Learned from lead-generation e2e runs (medical EMR, AI study app):
- routing.md: requested-N lists must be distinct entities that fit the
asked segment - branches/locations of one parent count once, parent
website domain is an ownership signal, fill shortfalls by widening
discovery instead of padding; place-shaped discovery routes to the
Maps specialist, open-web/enrichment to Search.
- google_maps prompt: exclusion criteria are strict (a satellite of an
excluded chain is still that chain); shared parent domains mean one
organization judged against the parent.
- google_search prompt: pure place-discovery tasks return blocked with
a next_step pointing to the Maps specialist.
Co-authored-by: Cursor <cursoragent@cursor.com>
Crawler engine: escalate thin JS-shell pages past static fetch, repair
currency-lossy extractions, emit categorized link records with anchor
provenance, and decode percent-encoded mailto:/tel: contacts; site crawls
reuse the connector ladder via Scrapling's spider engine with URL pattern
filters. Agent layer: read_run gains char_offset paging, search_run gains
match excerpts, new export_run turns stored runs into CSV workspace docs;
reddit search fair-shares the item budget across queries and dedupes
cross-query hits. Subagent prompts and routing teach crawl-after-search,
full-run coverage before summarizing, and executing own-tool next steps
instead of returning partial.
- Added `Run` model to track scraper invocations, including metadata such as status, input, and output.
- Implemented `ToolOutputSpill` model to store context-editing spills separately from user-facing logs.
- Updated middleware to handle spill placeholders and integrate with the new models.
- Enhanced REST API to record runs and expose run history through new endpoints.
- Adjusted tests to validate the new run logging functionality and ensure proper integration with existing capabilities.
MCP consolidation:
- Route all MCP-capable connectors (Slack, Jira, Linear, ClickUp, Airtable,
Notion, Confluence, interim Gmail/Calendar, custom MCP) through a single
`mcp_discovery` subagent. Drive/OneDrive/Dropbox stay native to enrich the KB.
- Deprecate Discord/Teams/Luma: no viable official MCP server.
Google-only web search:
- Remove the main-agent `web_search` tool and the SearXNG platform service;
all public web search now flows through the `google_search` subagent via task().
- Deprecate the Tavily/SearXNG/Linkup/Baidu search connectors (HTTP 410 on
create, "Deprecated" badge); guide heavy users to the custom MCP connector.
- Remove web search from anonymous chat (pure Q&A).
- Tear SearXNG out of docker compose + install scripts; drop tavily-python
and linkup-sdk deps and their config/env vars.
Fix:
- metrics._package_version() now swallows any metadata lookup failure. A
malformed editable-install distribution with no `Version` field raised
KeyError deep in importlib.metadata, and since it runs on every
record_subagent_invoke_duration call it was crashing every task()
delegation. Verified end-to-end against live GPT-5.4.
Co-authored-by: Cursor <cursoragent@cursor.com>
- Added a new `reddit` subagent to scrape structured data from Reddit posts, comments, and users.
- Introduced `reddit.scrape` capability for fetching data using URLs and search queries.
- Implemented tools for scraping and parsing Reddit data, including handling pagination and rate limits.
- Created input/output models for the Reddit scraper to define request and response structures.
- Added documentation for the new Reddit scraping functionality and its usage.
- Integrated the Reddit subagent into the existing multi-agent chat framework.
Resolve conflicts against the new native google-maps actor + repo-wide
ruff-format pass:
- Keep legacy webcrawler KB indexer + its test deleted (modify/delete).
- test_validators: keep WEBCRAWLER case removed (validator gone).
- test_fetch_resilience: keep platforms.youtube import path (our reorg).
- Relocate google_maps actor + tests scrapers/ -> platforms/ to match the
reorg convention (youtube already there); rewrite imports + fixture paths.
- Add missing __init__.py across the capabilities/ test subtree so duplicate
test basenames get unique module paths under importlib mode.
Note: google_maps fixture-backed tests error on ci_mvp too (fixtures/*.json
never committed upstream) - pre-existing, out of scope here.
web.crawl scrapes a single URL (maxCrawlDepth=0) or spiders a whole site,
backed by the proprietary site_crawler engine. Rewires the scraping subagent
tools and capability tests onto the new verb.
Wire captchatools as the StealthyFetcher-tier page_action to detect, harvest
(egressing from the crawl's own proxy IP), inject, and submit reCAPTCHA v2/v3
and hCaptcha tokens. Opt-in and off by default (zero attempts, zero cost).
Licensing split:
- Apache-2 app/utils/captcha/ holds the generic, vendor-agnostic config
(CaptchaConfig + captcha_enabled() = flag AND key present).
- Proprietary app/proprietary/web_crawler/captcha.py holds the bypass logic
(detect/harvest/inject) plus a process-wide solver latch that halts solving
on unrecoverable errors (no balance / bad key).
Crawler: CrawlOutcome gains captcha_attempts/captcha_solved, surfaced via a
per-call captcha_state dict threaded crawl_url -> _crawl_with_stealthy(_sync)
and stamped onto every stealth terminal outcome. The stealth tier captures the
proxy once and reuses it for both the fetch and the solver (IP-coherence).
Billing: WebCrawlCreditService gains captcha_billing_enabled,
captcha_solves_to_micros, charge_captcha, and a generic check_balance, sharing
a single _apply_debit path. The indexer accumulates attempts (even on failed
crawls), runs a combined crawl+captcha pre-flight, and posts a per-attempt
owner charge as usage_type="web_crawl_captcha". The captcha worst-case is only
reserved when solving is actually enabled, so a solving-off deployment is never
blocked for captcha that can never run. Both chat scrape tools fold attempts
into the current turn before the success/fail branch.
Fully config-driven prices; no migration. New unit tests cover the config,
factory (detection/latch/timeout/cap), credit service, indexer wiring, and the
chat fold.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add a config-driven crawl meter (1 USD / 1000 successes by default, retunable
via WEB_CRAWL_MICROS_PER_SUCCESS) mirroring EtlCreditService:
- WebCrawlCreditService (gate -> pre-check -> post-charge) on the unified wallet
- webcrawler indexer: pre-flight credit block + post-charge of the workspace
owner per success, with a web_crawl TokenUsage audit row
- chat scrape_webpage (main + research): fold each successful scrape into the
active turn's token accumulator so it settles with the chat turn
- WEB_CRAWL_CREDIT_BILLING_ENABLED / WEB_CRAWL_MICROS_PER_SUCCESS config + env
- unit tests (service) + integration tests (indexer wiring)
- scripts/e2e_phase3_crawl_billing.py: manual functional e2e (3a/3b/3c)
Co-authored-by: Cursor <cursoragent@cursor.com>
- Standardized the web crawler to use Scrapling exclusively, removing Firecrawl entirely.
- Updated the crawler's location to `app/proprietary/web_crawler/connector.py` under a non-Apache-2 license boundary.
- Refactored the `WebCrawlerConnector` to eliminate the Firecrawl API key dependency, simplifying the interface for crawling URLs.
- Adjusted related components to accommodate the new structure and ensure successful crawl outcomes are properly handled.
- Updated documentation to reflect these changes and the new implementation status.
- Updated main-agent middleware to clarify that both filesystem reads/writes and knowledge-base retrieval are handled by the `knowledge_base` subagent.
- Introduced `_forward_mention_pins` function to carry `@`-mention pins into subagent state.
- Revised system prompts to reflect the new retrieval method and ensure proper citation handling.
- Removed the `search_knowledge_base` tool and its related tests, consolidating functionality under the `task` tool.
- Enhanced documentation to guide usage of the new retrieval approach and citation practices.
The pull-based KB design (on-demand search_knowledge_base tool + pre-injected
workspace tree) fully replaced the old eager retrieval path. Remove its last
remnants:
- Delete KnowledgePriorityMiddleware (knowledge_search.py) and its tests.
- Drop the kb_priority state field + reducer default; trim
KbContextProjectionMiddleware to project only workspace_tree_text.
- Remove the now-dead feature flags enable_kb_priority_preinjection and
enable_kb_planner_runnable across backend (flags, route schema, tests,
env examples) and frontend (settings toggle, zod schema).
- Scrub <priority_documents> and stale KnowledgePriorityMiddleware references
from prompts, docstrings, and the ADR.
No functional change: nothing wrote kb_priority and neither flag gated live
behavior after the cutover. Full backend suite green (pre-existing unrelated
failures aside).
The model frequently writes citations glued to the preceding word
(docs[17]); the (?<!\w) lookbehind (added to dodge arr[1] array indexing)
silently skipped these, leaving raw [n] that fails to render and reads
like array access. Drop the lookbehind so glued citations resolve; genuine
code/array syntax stays protected by the existing code-region carve-out and
unresolved ordinals still drop harmlessly.
Rewrite the main-agent citation contract to a single [n] channel and sync
the orphaned system_prompt_composer surface to match; drop stale
[citation:chunk_id] / <chunk_index> references from dynamic_context and
provider hints. Reuse the shared hybrid search in the deliverables report
(citations omitted for now) and delete the orphaned report KB helper.
Remove the dead eager KnowledgePriorityMiddleware wiring (knowledge_priority
+ stack) and its legacy browse test. Update ADR 0001 to reflect the cutover.
web_search now registers each result as a WEB_RESULT (locator {url}) and
renders a <web_results> block of <document view="excerpt"> [n] passages,
returning Command(update={messages, citation_registry}) like
search_knowledge_base. Collapse the duplicate research-subagent web_search
into the shared tool and teach the prompts to cite web hits with [n].
The main agent's search_knowledge_base tool runs the hybrid spine, renders
a <retrieved_context> of numbered [n] passages, and persists the registry.
KB subagent prompts teach citing [n] from <document view="full"> reads
(evidence.chunk_ids -> evidence.citations). Delete the now-unused
search->read highlighting hand-off: the kb_matched_chunk_ids state field,
its reducer default, the tool's _matched_chunk_ids writer, and the dead
KnowledgePriorityMiddleware writes.
Add the checkpointed CitationRegistry (load/merge helpers + state field)
and a lightweight CitationStateMiddleware so subagents can register into
the same conversation registry. Resolve [n] -> [citation:<payload>] at
stream finalize from the registry, polymorphically by source type.
Add a shared document_render package that renders sources as
<document view="excerpt|full"> blocks with server-assigned [n] passage
labels (KB locator {document_id, chunk_id}, web locator {url}). Wire the
KB read backend (kb_postgres) and read_file to the new renderer and drop
the legacy per-document XML renderer (document_xml, retrieved_context) and
the old chunk_index / matched="true" / <chunk id> read format.
