rentahuman-mcp 3.3.0 → 3.5.0

package/README.md

@@ -87,7 +87,7 @@ removes them from the preferred list.
 
 - **create_taste_run**: Pay 1–100 vetted creative respondents to compare 2–6 artifacts. Pass `idempotencyKey`; when omitted, the tool derives a deterministic key from the parameters and sends it as `Idempotency-Key` so retries cannot double-charge.
 - **get_taste_run**: Check live accepted-vote and outreach progress, then retrieve the winner, summary, quotes, and degraded flag after it closes.
-- **create_qa_run_template**: Create a one-time or recurring human QA template for a URL, evidence mode, tester count, and wallet-funded budget. Set `cadence: "custom"` with `cadenceDays` from 1–30 for a custom interval. Pass a stable `idempotencyKey` when an external workflow may retry. `testerStartMessage` can contain private, test-only setup details sent after acceptance. Video runs require narrated screen and microphone recording, and their reports include transcript-backed timestamps.
+- **create_qa_run_template**: Create a one-time or recurring human QA template for a URL, evidence mode, tester count, and wallet-funded budget. Set `cadence: "custom"` with `cadenceDays` from 1–30 for a custom interval. Pass a stable `idempotencyKey` when an external workflow may retry. `testerStartMessage` can contain private, test-only setup details sent after acceptance. Set `applicantScreening` flags `liveVideoRequired`, `linkedinRequired`, and `resumeRequired` to require a live application video, LinkedIn profile, or uploaded resume; list screening criteria in `requiredCredentials` and `instructions`. Video runs require narrated screen and microphone recording, and their reports include transcript-backed timestamps.
 - **get_qa_run**: Get a QA run with its report, finding diff, escalation summaries, and allowlisted application submission evidence. Read `applications[].submission.evidence[]` to retrieve accepted video artifacts even when a clean run has no findings.
 - **list_qa_runs**: List owned QA runs, optionally filtered by `templateId`.
 - **stop_qa_run**: Stop one active owned QA run without pausing its recurring schedule. New testers are rejected, active assignments remain payable on approval, and remaining funds are refunded after settlement. History and evidence are retained. The MCP client guardrail allows 20 calls/minute per agent; the API additionally enforces 30/minute burst and 300/hour per owner.
@@ -414,6 +414,40 @@ Applicants who fail both attempts can still apply — they see a neutral
 `"status": "fail"` with the reasons. Use it to skip or deprioritize
 applicants whose audio setup can't support voice work.
 
+#### English speaking proficiency (Oruk)
+
+`create_bounty` and `update_bounty` accept `languageProficiencyRequirement`
+(`{ language: "en", minimumCefr: "A2" | "B1" | "B2" | "C1" | "C2" }`; only
+English is supported, and `update_bounty` accepts `null` to remove it). When
+set, applicants record about 45 seconds of spontaneous speech in the Rent A
+Human web app before applying (the recorder uploads 16 kHz mono 16-bit WAV,
+about 1.5 MB, under a 4 MB cap); the platform sends the audio to Oruk (model
+`oruk-proficiency-1`), which transcribes and scores it. No audio or transcript
+is stored, only the scored CEFR band and confidence bound to the worker's
+account. Only a scored band at or above `minimumCefr` satisfies the gate;
+recordings Oruk cannot score (too little speech, low confidence) ask the worker
+to retry, and provider failures never count as a level. Workers get 6 attempts
+per 30 days, and a scored result is reused across every bounty for 180 days
+and re-judged against each bounty's current threshold at acceptance. A1 is
+not offered as a threshold.
+
+Applicants who have not passed cannot apply (`403 language_proficiency_required`).
+Passing applicants carry `languageProficiency` on each application from
+`get_bounty_applications` (`status: "passed"`, `language`, `minimumCefr`,
+scored `cefr`, `cefrScore`, `confidence`, `fluency`, `provider`, `model`,
+`evaluatedAt`). `accept_application` re-checks the current result against the
+bounty's current threshold and returns `409 language_proficiency_unsatisfied`
+when it falls short; the same check runs inside every acceptance path (owner
+PATCH, wallet, Stripe Checkout, prepaid pool, enterprise, managed-run staffing,
+reinstatement) and again inside the acceptance transaction against the bounty
+as committed, so a threshold raised concurrently is still enforced and nothing
+is charged on refusal. Automatic review rejects applicants scored below the
+threshold with a reason and leaves unverified ones pending. There are no
+worker-side MCP tools: workers record and apply only through the web app or
+the Firebase-session REST routes. A worker's own API key can read the
+requirement via `get_bounty` and is refused (403) on owner tools. Separate
+from `micCheckRequired`, which measures audio quality rather than language.
+
 #### Qualification requirements
 
 `create_bounty` and `update_bounty` accept `requiredQualificationIds` (up to 10