CHAT_TURN, AGENT_REQUEST/CLAIM/STEP/FINISH

Fork V6 transactions for public chat memory and distributed agentic AI. See Quickstart › Agent Run for full lifecycle.

CHAT_TURN

A public signed message. Used standalone (regular chat) or as part of an agent run (agentRunId field).

Field	Type	Notes
type	"CHAT_TURN"	—
from	string	Sender
conversationId	string	sha256-derived conversation ID (see helpers)
role	string	"user", "assistant", "system", "tool"
content	string	Inline text, ≤ 16384 bytes (else use contentBlobId)
contentBlobId?	string	For large content via SOCIAL_BLOB
parentTurnHash?	string	For threading replies
agentRunId?	string	Link to AGENT_REQUEST if part of agent execution
attachments?	array	[{ type: "image"|"audio"|"file", blobId }]

Fee: implicit CHAT_TURN_FEE = 100 micro deducted from from balance.

Helpers:

• computeConversationId(creator: address, seed?: string) — sha256(creator + "|" + now + "|" + seed)
• computeContentHash(content: string) — sha256(utf8(content))

AGENT_REQUEST

Initiates an agent run. Escrows maxFee from submitter.

Field	Type	Notes
type	"AGENT_REQUEST"	—
from	string	Submitter
agentRunId	string	sha256(from + nonce + prompt[:256]) — helper: computeAgentRunId
conversationId	string	Thread this run belongs to
prompt	string	Initial user prompt
toolsWhitelist	array<string>	Tool names miner may use
maxFee	string (bigint)	Escrow in micro-OMBRA, typically MICRO_OMBRA = 1_000_000 (1 OMBRA)
maxSteps	number	Safety cap, typically 100
promptBlobId?	string	For large prompts
modelHint?	string	Optional preferred model

Validation:

• agentRunId == sha256(from + nonce + prompt[:256])
• state.balance(from) >= maxFee
• maxSteps in [1, 200]

Effect:

• state.balance(from) -= maxFee (escrow)
• state.agentRuns[agentRunId] = { status: "pending", submitter: from, ... }

AGENT_CLAIM

A miner claims the right to execute the run. First valid claim per agentRunId wins; rest are rejected.

Field	Type	Notes
type	"AGENT_CLAIM"	—
from	string	Miner wallet
agentRunId	string	—
minerAddress	string	Same as from
modelInfo	string	E.g., "llama-server:Qwen2.5-7B-Q4_K_M"
hardwareTier	string	"laptop", "desktop", "workstation", "server", "cluster"

Validation:

• state.agentRuns[agentRunId].status == "pending"
• Miner advertises all toolsWhitelist tools (off-chain check, advisory)

Effect:

• state.agentRuns[agentRunId].status = "claimed"
• state.agentRuns[agentRunId].minerAddress = from

AGENT_STEP

Per LLM iteration. Only the claimed miner can submit.

Field	Type	Notes
type	"AGENT_STEP"	—
from	string	Must equal state.agentRuns[agentRunId].minerAddress
agentRunId	string	—
stepIndex	number	0-based
outputTokens	number	Tokens generated in this step (for billing)
outputHash	string	sha256 of step output
thoughtBlobId?	string	Optional chain-of-thought blob
toolCalls?	array	[{ name, argsHash, resultHash, durationMs, ok }]
outputBlobId?	string	If output > 16K

Validation:

• from == state.agentRuns[agentRunId].minerAddress
• stepIndex < state.agentRuns[agentRunId].maxSteps
• stepIndex monotonically increasing

Effect:

• Append to state.agentSteps[agentRunId]
• state.agentRuns[agentRunId].status = "running"

AGENT_FINISH

Final settlement. Only claimed miner.

Field	Type	Notes
type	"AGENT_FINISH"	—
from	string	Claimed miner
agentRunId	string	—
totalTokensOut	number	Sum across all steps
status	string	"completed", "failed", "timeout", "insufficient_funds", "cancelled"
minerReward	string (bigint)	totalTokensOut × AGENT_REWARD_PER_TOKEN
userRefund	string (bigint)	maxFee − minerReward − networkFee
networkFee	string (bigint)	≈ steps × AGENT_NETWORK_FEE_PER_STEP
finalAnswerInline?	string	Final answer ≤ 16K
finalAnswerBlobId?	string	For larger answers

Validation:

• from == state.agentRuns[agentRunId].minerAddress
• minerReward + userRefund + networkFee == state.agentRuns[agentRunId].maxFee

Effect:

• state.balance(from) += minerReward
• state.balance(submitter) += userRefund
• state.totalBurned += networkFee
• state.agentRuns[agentRunId].status = status

Billing summary

totalTokensOut × AGENT_REWARD_PER_TOKEN (1)  → miner
steps          × AGENT_NETWORK_FEE_PER_STEP (100)  → burn
maxFee - minerReward - networkFee  → refund to submitter

Example: 1 OMBRA escrow, 5_000 output tokens across 8 steps:

• minerReward = 5_000 × 1 = 5_000 micro
• networkFee = 8 × 100 = 800 micro
• userRefund = 1_000_000 − 5_000 − 800 = 994_200 micro

Builders

buildChatTurnTx(from, { conversationId, role, content, ... }, nonce, privKey)
buildAgentRequestTx(from, { agentRunId, conversationId, prompt, toolsWhitelist, maxFee, maxSteps, ... }, nonce, privKey)
buildAgentClaimTx(from, { agentRunId, minerAddress, modelInfo, hardwareTier }, nonce, privKey)
buildAgentStepTx(from, { agentRunId, stepIndex, outputTokens, outputHash, ... }, nonce, privKey)
buildAgentFinishTx(from, { agentRunId, totalTokensOut, status, minerReward, userRefund, networkFee, ... }, nonce, privKey)

computeConversationId(creator, seed?)
computeAgentRunId(from, nonce, prompt)
computeContentHash(content)