@ramarivera/coding-agent-langfuse 0.1.41 → 0.1.43

package/README.md

@@ -3,9 +3,10 @@
 Universal coding-agent Langfuse backfiller and OTLP exporter helpers.
 
 It imports local histories from Codex, Claude Code, Grok, OpenCode, and Pi into
-Langfuse as session traces with child observations. LLM usage records are kept
-as observation metadata so historical imports do not create Langfuse billing or
-cost rows. Tool calls remain child spans under the same session.
+Langfuse as session traces with child observations. LLM generations include
+Langfuse canonical `usage_details` and `cost_details` attributes so historical
+backfills participate in Langfuse model-usage and cost dashboards. Tool calls
+remain child spans under the same session.
 
 ```sh
 coding-agent-langfuse-backfill --agents codex,claude,grok,pi,opencode
@@ -35,6 +36,48 @@ npx @ramarivera/coding-agent-langfuse@latest \
 The importer is fail-fast: the first failed OTLP POST stops the run, prints the
 real network cause, and preserves local state so reruns resume cleanly.
 
+## Cost calculation
+
+Backfill cost calculation follows Langfuse's OpenTelemetry mapping: generation
+spans receive `langfuse.observation.usage_details` and
+`langfuse.observation.cost_details` JSON attributes. If a source history already
+records a total cost, that recorded value wins. Otherwise, the importer
+calculates per-usage-type USD costs from a model catalog using rates in USD per
+1M tokens.
+
+The built-in catalog covers OpenAI GPT-5.5 API list pricing plus the toolbox/Pi
+models already used in local configuration, including Fireworks Kimi K2.6,
+Fireworks DeepSeek V4 Pro, MiniMax-M3, Together DeepSeek/Kimi/GLM/MiniMax, and
+Zai GLM. `gpt-5.5` is charged at current standard API list price by default:
+`$5.00` input, `$0.50` cached input, and `$30.00` output per 1M tokens. GPT-5.5
+Pro defaults to `$30.00` input and `$180.00` output per 1M tokens.
+
+Use an override only when you intentionally want a different accounting policy:
+
+```sh
+npx @ramarivera/coding-agent-langfuse@latest \
+  --agents codex \
+  --cost-rates-json '{"gpt-5.5":{"input":1,"output":2,"cacheRead":0.1,"cacheWrite":0}}'
+```
+
+You can also keep the policy in a JSON file and pass `--cost-rates PATH`, or set
+`CODING_AGENT_LANGFUSE_COST_RATES_PATH` /
+`CODING_AGENT_LANGFUSE_COST_RATES_JSON` for both manual backfills and generated
+services. A file can be either a direct model map or `{ "rates": { ... } }`:
+
+```json
+{
+  "rates": {
+    "gpt-5.5": {
+      "input": 1,
+      "output": 2,
+      "cacheRead": 0.1,
+      "cacheWrite": 0
+    }
+  }
+}
+```
+
 ## Follow as a host service
 
 Install a live follower directly from npm. The generated service keeps inference
@@ -112,7 +155,7 @@ npm run test:e2e
 The e2e suite verifies:
 
 - Codex full session traces with messages, reasoning, tool calls, tool results,
-  and usage metadata
+  usage details, and cost details
 - Follow mode picking up newly written Codex events
 - One CLI run posting reconstructable traces for Claude Code, Codex, Grok,
   OpenCode, and Pi

package/dist/backfill.d.ts

@@ -8,6 +8,7 @@ type Usage = {
     cacheWrite?: number;
     total?: number;
     cost?: number;
+    inputIncludesCache?: boolean;
 };
 type BackfillEvent = {
     agent: AgentName;
@@ -39,10 +40,24 @@ type BackfillOptions = {
     limit?: number;
     sinceMs?: number;
     untilMs?: number;
+    sessionIds: Set<string>;
     batchSize: number;
     maxRequestBytes: number;
     maxFieldBytes: number;
     postDelayMs: number;
+    costRates: CostCatalog;
+};
+type CostRates = {
+    input?: number;
+    output?: number;
+    reasoning?: number;
+    cacheRead?: number;
+    cacheWrite?: number;
+};
+type CostCatalog = Record<string, CostRates>;
+type OtlpOptions = {
+    maxFieldBytes?: number;
+    costRates?: CostCatalog;
 };
 type RunSummary = {
     discovered: Record<string, number>;
@@ -64,6 +79,7 @@ declare const allAgents: AgentName[];
 declare const defaultEndpoint = "https://langfuse.ai.roxasroot.net/otel/v1/traces";
 declare function parseArgs(argv: string[]): BackfillOptions;
 declare function codexEvents(homeDir: string, options?: {
+    sessionIds?: Set<string>;
     sinceMs?: number;
 }): BackfillEvent[];
 declare function claudeEvents(homeDir: string): BackfillEvent[];
@@ -75,9 +91,7 @@ declare function opencodeEvents(homeDir: string, options?: {
     untilMs?: number;
 }): BackfillEvent[];
 declare function fingerprint(event: BackfillEvent): string;
-declare function toOtlp(events: BackfillEvent[], options?: {
-    maxFieldBytes?: number;
-}): Record<string, unknown>;
+declare function toOtlp(events: BackfillEvent[], options?: OtlpOptions): Record<string, unknown>;
 declare function discoverEvents(options: BackfillOptions): BackfillEvent[];
 declare function run(options: BackfillOptions): Promise<RunSummary>;
 declare function follow(options: BackfillOptions): Promise<FollowSummary>;

package/dist/backfill.js

@@ -5,13 +5,13 @@ import { existsSync, mkdirSync, renameSync, readdirSync, readFileSync, statSync,
 import { hostname, homedir } from "node:os";
 import { dirname, join } from "node:path";
 const allAgents = ["claude", "codex", "grok", "opencode", "pi"];
-const importIdentityVersion = "v8-cached-input-token-split";
+const importIdentityVersion = "v9-cost-details";
 const importIdentityVersions = {
-    claude: "v11-tool-results",
-    codex: "v9-codex-conversation-events",
-    grok: "v11-chat-history-only",
-    opencode: "v10-opencode-message-parts",
-    pi: "v11-tool-results",
+    claude: "v12-cost-details",
+    codex: "v10-cost-details",
+    grok: "v12-cost-details",
+    opencode: "v11-cost-details",
+    pi: "v12-cost-details",
 };
 const defaultEndpoint = "https://langfuse.ai.roxasroot.net/otel/v1/traces";
 const deadRemoteEndpoint = "http://langfuse.ai.roxasroot.net:14318/v1/traces";
@@ -19,6 +19,79 @@ const defaultMaxRequestBytes = 12 * 1024 * 1024;
 const defaultMaxFieldBytes = 512 * 1024;
 const defaultStatePath = join(homedir(), ".local/state/coding-agent-langfuse/backfill-v6.json");
 const currentHost = hostname();
+const kimiFirepassRates = {
+    input: 2,
+    output: 8,
+    cacheRead: 0.3,
+    cacheWrite: 0,
+};
+const deepseekFireworksRates = {
+    input: 1.74,
+    output: 3.48,
+    cacheRead: 0.15,
+    cacheWrite: 0,
+};
+const miniMaxM3Rates = {
+    input: 0.6,
+    output: 2.4,
+    cacheRead: 0.12,
+    cacheWrite: 0,
+};
+const gpt55Rates = {
+    input: 5,
+    output: 30,
+    cacheRead: 0.5,
+    cacheWrite: 5,
+};
+const gpt55ProRates = {
+    input: 30,
+    output: 180,
+    cacheRead: 30,
+    cacheWrite: 30,
+};
+const defaultCostRates = {
+    "gpt-5.5": gpt55Rates,
+    "openai/gpt-5.5": gpt55Rates,
+    "gpt-5.5-pro": gpt55ProRates,
+    "openai/gpt-5.5-pro": gpt55ProRates,
+    "accounts/fireworks/routers/kimi-k2p6-turbo": kimiFirepassRates,
+    "fireworks-firepass/accounts/fireworks/routers/kimi-k2p6-turbo": kimiFirepassRates,
+    "kimi-for-coding": kimiFirepassRates,
+    "accounts/fireworks/models/deepseek-v4-pro": deepseekFireworksRates,
+    "fireworks/accounts/fireworks/models/deepseek-v4-pro": deepseekFireworksRates,
+    "MiniMax-M3": miniMaxM3Rates,
+    "minimax/MiniMax-M3": miniMaxM3Rates,
+    "together/deepseek-ai/DeepSeek-V4-Pro": {
+        input: 2.1,
+        output: 4.4,
+        cacheRead: 0.2,
+        cacheWrite: 0,
+    },
+    "together/zai-org/GLM-5.1": {
+        input: 1.4,
+        output: 4.4,
+        cacheRead: 0.2,
+        cacheWrite: 0,
+    },
+    "together/moonshotai/Kimi-K2.6": {
+        input: 1.2,
+        output: 4.5,
+        cacheRead: 0.2,
+        cacheWrite: 0,
+    },
+    "together/MiniMaxAI/MiniMax-M2.7": {
+        input: 0.3,
+        output: 1.2,
+        cacheRead: 0.06,
+        cacheWrite: 0,
+    },
+    "zai/glm-5.1": {
+        input: 1.4,
+        output: 4.4,
+        cacheRead: 0.2,
+        cacheWrite: 0,
+    },
+};
 function projectMetadata(cwd) {
     if (!cwd)
         return {};
@@ -40,11 +113,14 @@ Options:
   --home PATH             Home directory to scan (default: current user home)
   --since ISO_OR_MS       Only import events at or after this timestamp
   --until ISO_OR_MS       Only import events before or at this timestamp
+  --session-id ID         Only import one session id; repeat or comma-separate
   --limit N               Stop after N unsent events
   --batch-size N          OTLP spans per POST (default: 50)
   --max-request-bytes N   Split OTLP POSTs below this JSON byte size (default: ${defaultMaxRequestBytes})
   --max-field-bytes N     Truncate individual input/output fields above this byte size (default: ${defaultMaxFieldBytes})
   --post-delay-ms N       Delay after each successful OTLP POST (default: 0)
+  --cost-rates PATH       JSON model cost-rate overrides in USD per 1M tokens
+  --cost-rates-json JSON  Inline JSON model cost-rate overrides in USD per 1M tokens
   --follow                Keep scanning and sending newly written events
   --poll-interval-ms N    Delay between --follow scans (default: 5000)
   --idle-exit-after-ms N  Stop --follow after this much time without new sends
@@ -70,10 +146,12 @@ function parseArgs(argv) {
     let postDelayMs = Number.parseInt(process.env.LANGFUSE_BACKFILL_POST_DELAY_MS ?? "", 10);
     if (!Number.isFinite(postDelayMs))
         postDelayMs = 0;
+    let costRates = loadCostCatalogFromEnv();
     let follow = false;
     let pollIntervalMs = 5_000;
     let idleExitAfterMs;
     const agents = new Set(allAgents);
+    const sessionIds = new Set();
     for (let i = 0; i < argv.length; i++) {
         const arg = argv[i];
         const next = () => {
@@ -123,6 +201,12 @@ function parseArgs(argv) {
         else if (arg === "--post-delay-ms") {
             postDelayMs = Number.parseInt(next(), 10);
         }
+        else if (arg === "--cost-rates") {
+            costRates = mergeCostCatalog(costRates, loadCostCatalogFile(next()));
+        }
+        else if (arg === "--cost-rates-json") {
+            costRates = mergeCostCatalog(costRates, parseCostCatalogJson(next()));
+        }
         else if (arg === "--follow") {
             follow = true;
         }
@@ -138,6 +222,13 @@ function parseArgs(argv) {
         else if (arg === "--until") {
             untilMs = parseTime(next());
         }
+        else if (arg === "--session-id") {
+            for (const item of next().split(",")) {
+                const sessionId = item.trim();
+                if (sessionId.length > 0)
+                    sessionIds.add(sessionId);
+            }
+        }
         else {
             throw new Error(`Unknown argument '${arg}'`);
         }
@@ -176,11 +267,87 @@ function parseArgs(argv) {
         limit,
         sinceMs,
         untilMs,
+        sessionIds,
         batchSize,
         maxRequestBytes,
         maxFieldBytes,
         postDelayMs,
+        costRates,
+    };
+}
+function loadCostCatalogFromEnv() {
+    let catalog = { ...defaultCostRates };
+    const path = process.env.CODING_AGENT_LANGFUSE_COST_RATES_PATH ??
+        process.env.LANGFUSE_BACKFILL_COST_RATES_PATH;
+    const inlineJson = process.env.CODING_AGENT_LANGFUSE_COST_RATES_JSON ??
+        process.env.LANGFUSE_BACKFILL_COST_RATES_JSON;
+    if (path)
+        catalog = mergeCostCatalog(catalog, loadCostCatalogFile(path));
+    if (inlineJson)
+        catalog = mergeCostCatalog(catalog, parseCostCatalogJson(inlineJson));
+    return catalog;
+}
+function loadCostCatalogFile(path) {
+    return parseCostCatalogJson(readFileSync(path, "utf8"), path);
+}
+function parseCostCatalogJson(json, source = "inline JSON") {
+    let parsed;
+    try {
+        parsed = JSON.parse(json);
+    }
+    catch (error) {
+        throw new Error(`Invalid cost rates ${source}: ${describeError(error)}`);
+    }
+    const root = asRecord(parsed);
+    const nestedRates = asRecord(root.rates);
+    const entries = Object.keys(nestedRates).length > 0 ? nestedRates : root;
+    const catalog = {};
+    for (const [modelKey, rawRates] of Object.entries(entries)) {
+        const rates = normalizeCostRates(rawRates, modelKey, source);
+        if (rates)
+            catalog[modelKey] = rates;
+    }
+    return catalog;
+}
+function normalizeCostRates(value, modelKey, source) {
+    const record = asRecord(value);
+    const rates = {
+        input: asNumber(record.input) ??
+            asNumber(record.input_cost) ??
+            asNumber(record.inputPerMillion),
+        output: asNumber(record.output) ??
+            asNumber(record.output_cost) ??
+            asNumber(record.outputPerMillion),
+        reasoning: asNumber(record.reasoning) ??
+            asNumber(record.reasoning_cost) ??
+            asNumber(record.reasoningPerMillion),
+        cacheRead: asNumber(record.cacheRead) ??
+            asNumber(record.cache_read) ??
+            asNumber(record.cachedInput) ??
+            asNumber(record.input_cached_tokens),
+        cacheWrite: asNumber(record.cacheWrite) ??
+            asNumber(record.cache_write) ??
+            asNumber(record.inputCacheCreation) ??
+            asNumber(record.input_cache_creation),
     };
+    const values = Object.entries(rates).filter(([, rate]) => rate !== undefined);
+    if (values.length === 0)
+        return undefined;
+    const cleanRates = {};
+    for (const [name, rate] of values) {
+        if (rate === undefined || rate < 0) {
+            throw new Error(`Invalid ${name} cost rate for '${modelKey}' in ${source}; rates must be non-negative USD per 1M tokens.`);
+        }
+        cleanRates[name] = rate;
+    }
+    return cleanRates;
+}
+function mergeCostCatalog(base, override) {
+    const merged = { ...base };
+    for (const [modelKey, rates] of Object.entries(override)) {
+        merged[modelKey] = { ...(merged[modelKey] ?? {}), ...rates };
+    }
+    return merged;
 }
 function normalizeEndpoint(endpoint) {
     if (endpoint !== deadRemoteEndpoint)
@@ -302,10 +469,11 @@ function normalizeUsage(value) {
     const cache = asRecord(record.cache);
     const inputDetails = asRecord(record.input_tokens_details);
     const outputDetails = asRecord(record.output_tokens_details);
+    const directInput = asNumber(record.input);
+    const aggregateInput = asNumber(record.input_tokens) ??
+        asNumber(record.prompt_tokens);
     const usage = {
-        input: asNumber(record.input) ??
-            asNumber(record.input_tokens) ??
-            asNumber(record.prompt_tokens),
+        input: directInput ?? aggregateInput,
         output: asNumber(record.output) ??
             asNumber(record.output_tokens) ??
             asNumber(record.completion_tokens),
@@ -330,11 +498,16 @@ function normalizeUsage(value) {
             asNumber(record.total_cost) ??
             asNumber(record.cost),
     };
+    if (usage.input !== undefined) {
+        usage.inputIncludesCache = directInput === undefined;
+    }
     if (usage.total === undefined) {
+        const cacheRead = usage.inputIncludesCache === false ? (usage.cacheRead ?? 0) : 0;
         const total = (usage.input ?? 0) +
             (usage.output ?? 0) +
             (usage.reasoning ?? 0) +
-            (usage.cacheWrite ?? 0);
+            (usage.cacheWrite ?? 0) +
+            cacheRead;
         if (total > 0)
             usage.total = total;
     }
@@ -355,8 +528,8 @@ function usageDetails(usage) {
     if (!usage)
         return undefined;
     const details = {};
-    const cachedInput = usage.cacheRead ?? 0;
-    const cacheWrite = usage.cacheWrite ?? 0;
+    const cachedInput = usage.inputIncludesCache === false ? 0 : (usage.cacheRead ?? 0);
+    const cacheWrite = usage.inputIncludesCache === false ? 0 : (usage.cacheWrite ?? 0);
     const regularInput = usage.input === undefined
         ? undefined
         : Math.max(usage.input - cachedInput - cacheWrite, 0);
@@ -374,6 +547,62 @@ function usageDetails(usage) {
         details.total = usage.total;
     return Object.keys(details).length > 0 ? details : undefined;
 }
+function calculateCost(event, usage, costRates) {
+    if (!usage)
+        return undefined;
+    if (event.usage?.cost !== undefined) {
+        return {
+            details: { total: roundCost(event.usage.cost) },
+            source: "recorded",
+        };
+    }
+    const match = findCostRates(event, costRates);
+    if (!match)
+        return undefined;
+    const { rates, modelKey } = match;
+    const details = {};
+    setCostPart(details, "input", usage.input, rates.input);
+    setCostPart(details, "output", usage.output, rates.output);
+    setCostPart(details, "output_reasoning", usage.output_reasoning, rates.reasoning ?? rates.output);
+    setCostPart(details, "input_cached_tokens", usage.input_cached_tokens, rates.cacheRead ?? rates.input);
+    setCostPart(details, "input_cache_creation", usage.input_cache_creation, rates.cacheWrite ?? rates.input);
+    if (Object.keys(details).length === 0)
+        return undefined;
+    details.total = roundCost(Object.values(details).reduce((sum, value) => sum + value, 0));
+    return {
+        details,
+        source: "calculated",
+        modelKey,
+        rates,
+    };
+}
+function findCostRates(event, costRates) {
+    const modelName = normalizeModelName(event.model);
+    const candidates = [
+        event.provider && event.model ? `${event.provider}/${event.model}` : undefined,
+        event.provider && modelName ? `${event.provider}/${modelName}` : undefined,
+        event.model,
+        modelName,
+    ];
+    const seen = new Set();
+    for (const candidate of candidates) {
+        if (!candidate || seen.has(candidate))
+            continue;
+        seen.add(candidate);
+        const rates = costRates[candidate];
+        if (rates)
+            return { modelKey: candidate, rates };
+    }
+    return undefined;
+}
+function setCostPart(details, key, tokens, usdPerMillionTokens) {
+    if (tokens === undefined || usdPerMillionTokens === undefined)
+        return;
+    details[key] = roundCost((tokens * usdPerMillionTokens) / 1_000_000);
+}
+function roundCost(value) {
+    return Number(value.toFixed(12));
+}
 function isGenerationEvent(event) {
     return event.usage !== undefined && event.role !== "user" &&
         event.role !== "developer" && event.role !== "system";
@@ -391,6 +620,11 @@ function codexSessionPathTimeMs(path) {
     return Date.UTC(Number(year), Number(month) - 1, Number(day));
 }
 function fileMayContainCodexWindow(path, stat, options) {
+    if (options.sessionIds !== undefined &&
+        options.sessionIds.size > 0 &&
+        ![...options.sessionIds].some((sessionId) => path.includes(sessionId))) {
+        return false;
+    }
     if (options.sinceMs === undefined)
         return true;
     if (stat.mtimeMs >= options.sinceMs)
@@ -1174,6 +1408,7 @@ function limitEventPayload(event, maxFieldBytes) {
 }
 function toOtlp(events, options = {}) {
     const maxFieldBytes = options.maxFieldBytes ?? defaultMaxFieldBytes;
+    const costRates = options.costRates ?? loadCostCatalogFromEnv();
     const spansByTrace = new Map();
     for (const rawEvent of events) {
         const event = limitEventPayload(rawEvent, maxFieldBytes);
@@ -1244,6 +1479,7 @@ function toOtlp(events, options = {}) {
             const modelName = normalizeModelName(event.model);
             const generation = isGenerationEvent(event);
             const usage = usageDetails(event.usage);
+            const cost = generation ? calculateCost(event, usage, costRates) : undefined;
             const eventProject = projectMetadata(event.cwd);
             const attributes = [
                 attr("service.name", `agent.${event.agent}`),
@@ -1271,7 +1507,13 @@ function toOtlp(events, options = {}) {
                 attr("langfuse.observation.metadata.project_folder", eventProject.projectFolder),
                 attr("langfuse.observation.metadata.model", modelName ?? event.model),
                 attr("langfuse.observation.metadata.provider", event.provider),
+                attr("langfuse.observation.usage_details", generation ? usage : undefined),
+                attr("langfuse.observation.cost_details", cost?.details),
                 attr("langfuse.observation.metadata.usage_details", usage),
+                attr("langfuse.observation.metadata.cost_details", cost?.details),
+                attr("langfuse.observation.metadata.cost_source", cost?.source),
+                attr("langfuse.observation.metadata.cost_model_key", cost?.modelKey),
+                attr("langfuse.observation.metadata.cost_rates", cost?.rates),
                 attr("langfuse.observation.metadata.recorded_cost", event.usage?.cost),
                 attr("langfuse.observation.input", event.input),
                 attr("langfuse.observation.output", event.output),
@@ -1408,7 +1650,10 @@ function describeError(error) {
 function discoverEvents(options) {
     const providers = {
         claude: (inner) => claudeEvents(inner.homeDir),
-        codex: (inner) => codexEvents(inner.homeDir, { sinceMs: inner.sinceMs }),
+        codex: (inner) => codexEvents(inner.homeDir, {
+            sessionIds: inner.sessionIds,
+            sinceMs: inner.sinceMs,
+        }),
         grok: (inner) => grokEvents(inner.homeDir),
         opencode: (inner) => opencodeEvents(inner.homeDir, {
             rowLimit: inner.limit,
@@ -1422,6 +1667,7 @@ function discoverEvents(options) {
         .flatMap((agent) => providers[agent](options))
         .filter((event) => options.sinceMs === undefined || event.startMs >= options.sinceMs)
         .filter((event) => options.untilMs === undefined || event.startMs <= options.untilMs)
+        .filter((event) => options.sessionIds.size === 0 || options.sessionIds.has(event.sessionId))
         .sort((a, b) => a.startMs - b.startMs);
 }
 async function run(options) {
@@ -1446,6 +1692,7 @@ async function run(options) {
                 batchSize: options.batchSize,
                 maxRequestBytes: options.maxRequestBytes,
                 maxFieldBytes: options.maxFieldBytes,
+                costRates: options.costRates,
             });
         }
         catch (error) {
@@ -1471,6 +1718,7 @@ async function run(options) {
             try {
                 await postOtlp(options.endpoint, batch, {
                     maxFieldBytes: options.maxFieldBytes,
+                    costRates: options.costRates,
                 });
                 for (const event of batch) {
                     state.sent[fingerprint(event)] = new Date().toISOString();

package/dist/service.d.ts

@@ -13,6 +13,8 @@ type ServiceOptions = {
     batchSize: number;
     pollIntervalMs: number;
     postDelayMs: number;
+    costRatesPath?: string;
+    costRatesJson?: string;
     npxPath: string;
     since?: string;
     dryRun: boolean;

package/dist/service.js

@@ -24,6 +24,8 @@ Service options:
   --batch-size N          OTLP spans per POST (default: 10)
   --poll-interval-ms N    Delay between --follow scans (default: 5000)
   --post-delay-ms N       Delay after each successful OTLP POST (default: 0)
+  --cost-rates PATH       JSON model cost-rate overrides in USD per 1M tokens
+  --cost-rates-json JSON  Inline JSON model cost-rate overrides in USD per 1M tokens
   --since ISO_OR_MS       Optional lower bound for events the follower may send
   --working-directory DIR Directory the service starts in (default: --home)
   --path VALUE            PATH value injected into the service environment
@@ -49,6 +51,10 @@ function parseServiceArgs(argv) {
     let batchSize = 10;
     let pollIntervalMs = 5_000;
     let postDelayMs = 0;
+    let costRatesPath = process.env.CODING_AGENT_LANGFUSE_COST_RATES_PATH ??
+        process.env.LANGFUSE_BACKFILL_COST_RATES_PATH;
+    let costRatesJson = process.env.CODING_AGENT_LANGFUSE_COST_RATES_JSON ??
+        process.env.LANGFUSE_BACKFILL_COST_RATES_JSON;
     let since;
     let dryRun = false;
     let start = true;
@@ -95,6 +101,12 @@ function parseServiceArgs(argv) {
         else if (arg === "--post-delay-ms") {
             postDelayMs = parseNonNegativeInt(arg, next());
         }
+        else if (arg === "--cost-rates") {
+            costRatesPath = next();
+        }
+        else if (arg === "--cost-rates-json") {
+            costRatesJson = next();
+        }
         else if (arg === "--since") {
             since = next();
         }
@@ -134,6 +146,8 @@ function parseServiceArgs(argv) {
         batchSize,
         pollIntervalMs,
         postDelayMs,
+        costRatesPath,
+        costRatesJson,
         since,
         dryRun,
         start,
@@ -282,6 +296,10 @@ function buildFollowCommand(options) {
     ];
     if (options.since)
         command.push("--since", options.since);
+    if (options.costRatesPath)
+        command.push("--cost-rates", options.costRatesPath);
+    if (options.costRatesJson)
+        command.push("--cost-rates-json", options.costRatesJson);
     return command;
 }
 function renderSystemdUnit(options, command) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ramarivera/coding-agent-langfuse",
-  "version": "0.1.41",
+  "version": "0.1.43",
   "description": "Universal coding-agent Langfuse backfiller and live OTLP helpers",
   "type": "module",
   "license": "MIT",