Local-first web research

Mono-agent’s Pi runtime exposes two complementary public-web tools:

• WebSearch discovers and ranks candidate URLs.
• WebFetch retrieves one URL and converts its content into compact, model-readable text.

Both tools return a compact JSON envelope with status (ok/partial/blocked/error), a host-written summary, untrusted content or results, source and coverage metadata, and typed next_actions with schema-valid tool arguments. partial means usable but incomplete output, blocked means policy/access/budget prevents progress, and error means execution failure. Untrusted provider and page text is listed in untrusted_fields; host-written summaries, coverage, and next actions never quote page bodies, headers, or credentials.

Both tools run inside one ephemeral controller per model run. Identical calls share in-flight work and a bounded in-memory cache; the controller and any browser namespace close at the end of the run. Successful searches also share a bounded process cache for 15 minutes. Host coordination persists operational limits only; it creates no durable search history, cookie jar, or browser profile.

Recommended configuration

The framework defaults to ["parallel", "ollama"] search (Parallel, then local Ollama) and "local" static fetch extraction. Keyless engines are opt-in. For several agents running under the same OS user, opt into host coordination:

{
  "tools": {
    "web": {
      "coordination": "host",
      "search": {
        "backend": ["parallel", "ollama"],
        "maxRequestsPerRun": 4,
        "ollama": { "baseUrl": "http://127.0.0.1:11434" }
      },
      "fetch": {
        "provider": "local",
        "render": "never",
        "browserCommand": "agent-browser"
      }
    }
  }
}

No provider block is required for anonymous Parallel or local Ollama, including when Ollama appears in a chain. To opt into SearXNG, select it explicitly, for example ["searxng", "parallel"], and configure searxng.endpoint. The legacy tools.web.search.endpoint spelling remains a migration alias. When present, it must be an unauthenticated loopback http:// URL; remote endpoints, URL credentials, queries, and fragments are rejected during config loading. The companion service is deliberately operator-owned—mono-agent probes it but never starts, stops, or upgrades it.

Set tools.web.fetch.render to auto only when this agent regularly needs JavaScript-heavy pages and agent-browser 0.33.1 or newer is installed. Static extraction remains the first choice even in auto mode.

Environment equivalents:

Config key	Environment variable	Default
tools.web.coordination	MONO_AGENT_WEB_COORDINATION	process
tools.web.search.backend	MONO_AGENT_WEB_SEARCH_BACKEND	parallel,ollama
tools.web.search.maxRequestsPerRun	MONO_AGENT_WEB_SEARCH_MAX_REQUESTS_PER_RUN	4
tools.web.search.searxng.endpoint	MONO_AGENT_WEB_SEARCH_SEARXNG_ENDPOINT	unset
legacy tools.web.search.endpoint	MONO_AGENT_WEB_SEARCH_ENDPOINT	unset
tools.web.search.ollama.baseUrl	MONO_AGENT_WEB_SEARCH_OLLAMA_BASE_URL	http://127.0.0.1:11434 when Ollama is selected
tools.web.search.ollama.apiKeyEnv	MONO_AGENT_WEB_SEARCH_OLLAMA_API_KEY_ENV	unset
tools.web.search.ollama.trustPublicUrl	MONO_AGENT_WEB_SEARCH_OLLAMA_TRUST_PUBLIC_URL	false
tools.web.search.codex.model	MONO_AGENT_WEB_SEARCH_CODEX_MODEL	gpt-5.6-luna
tools.web.search.parallel.apiKeyEnv	MONO_AGENT_WEB_SEARCH_PARALLEL_API_KEY_ENV	unset (anonymous)
tools.web.fetch.provider	MONO_AGENT_WEB_FETCH_PROVIDER	local
tools.web.fetch.parallel.apiKeyEnv	MONO_AGENT_WEB_FETCH_PARALLEL_API_KEY_ENV	unset (anonymous)
tools.web.fetch.render	MONO_AGENT_WEB_FETCH_RENDER	never
tools.web.fetch.browserCommand	MONO_AGENT_WEB_BROWSER_COMMAND	agent-browser

WebSearch

A provider name is strict; a non-empty array is an ordered fallback chain. Duplicate or unknown names are config errors. Providers returning ordinary failures, empty, irrelevant or out-of-domain results advance the chain; observed Hound access/robots refusals do not. Cooldowns and deferrals skip work. Budget exhaustion or unsafe host coordination stops the whole call. SearXNG requires an endpoint at load time; Ollama needs no block.

Provider name	Behavior
parallel	Anonymous Parallel Search MCP, with optional named API-key variable.
searxng	Configured loopback JSON endpoint.
ollama	Local Ollama by default; explicitly configured hosted search is also supported.
codex	ChatGPT-authenticated codex app-server with the configured model and web-search capability.
keyless	Virtual opt-in group: DuckDuckGo HTML, then Startpage.
duckduckgo, startpage	Select either HTML engine individually.
hound	Built-in native Node metasearch: fixed DuckDuckGo HTML, Brave HTML, and Mojeek HTML pool. No service endpoint.

Migrating from auto

Search auto was removed, including MONO_AGENT_WEB_SEARCH_BACKEND=auto. Config loading prints the equivalent previous chain for that configuration: explicit Ollama block first, then a configured SearXNG endpoint, then Codex and keyless. For example, an old SearXNG configuration reports:

tools.web.search.backend "auto" was removed; use ["searxng","codex","keyless"] (the previous auto order for this configuration)

Use that array to preserve old behavior, or omit backend to adopt the new Parallel → local Ollama default. Environment arrays use comma-separated names, for example MONO_AGENT_WEB_SEARCH_BACKEND=searxng,codex,keyless.

Parallel Search MCP

The fixed streamable HTTP endpoint is https://search.parallel.ai/mcp. Anonymous access has provider-controlled limits; no unlimited quota is promised. Queries, advisory objectives, and fetched URLs go to parallel.ai. A local fallback does not make the first Parallel request private. Select strict local Ollama/SearXNG instead when remote query disclosure is inappropriate.

Primary and up to three alternate queries are sent unchanged in one web_search call, costing one answered-search request. Language, country, time range, and domain preferences also inform an advisory objective; country support is reported as advisory rather than an enforced geographic filter. Local relevance and domain gates still apply. Published dates are shown when supplied. Excerpts use the normal bounded snippets, not an unbounded provider payload.

Optional search.parallel.apiKeyEnv and fetch.parallel.apiKeyEnv name an environment variable. Omit them for anonymous access. A named missing/empty variable is an error, not anonymous downgrade. Credentials are read at call time, never logged, and only their digest affects cache identity. No OAuth or paid REST API integration is involved.

A session id hashes the runtime run id (mono- plus SHA-256); without a run id, a random identity is memoized on the shared run state. Search and fetch share it only within that run. It is never persisted, logged, or added to shared result-cache keys. Each attempt closes its MCP client/transport, propagates abort, rejects redirects, and bounds streamed responses to 2 MiB for search or 20 MiB for fetch. Both the endpoint and fetch target must pass sandbox policy. Remote extraction cannot attest redirects that Parallel performs internally; use local when policy requires enforcement at every source redirect hop.

HTTP 429 or MCP rate-limit errors open a cooldown; malformed/protocol/auth responses are backend_unavailable, never a fabricated No results. Strict Parallel doctor/validate liveness uses only tools/list, without a query or extraction, and reports anonymous versus configured apiKeyEnv access.

Native Hound (opt-in)

Hound runs in-process in Node: no Python, external service, runtime download, MCP transport, browser impersonation, proxies, archive fallback, or key rotation. It ports selected Hound 13.2.0/ddgs engine parsers, consensus/snippet merging, canonical deduplication and host diversity, not full upstream feature parity. The package ships their licenses and precise source attribution.

{
  "tools": { "web": {
    "search": { "backend": "hound" },
    "fetch": { "provider": "hound", "render": "never" }
  } }
}

The fixed search pool discloses the query to DuckDuckGo, Brave, and Mojeek. Every actual engine/robots URL is policy-gated before host admission and shares ordinary per-engine process pacing/cooldowns. There are at most three concurrent engine tasks, 30 candidates per engine, 2 MiB per engine response, and a 15-second aggregate deadline. There is no automatic engine retry or extra-engine fallback. Partial allowlists can produce useful partial results; bounded engineOutcomes report each engine’s actual coverage independently of fatal provider failures. Useful results plus a sibling denial or exhausted dispatch ceiling remain partial; exhaustion without results, cancellation, and coordinator failure are terminal.

A query reserves one answered-search request on its first actual send. Every robots/engine send counts against the existing non-refundable dispatch ceiling (maxRequestsPerRun * 4); only the answered reservation is refunded on overall failure. A cold full-pool query normally sends six HTTP requests. Domain and relevance gates remain local. Language is advisory. Day/month/year use native DuckDuckGo and Brave parameters; Mojeek is explicitly skipped for time filters. When country is requested, only DuckDuckGo has a reviewed country transport; Brave and Mojeek are skipped before robots, admission, or budget and are reported as unsupported_country_filter. The resulting partial search therefore sends at most the DuckDuckGo robots and search requests.

Hound search and fetch proactively honor robots rules, with a five-second, 64 KiB robots bound and at most two robots redirects. Missing robots (404/410) permit access; unavailable, malformed, denied, or challenged rules fail closed. Only completed rules are cached (up to 64 origins for one hour per owner/policy/ user agent); no cross-owner in-flight coalescing or shielded requests are used. Positive Crawl-delay is conservatively deferred: fixed host pacing is not an attestation for arbitrary site-specific delays. Access/robots/auth/429 refusals stop fallback and alternate-query attempts; the fixed, independently governed sibling engines may still supply already-requested evidence.

Fetch is HTTP-only common local acquisition/extraction, including raw format, allowed request headers, focus, pagination and static page links. Source redirects are policy- and robots-checked at every hop. It never invokes browser rendering or repeats an equivalent local/Hound acquisition later in the same provider chain. Default local/browser fetch keeps its existing robots posture; it gains no new proactive robots prerequisite. Observed access refusals remain terminal for all fetch providers. Doctor and inspectHoundWeb() report native capability only, not public engine availability or a successful extraction probe.

Migrating the retired endpoint settings

Remove tools.web.search.hound.endpoint, tools.web.fetch.hound.endpoint, MONO_AGENT_WEB_SEARCH_HOUND_ENDPOINT, and MONO_AGENT_WEB_FETCH_HOUND_ENDPOINT. Their presence is rejected, even if Hound is unselected, a higher-priority layer would override the setting, or an answer is cached. Error messages do not echo endpoint values. Selecting hound without endpoint settings is now sufficient. Existing search/fetch defaults and keyless expansion are unchanged.

Ollama Web Search

Local and signed-in self-hosted Ollama default to http://127.0.0.1:11434:

{ "tools": { "web": { "search": { "backend": "ollama" } } } }

Hosted search is bound to the exact official origin and an explicitly named environment variable:

{
  "tools": { "web": { "search": {
    "backend": "ollama",
    "ollama": { "baseUrl": "https://ollama.com", "apiKeyEnv": "OLLAMA_API_KEY" }
  } } }
}

Mono-agent posts to /api/experimental/web_search locally and retries /api/web_search at the same origin only for 404 or 405. Hosted search uses only https://ollama.com/api/web_search with bearer auth. The credential is never sent to local, private, or custom origins; apiKeyEnv is rejected for those origins. A custom public origin requires HTTPS and trustPublicUrl: true, remains unauthenticated, and never receives an Ollama hosted key. Redirects are rejected. Language and time range are advisory for Ollama, and strict Ollama never falls back to another provider. Both Ollama endpoint variants receive the caller’s effective 1–10 result limit as max_results; a compatibility retry does not reset it.

The tool accepts one query, up to three alternate_queries, a result limit from 1–10, domains, exclude_domains, language, an optional country, and a time_range of day, month, or year. country is a case-insensitive, two-letter ISO 3166-1 alpha-2 code such as PL, GB, or US; output and cache identity use uppercase. Invalid/unassigned codes fail before network, admission, or search budget. For sequential providers, the primary query runs first. Supplied alternates run in order only while no relevant result has been accepted. A transport failure, quota skip or block ends that stage immediately; alternate wording cannot repair it. Codex gets at most one exact-query turn. Quotes and site: operators are never stripped or relaxed. Results are normalized, tracking parameters are removed, duplicates are fused with reciprocal-rank fusion, and include/exclude domain filters plus a deterministic query-term/quoted-phrase relevance gate are enforced before a provider can end the chain. Parallel batches the primary and alternates once; Codex receives only the primary query.

Country localization

country is a search-region/localization preference, separate from language. It does not guarantee that every result is hosted in or geographically located in that country. Provider ranking may still reflect the request’s source IP. Omitting country sends no country preference to advisory providers and selects DuckDuckGo’s documented wt-wt No region mode; it does not promise that provider-side or IP-based localization disappears.

Per-call support is explicit:

Provider	Country behavior
parallel	Adds the normalized country to the natural-language objective; filterSupport.country is advisory.
duckduckgo	Sends DuckDuckGo’s documented kl region value; wt-wt is used when omitted.
hound	Sends Hound’s reviewed DuckDuckGo l region field; Brave and Mojeek are skipped and reported for country-filtered calls.
keyless	Can continue to country-capable DuckDuckGo; Startpage is skipped rather than queried without the requested filter.
searxng, ollama, codex, startpage	This adapter has no reviewed per-call country transport, so it skips before dispatch/budget. A strict selection returns actionable unsupported_country_filter.

DuckDuckGo country support is bounded by its documented region values: AR, AU, AT, BE, BR, BG, CA, CL, CN, CO, HR, CZ, DK, EE, FI, FR, DE, GR, HK, HU, IN, ID, IE, IL, IT, JP, KR, LV, LT, MY, MX, NL, NZ, NO, PE, PH, PL, PT, RO, RU, SG, SK, SI, ZA, ES, SE, CH, TW, TH, TR, UA, GB, US, VE, and VN. Valid ISO countries outside that list are unsupported for DuckDuckGo/Hound and never fall back to global or US results. Where DuckDuckGo documents several locales (BE, CA, ID, MY, PH, CH, and US), a matching language selects that locale; otherwise the listed primary locale is used. GB maps to DuckDuckGo’s uk-en token, while country remains GB in public metadata.

Startpage documents a saved region preference, but this stateless adapter does not invent an undocumented POST field or mutate persistent settings. SearXNG’s documented Search API exposes language and time_range, not a country field; Ollama’s documented Web Search request exposes query and max_results. Codex subscription search uses the app-server path, not the separate OpenAI Responses API user_location contract.

Every backend shares the same model-facing output bounds. A result title is at most 500 characters and its snippet is at most 4,000 characters, including the visible marker [snippet truncated; use WebFetch for full source]. The ranked result entries reuse the 64 KiB UTF-8 body allocation; envelope framing (summary, coverage, and next actions) stays outside it. Under pressure, lower-ranked snippets are shortened before a whole result is omitted, and lossy truncation is reported as partial. Result entries carry the source citations (title/url/published); snippets are untrusted discovery leads. Use WebFetch on a result URL when the marker says the snippet is incomplete. Successful searches offer up to three typed WebFetch next actions for the strongest returned URLs. Suggestions are filtered against the resolved network policy at creation and again at delivery (including shared-cache hits), and WebFetch suggestions are only delivered when WebFetch is exposed to the run. A denied URL stays visible as a discovery lead but never becomes an action. No next action is offered for genuine no-results, rate limits, budgets, or other terminal failures, and next actions never repeat page prose or suggest bypasses, cooldown waits, or provider changes to evade access gates.

Start research with one broad, high-yield query that covers the decision’s main constraints. Treat snippets as leads and use WebFetch on the strongest returned URLs before searching again. Supply alternate queries only when a material evidence gap remains; do not split a topic into many narrow searches.

maxRequestsPerRun is a hard integer limit from 1 through 20 on answered provider searches in one logical runtime run. It defaults to 4 and is shared by runtime route retries. Each child agent and later run receives a fresh budget. A successful provider response costs one request, including a well-formed empty answer. Failed attempts are refunded; cache hits, in-flight followers, provider cooldown skips, sandbox denials, and Codex quota skips consume zero requests. Reservations are synchronous, so concurrent searches cannot oversubscribe the budget while responses are pending.

A separate, non-refundable ceiling of maxRequestsPerRun * 4 provider dispatches (16 by default) bounds network work even when every attempt fails. A local Ollama compatibility probe across both supported paths costs one answered search if it succeeds, but two dispatches. Outcomes expose dispatchesUsed, maxDispatches, and dispatchesRemaining alongside the existing request counters.

When either limit refuses a dispatch, WebSearch returns search_budget_exhausted, requestsUsed, requestsRemaining, retryInRun: false, and nextAction: "use_available_evidence" without sending that request. The message includes a bounded, deduplicated summary of actual provider failures in the run and known retry timing, not provider URLs or raw error bodies. If the dispatch ceiling was reached, it explicitly says the run spent its dispatches on failing providers; requestsRemaining may still be positive in that case.

A provider that returns a rate limit is deferred for the rest of that run. An ordered chain advances immediately to the next eligible provider; single names stay strict. The result reports retryAfterMs when known, an absolute retryAt, the provider disposition, and whether another search attempt in the run can help. Do not sleep, retry, or delegate to wait out a cooldown. Fetch URLs already returned, or answer from available evidence and state the limitation.

ChatGPT-subscription Codex search

Codex search uses the installed codex app-server and the operator’s existing ChatGPT sign-in. It does not read, export, log, or persist OAuth tokens, and it does not use an OpenAI API key or API-billed Responses request. Readiness requires all three of:

• account/read reports ChatGPT authentication;
• modelProvider/capabilities/read reports webSearch: true;
• model/list includes tools.web.search.codex.model.

Each fallback executes one ephemeral, low-effort search turn in a private scratch working directory. MCP servers, environments, dynamic tools, project instructions, and capability roots are empty. Mono-agent consumes only the one completed structured webSearch.results item; assistant prose and any URLs it contains are ignored. A server interaction, a second search item, or any non-search tool item interrupts and rejects the fallback. Concurrent subscription searches are serialized process-wide (and across opted-in host processes), while the ordinary successful-result cache still prevents repeated calls for the same request.

Search reads account/rateLimits/read or its update notification and caches the snapshot for at most 60 seconds. It preserves a 10% allowance reserve: if either reported Codex window is at least 90% used, it skips the turn until quota is available. Missing, invalid, or stale/unrefreshable quota also skips Codex. This uses subscription allowance, not unlimited free search. No automatic credit purchase or account rotation is involved. Language and time-range preferences are sent separately from the unchanged query; they are advisory for Codex. SearXNG supports both filters, DuckDuckGo receives its date parameter and a language hint, and Startpage receives an advisory date parameter. These HTML endpoints do not guarantee freshness. outcome.filterSupport reports these limitations; verify dates in fetched sources.

An empty result set is a successful answer (No results.) only when the backend that produced it was actually working. A tool error means every eligible backend failed or policy blocked every request; the result keeps that distinction so the model does not waste another reasoning round repeating the same call. Because No results. is a claim about the web rather than about the infrastructure, every way a backend can be blocked while still answering 200 is classified as an error instead — see the two sections below.

SearXNG engine health

A SearXNG instance whose engines are all rate-limited or captcha’d still answers HTTP 200 with an empty results array. The response’s unresponsive_engines field is the only thing that separates that from a query nothing matched:

SearXNG response	Treated as
results present	success
empty results, no unresponsive engines	genuine No results.
empty results, one or more unresponsive engines	rate_limited or backend_unavailable, naming each engine and its reason

The error text names every failed engine (duckduckgo: CAPTCHA; brave: too many requests), so a blocked instance is diagnosable from the tool output without reading container logs. An explicitly configured chain can continue to its next provider after it.

The stock SearXNG engine set may not be usable from an ordinary residential IP: engines can answer with a CAPTCHA or require an API key. Configure at least one engine that works from the operator network and use the copyable loopback, JSON-shape, and engine-health checks in the local-first web research playbook.

Keyless rate limiting

The keyless engines are free HTML endpoints that throttle by source IP, and they announce it in ways that look like success or like a network fault:

Signal	Engine	Treated as
HTTP 202, 403, or 429	DuckDuckGo	rate_limited
3xx to a captcha or block page	Startpage	rate_limited
2xx that parses to nothing but carries challenge markers	either	rate_limited
2xx proof-of-work interstitial (Anubis, “Verifying your request…”)	Startpage	rate_limited
2xx that parses to nothing	either	genuine No results.

No credentials are ever sent to these endpoints, so a 403 can only mean “blocked”, never “unauthorized”.

Redirects are never followed for search: on these engines a redirect is the block, so following it only costs a round trip and still yields no results.

The default process mode retains the process-wide keyless bounds: three requests in flight, 1.5 seconds between starts to the same engine, and a five-minute throttle cooldown. host adds admission shared by every opted-in agent and subagent under the same OS user:

Backend scope	Concurrent requests	Minimum start spacing
SearXNG endpoint	1	2 seconds
Ollama origin	1	2 seconds
Parallel MCP (search and fetch)	1	2 seconds
Other registered provider kinds	1	2 seconds
DuckDuckGo / Startpage, separately	1 each	3 seconds
Codex subscription	1	serialized
Fetch origin (HTTP and renderer admission)	2	500 ms

Host mode honors Retry-After; without it, throttled searches cool down for five minutes and fetch origins for one minute. Repeated throttling doubles that delay up to an hour. Two infrastructure failures open a one-minute cooldown. Only one probe is admitted when a cooldown expires. A later successful probe resets the failure streak. Cooldown skips make no provider request.

A search has a 60-second deadline including admission, startup and I/O. SearXNG admission and execution get a three-second stage budget when it is not the only provider in the chain. Strict SearXNG retains its 15-second per-request timeout within the total budget. Cancellation closes active Codex transport before releasing admission; process shutdown may add its bounded cleanup time.

Successful searches are cached process-wide for 15 minutes, keyed by the query parameters and the backend configuration, so sibling subagents and later turns reuse a result instead of re-querying. Failures are never cached.

outcome.rateLimited and outcome.cooldownBackends report throttling even when a fallback backend rescued the query, so a silent degradation stays visible.

Host state and recovery

~/.mono-agent/web-control/state.json stores hashed backend keys, cooldowns, PID/incarnation leases, and quota counters. Owner-private locking and atomic replacement coordinate processes without a daemon. The directory is 0700, state is 0600, capped at 256 KiB and 512 buckets. Query text, fetched content, headers and credentials are never written there. Expired leases and proven-dead owners are reclaimed during admission. Unsafe or corrupt state fails closed; there is no uncoordinated network fallback.

mono-agent web-control status --json
mono-agent web-control reset --json

Status reports only operational metadata. Reset clears validated state only when there are no active requests. It does not repair unsafe permissions or corrupt JSON; stop opted-in consumers and inspect the private directory before manual recovery. Ordinary session reset and restart do not clear host cooldowns.

SearXNG remains dependent on upstream engine limits. A VPN changes the network path, but does not expand account quota or provide a reliable search budget; shared exits can themselves be blocked. Prefer fewer queries, cached results, working operator-selected engines and respected cooldowns. The framework does not rotate accounts, proxies or VPN exits.

WebFetch

tools.web.fetch.provider defaults to "local". Select "parallel" for strict remote extraction or an ordered array such as ["local", "parallel"] for explicit fallback. Select "hound" for HTTP-only native extraction with proactive robots enforcement (see Native Hound above). Local/Hound do not repeat an equivalent acquisition in the same chain. No remote fetch fallback is enabled by default.

Parallel calls web_fetch for one URL with full_content: true, preferring full content. If only excerpts are returned, document metadata explicitly says [excerpts only]; continuation ranges refer only to the available extracted text. It supports Markdown/plain text and the usual line/output bounds.

Parallel-only calls reject format: "raw", any custom request headers, and explicit render: "auto" or "always" with unsupported_parameter, naming the option, before any connection. Config fetch.render: "auto" requires a local provider. In a chain, incompatible options skip Parallel and are handled by local extraction; they are never silently discarded or sent remotely.

Fetch advances to the next selected provider only for ordinary unusable_content (sparse loading shell), backend_unavailable, non-refusal HTTP errors, or retryable request_failed/timeout. Ordinary local transient retries complete first; Hound does not retry. Access challenges, 401/403/407/429, robots refusal, sandbox, invalid parameters, unsupported content, byte limits, cancellation, and unsafe coordination are terminal. Successful local extraction never triggers Parallel. Source failures in Parallel’s errors[] preserve their HTTP status. Rendering and alternate providers do not grant permission to bypass site policy or access controls.

WebFetch accepts http:// and https:// URLs and returns one of:

format	Result
markdown	Default. Article-shaped Markdown for HTML/RSS, pretty JSON, PDF text, or decoded plain text.
text	Readable plain text with Markdown decoration removed.
raw	Decoded response body; requires render: "never".

Static extraction is local and content-aware. Its HTML fallback and link classification algorithms adapt Hound’s MIT-licensed code using native Node parsers, without a Python runtime or Hound service for the local provider. This does not add a robots.txt preflight to existing local or browser calls:

1. Follow at most five redirects, re-checking sandbox network policy at every hop.
2. Bound transport at 20 MiB and structured parsing at 8 MiB.
3. Decode by BOM, HTTP charset, HTML meta/XML declaration, then UTF-8, reporting replacement characters and rejecting unsupported declared charsets.
4. Parse HTML with Defuddle, then Readability plus Turndown, then main-content or cleaned-body Turndown fallback. The native Hound-derived pipeline retains title metadata, rejects tiny candidates when a substantially larger article exists, and preserves links and code in fallback Markdown tables. Relative links become safe absolute HTTP(S) links.
5. Strictly parse declared JSON/XML, extract RSS/Atom entries and PDF text, or decode ordinary text.
6. Apply the normal tool-output cap and return the result inside the JSON envelope’s untrusted content field.

HTTP 429 is terminal for the current fetch; its Retry-After metadata is not shortened into an automatic retry. Observed access/authentication challenges (including challenge pages served as HTTP 503) and HTTP 403 also stop provider fallback. An explicit provider chain does not authorize bypassing a refusal. Ordinary transient transport/service outages retain bounded retry behavior.

Request headers are limited to Accept, Accept-Language, Range, and User-Agent. Cookie, authorization, proxy, forwarding, and arbitrary custom headers are rejected, as are credentials embedded in a URL.

Transient transport failures and HTTP 408/425/429/5xx responses receive up to two bounded retries. Retry-After is honored up to five seconds. Non-success HTTP responses, unsupported content, and policy denials are returned as structured tool failures; browser rendering never runs for those responses.

Read a bounded page slice

Use start_line (one-based) and max_lines (1–10000, default 200 when slicing). Omitting both preserves the normal capped document output. max_output_chars bounds the returned page content to exactly that many page characters; the truncation notice and any saved-artifact path appear in the summary, never inside the content. Envelope framing (summary, coverage, and next actions) always sits outside that budget and is never cut to fit it. A view that starts after line 1 omits earlier lines and is partial, even the final page; a start_line beyond the total reports no content as partial rather than success. The result reports startLine, endLine, totalLines and nextLine in coverage, plus a typed WebFetch continuation next action that preserves the call’s format, focus, and link options. Any returned view that omits lines or truncates to the character budget is partial, even when the requested slice itself was satisfied; coverage (truncated, line coordinates, focus block counts, link availability) distinguishes the cause. A line too large for the budget requires a larger character cap or reading the saved output artifact; it is never silently skipped.

{ "url": "https://example.com/guide", "start_line": 201, "max_lines": 100 }

Focused views and page links

WebFetch accepts an optional focus string (at most 500 characters) and an optional include_links boolean. Both are deterministic post-extraction views over the cached document: they never change transport or cache identity and add no requests.

• focus keeps the blank-line-separated blocks relevant to the focus terms, preserving document order and provenance. Filtering happens before pagination, so continuations stay in focused coordinates while the focus string is preserved. A focused subset is reported as partial; a focus with no matching blocks reports focus_no_match with no content rather than pretending full success.
• include_links lists up to 20 deduplicated absolute HTTP(S) links from the already-downloaded static HTML, labeled main-content or page. Content citations take priority over navigation/header/footer links before the cap; duplicate fragment targets collapse to one URL. Rendered, remote, raw, and non-HTML documents report the capability as unavailable with an explicit reason instead of empty success. Parallel remote extraction rejects include_links as unsupported_parameter before any connection.

The run caches at most 64 extracted documents and 32 MiB of document text. Changing the slice reuses extraction without refetching or rerendering. Cache keys retain headers, extraction/render settings and resolved network policy. Cache eviction or run completion requires a fresh fetch. One 45-second deadline covers admission, redirects, retry waits and rendering, plus bounded cleanup. In host mode a throttle starts an origin cooldown, so a retry cannot bypass it.

Browser rendering

The tool call may request render: "never", "auto", or "always", but the agent config is the authority:

• Config never is a capability ceiling: every call stays static, even if model input requests always.
• Config auto lets individual calls request or automatically trigger browser rendering.
• Call always is browser-first and strict: Node static fetch is not attempted, and a rendering failure is a tool error.
• Call auto escalates only after a successful HTML response is classified as a sparse application shell. If rendering then fails, the loading shell is not returned as success; cancellation is never returned as static success.

Automatic rendering is attempted only for successful HTML whose extracted text is sparse and whose markup looks like a client-rendered application. JSON, PDFs, feeds, plain text, and HTTP errors never launch a browser.

Each render uses one 20-second budget, a random agent-browser namespace and session, an empty locked config file, origin-scoped --allowed-domains, untrusted-content boundaries, and no profile, restore state, remote CDP attachment, auto-connect, or state autosave. It opens the requested URL, waits for DOMContentLoaded, validates the browser’s final URL against the sandbox and domain policy, reads agent-oriented page content, then closes the browser and removes its temporary config. The executable is invoked directly—browserCommand is not evaluated by a shell.

Clear authentication pages and CAPTCHA/access challenges return authentication_required or access_challenge. The renderer does not click, type, solve challenges, reuse a profile, defeat Cloudflare, or bypass robots, access controls, authentication, or site policy.

Sandbox and validation

Tool policy controls whether WebSearch / WebFetch exist. The native sandbox separately controls which network destinations they may contact:

• network.mode: "none" blocks every web request.
• localhost admits local SearXNG or Ollama but blocks public keyless search and public fetches.
• an allowlist must include the local endpoint, chatgpt.com for Codex search, plus every public destination the agent is authorized to search or fetch.
• all permits public egress while retaining filesystem enforcement.

mono-agent validate adds a Web search & fetch section. With liveness enabled it sends a bounded JSON query to the selected SearXNG or Ollama endpoint. Strict codex mode also verifies ChatGPT login, web-search capability, and model availability; auto checks that fallback lazily only if a search reaches it. When rendering is enabled, validation checks that browserCommand --version reports agent-browser 0.33.1 or newer. liveness: false skips external probes without changing structural validation.

Security and observability

Search snippets, the actual search query, and fetched pages are always labelled untrusted through untrusted_fields. WebSearch output includes bounded backend/query/provenance coverage so fallback behavior is inspectable, while sanitized failures expose only a backend and stable category. Fetch failures carry a stable code and a blocked (network_denied, access_challenge, authentication_required, rate limits including local http_429, exhausted search budget, unavailable coordination) versus error (execution/provider failure) status. Timing events retain only bounded operational fields such as status, error code, backend, attempt count, request budget and remaining count, absolute retry time, next action, byte count, HTTP/exit status, timeout, rendered, cache-hit, truncation flags, queue wait, backend time, cooldown skips and quota skips; request headers and command arguments stay out of them. WebFetch additionally reports bounded content-kind, charset, extraction-stage, parser-failure, rendering-reason, and browser-recommendation metadata without source URLs or page content.

The tools are public-web readers, not an authenticated browsing surface. Codex uses an existing ChatGPT subscription only as the search transport; neither search results nor model-visible output receives account data. The tools do not expose browser profiles, cookies, login state, file downloads, arbitrary headers, or remote SearXNG credentials. Browser rendering is a retrieval mode, not an anti-bot or authenticated browsing feature.

Adding a provider

Providers are source-level modules, not dynamically loaded config plugins:

1. Add one module under packages/agent-runtime/src/agent/tools/web-search-providers/ implementing the JSDoc SearchProvider contract in registry.js.
2. Register it once in that registry. Declare its name, configure, eligibility/requirements, admission kind/key/process policy, networkTargets, filterSupport, batchesQueries, and search function. Optional primaryOnly and chainDeadlineMs cover restricted providers.
3. Add its name/config validation, config-view/reference, docs, and tests. Do not add branches to the search chain body. Unknown coordination kinds receive the generic one-request/two-second host limit.

The chain owns admission, deferrals, refund settlement, local relevance/domain gates, reciprocal-rank fusion, and bounded untrusted output. An adapter claims its request immediately before dispatch using web-search-state.js; failed answers are refunded, while dispatches remain counted. Every actual network destination must be sandbox-gated. Never put queries, targets, credentials, or response bodies into timing/coordination metadata. The registry tests demonstrate both a fake standalone provider and ["fake-fail", "keyless"] fallback without orchestrator edits.