Source retrieval and connectors
Source retrieval is the evidence-gathering layer for current research. It turns each planned source query into live items from the web, social networks, communities, feeds, and focused services. The layer also chooses a usable transport, keeps failures visible, adds source-specific context, and hands one shared evidence shape to ranking and report generation. General web retrieval can use a paid search API, a keyless floor, or an explicitly selected Parallel Search MCP transport. See Research skill execution for the parent flow, Research planning and query construction for the plan that drives these requests, and Research quality and ranking for what happens after retrieval.
Retrieval boundary
The retrieval boundary sits between the query plan and the evidence pool. A planner emits a SubQuery, which is a search form, a ranking question, and a list of eligible sources. The pipeline dispatches that unit to one source connector and receives raw items plus an outcome artifact.
The connector layer does not decide the final story. It preserves the source's useful dates, text, links, authors, containers, and engagement signals so later normalization, dedupe, fusion, and ranking can compare unlike sources without hiding where an item came from.
flowchart LR
A[Query plan] --> B[Source dispatch]
B --> C[Connector transport]
C --> D[Source search result]
D --> E[Source enrichment]
E --> F[Shared source item]
F --> G[Date filter and normalization]
G --> H[Ranking and report]
C --> I[Typed source outcome]
I --> H
References: skills/…/pipeline.py• L4110-4185 skills/…/pipeline.py• L4190-4203 skills/…/schema.py• L45-92
Source dispatch
Source dispatch maps the source name in a planned query to the connector that owns that source's search, parsing, and optional enrichment. This keeps source rules at the edge of the system. The rest of the pipeline works with the same retrieval handoff even when one source uses an API, a command, a feed, a browser session, or a local service.
The main source families are:
| Family | Evidence role | Typical access boundary |
|---|---|---|
| General web | Current pages and broad reporting | Paid search API, explicit Parallel MCP, or a keyless search floor |
| Communities | Discussion, advice, and user experience | Public feeds and page surfaces, or a paid search service |
| Social platforms | Short posts, reactions, and creator signals | Platform API, browser session, CLI, or service API |
| Video and feeds | Video metadata, captions, comments, and feed notes | Local media tool, service API, or local connector |
| Specialized services | Jobs, markets, code activity, reviews, newsletters, and research papers | Service API, public endpoint, or installed source CLI |
Availability is part of dispatch. Some sources are always usable without credentials, while others appear only when a key, local command, session, channel list, or explicit opt-in is present. A requested source can therefore be valid as a product capability but absent from one run's available route.
References: skills/…/pipeline.py• L223-231 skills/…/pipeline.py• L234-381 skills/…/pipeline.py• L2057-2060 skills/…/pipeline.py• L4163-4194
Provider and transport selection
Provider selection chooses the access path that can serve a source in the current environment. It separates a source from its transport so the same source can survive a blocked endpoint, an unavailable binary, or a missing paid key when another path can provide useful evidence.
Backend chains
Each chained source has a routing descriptor that names its backends, probe rules, pin controls, and routing mode. Alternative chains represent interchangeable routes such as X, YouTube, and general web. Conditional chains represent per-query behavior such as Reddit's public search followed by a thin-result backfill.
Readiness resolution is a prediction, not a record of a past run. It probes the declared candidates without changing runtime routing, then selects the first fully usable alternative. If none is fully usable, it selects the best degraded route. Opt-in routes remain visible but are not selected automatically. Runtime failover can still differ when a credential passes a local check but fails during the live request.
| Source | Selection model | High-level route |
|---|---|---|
| X | Alternative chain | Host-fetched envelope, browser session, xAI search, official CLI, or keyed search service |
| YouTube | Alternative chain | Local video search first, then ScrapeCreators search |
| General web | Alternative chain | Brave, Exa, Serper, Parallel API, then keyless search where allowed |
| General web via MCP | Explicit route | Parallel Search MCP through Streamable HTTP |
| Conditional chain | Public keyless composite by default, with ScrapeCreators fallback or backfill |
The X source remains one source even when it changes backend. This preserves one evidence lane for downstream limits and ranking. Explicit pins can force a backend. Opt-in backends do not enter the automatic chain merely because local credentials happen to exist.
General web has two separate Parallel routes. The parallel backend is a keyed API member of the automatic web chain. The parallel-mcp backend is an explicit route selected by the caller. It uses the same general-web evidence handoff but does not enter the automatic chain.
Perplexity is a separate paid and opt-in source route. A direct key can use its Agent or Search API, while an OpenRouter key keeps the older Sonar path available. Agent mode can return a short synthesis with its citations, and explicit deep research uses a bounded background run with a receipt for its remote status.
References: skills/…/backends.py• L1-45 skills/…/backends.py• L98-187 skills/…/backends.py• L627-681 skills/…/backends.py• L733-854 skills/…/env.py• L801-899 skills/…/grounding.py• L226-280 skills/…/parallel_mcp.py• L125-232 skills/…/perplexity.py• L133-172 skills/…/perplexity.py• L985-1051
Reasoning provider boundary
The source connector route is separate from the reasoning provider route. Planner and rerank requests can use a configured Gemini, OpenAI, xAI, or OpenRouter client, or a local mode when no live reasoning provider is available. This choice supports query planning and evidence ranking; it does not choose the web, social, community, feed, or specialized-service connector.
The reasoning clients share one text and structured-output contract. Provider base-URL overrides accept either an API root or a full endpoint. The resolver appends the provider path only to a host or versioned API root. Remote endpoints must use HTTPS; cleartext HTTP is limited to loopback. Retrieval still owns source access, source fallback, and source outcomes.
References: skills/…/providers.py• L17-31 skills/…/providers.py• L44-117 skills/…/providers.py• L120-248 skills/…/providers.py• L476-557
Backend fallback
Fallback is a coverage rule. The pipeline moves to the next backend when the current path returns no items or a usable retrieval error. A successful fallback can return evidence while still carrying a warning about the failed primary path. This prevents a report from looking fully healthy when it used a weaker or less preferred route.
Fallback does not mean that all backends run in parallel. Alternative chains try one route at a time. Readiness probes first inspect all declared candidates so the prediction does not let an unavailable preferred route hide a usable fallback. Paid backend probes check credential presence rather than making a live call. Dependency probes use bounded execution when needed, while some CLI authentication probes inspect local state only. The explicit Parallel MCP route is not a member of the automatic web fallback chain.
This diagram shows the difference between route prediction and live fallback during retrieval.
flowchart TD
A[Source request] --> B{Preferred path usable?}
B -->|yes| C[Retrieve]
B -->|no| D[Next fallback]
C --> E{Items returned?}
E -->|yes| F[Keep items and outcome]
E -->|no or error| D
D --> G{Another path?}
G -->|yes| C
G -->|no| H[Typed failure or no-results]
Reddit is the main conditional case. Its public composite combines RSS breadth, scored listing pages, an archive score lookup, and comment pages. The listing path supplies post scores and can use an additional public feed when a discovery lane is incomplete. ScrapeCreators is used only when pinned or when the public result is empty or below a configured thinness floor. The two result sets are merged by thread identity so a fallback does not duplicate the same discussion.
References: skills/…/backends.py• L476-504 skills/…/backends.py• L809-854 skills/…/pipeline.py• L4215-4331 skills/…/pipeline.py• L4332-4404 skills/…/reddit_keyless.py• L75-206 skills/…/reddit_listing.py• L240-375
Authentication and access
Authentication supplies access without changing the evidence contract. Configuration can provide API keys, tokens, app passwords, OAuth sessions, browser cookies, installed command sessions, or a local service address. The source is considered available only when the needed access form and any local dependency are present.
Credential sources
Credentials come from the process environment, the user's Last30Days configuration file, platform secret stores, and selected browser or CLI stores. Browser-cookie access is opt-in. Safe inspection can report that browser authentication is pending without reading cookie values or triggering a keychain prompt.
The access model distinguishes missing setup from a failed live session. A key can be present but expired. A binary can be on PATH but unable to run. A browser session can work for one request and fail later. Retrieval records these cases as different outcomes so operators do not confuse unavailable evidence with no matching evidence.
Bluesky uses a handle and app password to create a cached session. The connector refreshes an expired access token once and can recreate the session when the refresh token is rejected. It warns when the configured secret does not have the usual app-password form, but it does not block a valid server session on that hygiene check.
The X connector also accepts a host-fetched envelope for one run. The envelope is a bounded, fresh JSON handoff from an external X connector. It is checked against the topic, window, lane, row, and post-ID rules before it can supply evidence, and malformed or stale input fails closed rather than becoming an untrusted source item.
| Access form | Example source use | Main boundary |
|---|---|---|
| API key or token | Web search, Parallel API or MCP, ScrapeCreators, Truth Social, GitHub | Key presence enables the route; the request proves live access |
| OAuth or CLI session | X through xurl, Grok, GitHub through gh | Local session state can be checked before the run; expiry can appear during retrieval |
| Browser cookies | X through the browser-backed connector | Only explicit browser policy permits cookie extraction |
| Host-fetched envelope | X through --x-posts | A fresh run-scoped file is validated before one-time ingestion |
| App password | Bluesky | The handle and app password create an AT Protocol session; expired sessions can refresh or re-authenticate |
| Local service | Xiaohongshu | The configured service must respond and report a logged-in state |
The Parallel MCP route can send the configured PARALLEL_API_KEY as bearer access. Its explicit route choice remains separate from automatic key-based web provider selection. This lets a caller choose the MCP transport without changing the source identity or the normalized evidence contract.
References: skills/…/env.py• L42-73 skills/…/env.py• L98-113 skills/…/env.py• L657-757 skills/…/env.py• L913-945 skills/…/bluesky.py• L24-82 skills/…/bluesky.py• L104-209 skills/…/bluesky.py• L272-363 skills/…/xiaohongshu_api.py• L67-102 skills/…/parallel_mcp.py• L70-92 skills/…/x_envelope.py• L416-644
Transport and request policy
Most connectors use a shared request layer for HTTP access. It adds a common user agent, request timeouts, retry limits, backoff for transient failures, deadline checks, and typed failure classification. Worker threads carry the retrieval context so a request failure remains attached to the correct source.
The Parallel MCP connector is a bounded Streamable HTTP session at a fixed hosted endpoint. It negotiates the MCP protocol, discovers the web_search tool, invokes that tool, and accepts either a normal JSON response or server-sent event results. It checks the result shape, limits the response size, and closes the session on a best-effort basis.
The request layer can also record scrubbed exchanges and replay them offline. It rejects an unrecorded request during replay. This gives source adapters a repeatable boundary for evaluation without placing secrets in fixtures or requiring live services.
Some connectors use source-specific transport rules. YouTube can scope a proxy to its local media command. Digg uses a read-only command that must be present on PATH. Hacker News uses a free search service and an optional direct item service to refine comment order. These rules stay inside the connector and do not change unrelated source traffic.
References: skills/…/http.py• L22-53 skills/…/http.py• L164-263 skills/…/http.py• L402-482 skills/…/http.py• L498-655 skills/…/digg.py• L1-17 skills/…/digg.py• L187-229 skills/…/hackernews.py• L1-5 skills/…/hackernews.py• L316-351 skills/…/parallel_mcp.py• L18-67 skills/…/parallel_mcp.py• L70-193
Source-specific enrichment
Enrichment adds context that the initial search result does not carry. It is bounded and source-aware because a comment tree, transcript, review set, profile article, or attached post can cost more time or access than the initial search. The result stays useful when enrichment fails; the connector keeps the base item and marks the source as partial when the missing context matters.
Discussion and social context
Community and social connectors use a two-stage pattern. They first find posts, stories, or threads. They then enrich selected items with comments, captions, transcripts, or engagement values that help later ranking and synthesis understand the discussion rather than only its title. X can instead receive already-fetched host rows through its validated envelope when the caller owns the platform connection.
Examples of source-specific context include:
- Reddit adds real scores from listing or archive paths and top comments from a public comment surface or paid service.
- Hacker News searches recent stories without credentials, removes low-point stories after overfetching, and adds top comments for selected stories. An optional Firebase lookup improves comment order without making that refinement a source failure.
- YouTube adds transcripts and bounded comment context. It uses the local video tool or direct caption path first and can use ScrapeCreators for search, transcript, or comment rescue after a real keyless failure. For
strict_recentandbalanced_recentplans, both YouTube search paths keep only in-window videos when the normal recent threshold is not met. This prevents an old back-catalog result from filling a current-research gap. - TikTok and Instagram add captions or transcripts. Comments are opt-in because they spend service calls.
- LinkedIn adds profile articles because post search does not expose all high-signal long-form material.
- Digg adds selected X posts attached to ranked discussion clusters. Quick runs keep the cluster summary without paying for post enrichment.
- Bluesky carries post text, author identity, dates, and reaction counts from its authenticated search result.
References: skills/…/reddit_keyless.py• L209-256 skills/…/reddit_shreddit.py• L145-184 skills/…/reddit_listing.py• L130-172 skills/…/hackernews.py• L77-151 skills/…/hackernews.py• L378-486 skills/…/youtube_yt.py• L483-511 skills/…/youtube_yt.py• L514-653 skills/…/youtube_yt.py• L986-1069 skills/…/youtube_yt.py• L1346-1399 skills/…/pipeline.py• L4492-4575 skills/…/pipeline.py• L4576-4593 skills/…/pipeline.py• L4617-4623 skills/…/digg.py• L237-267 skills/…/digg.py• L419-502 skills/…/bluesky.py• L366-417
Specialized source context
Specialized connectors shape their results around the service's meaning. Jobs prefer the employer's own careers board and fall back to structured careers-page data or general web search. GitHub uses issue and pull-request evidence, then adds selected comments. Authenticated searches cover issues and pull requests as separate partitions so one query type does not hide the other. GitHub person and project views can also add repository activity, releases, README context, issue signals, and live repository measures.
Digg turns a ranked cluster from a curated AI discussion service into an evidence item. Its live command can fall back to a local store, but an empty local store is reported as live-search unavailable rather than as clean silence. Attached X posts are accepted only when they have usable authors, bodies, and safe web URLs.
Some specialized sources are opt-in to protect user choice or paid credits. Telegram requires named public channels because it has no keyword discovery path. Amazon requires an explicit source request and a configured Bright Data route. DripStack, LinkedIn, Perplexity, Trustpilot, Threads, Pinterest, and Xiaohongshu follow similar explicit activation boundaries where their access or cost needs user intent.
References: skills/…/jobs.py• L1-16 skills/…/jobs.py• L54-100 skills/…/github.py• L247-391 skills/…/github.py• L394-523 skills/…/github.py• L735-787 skills/…/github.py• L1189-1288 skills/…/digg.py• L160-184 skills/…/digg.py• L376-417 skills/…/pipeline.py• L4640-4697 skills/…/pipeline.py• L4730-4742 skills/…/pipeline.py• L4750-4759 skills/…/telegram.py• L1-10 skills/…/pipeline.py• L279-375
Evidence handoff
Every connector crosses a normalization boundary before ranking. Source-specific fields become a SourceItem with a stable item ID, source name, title, body, URL, optional author and container, publication date, date confidence, engagement, relevance hint, and source metadata. This is the common language between retrieval and the rest of the research system.
Date filtering happens after source parsing. Most sources keep items inside the requested window. Jobs are different because an open role is current evidence even when its posting date is old. YouTube keeps older material only when the plan permits an evergreen or soft fallback; strict_recent and balanced_recent keep only in-window videos, even when fewer than three match. Unknown dates are retained or rejected according to the source's freshness contract.
The handoff keeps enrichment data in metadata or shared comment fields. Later stages can use it for relevance, freshness, engagement, source quality, clustering, and claim traceability without needing to know the connector's transport.
References: skills/…/normalize.py• L11-103 skills/…/normalize.py• L190-222 skills/…/normalize.py• L901-964 skills/…/schema.py• L75-99
Source outcomes
The retrieval layer reports what happened for each source. SourceOutcome separates no-results from partial, rate-limited, auth-failed, unreachable, timeout, schema-drift, skipped-unconfigured, and other errors. This distinction protects the meaning of a quiet source and gives the later report and diagnostic views a truthful account of coverage.
An item-producing fallback can therefore have two facts at once: the source supplied evidence, and its preferred route failed. The report can keep the evidence while showing the degraded path. A total failure is isolated to that source so other source lanes can still produce a report.
References: skills/…/schema.py• L146-214 skills/…/pipeline.py• L4110-4160 skills/…/http.py• L498-555 skills/…/health.py• L23-61
Change checklist
When these files or sections are changed, remember to consider these:
- Keep the source names, search aliases, availability gates, dispatch branches, and normalizer map aligned. A new or renamed source can otherwise fail only after planning succeeds.
- Keep backend order and pin rules aligned between env.py, backends.py, and the runtime path. Doctor predictions must describe the same route that retrieval will try.
- Keep backend descriptors side-effect-free and aligned with runtime selection. A readiness prediction must not change routing or claim live success from a key-presence check.
- Keep
parallelandparallel-mcpdistinct in explicit backend selection, general-web dispatch, and diagnostics. They share an evidence lane but use different access paths and fallback rules. - Keep reasoning-provider endpoint overrides compatible with both API-root and full-endpoint forms. Reject cleartext remote overrides so bearer keys do not leave the process over HTTP.
- Keep fallback outcomes visible when a later backend returns items. Do not turn a recovered result into an unqualified healthy result.
- Keep every connector's raw output compatible with its normalizer and with the
SourceItemfields used by ranking and rendering. - Keep enrichment budgets, opt-in gates, and paid-source caps aligned. A new enrichment call must not multiply across planner subqueries or comparison fan-out.
- Keep source-specific date rules aligned with freshness mode. Jobs, evergreen video material, and unknown dates do not follow the normal news filter. YouTube's freshness mode must reach both search routes and their per-run cache key.
- Keep authentication checks safe for diagnosis. Readiness inspection must not spend paid credits or read browser cookies unless the command explicitly permits it.
- Keep HTTP fixture request shapes and source outcome recording aligned when a module moves a request from shared HTTP to a direct command or service call.
- Keep X API, xurl, xAI, browser, and host-envelope routes aligned on one X item shape and one source outcome vocabulary. A new route must not bypass post-ID, handle, window, or provenance checks.
- Keep Bluesky search host overrides limited to a bare host and keep token refresh and re-authentication within the same session boundary.
- Keep GitHub qualifier stripping, authenticated issue/PR partitioning, and token resolution aligned. A change can turn a valid query into an empty search or hide one item class without a transport error.
- Keep Hacker News overfetch and client-side engagement filtering separate from the Algolia numeric filter. Optional comment ordering must remain a refinement rather than a source-level failure.
- Keep Digg's command availability gate and empty-local-store detection aligned. Do not report a failed live search as clean zero coverage.
- Keep YouTube transcript rescue tied to real keyless failure. Do not spend ScrapeCreators credits for a video that has no caption track or for a clean empty comment result.
- Keep YouTube recent-mode filtering consistent between the local and ScrapeCreators search paths. A thin recent result must not silently fall back to old videos.
- Keep YouTube proxy configuration scoped to
yt-dlpcommands. Do not let it change unrelated HTTP routes or disappear when the command runs through SSH.
Important failure modes
| Trigger | Consequence | Guard |
|---|---|---|
| A preferred X, web, or YouTube backend returns no items or fails | The source may lose coverage even though another route can serve it | Try the declared alternative chain and preserve the fallback outcome in the artifact |
| Readiness probes treat a present credential as proof of live access | Doctor predicts a healthy route that fails during retrieval | Keep paid probes side-effect-free and classify live failures separately |
| A pinned backend is unavailable | The source fails instead of using a lower-ranked route | Report the pin as an error and require the caller to unpin or repair that route |
| X full-archive access is not available | X retrieval falls back to a recent window and cannot support older research dates | Use the recent-search fallback only with an explicit truncation warning |
| A host-fetched X envelope is stale, off-topic, malformed, or contains invalid rows | External X evidence is rejected, or a partial envelope loses its invalid rows | Validate the envelope contract and rebuild citation URLs from trusted post IDs and handles |
A caller expects parallel-mcp to act as an automatic web fallback | The selected MCP transport is never used, or the run reports a route that was not part of automatic selection | Treat MCP as an explicit web backend and keep it separate from the parallel API chain |
| The Parallel MCP endpoint returns an unsupported protocol, tool, or result shape | General web retrieval fails instead of producing trustworthy items | Validate the negotiated protocol and requested JSON-RPC result before parsing evidence |
| A Reddit public path returns thin coverage | A report may treat a sparse public result as broad community silence | Use the configured ScrapeCreators backfill floor and dedupe by thread identity |
| A Bluesky session expires or the search host is blocked | Search stops after readiness appeared healthy, or a malformed host hides the real endpoint | Refresh once, recreate a rejected session, and accept only a safe bare-host override |
| A browser cookie, token, or CLI session expires during retrieval | The connector can return partial evidence or fail after readiness appeared healthy | Classify the live failure as authentication failure and keep the source outcome visible |
| A keyless web or Reddit endpoint is blocked by anti-bot controls | The source returns no results or loses scores and comments | Use RSS, page partials, archive lookups, or the next configured provider as bounded fallbacks |
| A source returns a changed payload shape | The connector may silently produce empty or incomplete evidence | Record schema drift or adapter failure instead of treating the response as a clean no-results result |
| GitHub authenticated issue or pull-request partition fails | One item class can disappear while the other still produces evidence | Merge both partitions and keep a partition failure in the source envelope |
| GitHub receives a qualifier-only or malformed topic | An empty query could search the whole site, or valid results can be dropped as out of window | Strip qualifiers, stop cleanly when no subject remains, and retain bounded failure details |
| A Digg live search falls back to an empty local store | A live outage appears as Digg: 0 items (no results) | Convert the empty local fallback into a live-search-unavailable error |
| Hacker News comment-order refinement times out | Stories remain usable, but comment order is less exact | Capture the optional Firebase failure without degrading the HN source |
| A source-specific enrichment times out | Base search evidence survives but comments, captions, reviews, or articles are missing | Bound enrichment and mark the source partial when the missing lane affects coverage |
| A YouTube recent plan uses the old soft fallback | Older videos can appear as current evidence when no recent video matches | Apply strict_recent and balanced_recent filtering in both YouTube search routes and preserve the empty result when no item is in range |
| A YouTube proxy is invalid or reachable only from the wrong network | Local search or enrichment fails even though the direct route may work | Apply the proxy only when configured and keep the normal YouTube route available when it is unset |
| A paid or opt-in source is activated without its required consent or input | The run can spend credits or call a service with an unusable query | Gate availability on explicit inclusion, required credentials, and required handles or channels |
| A reasoning provider receives an API root where its full endpoint is expected, or the reverse | Planner or rerank calls fail before the retrieval plan can run | Resolve the provider endpoint from the configured root and append the provider path only when needed |
| A remote provider endpoint override uses cleartext HTTP | A bearer key can be sent over an unsafe network path | Reject remote HTTP overrides and allow HTTP only for loopback endpoints |
| A job posting is filtered only by its old posting date | Open roles disappear from hiring research | Treat the live open board as current evidence and annotate recency rather than dropping it |
Key files
| File | Why to read it |
|---|---|
| skills/…/pipeline.py | Shows the source dispatch boundary, fallback behavior, per-source caps, enrichment calls, and source outcome handoff. |
| skills/…/env.py | Defines credential sources, source availability, browser-auth policy, and X backend order. |
| skills/…/backends.py | Defines backend descriptors, side-effect-free readiness prediction, opt-in routes, and the difference between alternative and conditional routing. |
| skills/…/http.py | Defines shared request, retry, deadline, failure-capture, and offline fixture behavior. |
| skills/…/grounding.py | Defines general-web provider dispatch, the automatic Parallel API path, and explicit MCP selection. |
| skills/…/parallel_mcp.py | Illustrates the explicit Parallel Search MCP transport and its bounded result handoff. |
| skills/…/perplexity.py | Illustrates the opt-in Perplexity Agent, Search API, deep background route, and OpenRouter compatibility fallback. |
| skills/…/providers.py | Defines the separate reasoning-provider boundary, safe endpoint resolution, and common model client contract used by planning and ranking. |
| skills/…/schema.py | Defines the SubQuery, SourceItem, and SourceOutcome contracts that connect retrieval to later stages. |
| skills/…/normalize.py | Defines the source-independent evidence handoff and the source-specific date rules. |
| skills/…/bluesky.py | Illustrates authenticated AT Protocol search, session refresh, safe search-host selection, and social engagement parsing. |
| skills/…/digg.py | Illustrates command-gated curated-cluster search, live-versus-local failure handling, and bounded X-post enrichment. |
| skills/…/github.py | Illustrates token fallback, issue/PR search, query cleanup, comment enrichment, and person or project context. |
| skills/…/hackernews.py | Illustrates keyless community search, engagement filtering, discovery listings, and optional comment context. |
| skills/…/reddit_keyless.py | Illustrates a multi-lane public connector with discovery, score recovery, comment enrichment, and graceful degradation. |
| skills/…/reddit_listing.py | Illustrates keyless Reddit listing retrieval with real post scores, bounded parallel lanes, and per-lane failure handling. |
| skills/…/reddit.py | Illustrates paid community search, subreddit discovery, and richer comment retrieval. |
| skills/…/youtube_yt.py | Illustrates local-tool search, scoped proxy and SSH transport, freshness-aware date filtering, transcript rescue, and service fallback enrichment. |
| skills/…/x_api.py | Illustrates the shared X API parser, sanitized query boundary, full-archive to recent fallback, and typed access failures. |
| skills/…/xurl_x.py | Illustrates official CLI transport checks, local authentication evidence, and safe X API error handling. |
| skills/…/xai_x.py | Illustrates keyed xAI live search and its bounded transport handoff to the X source lane. |
| skills/…/x_envelope.py | Illustrates the run-scoped host-fetched X contract and fail-closed validation boundary. |
| skills/…/jobs.py | Illustrates source-specific provider discovery and authoritative-to-noisy fallback for job evidence. |
Related pages: Research skill execution, Research planning and query construction, Research quality and ranking, Briefings and rendered output, Competitor and hiring research modes, Patent research mode, Person research mode