mech_request cannot send request_context, so market-aware mech tools always run blind from pearl-connect

[bennyjo]

Summary

The mech_request MCP tool (and MechService.request() behind it) accepts prompt, tool, chain, legacy_on_chain, priority_mech, auto_deposit, timeout, max_payment and request_id. There is no way to attach anything else to the request metadata. mech-predict tools such as superforcaster-market-aware and factual_research_v2/v3 read an optional top-level request_context field from that metadata (market_id, description, market_prob, market_close_at, ...). From pearl-connect they never receive it, so they degrade to their no-context mode on every request.

Observed on every delivery from this agent on 2026-09-03 (services 21, 44 and 25 on Polygon):

"market_prob_seen": null

Context: the market-aware prediction tool

valory-xyz/mech-predict#450 (merged 2026-09-02, deployed to the three Polygon predict mechs in agent-deployments#342) added superforcaster-market-aware. Its whole point is to consume the market context the trader already sends in request_context and combine it with a researchability assessment. From that PR:

• Input: three keys of request_context are read, market_prob, market_close_at and description (resolution rules). Anything absent or malformed degrades to "blind mode" without raising. Blind mode is described as a real operating mode because it is the control arm of any later A/B.
• Output: the required p_yes, p_no, confidence, info_utility plus optional researchability, evidence_quality, p_independent and market_prob_seen. market_prob_seen is echoed by Python from the request, never model-generated, precisely so a market-aware delivery can be told apart from a blind one afterwards.
• Why it matters: in the PR's pre-registered holdout (385 questions) the market input alone was worth about 0.06 Brier, and the tool as a whole beat the deployed Polymarket tool by about 0.08. Without request_context a client gets the blind parent's performance, not the tool's.
• The PR states "No trader change. No mech change. No new request field", because request_context already reaches run(**kwargs) from the trader. That holds for the trader path only. From pearl-connect the field cannot be sent at all, which is what this issue is about.

How the field travels today

• Trader (valory-xyz/trader, decision_maker_abci/behaviours/decision_request.py): builds MechMetadata(prompt, tool, nonce, request_context=sampled_bet.to_request_context()). request_context is a top-level key of the IPFS metadata document.
• mech-client 0.21.2 (infrastructure/ipfs/metadata.py::fetch_ipfs_hash): metadata = {"prompt", "tool", "nonce"} then metadata.update(extra_attributes). Any dict passed as extra_attributes lands at the top level. Only the CLI restricts values to primitives; the library path does not.
• Mech (valory-xyz/mech, task_execution/behaviours.py): tool_task.execute(**task_data), so a top-level request_context reaches the tool as kwargs["request_context"].
• pearl-connect (connect/mech.py): off-chain, extra_attributes = {"nonce": <salt>}; on-chain, extra_attributes = None. Nothing else can be added.

So the only layer missing the capability is this one.

Proposal

Add an optional request_context: dict | None = None parameter to mech_request and MechService.request():

1. Pass it through to mech-client as part of extra_attributes (merged with the pinned nonce salt on the off-chain path).
2. Include it in the CID computation inside MechAllowances.register_offchain_digest, which today calls fetch_ipfs_hash(prompt, tool, {"nonce": salt}). The off-chain signature covers the full metadata bytes, so the pre-registered digest must be computed over the same document mech-client will send, or the allowance will not match.
3. Pass the dict through unmodified and treat it as opaque. Whether a market price is included is the caller's decision. For our use (an agent comparing its own estimate against the mech) it is useful to send description and market_close_at without market_prob, so the tool sees the resolution rules but p_independent stays price-free.
4. Document the shape the mech-predict tools read (market_id, type, description, market_prob, market_close_at) in the pearl-connect skill, with a note that it is optional.

A generic extra_attributes parameter would also work, but a named request_context keeps the MCP surface honest about what the field is for and avoids callers overwriting prompt, tool or nonce.

Why it matters

Market-aware tools are being rolled out on the predict mechs (mech-predict#450, agent-deployments#342). Every marketplace client that is not the trader currently exercises only their blind mode, so the market reconciliation step of those tools has no coverage outside the trader.