEnvelope(prefix) {
+    // JSON property order is not semantic. Scan string/object depth so a nested
+    // payload type can never be mistaken for the top-level event type. If the
+    // bounded prefix does not prove the envelope, return unknown and let the
+    // complete line go through JSON.parse rather than silently prefiltering it.
+    const { topLevelType, payloadType } = scanCodexEnvelopeTypes(prefix);
+    if (!topLevelType)
+        return "unknown";
+    if (topLevelType === "session_meta" || topLevelType === "turn_context") {
+        return "financial";
+    }
+    if (topLevelType !== "event_msg")
+        return "nonfinancial";
+    if (!payloadType)
+        return "unknown";
+    return payloadType === "token_count" || payloadType === "task_started"
+        ? "financial"
+        : "nonfinancial";
+}
+function scanCodexEnvelopeTypes(prefix) {
+    let objectDepth = 0;
+    let payloadDepth;
+    let topLevelType;
+    let payloadType;
+    for (let index = 0; index < prefix.length; index += 1) {
+        const character = prefix[index];
+        if (character === "{") {
+            objectDepth += 1;
+            continue;
+        }
+        if (character === "}") {
+            if (payloadDepth === objectDepth)
+                payloadDepth = undefined;
+            objectDepth = Math.max(0, objectDepth - 1);
+            continue;
+        }
+        if (character !== '"')
+            continue;
+        const token = readJsonStringToken(prefix, index);
+        if (!token)
+            break;
+        index = token.end;
+        let cursor = skipJsonWhitespace(prefix, token.end + 1);
+        if (prefix[cursor] !== ":")
+            continue;
+        cursor = skipJsonWhitespace(prefix, cursor + 1);
+        if (token.value === "payload" && objectDepth === 1 && prefix[cursor] === "{") {
+            payloadDepth = objectDepth + 1;
+            continue;
+        }
+        if (token.value !== "type" || prefix[cursor] !== '"')
+            continue;
+        const value = readJsonStringToken(prefix, cursor);
+        if (!value)
+            break;
+        if (objectDepth === 1)
+            topLevelType = value.value;
+        if (payloadDepth === objectDepth)
+            payloadType = value.value;
+        index = value.end;
+    }
+    return { topLevelType, payloadType };
+}
+function readJsonStringToken(input, start) {
+    if (input[start] !== '"')
+        return undefined;
+    let escaped = false;
+    for (let index = start + 1; index < input.length; index += 1) {
+        const character = input[index];
+        if (escaped) {
+            escaped = false;
+            continue;
+        }
+        if (character === "\\") {
+            escaped = true;
+            continue;
+        }
+        if (character !== '"')
+            continue;
+        try {
+            const value = JSON.parse(input.slice(start, index + 1));
+            return typeof value === "string" ? { value, end: index } : undefined;
+        }
+        catch {
+            return undefined;
+        }
+    }
+    return undefined;
+}
+function skipJsonWhitespace(input, start) {
+    let index = start;
+    while (index < input.length && /\s/.test(input[index]))
+        index += 1;
+    return index;
+}
+function createCodexReverseProof(isSubagent, forceFullScan) {
+    return {
+        isSubagent,
+        forceFullScan,
+        totalSeen: false,
+        turnSeen: false,
+        rateLimitsSeen: false,
+        modelSeen: false
+    };
+}
+function observeCodexReverseProof(proof, entry) {
+    if (proof.forceFullScan || proof.isSubagent)
+        return false;
+    const payload = isRecord(entry.payload) ? entry.payload : undefined;
+    const info = payload?.type === "token_count" && isRecord(payload.info)
+        ? payload.info
+        : undefined;
+    const hasTotal = Boolean(info && isRecord(info.total_token_usage));
+    if (hasTotal)
+        proof.totalSeen = true;
+    if (info && isRecord(info.last_token_usage))
+        proof.turnSeen = true;
+    if (parseCodexRateLimits(payload?.rate_limits, stringOf(entry.timestamp))) {
+        proof.rateLimitsSeen = true;
+    }
+    if (entry.type === "turn_context" && stringOf(payload?.model)) {
+        proof.modelSeen = true;
+    }
+    return !proof.isSubagent &&
+        proof.totalSeen &&
+        proof.turnSeen &&
+        proof.rateLimitsSeen &&
+        proof.modelSeen;
+}
+async function shouldStreamFile(file, sinceMs, agent, scan, diagnostics) {
+    let fileStat;
+    try {
+        fileStat = await lstat(file);
+    }
+    catch (error) {
+        recordUnreadableFile(agent, scan, diagnostics, error);
+        return false;
+    }
+    // Refuse a path swapped to a symlink/non-file after directory discovery.
+    if (!fileStat.isFile()) {
+        recordUnreadableFile(agent, scan, diagnostics, newErrorWithCode("EINVAL"));
+        return false;
+    }
+    if (typeof sinceMs !== "number")
+        return true;
+    // ctime catches a file whose mtime was restored after recent replacement;
+    // birthtime catches a newly copied file with a preserved old mtime. Skip
+    // only when all available metadata proves the file predates the window.
+    const newestFileEvidence = Math.max(fileStat.mtimeMs, fileStat.ctimeMs, fileStat.birthtimeMs);
+    if (!Number.isFinite(newestFileEvidence) || newestFileEvidence >= sinceMs) {
+        return true;
+    }
+    scan.filesSkippedBeforeWindow = (scan.filesSkippedBeforeWindow ?? 0) + 1;
+    return false;
+}
+function newErrorWithCode(code) {
+    const error = new Error(code);
+    error.code = code;
+    return error;
+}
+function parseClaudeFinancialEntry(entry, filePath, sinceMs, seen, onDiagnostic) {
+    if (entry.type !== "assistant")
+        return undefined;
+    const message = isRecord(entry.message) ? entry.message : undefined;
+    const usage = message && isRecord(message.usage) ? message.usage : undefined;
+    if (!message || !usage || stringOf(message.model) === "<synthetic>")
+        return undefined;
+    const dedupeKey = `${stringOf(message.id) ?? ""}:${stringOf(entry.requestId) ?? ""}`;
+    if (dedupeKey !== ":" && seen.has(dedupeKey))
+        return undefined;
+    seen.add(dedupeKey);
+    const timestamp = toIso(stringOf(entry.timestamp)) ?? new Date(0).toISOString();
+    if (typeof sinceMs === "number" && Date.parse(timestamp) < sinceMs)
+        return undefined;
+    const parsedUsage = parseClaudeFinancialUsage(usage, onDiagnostic);
+    const workingDirectory = absoluteWorkingDirectory(stringOf(entry.cwd));
+    return {
+        agent: "claude-code",
+        model: stringOf(message.model) ?? "claude-code",
+        timestamp,
+        project: projectFromCwd(workingDirectory) ?? projectFromTranscriptPath(filePath),
+        workingDirectory,
+        sessionId: stringOf(entry.sessionId),
+        ...(parsedUsage.latestTurnUsage
+            ? { latestTurnUsage: parsedUsage.latestTurnUsage }
+            : {}),
+        usageScope: "turn",
+        ...(parsedUsage.usageSupport ? { usageSupport: parsedUsage.usageSupport } : {}),
+        ...(parsedUsage.reportedTotalTokens !== undefined
+            ? { reportedTotalTokens: parsedUsage.reportedTotalTokens }
+            : {}),
+        usage: parsedUsage.usage
+    };
+}
+function createCodexFinancialStreamState() {
+    return {
+        rootSessionMetaSeen: false,
+        rootTaskStarted: false,
+        isSubagent: false
+    };
+}
+function consumeCodexFinancialEntry(state, entry) {
+    const payload = isRecord(entry.payload) ? entry.payload : undefined;
+    if (entry.type === "session_meta" && payload && !state.rootSessionMetaSeen) {
+        state.rootSessionMetaSeen = true;
+        state.sessionId = stringOf(payload.id);
+        state.rootCwd = stringOf(payload.cwd);
+        state.startedAt = toIso(stringOf(payload.timestamp) ?? stringOf(entry.timestamp));
+        state.rootStartedAtMs = timestampMilliseconds(payload.timestamp ?? entry.timestamp);
+        state.isSubagent = stringOf(payload.thread_source) === "subagent" ||
+            isRecord(payload.source) && "subagent" in payload.source;
+    }
+    if (entry.type === "turn_context" && payload) {
+        state.model = stringOf(payload.model) ?? state.model;
+        state.rootCwd ??= stringOf(payload.cwd);
+    }
+    if (state.isSubagent &&
+        !state.rootTaskStarted &&
+        payload?.type === "task_started" &&
+        isRootSpecificTaskStart(payload.started_at, state.rootStartedAtMs)) {
+        state.inheritedUsageBaseline = state.lastTotal;
+        state.lastTotal = undefined;
+        state.rootTaskStarted = true;
+        state.model = undefined;
+        state.lastTurn = undefined;
+        state.lastRateLimits = undefined;
+        state.lastActivityAt = toIso(stringOf(entry.timestamp)) ?? state.startedAt;
+    }
+    if (entry.type !== "event_msg" || payload?.type !== "token_count")
+        return;
+    const eventTimestamp = toIso(stringOf(entry.timestamp)) ??
+        state.lastActivityAt ??
+        state.startedAt;
+    const info = isRecord(payload.info) ? payload.info : undefined;
+    const total = info && isRecord(info.total_token_usage)
+        ? info.total_token_usage
+        : undefined;
+    const turn = info && isRecord(info.last_token_usage)
+        ? info.last_token_usage
+        : undefined;
+    if (total) {
+        state.lastTotal = total;
+        state.lastActivityAt = eventTimestamp;
+    }
+    if (turn) {
+        state.lastTurn = turn;
+        state.lastActivityAt = eventTimestamp;
+    }
+    const rateLimits = parseCodexRateLimits(payload.rate_limits, eventTimestamp);
+    if (rateLimits)
+        state.lastRateLimits = rateLimits;
+}
+function finishCodexFinancialStream(state, onDiagnostic) {
+    if (!state.lastTotal || state.isSubagent && !state.rootTaskStarted)
+        return undefined;
+    const parsedUsage = parseCodexCumulativeUsage(state.lastTotal, state.inheritedUsageBaseline);
+    const parsedTurn = state.lastTurn
+        ? parseCodexTurnUsage(state.lastTurn)
+        : { supported: true };
+    const usageSupport = parsedUsage.supported && parsedTurn.supported
+        ? "complete"
+        : "unsupported_token_shape";
+    if (usageSupport === "unsupported_token_shape") {
+        onDiagnostic?.({ code: "unsupported_token_shape", count: 1 });
+    }
+    const workingDirectory = absoluteWorkingDirectory(state.rootCwd);
+    return {
+        agent: "codex",
+        model: state.model ?? "codex",
+        timestamp: state.lastActivityAt ?? state.startedAt ?? new Date(0).toISOString(),
+        startedAt: state.startedAt,
+        project: projectFromCwd(workingDirectory),
+        workingDirectory,
+        sessionId: state.sessionId,
+        rateLimits: state.lastRateLimits,
+        ...(usageSupport === "complete" && parsedTurn.usage
+            ? { latestTurnUsage: parsedTurn.usage }
+            : {}),
+        usageScope: "session_cumulative",
+        usageSupport,
+        ...(parsedUsage.reportedTotalTokens !== undefined
+            ? { reportedTotalTokens: parsedUsage.reportedTotalTokens }
+            : {}),
+        usage: parsedUsage.usage
+    };
+}
 /** Aggregate per-call usage into one UsageRecord per day+agent+model+project. */
 export function aggregateCalls(calls) {
     const groups = new Map();
@@ -969,19 +1727,6 @@ function topicTokens(value) {
         .replace(/\b(?:attached|attachment|clipboard|image|images|photo|picture|screenshot|screenshots)\b/gi, " ");
     return promptTokens(sanitizeLocalActivityText(withoutAbsolutePaths));
 }
-function codexTurnUsage(value) {
-    const rawInput = numberOf(value.input_tokens) ?? 0;
-    const cached = Math.min(rawInput, numberOf(value.cached_input_tokens) ?? 0);
-    const output = numberOf(value.output_tokens) ?? 0;
-    return {
-        inputTokens: Math.max(0, rawInput - cached),
-        outputTokens: output,
-        cacheReadTokens: cached,
-        contextTokens: rawInput,
-        totalTokens: numberOf(value.total_tokens) ?? rawInput + output,
-        source: "transcript_last_token_usage"
-    };
-}
 function toTurnUsage(usage, source) {
     const contextTokens = usage.inputTokens +
         (usage.cacheReadTokens ?? 0) +
@@ -1096,6 +1841,10 @@ function sum(calls, pick) {
 function numberOf(value) {
     return typeof value === "number" && Number.isFinite(value) ? value : undefined;
 }
+function tokenComponentOf(value) {
+    const parsed = numberOf(value);
+    return parsed !== undefined && parsed >= 0 ? parsed : undefined;
+}
 function stringOf(value) {
     return typeof value === "string" && value.length > 0 ? value : undefined;
 }