Brands with a black logo and black calls to action, with color only in links
or artwork, had their primary color reported as whatever link or artwork color
the model saw, because the merge step rejects near-black on light pages to keep
dark nav bars out. When the chosen primary button is a neutral near-black on a
light page, and the chosen color doesn't also fill a button, the primary color
is now that black and the colored pick moves to accent. Navy buttons, dark
pages, and brands that pair a black button with colored ones are unchanged.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(keyless): point suspicious IPs to Google/GitHub signup
* fix(keyless): say VPN or proxy instead of suspicious
---------
Co-authored-by: tomsideguide <tom@sideguide.dev>
* feat(branding): find more of the page's real call-to-action buttons
The in-page scan missed many pages' main call to action before any model saw
it:
- Sampling took the first 100 button matches and first 100 links in document
order, so menus and footers used up the budget before the hero. It now
considers every button and button-like link and keeps the visible ones
nearest the top of the page.
- Buttons whose fill lives on an inner element or a ::before/::after layer
read as transparent and were dropped. The fill is now taken from there.
- Filled or outlined buttons inside the header nav were discarded as
navigation. They stay buttons; plain menu links still don't count.
- Hidden copies (closed menus, dialogs) are left out, snapshots carry their
page position and visibility, and buttons on the first screen rank higher.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(branding): take the text color from the page's text
The text color was the first dark color in a vote over every sampled color,
weighted toward large backgrounds and button fills. Sampling more buttons let
dark button fills outvote the paragraphs, and on dark pages it often returned
the background. It is now the most common text color of non-button elements
that stands out from the background, with the old pick as a fallback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(branding): tighten visibility and position in the button scan
- An element inside a faded-out ancestor (opacity 0) no longer counts as
visible: opacity doesn't inherit, so ancestors are checked too.
- A ::before/::after layer at opacity 0 (hover-only) is no longer a fill.
- Fixed and sticky elements keep their on-screen position instead of adding
the scroll offset, so a header button isn't pushed below the first screen.
- Hidden snapshots don't vote on the text color.
- The hidden-copy test gives the hidden button a real box, so it exercises
display: none, and covers a faded-out parent.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
ZDR browser/Interact sessions now bill 240/hr code-only (4/min) and
540/hr with prompts (9/min). The surcharge applies to the create-time
estimate, the first-prompt reservation, and settlement.
Claude-Session: https://claude.ai/code/session_013R47uwW8MefsRuLhpzaiq2
Co-authored-by: Claude <noreply@anthropic.com>
* fix(branding): keep fallback and duplicate fonts out of Jev's font list
Jev's font answer kept every family a page lists, so system fallbacks
(Arial, Segoe UI, Roboto, Times New Roman, ...) and unplaced extras showed up
next to the real typefaces.
- Fallback fonts are dropped unless the page's typography uses them for text
or nothing else is left.
- Once headings or body are covered, fonts with no known role are dropped.
- A font stack reports its first family, "Variable"/"VF" suffixes are
stripped, and names that differ only in spacing merge into one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(branding): match font roles on the same key families merge on
Merging "CormorantGaramond" into "Cormorant Garamond" kept the spaced name, so
the page's typography no longer matched it and the font lost its heading or
body role. Roles and merging now share one key (lowercase, whitespace removed),
and hyphenated names are no longer merged with spaced ones.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(branding): keep typographyRole's doc comment on typographyRole
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Page text can contain lone UTF-16 surrogates: the branding script truncates
text by code unit, which can split a character outside the BMP (emoji,
mathematical bold letters), and clip() could do the same. TypeSafe rejects
such requests with "400 Request contains invalid Unicode text", so the call
fell back to the LLM.
clip() no longer ends on the first half of a surrogate pair, button labels
and file names go through it, and every string in the Jev state has any
remaining lone surrogate replaced with U+FFFD before sending.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(api): the job log is Pub/Sub only, no PostgreSQL copy
log_job publishes each row and stops there: the PostgreSQL insert, its
retry flag, and the Drizzle definitions of the seventeen job tables go.
E2E helpers and tests that read the job tables read the ClickHouse job
log instead; the feedback tests seed and back-date feedback jobs in
Bigtable rather than searches rows.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* test(api): share job log polling and skip job log snips without ClickHouse
- jobLogJson returns null for empty or malformed JSON text.
- getCostTrackingCalls reads through jobLogRows/jobLogJson and keeps polling
until the scrape row carries parseable cost tracking.
- zdr-helpers polls through a shared waitForJobLogRows instead of its own
query loop; waitForJobLogRow builds on it.
- HAS_JOB_LOG gates every snip that reads the job log (developer, research,
parse, zdr, lockdown, branding and X cost tracking).
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* feat(branding): stop calling Jev for a minute after repeated failures
After 5 Jev calls in a row fail (timeouts or upstream errors), branding skips
Jev for 60 seconds and answers with the LLM straight away instead of waiting
out the timeout on every request. Calls go through again after the cooldown,
and one more failure reopens it. Unusable answers don't count, and the shadow
pauses while it is open.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(branding): keep the Jev breaker armed when a late call succeeds
A call that started before the breaker opened and succeeded afterwards reset
the failure count, so the first failure after the cooldown no longer reopened
it. Successes only reset the count while the breaker is closed. The breaker
tests now fake Date with timers restored in afterEach, so a failing assertion
can't leave the clock frozen for later tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(api): run interact sessions from the scrape's location country
Carry the scrape's location.country and proxy type on BrowserOptions and
map them to Hangar's proxy selection, so an interact session egresses from
the country (and pool) the scrape used.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* feat(api): accept location on browser session create
/v2/browser and /v2/interact take location.country (ISO 3166-1 alpha-2)
and pass it to Hangar's proxy selection. Add the option to the JS, Python,
Go, Java, PHP and Elixir SDKs.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(sdks): cover omitted Elixir browser location; add profile and location to the PHP facade annotation
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(php-sdk): type the facade browser() array parameters
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* Honor Exchange terms overrides without changing universal scrape
* Preserve revocation recovery and acceptance compatibility
* Run Exchange terms compatibility regressions in CI
* feat(api): make checkPromptInjection a top-level scrape option
The json format's checkPromptInjection is lifted into the new top-level
flag at parse time. The guard runs once per scrape on the page markdown,
before any LLM-backed format, and bills +4 regardless of format.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(api): project the prompt injection guard fee for keyless scrapes; cover SmartScrape page scans
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* docs(sdks): say checkPromptInjection works with any format
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* chore(js-sdk): move tests from Jest to Vitest
Replace jest.config.js with a minimal vitest.config.ts, port the unit and e2e suites to vitest imports and vi, and drop jest, ts-jest, @jest/globals, @types/jest and @types/mocha along with the pnpm overrides and build allowances that only Jest's tree needed. Also drop the remaining Jest mentions from repo docs, API test comments and the v1 client's test-env check.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* docs: run vitest through pnpm exec so the harness waits for the API
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(security): drop braces, sprintf-js and csv-parse from api, js-sdk and test-suite
knip 6 and lint-staged 16.4 no longer pull micromatch/braces in apps/api.
jest 30.5 drops micromatch in the JS SDK, and load-nyc-config's js-yaml is
moved to 4.x so argparse@1 and sprintf-js leave the tree. The artillery
load test in apps/test-suite targeted the retired v0 staging host and
nothing ran it, so it is removed with its npm project and CI wiring.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* chore(ui): remove the unmaintained ingestion-ui template
Nothing builds or deploys apps/ui/ingestion-ui; it has had no functional
change since 2024 and only receives dependency patches. Removing it takes
tailwindcss 3 and postcss-selector-parser 6 out of the repo.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* feat(api): point scrapes of search results pages to /search
Scrapes and batch scrapes of a search engine results page
(google.*/search?q=, bing.com/search?q=, DuckDuckGo, Brave, Yahoo,
Baidu, Yandex, Ecosia, Startpage) get a warning that /search returns
ranked results for the query directly. The warning is off by default:
SERP_SCRAPE_WARNING turns it on for every team,
SERP_SCRAPE_WARNING_TEAM_IDS for listed teams, and
SERP_SCRAPE_WARNING_ROLLOUT_PERCENT for a stable share of teams. Every
results-page scrape is logged with whether the warning was shown, so
warned and unwarned teams can be compared.
Hint-enabled /v2/scrape responses for a results page get a next-step
hint with a ready firecrawl_search call, and provider rules can match
the new page_serp_engine and page_serp_query signals.
Also drops export from four agent-hint constants and types that only
their own module uses, which knip reports.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(api): tighten DuckDuckGo and Yahoo results-page matching
DuckDuckGo now matches only its result paths (/, /html, /lite) and Yahoo
only /search, /search/... and /search;... so pages such as
duckduckgo.com/about?q= and /searchfoo are not treated as results pages.
Tests now pin that the requested URL decides when the final URL differs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Agent hint content now comes only from provider-supplied response
guidance rules. Without a provider, responses carry no agent_hints.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
The crawl start URL was scraped first and only checked against robots.txt
afterwards, so a disallowed start URL was fetched and then discarded. Check
it against the crawl's robots.txt before the scrape starts, honoring
ignoreRobotsTxt.
Also record robots.txt blocks the crawl already found but dropped:
- sitemap and index URLs that filterLinks denies for robots.txt
- redirect targets robots.txt disallows, recorded as the target URL, with
nothing crawled from a disallowed start URL target
CrawlDenialError carries the blocked URL as robotsBlockedUrl instead of a
robots flag, and filterLinks returns robotsBlocked.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* feat(api): log every auth denial as one canonical auth/denied line
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(api): redact the admin secret from auth/denied routes
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(security): patch open Dependabot alerts in examples, Rust SDK and Go html-to-md
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* docs(examples): require Python 3.10+ where pins now need it
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* feat(js-sdk): abortable crawl() that cancels its job; cancelCrawl returns false on 409
cancelCrawl now resolves false when the API answers 409 because the crawl
is already completed. Other errors still throw.
crawl() accepts an optional AbortSignal. When it aborts, polling stops,
the SDK sends a best-effort cancel request for the job, and the promise
rejects with signal.reason. A signal that is already aborted rejects
before any request. waitForCrawlCompletion also accepts the signal, but
only stops polling and does not cancel the job.
Bumps the JS SDK to 4.43.0.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5YqtAzmu8fHaWAUDQLZa4
* chore(js-sdk): bump to 4.45.0 for the crawl() abort signal
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(js-sdk): reject an aborted crawl() without waiting for the cancel request
The best-effort cancel now runs in the background, so a slow or hanging
DELETE no longer delays the abort rejection.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* test(js-sdk): read the abort reason after the abort in the hanging-cancel test
The test passed `controller.signal.reason` to `rejects.toBe()` before the
abort happened, so the expected value was `undefined`. Catch the error
first, then compare it with the reason.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5YqtAzmu8fHaWAUDQLZa4
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: mogery <mogery@sideguide.dev>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(python-sdk): cancel the crawl when async crawl() is cancelled; cancel_crawl returns False on 409
cancel_crawl now returns False when the API answers 409, which it does
when the crawl already completed. Before, it raised FirecrawlError. Other
error codes still raise. This applies to the sync and async clients.
The async crawl() starts a crawl and waits for it. If the task that
awaits it is cancelled (asyncio.timeout, asyncio.wait_for, Task.cancel),
crawl() now sends a best-effort cancel for the started job, then
re-raises the CancelledError unchanged. The cancel runs in a shielded
task with a 10 second limit, and its errors are ignored. start_crawl and
wait_crawl do not change, and the SDK timeout parameter still raises
CrawlJobTimeoutError without a cancel.
Bump python-sdk to 4.46.1.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5YqtAzmu8fHaWAUDQLZa4
* fix(python-sdk): address review on crawl cancellation
- README and the async crawl() docstring: the best-effort cancel runs
only after start_crawl returns the job id. Task.cancel() gives
CancelledError, while asyncio.timeout and asyncio.wait_for give
TimeoutError.
- cancel_crawl docstrings: also list the transport error that the HTTP
client can raise (httpx.HTTPError async, requests.RequestException
sync).
- _cancel_abandoned_crawl: add a done callback that reads the cancel
task's outcome. A second cancellation of the caller no longer leaves
a "Task exception was never retrieved" log.
- Tests: time the slow-cancel test (under 1 s), and add a test that
cancels twice and checks that the loop reports no error.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5YqtAzmu8fHaWAUDQLZa4
* chore(python-sdk): bump to 4.49.3
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: mogery <mogery@sideguide.dev>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(api): record robots.txt denials on crawls
The crawl status robots warning and the robotsBlocked list on
/crawl/:id/errors matched one exact message text. Only the flag-gated
scrape robots check used that text, so default crawls recorded nothing.
Discovered links never reached the robots check either, because link
extraction dropped them first.
- Add isRobotsDenialReason() for the crawler robots denial.
- Give CrawlDenialError a robots marker that survives serialization.
The scrape robots check and the start-URL robots check set it.
- Skip robots in link extraction. filterLinks applies robots after it
and reports the denial.
- Reword the warning to point at robotsBlocked on the errors endpoint.
- Add a test-site robots fixture, snips tests and unit tests.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5YqtAzmu8fHaWAUDQLZa4
* fix(api): batch robots.txt block records and count them with SCARD
Record every robots.txt-blocked link from a page with one SADD instead of
one SADD and EXPIRE per link, and read the warning count with SCARD so
status polls do not load the whole set. Reuse the snips crawlErrors helper
in the crawl robots tests.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(api): chunk robots.txt block records under the Redis argument limit
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: mogery <mogery@sideguide.dev>
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(api): report crawl completedAt correctly when no page succeeds
The crawl status took completedAt from the newest entry in
jobs_donez_ordered. That set only holds successful jobs. When no page
succeeded, completedAt fell back to createdAt and duration was 0.
addCrawlJobDone now also stores the finish time of the latest job of any
outcome in crawl:<id>:last_job_done_at. getLastDoneJobTimestamp returns
the later of that value and the newest successful job. The semantics of
jobs_donez_ordered do not change.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5YqtAzmu8fHaWAUDQLZa4
* fix(api): keep the latest crawl job finish time with an atomic max
- Write crawl:<id>:last_job_done_at with a small Lua script. The script
sets the value only if it is greater than the stored value, and keeps
the 24h TTL. Jobs can finish out of order, so a plain SET could replace
a newer time with an older one.
- Refresh the TTL of last_job_done_at in getLastDoneJobTimestamp, next to
the refresh of jobs_donez_ordered.
- In the snips test, replace the comparison with the /errors timestamp
with a wall-clock bound taken after the crawl returns.
- Remove 4 unused exports that knip reports on main.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01H5YqtAzmu8fHaWAUDQLZa4
---------
Co-authored-by: Claude <noreply@anthropic.com>
* chore(api): drop unused exports flagged by knip
MAX_RULE_HINTS, AGENT_HINTS_PROVIDER_CONTRACT_VERSION, AgentHintRuleScalar and
AgentHintRuleOp are only used in their own modules.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(api): record token usage and team on TypeSafe spans
Every typesafe.systemone span now carries the call's input and output tokens
and the team, so LLM spend tracking can price Jev calls and attribute them:
- Safe search spans gain typesafe.usage.input_tokens / output_tokens and
teamId (removeExplicitResults takes the team).
- Branding spans already had input tokens; they gain output tokens and teamId.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(api): name the option that keeps a provider out of a blocklisted URL
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(api): raise the unsupported-option refusal from engine selection too
Policy-forced zero data retention is named in plain words and the message no longer tells the caller to drop it.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* feat(api): report the serving provider as metadata.provider on Exchange scrapes
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(api): fall back to the direct provider for empty waterfalls or invalid step costs
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(sdks): surface success:false scrape errors in Go, Java, .NET, Ruby and PHP
* fix(dotnet-sdk): branch on success flag and keep details; assert scrape request in Java test
---------
Co-authored-by: mogery <mogery@sideguide.dev>
* fix(sdks): match the Alexandria client timeout to the API's 120 s cap
The JS and Python SDKs capped the Alexandria execution timeout at 50 s.
With the 30 s response margin, the client gave up at 80 s. The API now
defaults and caps the timeout at 120 s, so the server continued and
billed a late success that the caller never got.
Default and cap the client timeout at 120 s with named constants, and
keep the 30 s margin (150 s transport timeout). Bump JS to 4.44.1 and
Python to 4.49.2.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PHiJVRAaJ4rUWemp1cQBqq
* fix(sdks): apply the 120 s Alexandria timeout cap to Go and Rust
The Go SDK waited for the caller's timeout plus 5 s, or its 5 min
client default. The Rust SDK waited for the caller's timeout plus 5 s,
or used no timeout. Use the same rule as the JS and Python SDKs: cap
the execution timeout at the API's 120 s default and add a 30 s
margin for response delivery (150 s when no timeout is given).
Bump Go to 1.16.4 and Rust to 2.21.4.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PHiJVRAaJ4rUWemp1cQBqq
---------
Co-authored-by: Claude <noreply@anthropic.com>
* feat(api): evaluate provider-supplied response guidance rules
Compute a fixed set of response signals locally, evaluate provider rules
with a small generic matcher, and fill placeholders in the API. The
built-in hints remain the fallback when no provider rules are available.
Bumps the provider contract to version 2; version 1 hints keep working.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(api): tighten rule placeholder limits and document rule caps
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* chore(api): remove AGENT_HINTS.md
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* fix(api): never let agent hint computation change a response
Any error while computing hints now sends the original body unchanged,
logs a warning and counts an eval_error. Provider rule sets are
validated as a whole: a set with any invalid rule is treated as a failed
lookup and negative-cached.
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
---------
Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
* feat(elixir-sdk): agent thread_id, mode and exchange from the docs spec
Regenerated from firecrawl-docs fire-98/docs-agent-exchange. start_agent
gains thread_id, mode and exchange; also picks up docs main drift
(gov_search, gov_search_post, create_browser_session block_ads).
Bumps 1.11.3 -> 1.12.0.
* fix(elixir-sdk): validate closed objects' keys and send an empty one as {}
Objects with additionalProperties: false (auditMetadata, exchange) now get
nested NimbleOptions keys and a nested key mapping, replacing the
auditMetadata special case. exchange: [] sends {} instead of [].
AgentOptions gains exchange (AgentExchangeOptions with nested approve and
decline), threadId and mode. AgentResponse exposes threadId/threadTurn and
AgentStatusResponse exposes thread fields, the chat message, pendingApproval
and the exchange run summary. Bumps the SDK to 1.16.2.
* feat(ruby-sdk): add agent exchange options, threadId and mode
* fix(ruby-sdk): split terms and calls approval fixtures, release as 1.17.3
* test(ruby-sdk): keep the terms approval unresolved in the fixture
* feat(java): add exchange, threadId and mode to agent options
Adds AgentExchangeOptions (with Approve/Decline) plus threadId and mode to
AgentOptions, threadId/threadTurn to AgentResponse, and threadId, threadTurn,
mode, message, pendingApproval and exchange to AgentStatusResponse. Bumps the
SDK to 1.18.3.
* fix(java): require approvalId on Approve/Decline, null-guard README exchange example
