scoutline 0.7.0 → 0.10.0

package/dist/providers/exa/adapter.js

@@ -0,0 +1,894 @@
+/**
+ * Exa Provider Adapter (tech-plan §7, Control Mapping, Field Normalization,
+ * Failure Normalization).
+ *
+ * Implements the Exa Provider Descriptor with Search, Reader, Research,
+ * and Diagnostics capabilities on top of the direct-HTTP transport
+ * (`./client.ts`). The Adapter owns credentials, transport lifecycle,
+ * Provider field mapping, and failure normalization; shared execution
+ * owns cache and retry policy.
+ *
+ * Boundary rules (ARCHITECTURE.md §2):
+ *   - May import capability types, normalized errors, Provider identity
+ *     types, and the Adapter-local credential and transport Modules.
+ *   - Must NOT import command presentation, output mode, or another
+ *     Provider's Adapter.
+ *
+ * Field mapping (tech-plan §7 Exa mapping):
+ *   Search results[].title          -> title
+ *   Search results[].url            -> url
+ *   Search results[].highlights[]   -> summary (join with " ")
+ *   Search results[].author         -> source
+ *   Search results[].publishedDate  -> date
+ *   Search results[].score          -> (dropped)
+ *
+ * Control mapping (SearchControls → Exa-native API params):
+ *   domain      -> includeDomains: [domain]
+ *   recency     -> startPublishedDate (oneDay→now-1d ISO, etc.)
+ *   contentSize -> type (medium/omitted→"auto", high→"deep")
+ *   topic       -> category (general→omit, news→"news",
+ *                  finance→"financial report")
+ *   location    -> REJECTED (UnsupportedOptionError)
+ */
+import crypto from "node:crypto";
+import { decodeReaderFetchResult } from "../../capabilities/reader.js";
+import { decodeResearchResult } from "../../capabilities/research.js";
+import { computeAsyncJobStateHash, createProductionAsyncJobStateFile, } from "../../lib/async-job-state.js";
+import { asyncJobStateDir } from "../../lib/cache.js";
+import { ApiError, AuthError, ConfigurationError, NetworkError, QuotaError, TimeoutError, UnsupportedOptionError, ValidationError, } from "../../lib/errors.js";
+import { requireExaApiKey, isExaConfigured } from "./credentials.js";
+import { fetchExaSearch, fetchExaContents, createExaAgentRun, pollExaAgentRun, } from "./client.js";
+import { createExaDiagnosticsCapability } from "./diagnostics.js";
+// ---------------------------------------------------------------------------
+// Provider-owned credential fingerprint
+// ---------------------------------------------------------------------------
+function credentialFingerprint(apiKey) {
+    return crypto.createHash("sha256").update(apiKey).digest("hex");
+}
+function resolveApiKey(env) {
+    return requireExaApiKey(env);
+}
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+function isPlainObject(value) {
+    return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+// ---------------------------------------------------------------------------
+// Control mapping (SearchControls → Exa-native API params)
+// ---------------------------------------------------------------------------
+/**
+ * Map a recency filter to an Exa `startPublishedDate` (ISO 8601). Exa
+ * accepts a date lower bound; `noLimit` omits the filter. The cutoff is
+ * computed relative to `now` so the Adapter can produce a deterministic
+ * value in tests.
+ */
+function mapRecencyToStartPublishedDate(recency, now) {
+    const MS_PER_DAY = 24 * 60 * 60 * 1000;
+    switch (recency) {
+        case "oneDay":
+            return new Date(now.getTime() - 1 * MS_PER_DAY).toISOString();
+        case "oneWeek":
+            return new Date(now.getTime() - 7 * MS_PER_DAY).toISOString();
+        case "oneMonth":
+            return new Date(now.getTime() - 30 * MS_PER_DAY).toISOString();
+        case "oneYear":
+            return new Date(now.getTime() - 365 * MS_PER_DAY).toISOString();
+        case "noLimit":
+            return undefined;
+        default:
+            return undefined;
+    }
+}
+/**
+ * Map a topic hint to an Exa `category`. Exa has no `general` category;
+ * general is omitted so the search is unscoped. The Tavily adapter
+ * passes `topic` natively; Exa remaps to `category`.
+ */
+function mapTopicToCategory(topic) {
+    switch (topic) {
+        case "news":
+            return "news";
+        case "finance":
+            return "financial report";
+        case "general":
+            return undefined;
+        default:
+            return undefined;
+    }
+}
+/**
+ * Map Provider-neutral `SearchControls` to Exa-native search params.
+ * `now` defaults to the current time; tests pass a fixed date so the
+ * `startPublishedDate` cutoff is deterministic.
+ */
+function mapSearchControls(controls, now = new Date()) {
+    if (!controls)
+        return undefined;
+    const params = {};
+    if (controls.domain) {
+        params.includeDomains = [controls.domain];
+    }
+    if (controls.recency) {
+        const start = mapRecencyToStartPublishedDate(controls.recency, now);
+        if (start)
+            params.startPublishedDate = start;
+    }
+    if (controls.contentSize) {
+        params.type = controls.contentSize === "high" ? "deep" : "auto";
+    }
+    if (controls.topic) {
+        const category = mapTopicToCategory(controls.topic);
+        if (category)
+            params.category = category;
+    }
+    return params;
+}
+// ---------------------------------------------------------------------------
+// Response normalization
+// ---------------------------------------------------------------------------
+/**
+ * Normalize a raw Exa search response into `SearchSource[]`.
+ *
+ *   results[].title          -> title
+ *   results[].url            -> url
+ *   results[].highlights[]   -> summary (join with " "; empty→"")
+ *   results[].author         -> source
+ *   results[].publishedDate  -> date
+ *   results[].score          -> (dropped)
+ *
+ * `title` and `url` must be strings; any malformed shape is a retryable
+ * `ApiError` 500. `highlights` is optional — when missing or empty,
+ * `summary` is the empty string (the result is still returned with its
+ * title/url). `author` and `publishedDate` are optional; when present
+ * they must be strings.
+ */
+function normalizeExaSearchResults(raw) {
+    if (!isPlainObject(raw)) {
+        throw new ApiError("Exa search returned a malformed response", 500);
+    }
+    const results = raw.results;
+    if (!Array.isArray(results)) {
+        throw new ApiError("Exa search returned a malformed response", 500);
+    }
+    const out = [];
+    for (const entry of results) {
+        if (!isPlainObject(entry)) {
+            throw new ApiError("Exa search returned a malformed response", 500);
+        }
+        const title = entry.title;
+        const url = entry.url;
+        if (typeof title !== "string" || typeof url !== "string") {
+            throw new ApiError("Exa search returned a malformed response", 500);
+        }
+        // highlights is an optional string array; join with " ".
+        const highlights = entry.highlights;
+        let summary;
+        if (highlights === undefined) {
+            summary = "";
+        }
+        else if (Array.isArray(highlights)) {
+            summary = highlights.filter((h) => typeof h === "string").join(" ");
+        }
+        else {
+            throw new ApiError("Exa search returned a malformed response", 500);
+        }
+        const source = { title, url, summary };
+        if (typeof entry.author === "string") {
+            source.source = entry.author;
+        }
+        if (typeof entry.publishedDate === "string") {
+            source.date = entry.publishedDate;
+        }
+        out.push(source);
+    }
+    return out;
+}
+// ---------------------------------------------------------------------------
+// Failure normalization: stable public codes, no raw payloads (NFR-006)
+// ---------------------------------------------------------------------------
+/**
+ * Resolve a stable HTTP-style status code for retry classification.
+ * Explicit terminal client errors (400, 404, 410, 422) map to their
+ * real codes; transient failures map to a representative status in the
+ * retryable set (429 or any 5xx 500..599 inclusive). Unknown failures
+ * default to 500 (transient). When the caller already carries a numeric
+ * status (a typed ApiError), that status is honoured directly.
+ */
+function inferStatusCode(lower, known) {
+    if (typeof known === "number" && Number.isFinite(known))
+        return known;
+    if (lower.includes("404") || lower.includes("not found"))
+        return 404;
+    if (lower.includes("400") || lower.includes("bad request"))
+        return 400;
+    if (lower.includes("410") || lower.includes("gone"))
+        return 410;
+    if (lower.includes("422") || lower.includes("unprocessable"))
+        return 422;
+    if (lower.includes("500") || lower.includes("internal"))
+        return 500;
+    if (lower.includes("502") || lower.includes("bad gateway"))
+        return 502;
+    if (lower.includes("503") || lower.includes("service unavailable"))
+        return 503;
+    if (lower.includes("504") || lower.includes("gateway timeout"))
+        return 504;
+    return 500;
+}
+/**
+ * Status-keyed outward message for rewrapped Exa ApiErrors. The rewrap
+ * does not echo upstream `error.message` — a future change embedding a
+ * raw Provider body in an ApiError message would leak through
+ * normalization, the cache, and stdout. Curated constants only.
+ */
+function exaApiErrorMessage(statusCode) {
+    if (statusCode === 429)
+        return "Exa rate limit exceeded";
+    return "Exa request failed";
+}
+/**
+ * Normalize a Provider failure with sanitized messages. Raw response
+ * bodies never cross the adapter boundary. Mirrors the Tavily adapter's
+ * `normalizeTavilyError` pattern.
+ */
+function normalizeExaError(error) {
+    // QuotaError pass-through — terminal retry guarantee preserved.
+    if (error instanceof QuotaError)
+        return error;
+    // Configuration/option/validation errors carry clean, human-authored
+    // messages and are safe to surface verbatim.
+    if (error instanceof ValidationError ||
+        error instanceof UnsupportedOptionError ||
+        error instanceof ConfigurationError) {
+        return error;
+    }
+    // Re-wrap typed transport errors with sanitized messages so a raw
+    // Provider response body embedded upstream never survives. Code +
+    // statusCode (retry signal) are preserved.
+    if (error instanceof AuthError) {
+        return new AuthError("Exa authentication failed", "EXA_API_KEY");
+    }
+    if (error instanceof NetworkError) {
+        return new NetworkError("Exa network error");
+    }
+    if (error instanceof TimeoutError) {
+        return new TimeoutError(error.durationMs, "Try again or increase timeout with EXA_TIMEOUT env var");
+    }
+    if (error instanceof ApiError) {
+        const statusCode = inferStatusCode("", error.statusCode);
+        return new ApiError(exaApiErrorMessage(statusCode), statusCode);
+    }
+    const message = error instanceof Error ? error.message : String(error);
+    const lower = message.toLowerCase();
+    if (lower.includes("401") ||
+        lower.includes("403") ||
+        lower.includes("unauthorized") ||
+        lower.includes("forbidden")) {
+        return new AuthError("Exa authentication failed");
+    }
+    if (lower.includes("timeout") || lower.includes("timed out") || lower.includes("etimedout")) {
+        return new TimeoutError(30000);
+    }
+    if (lower.includes("econnrefused") ||
+        lower.includes("econnreset") ||
+        lower.includes("network") ||
+        lower.includes("enotfound") ||
+        lower.includes("fetch failed")) {
+        return new NetworkError("Exa network error");
+    }
+    if (lower.includes("429") || lower.includes("rate limit")) {
+        return new ApiError("Exa rate limit exceeded", 429);
+    }
+    return new ApiError("Exa request failed", inferStatusCode(lower));
+}
+function createExaSearchCapability(options) {
+    const { env, transport } = options;
+    const capability = {
+        validate(request) {
+            if (!request || typeof request.query !== "string" || request.query.trim() === "") {
+                throw new ValidationError("Search query must contain at least one non-whitespace character");
+            }
+            // Exa supports domain, recency, contentSize, and topic natively.
+            // location is Z.AI-specific and rejected before any transport call.
+            if (request.controls?.location !== undefined) {
+                throw new UnsupportedOptionError("exa", "search", "location");
+            }
+        },
+        cacheIdentity(request) {
+            const apiKey = resolveApiKey(env);
+            const identityRequest = {
+                query: request.query,
+            };
+            if (request.controls) {
+                identityRequest.controls = request.controls;
+            }
+            return {
+                provider: "exa",
+                capability: "search",
+                credentialFingerprint: credentialFingerprint(apiKey),
+                request: identityRequest,
+                // Exa never probes legacy keys — no legacyCandidates.
+            };
+        },
+        async invoke(request) {
+            capability.validate(request);
+            const apiKey = resolveApiKey(env);
+            try {
+                const params = mapSearchControls(request.controls);
+                const raw = await fetchExaSearch(apiKey, request.query, params, transport);
+                return normalizeExaSearchResults(raw);
+            }
+            catch (error) {
+                throw normalizeExaError(error);
+            }
+        },
+    };
+    return capability;
+}
+// ---------------------------------------------------------------------------
+// Reader validation helpers
+// ---------------------------------------------------------------------------
+/** Options the Exa Reader does NOT accept (Z.AI-only). `retainImages`
+ * is accepted but silently ignored (Exa has no equivalent param); the
+ * read command handler sends it as `true` by default. */
+const UNSUPPORTED_READER_OPTIONS = [
+    "withLinksSummary",
+    "noGfm",
+    "keepImgDataUrl",
+    "withImagesSummary",
+];
+function assertHttpUrl(url) {
+    if (typeof url !== "string" || url.length === 0) {
+        throw new ValidationError("Exa reader URL must be a non-empty string");
+    }
+    if (!/^https?:\/\//.test(url)) {
+        throw new ValidationError("URL must start with http:// or https://");
+    }
+}
+function assertNoUnsupportedReaderOptions(request) {
+    for (const key of UNSUPPORTED_READER_OPTIONS) {
+        if (request[key] === true) {
+            throw new UnsupportedOptionError("exa", "reader", key);
+        }
+    }
+}
+// ---------------------------------------------------------------------------
+// Markdown stripping (format: text — best-effort)
+// ---------------------------------------------------------------------------
+/**
+ * Best-effort markdown-to-text conversion. Exa always returns text
+ * content; when the caller requests `format: "text"`, this strips
+ * common markdown markers so the output is closer to plain text.
+ * Code blocks and tables degrade (their content is kept but
+ * formatting is lost); that is an acceptable edge case for a "rough
+ * text" mode. This is adapter-local — no shared stripper exists.
+ */
+function stripMarkdown(input) {
+    let result = input;
+    // Remove fenced code blocks (keep the inner text).
+    result = result.replace(/```[\s\S]*?```/g, (block) => block.replace(/```\w*\n?/g, "").replace(/```$/g, ""));
+    // Remove inline code backticks.
+    result = result.replace(/`([^`]+)`/g, "$1");
+    // Images: ![alt](url) → alt.
+    result = result.replace(/!\[([^\]]*)\]\([^)]+\)/g, "$1");
+    // Links: [text](url) → text.
+    result = result.replace(/\[([^\]]*)\]\([^)]+\)/g, "$1");
+    // Headers: leading # markers.
+    result = result.replace(/^#{1,6}\s+/gm, "");
+    // Emphasis: **bold**, __bold__, *italic*, _italic_, ~~strike~~.
+    result = result.replace(/\*\*(.+?)\*\*/g, "$1");
+    result = result.replace(/__(.+?)__/g, "$1");
+    result = result.replace(/\*(.+?)\*/g, "$1");
+    result = result.replace(/_(.+?)_/g, "$1");
+    result = result.replace(/~~(.+?)~~/g, "$1");
+    // Blockquotes: leading > markers.
+    result = result.replace(/^>\s+/gm, "");
+    // Horizontal rules: ---, ***, ___.
+    result = result.replace(/^[-*_]{3,}\s*$/gm, "");
+    // List markers: -, *, +, 1.
+    result = result.replace(/^[\s]*[-*+]\s+/gm, "");
+    result = result.replace(/^[\s]*\d+\.\s+/gm, "");
+    return result.trim();
+}
+// ---------------------------------------------------------------------------
+// Per-URL status total function (the load-bearing mechanism)
+// ---------------------------------------------------------------------------
+/**
+ * Known error-tag → HTTP status mapping. When `error.httpStatusCode` is
+ * present on the response, it is preferred over this table for retry
+ * classification. Unknown tags default to 500 (retryable).
+ */
+const CONTENTS_ERROR_STATUS = {
+    CRAWL_NOT_FOUND: 404,
+    UNSUPPORTED_URL: 400,
+    SOURCE_NOT_AVAILABLE: 403,
+    CRAWL_TIMEOUT: 504,
+    CRAWL_LIVECRAWL_TIMEOUT: 504,
+    CRAWL_UNKNOWN_ERROR: 500,
+};
+/**
+ * Normalize a raw Exa `/contents` response into a `ReaderFetchResult`.
+ *
+ * **Critical mechanism — per-URL status inspection (total function).**
+ * `/contents` returns HTTP 200 even when an individual URL fails. The
+ * adapter MUST resolve the status entry whose `statuses[].id` matches
+ * the requested URL, then apply a total mapping. Match by `id`, never
+ * assume `results[0]` is the requested URL.
+ *
+ * Field mapping:
+ *   results[].text   -> content   (the entry whose id/status matches)
+ *   results[].url    -> finalUrl
+ *   results[].title  -> title     (coerce blank → null)
+ *   request.url      -> url
+ *   request.format   -> contentFormat (after any text stripping)
+ *
+ * On `format: "text"`, the content is run through {@link stripMarkdown}
+ * (best-effort) and `contentFormat` is set to `"text"`.
+ */
+function normalizeExaContentsResult(raw, request) {
+    if (!isPlainObject(raw)) {
+        throw new ApiError("Exa contents returned a malformed response", 500);
+    }
+    // Step 1: find the status entry matching the requested URL. Never
+    // assume results[0] is the match — the API returns HTTP 200 even on
+    // per-URL failure. For a single-URL fetch, accept the sole entry
+    // even if its id doesn't exactly match (Exa may normalize URLs).
+    const statuses = raw.statuses;
+    if (!Array.isArray(statuses)) {
+        throw new ApiError("Exa contents returned a malformed response", 500);
+    }
+    let statusEntry = statuses.find((s) => isPlainObject(s) && typeof s.id === "string" && s.id === request.url);
+    // Single-URL fallback: if no exact id match but exactly one status
+    // entry exists, accept it. This mirrors the results[] leniency and
+    // guards against URL normalization differences.
+    if (!statusEntry && statuses.length === 1 && isPlainObject(statuses[0])) {
+        statusEntry = statuses[0];
+    }
+    if (!statusEntry) {
+        throw new ApiError("Exa contents returned a malformed response", 500);
+    }
+    const statusValue = statusEntry.status;
+    // Step 2: error path — map the tag + httpStatusCode to a sanitized ApiError.
+    if (statusValue !== "success") {
+        const errorObj = isPlainObject(statusEntry.error) ? statusEntry.error : {};
+        const tag = typeof errorObj.tag === "string" ? errorObj.tag : undefined;
+        const httpStatusCode = typeof errorObj.httpStatusCode === "number" ? errorObj.httpStatusCode : undefined;
+        const statusCode = httpStatusCode ?? CONTENTS_ERROR_STATUS[tag ?? ""] ?? 500;
+        throw new ApiError("Exa contents request failed", statusCode);
+    }
+    // Step 3: success path — find the matching result in results[].
+    const results = raw.results;
+    if (!Array.isArray(results)) {
+        throw new ApiError("Exa contents returned a malformed response", 500);
+    }
+    // For a single-URL fetch, the result entry's id or url should match.
+    const result = results.find((r) => isPlainObject(r) &&
+        ((typeof r.id === "string" && r.id === request.url) ||
+            (typeof r.url === "string" && r.url === request.url)));
+    // Fall back to the first result if no URL match (single-URL fetch).
+    const entry = result ?? (results.length > 0 && isPlainObject(results[0]) ? results[0] : null);
+    if (!entry) {
+        throw new ApiError("Exa contents returned a malformed response", 500);
+    }
+    // Step 4: field mapping.
+    const content = entry.text;
+    if (typeof content !== "string" || content.length === 0) {
+        throw new ApiError("Exa contents returned a malformed response", 500);
+    }
+    const finalUrl = typeof entry.url === "string" && entry.url.length > 0 ? entry.url : request.url;
+    const rawTitle = typeof entry.title === "string" ? entry.title.trim() : "";
+    const title = rawTitle.length > 0 ? rawTitle : null;
+    const requestedFormat = request.format ?? "markdown";
+    if (requestedFormat === "text") {
+        return {
+            schemaVersion: 1,
+            url: request.url,
+            finalUrl,
+            title,
+            content: stripMarkdown(content),
+            contentFormat: "text",
+        };
+    }
+    return {
+        schemaVersion: 1,
+        url: request.url,
+        finalUrl,
+        title,
+        content,
+        contentFormat: "markdown",
+    };
+}
+/**
+ * Map a `ReaderFetchRequest` to Exa-native contents params. The CLI
+ * `--timeout` is in seconds; Exa's `livecrawlTimeout` is in
+ * milliseconds. The conversion (`* 1000`) is validated here so a direct
+ * pass-through doesn't send 20ms instead of 20s.
+ */
+function mapReaderControls(request) {
+    if (typeof request.timeout !== "number" ||
+        !Number.isFinite(request.timeout) ||
+        request.timeout <= 0) {
+        return undefined;
+    }
+    const livecrawlTimeout = Math.round(request.timeout * 1000);
+    if (!Number.isFinite(livecrawlTimeout) || livecrawlTimeout <= 0) {
+        return undefined;
+    }
+    return { livecrawlTimeout };
+}
+function createExaReaderCapability(options) {
+    const { env, transport } = options;
+    const fetchOp = {
+        kind: "reader-fetch",
+        validate(request) {
+            assertHttpUrl(request.url);
+            assertNoUnsupportedReaderOptions(request);
+        },
+        cacheIdentity(request) {
+            const apiKey = resolveApiKey(env);
+            return {
+                provider: "exa",
+                capability: "reader",
+                operation: "reader-fetch",
+                credentialFingerprint: credentialFingerprint(apiKey),
+                request,
+                legacyCandidates: [],
+            };
+        },
+        decodeCached(value) {
+            return decodeReaderFetchResult(value);
+        },
+        async invoke(request) {
+            fetchOp.validate(request);
+            const apiKey = resolveApiKey(env);
+            try {
+                const params = mapReaderControls(request);
+                const raw = await fetchExaContents(apiKey, request.url, params, transport);
+                return normalizeExaContentsResult(raw, request);
+            }
+            catch (error) {
+                throw normalizeExaError(error);
+            }
+        },
+    };
+    return { fetch: fetchOp };
+}
+// ---------------------------------------------------------------------------
+// Research Capability (the hardest mechanism — tech-plan §3, §7)
+// ---------------------------------------------------------------------------
+const DEFAULT_RESEARCH_POLL_INTERVAL_MS = 5000;
+function resolvePollIntervalMs(env) {
+    const raw = env?.EXA_RESEARCH_POLL_INTERVAL_MS;
+    const parsed = parseInt(raw ?? "", 10);
+    return Number.isFinite(parsed) && parsed >= 0 ? parsed : DEFAULT_RESEARCH_POLL_INTERVAL_MS;
+}
+/**
+ * Build an abortable `sleep(ms)` from the injected timers. Copied from
+ * the Tavily adapter — same mechanism, different transport type.
+ */
+function makeSleep(deps, signal) {
+    const setT = deps?.setTimeout ?? setTimeout;
+    const clearT = deps?.clearTimeout ?? clearTimeout;
+    return (ms) => new Promise((resolve, reject) => {
+        if (signal?.aborted) {
+            reject(new TimeoutError(0, "Research polling aborted"));
+            return;
+        }
+        if (ms <= 0) {
+            setImmediate(() => {
+                if (signal?.aborted) {
+                    reject(new TimeoutError(0, "Research polling aborted"));
+                    return;
+                }
+                resolve();
+            });
+            return;
+        }
+        const onAbort = () => {
+            clearT(id);
+            reject(new TimeoutError(0, "Research polling aborted"));
+        };
+        const id = setT(() => {
+            signal?.removeEventListener("abort", onAbort);
+            resolve();
+        }, ms);
+        signal?.addEventListener("abort", onAbort);
+    });
+}
+function isEexistError(err) {
+    return (typeof err === "object" &&
+        err !== null &&
+        "code" in err &&
+        err.code === "EEXIST");
+}
+/**
+ * Map `model` → Exa Agent `effort`. The result echoes the REQUESTED
+ * model (not the effort string) so the contract is identical across
+ * Tavily and Exa. Exa accepts effort values `low|medium|high|xhigh|
+ * auto`; only `low` (mini), `high` (pro), and `auto` are reachable
+ * from the Normal command.
+ */
+function mapModelToEffort(model) {
+    switch (model) {
+        case "mini":
+            return "low";
+        case "pro":
+            return "high";
+        case "auto":
+        default:
+            return "auto";
+    }
+}
+/**
+ * Validate a `ResearchRequest` for Exa. Exa supports `query` and
+ * `model` natively; `outputLength`, `citationFormat`, and `domain` are
+ * concepts the Agent lacks and are rejected before transport.
+ *
+ * **OD1 note:** `domain` is rejected for now. It MAY be revalidatable
+ * against the pinned Agent's internal search-tool config — track as a
+ * follow-up if the config accepts `includeDomains`.
+ */
+function assertNoUnsupportedResearchOptions(request) {
+    if (request.outputLength !== undefined) {
+        throw new UnsupportedOptionError("exa", "research", "outputLength");
+    }
+    if (request.citationFormat !== undefined) {
+        throw new UnsupportedOptionError("exa", "research", "citationFormat");
+    }
+    if (request.domain !== undefined) {
+        throw new UnsupportedOptionError("exa", "research", "domain");
+    }
+}
+/**
+ * Normalize a completed Exa Agent run's `output` into a
+ * `ResearchResult`.
+ *
+ *   output.text                       -> report
+ *   output.grounding[].citations[]    -> sources[] (flatten {title, url};
+ *                                        drop incomplete)
+ *   output.structured                 -> ignored (distinct capability)
+ *   request.model                     -> model (echoed, NOT effort string)
+ */
+function normalizeExaResearchResult(poll, request) {
+    const output = poll.output;
+    if (!isPlainObject(output)) {
+        throw new ApiError("Exa research returned a malformed response", 500);
+    }
+    const text = output.text;
+    if (typeof text !== "string") {
+        throw new ApiError("Exa research returned a malformed response", 500);
+    }
+    const sources = [];
+    const grounding = output.grounding;
+    if (Array.isArray(grounding)) {
+        for (const entry of grounding) {
+            if (!isPlainObject(entry))
+                continue;
+            const citations = entry.citations;
+            if (!Array.isArray(citations))
+                continue;
+            for (const citation of citations) {
+                if (!isPlainObject(citation))
+                    continue;
+                if (typeof citation.title === "string" && typeof citation.url === "string") {
+                    sources.push({ title: citation.title, url: citation.url });
+                }
+            }
+        }
+    }
+    return {
+        schemaVersion: 1,
+        query: request.query,
+        model: request.model ?? "auto",
+        report: text,
+        sources,
+    };
+}
+/**
+ * True when an error from the poll GET is safe to retry. The poll is
+ * idempotent — retrying it never creates a new run or charges the
+ * account. Only transient failures (429, 5xx, network, timeout) qualify;
+ * auth/quota/validation errors are terminal and propagate immediately.
+ */
+function isTransientPollError(err) {
+    if (err instanceof ApiError && typeof err.statusCode === "number") {
+        return err.statusCode === 429 || (err.statusCode >= 500 && err.statusCode <= 599);
+    }
+    if (err instanceof NetworkError || err instanceof TimeoutError)
+        return true;
+    return false;
+}
+/**
+ * POST /agent/runs to create a task, then persist its run ID in the
+ * state file atomically. On EEXIST (a concurrent invocation already
+ * created a task for this request), read the existing state file and
+ * return its run ID instead — the concurrent task is polled, not
+ * duplicated.
+ *
+ * **OD1 limitation:** the POST happens before the `wx` write, so two
+ * callers that both read "absent" can both POST before either writes.
+ * This reduces (Ctrl-C+retry) but does NOT eliminate duplicate runs.
+ * Same pre-existing Tavily issue Exa inherits.
+ */
+async function createResearchTask(apiKey, request, identityHash, stateFile, transport) {
+    const agentParams = {
+        query: request.query,
+        effort: mapModelToEffort(request.model),
+    };
+    const created = await createExaAgentRun(apiKey, agentParams, transport);
+    const runId = created.id;
+    const state = {
+        requestId: runId,
+        identityHash,
+        createdAt: new Date().toISOString(),
+        status: "pending",
+    };
+    try {
+        await stateFile.write(identityHash, state);
+    }
+    catch (err) {
+        if (isEexistError(err)) {
+            const existing = await stateFile.read(identityHash);
+            if (existing !== null) {
+                return existing.requestId;
+            }
+        }
+        else {
+            throw err;
+        }
+    }
+    return runId;
+}
+function createExaResearchCapability(options) {
+    const { env, transport, researchStateFile } = options;
+    const run = {
+        kind: "research-fetch",
+        validate(request) {
+            if (!request || typeof request.query !== "string" || request.query.trim() === "") {
+                throw new ValidationError("Research query must contain at least one non-whitespace character");
+            }
+            assertNoUnsupportedResearchOptions(request);
+        },
+        cacheIdentity(request) {
+            const apiKey = resolveApiKey(env);
+            return {
+                provider: "exa",
+                capability: "research",
+                credentialFingerprint: credentialFingerprint(apiKey),
+                request,
+            };
+        },
+        decodeCached(value) {
+            return decodeResearchResult(value);
+        },
+        async invoke(request, signal) {
+            run.validate(request);
+            const apiKey = resolveApiKey(env);
+            const credFingerprint = credentialFingerprint(apiKey);
+            const identityHash = computeAsyncJobStateHash({
+                provider: "exa",
+                capability: "research",
+                credentialFingerprint: credFingerprint,
+                request,
+            });
+            const pollIntervalMs = resolvePollIntervalMs(transport?.env);
+            const sleep = makeSleep(transport, signal);
+            try {
+                // 1. Check for an in-flight task (resume after Ctrl-C / crash).
+                const existingState = await researchStateFile.read(identityHash);
+                let runId;
+                if (existingState !== null) {
+                    runId = existingState.requestId;
+                }
+                else {
+                    // 2. No in-flight task: POST to create one. NO retry — a
+                    //    transient POST failure is terminal (double-charge
+                    //    prevention on a usage-based endpoint).
+                    runId = await createResearchTask(apiKey, request, identityHash, researchStateFile, transport);
+                }
+                // 3. Poll loop until terminal status.
+                //    The zero-retry policy wraps the whole invoke() and protects
+                //    the POST (create). The GET (poll) is idempotent and safe to
+                //    retry — a transient 429/5xx/network error on poll MUST NOT
+                //    terminate a paid research run that is still active
+                //    server-side. So we catch transient poll errors and retry
+                //    the GET (bounded by MAX_POLL_RETRIES) before propagating.
+                const MAX_POLL_RETRIES = 3;
+                let consecutivePollFailures = 0;
+                for (;;) {
+                    if (signal?.aborted) {
+                        throw new TimeoutError(0, "Research polling aborted");
+                    }
+                    let poll;
+                    try {
+                        poll = await pollExaAgentRun(apiKey, runId, transport);
+                        consecutivePollFailures = 0;
+                    }
+                    catch (pollErr) {
+                        if (isTransientPollError(pollErr) && consecutivePollFailures < MAX_POLL_RETRIES) {
+                            consecutivePollFailures++;
+                            await sleep(pollIntervalMs);
+                            continue;
+                        }
+                        throw pollErr;
+                    }
+                    if (poll.status === "completed") {
+                        await researchStateFile.remove(identityHash);
+                        return normalizeExaResearchResult(poll, request);
+                    }
+                    if (poll.status === "failed") {
+                        await researchStateFile.remove(identityHash);
+                        throw new ApiError("Exa research task failed", 500);
+                    }
+                    if (poll.status === "cancelled") {
+                        // Exa-specific: cancelled is terminal (treated as failure).
+                        await researchStateFile.remove(identityHash);
+                        throw new ApiError("Exa research task was cancelled", 500);
+                    }
+                    if (poll.status === "not_found") {
+                        // 404 — server-side run expired. Delete stale state and
+                        // create a fresh run.
+                        await researchStateFile.remove(identityHash);
+                        runId = await createResearchTask(apiKey, request, identityHash, researchStateFile, transport);
+                        continue;
+                    }
+                    // queued or running: sleep and poll again.
+                    await sleep(pollIntervalMs);
+                }
+            }
+            catch (error) {
+                throw normalizeExaError(error);
+            }
+        },
+    };
+    return { run };
+}
+// ---------------------------------------------------------------------------
+// Descriptor factory
+// ---------------------------------------------------------------------------
+/**
+ * Build the Exa Provider Descriptor. The descriptor advertises the Exa
+ * capability set (search, reader, research, diagnostics) and constructs
+ * an Adapter whose Capabilities own credentials, transport, Provider
+ * field mapping, and failure normalization. Construction is
+ * side-effect-free; the transport is invoked per Capability call. Tests
+ * pass `transport` (typically a fake-fetch wrapper); production uses
+ * the no-argument factory which resolves to the global `fetch` and
+ * timers inside the transport Module.
+ */
+export function createExaDescriptor(dependencies) {
+    const transport = dependencies?.transport;
+    const researchStateFile = dependencies?.researchStateFile ??
+        createProductionAsyncJobStateFile(asyncJobStateDir("research"));
+    return {
+        id: "exa",
+        isConfigured(env) {
+            return isExaConfigured(env);
+        },
+        capabilities() {
+            return new Set(["search", "reader", "research", "diagnostics"]);
+        },
+        create(context) {
+            const search = createExaSearchCapability({
+                env: context.env,
+                transport,
+            });
+            const reader = createExaReaderCapability({
+                env: context.env,
+                transport,
+            });
+            const research = createExaResearchCapability({
+                env: context.env,
+                transport,
+                researchStateFile,
+            });
+            const diagnostics = createExaDiagnosticsCapability({
+                env: context.env,
+                transport,
+            });
+            return { id: "exa", search, reader, research, diagnostics };
+        },
+    };
+}
+//# sourceMappingURL=adapter.js.map