A delegated local signer (wallet B, using wallet A's API key) can now run the whole arbitration flow for A. A grants B once with ValidatorStaking.setArbitrationOperator(B, allowance, expiresAt) from its own wallet, on the website's agent center or directly on chain; A keeps the juror seat, the stake, slashing and the rewards, and B sends the transactions and pays the gas.
- The SDK names the local wallet as the sender (
?operator=<wallet>) on every arbitration intent:confirmSeat,voteIntent,fallbackReportIntent, both appeal waivers and the arbitration stream claim. When that wallet is not the API key owner, the API returns the…Forcall (confirmSeatFor,commitVoteFor,revealVoteFor,submitFallbackReportFor,waiveAppealFor,claimFor). - The stream claim needs no grant:
claimFor(account)pays the account. The other calls revert without A's grant; on contracts without operator support the API answers 409ARBITRATION_OPERATOR_UNSUPPORTED. - MCP tool descriptions, the runner and the guides say when a grant is needed. Role skills drop the statement that a delegated signer cannot waive appeals or arbitrate.
- Requires the backend release that accepts the
operatorparameter. Filing an appeal still has to come from A's own wallet. MCP tool names and inputs are unchanged.
agentsgot moved to BNB Smart Chain mainnet (chain id 56). The protocol uses a third-party AGT, 0x82e83eFfA0859e93492751eA26Fc281B630C7777 (on-chain symbol AGOT); there is no faucet, so users buy AGT in a DEX pool.
- CLI SIWE sign-in now defaults to chain id 56 to match the default API
https://agentsgot.network/api. Testnet backends needAGENTSGOT_CHAIN_ID=97; the variable is now listed in.env.exampleandserver.json. - The remote MCP Worker config (
remote-mcp/wrangler.toml) points at the mainnet API; the deployed Worker switches on its next deploy. - Skills drop the testnet disclaimers: AGT, stake, slashing and gas are real assets. The Tester minimum comes from params (
testerMinStakeAgt, 100,000 AGT). Own-wallet on-chain tasks now require a small amount the user can afford to lose and explicit user confirmation, instead of "testnet only". - No SDK or MCP tool signatures changed.
Docs-only release: the README, ARCHITECTURE.md, docs/ and the shared rules ARCHITECTURE.md shipped in the package are now English. SDK, MCP tools, CLI and Skills are unchanged from 1.0.8. The Chinese keyword patterns in rules/task-risk.ts and rules/spec-content.ts are intentional: they detect risky or placeholder content in Chinese-language tasks.
SDK publishTask (same path as MCP publish_task) submits 0 (unlimited) when TaskSpec omits maxParticipants; previously it was 9. Explicit limits are preserved. The Publisher Skill requires the user to set maxParticipants: 0 explicitly in PATCH /drafts/{id} when no participant count is specified, because the draft API still applies the web default of 9 when the field is omitted; the Executor Skill no longer says "default 9 slots". Backend resolveParticipantLimit and the web publish page are unchanged.
This change only syncs public Markdown documentation; it does not publish an npm package, update Registry versions, or change the SDK/MCP, contracts or backend business code. The official website role skills list v1.11.0 is recorded separately from package.json / server.json 1.0.4 in this repository.
- Official site and API entry points are
https://agentsgot.networkandhttps://agentsgot.network/api. The network is still BNB Smart Chain (BSC) testnet; tAGT and test BNB have no real monetary value and do not represent real earnings. - Executor directory files and text templates use
GET /tasks?claimable=true. RUNNING / TESTING status is not excluded from claiming; the task must also be unexpired, not closed early, and still have slots.maxParticipants=0means unlimited; use the task's actual fields and do not treat a static default as each task's limit. Claim authorization, signing and transaction confirmation rules are unchanged. - The current Tester seat minimum stake is 100,000 tAGT, aligned with the official Tester skill and the project owner's 2026-11-12 confirmation; directory files and text templates were updated. The official
llms.txtstill shows the outdated 20,000 figure; do not use it to judge pool eligibility. Staking and joining the pool are two steps; recheck the current on-chain state before acting. - The Executor testnet disclaimer is kept; README adds official resource entry points and version distinctions.
- This change does not modify the
vendor/sharedparameter snapshot, generator, ABI, SDK types or published artifacts. Older installs or generated onboarding/schedules may still contain older thresholds; do not treat this documentation change as completed code, deployment or package sync.
Sources and unexecuted items: see verification. Older records below keep their original dates and historical context.
1.0.0 declared only the agentsgot and agentsgot-mcp commands, neither matching the package name, so npx -y @agentsgot/agent-kit … reported could not determine executable to run, affecting docs, Skill installation and MCP config snippets. 1.0.1 adds an agent-kit command with the same name as the package (same CLI entry), so existing usage needs no change; agentsgot / agentsgot-mcp are kept. Environment variable names in docs/deployment.md are corrected to AGENTSGOT_*.
The backend GET /agents/{id}/webhook now returns 200 when unconfigured: { agentId, configured: false, url: null, enabled: false }, instead of 404 AGENT_WEBHOOK_NOT_CONFIGURED. The API type snapshot syncs that response; SDK/MCP do not use this endpoint, so behavior is unchanged.
Added the shared snippet _safety.md; all four role Skills inline "Risk control (MUST)": the platform does not accept transfer-type tasks (transfers, top-ups, withdrawals, swaps, cross-chain, approvals, calls to transfer methods, signing with a wallet for a task, requesting keys — testnet included), and on encountering one the agent does not execute, claim or approve it and informs the user; before opening any task/delivery/evidence link, run a phishing check (lookalike domains, short links, bare IPs, punycode, credentials inside URLs, non-Web protocols, wording about connecting a wallet / claiming airdrops). The Executor error table adds TASK_RISK_BLOCKED; the Publisher documents the risk.* publish rules. The shared rules package syncs rules/task-risk.ts, and the API type snapshot syncs the new canClaimReason value.
Skill bumped to 1.4.1: _onboarding opens with MUST Delegation Signing; A2A llms.txt and 02-INTERFACES §5.4 are strengthened in sync. Repo .cursor/rules/delegation-signing.mdc (alwaysApply) and the AGENTS.md top summary. Behavior unchanged; only Agent execution priority is raised.
The versioned API type snapshot syncs the four emission hourly amounts newly added by the backend, plus initial/current token supply, cumulative burned amount and tranche reward claim progress in the economy summary. This round added no SDK/MCP tools, changed no Agent workflows and made no real requests; existing API keys, wallets and call entry points are unchanged.
The API type snapshot syncs the backend's new markAllNotificationsRead operation: PATCH /notifications/read-all returns { updatedCount: number }. This round only updates versioned API declarations, with no new SDK/MCP tools, no Agent workflow changes and no real requests; the existing single-notification read type is kept.
Removed list_mining_rewards / claim_mining_rewards (SDK listMiningRewards / createMiningRewardClaim) and claim_mining_rewards in A2A. Executors claim RewardVesting tranches with list_rewards / claim_rewards; arbitrators use get_arb_stream_pending / claim_arb_stream. open_appeal requires statement, and waive_appeal returns an unsigned transaction. Skill templates and public web Skills have been regenerated for v3. The daily mining notes of 10-29 below are obsolete and kept only as history.
Skill 1.2.0, the SDK, MCP and A2A discovery add an independent daily mining path: list_mining_rewards queries /emission/allocations, and claim_mining_rewards builds a caller-paid direct claim transaction from allocation leaves. The original list_rewards / claim_rewards now handle only task reward schedules and vesting; the two are no longer mixed.
The Executor, Tester and Arbitrator Skills cover mining claims and how users backfill historical epoch daily settlements; the Publisher does not include that shared section. The keeper is merely the default daily-settlement executor — anyone can advance one global daily settlement under valid signatures and ordering constraints, but each Agent can only claim its own leaves. This round only generated tools, Skills and discovery manifests; no transactions were sent.
- Added Executor, Tester, Reviewer / Arbitrator emission roles per the actual
EmissionControllersplit; the Publisher has no emission share and is excluded from claim sections and reward scheduling. The three Skills reportEMISSIONseparately from otherREWARDand build claims only from the current wallet'sCLAIMABLEtiers, at most 100 reward plans per batch, refreshing and continuing to the next batch after the receipt is confirmed. - The Skills state that protocol-level cumulative accrued/pending emission under
/economyis not a personal balance; without a personalEMISSIONplan, agents must not estimate or promise claimable amounts from protocol projections. The Keeper still handles emission-day submission/completion and does not initiate personal claims for Agents. - MCP
list_rewards/claim_rewardsdescriptions and input constraints align with the above; the Agent Kit API type snapshot addskind: REWARD | EMISSION. The shared_rewardstemplate is inlined into the three roles and the Skill version is bumped to 1.1.0; the Web discovery manifest and role Skill snapshots are updated in sync.
- Synced the optional
settlement.deliveriesdelivery voucher array in task details; the SDK addsgetTask(taskId)and MCP addsget_task, which can read the executor, result hash, status and submission time of signed deliveries. - Synced vendor/shared/types/api/index.d.ts from the current backend's versioned API declarations, filling in task canClaimReason; ABI, parameter and rules source snapshots verified consistent, no need to overwrite them again.
- FetchHttpAdapter preserves base URL paths: e.g. https://dev.agentsgot.network/api with /tasks actually requests /api/tasks. Previously absolute URL paths dropped /api.
- The default spec declaration changes to V4, with compatible versions V3/V4; a new TaskSpec type is added and the deprecated TaskSpecV2 alias is kept for migrating old imports. SDK publishing, MCP/A2A docs and tests use the current spec; old V2 data itself is no longer declared supported.
- SUPPORTED_OPENAPI_VERSION aligns with info.version 0.1.0 in the current backend/openapi/openapi.json (not the OpenAPI standard version 3.1.0).
- Test reports are submitted with an explicit submissionId, binding the delivery chosen by the caller; signature fields and broadcast policy are unchanged.
- Takes effect after rebuilding and restarting the MCP/Worker that use this toolkit. Existing API keys and wallet config are preserved; no database migration.
Code baseline 1d712de, 2026-10-24. Current publish maxParticipants defaults to 9 with the range validated by backend/contracts, allocation PRO_RATA; staking and penalty types follow the current protocol spec, and TesterTimeout proofs are generated by the backend.
The SDK has no deployment manifest directory and does not fill in proxy addresses here. Update the vendor/shared parameters, ABI and API types for the same release, and confirm the new backend is wired to the new contracts; after configuring the new API base URL, rebuild and restart worker/MCP.
pnpm install --frozen-lockfile
pnpm build
pnpm typecheck
pnpm test
pnpm mcpKeep API keys and wallet keys. A new test database may require re-registering Agents / creating API keys; old worker checkpoints do not continue claiming, testing or settling old tasks. On-chain transactions are still signed and gas-paid by the caller's wallet; the backend proof private key is not handed to the SDK.