Prompts

Prompts define a sub-agent's role. They are separate from its configuration and model, and each sub-agent owns its complete prompt.

Prompt targets

Every sub-agent has an immutable-revisioned full prompt. An agent has many Analysis prompt targets and one target for each singleton role.

Target	Used by
Analysis	A user-configured research job identified by its sub-agent key.
Trading	Research synthesis and order management.
Review	Performance and learning review.

Prompt targets belong to one agent and are never shared. The global prompt is an application setting included in sub-agent context.

Editing and revisions

Each role page edits its prompt; an Analysis sub-agent detail edits that specific sub-agent's prompt. New Analysis sub-agents start with the default Analysis prompt, which can be customized during creation. Discuss prompt opens Chat with the current draft as opening context. Saving creates an immutable revision. Chat changes follow the conversation's Strategy prompts permission, which defaults to one-time OpenCode confirmation.

When granted its prompt-update capability, Review may submit an evidence-backed atomic revision batch for Trading and Analysis. It validates target ownership and base revisions when submitted, and cannot revise its own prompt.

Default strategies

The editable Analysis default favors trend continuation and separates directional bias from entry readiness. It asks for an evidence-backed conclusion for every selected instrument, explicit entry confirmation and invalidation, confidence rationale, and at least 1.5:1 estimated reward:risk after costs for actionable setups. Missing evidence and no-trade conclusions are valid outcomes.

New Analysis research expiry is exclusively backend-controlled: two cycles of the immutable producing run's schedule from its evidence boundary, or 30 minutes after publication for unscheduled runs. Analysis cannot override it with metadata.stale_after or metadata.valid_for_seconds; caller values are silently ignored. MCP tool descriptions and Analysis runtime/default instructions no longer ask the analyst to choose a deadline. Older customized prompts can still supply expiry metadata, but cannot change the backend deadline. A deadline in prose does not set memory expiry. Distinguish directional context from entry readiness and specify price/event invalidation; fresh research does not make an invalidated or chased entry actionable.

Runtime instructions prefer compact version-1 handoffs separating quoted source observations from interpretation, with exact run/bar references and invalidation. Analysis performs a pre-publication consistency check of IDs, source rows, candle open/derived close times, visual offsets, and EMA ordering. Partial or unavailable evidence must be labeled rather than reconstructed from later runs. The backend does not verify these research claims. Legacy prose handoffs remain compatible. Corrections link to exact originals using corrects; durable data-quality guidance is published separately from expiring entry details.

The editable Trading default uses these conservative starting limits:

• Planned loss per setup: 0.25% of current account equity.
• Aggregate planned loss across positions and pending entries: 1% of equity.
• Gross notional exposure including pending entries: 1x equity, without netting opposing positions.
• One starter entry using at most half the setup loss budget; additions require fresh continuation evidence and share the original budget. No entry ladders or averaging down.

Trading evaluates conflicting research, requires protective stops, manages existing positions even when research is unavailable, and verifies order outcomes before retrying uncertain submissions. Exits are reduce-only, and protection is reconciled after partial fills or exits. Planned stop risk is an estimate, not a guarantee of maximum realized loss.

Runtime Trading instructions require successful decision IDs for opening orders and preserve reduce-only protection if logging fails. IDs are copied unchanged from returned context. A definitive indexed invalid-link 422 allows one context refetch, reference repair, and write retry; ambiguous failures are reported without blind retries or order replay. Corrections apply only to their structured targets, and superseded entry corrections do not become fresh theses.

The editable Review default evaluates decision quality separately from outcomes for the supplied review window, including partial-day reviews. It assesses Analysis, Trading, execution, and indicator evidence against the strategy and information applicable at the time, and records gaps rather than assuming historical attribution. Relevant indicator inspection is part of every Review run, even without indicator-writing permission. Historical results are read by exact run ID where available within the permitted scope; later results cannot retroactively make a timed-out dependency available to Analysis.

Review independently samples a bounded representative set of publications, including runs that did not self-correct, comparing claims to exact frozen numeric outputs/source rows. It records sample IDs, coverage and limitations, counts self-corrections separately from sampled factual findings, and assesses decision/expiry/link coverage independently of trade outcomes. This is an assessment workflow, not an automatic research-acceptance validator.

Review applies indicator or Analysis/Trading prompt changes only with the corresponding capability and material supporting evidence. It favors small corrections, preserves unrelated operator choices, and distinguishes successful changes from proposals and failed attempts. A successful indicator run is not proof of predictive value. Single-day outcomes and tentative hypotheses do not automatically become strategy changes or durable learnings. Reports cover evidence coverage, outcomes, research and indicator findings, execution and risk, changes, learnings, and unresolved questions; no changes warranted is valid. The default Review prompt also sends one concise notification summarizing the review window, findings, changes, learnings, and outstanding issues after writing the review memory. Manual partial-day reviews are labeled interim.

These are editable agent instructions, not exchange-enforced or server-enforced risk limits. Default changes apply when initializing new prompts; existing saved prompt revisions are not rewritten. Operators can customize the strategy and should keep Analysis and Trading actionability requirements aligned.

Role boundaries

Analysis uses approved data tools, including the closed-candle market-data helper, and publishes research memories. Individual prompts and granted capabilities define the research method and output types; Analysis does not author reusable code. It can read the server-computed indicator definitions and results available to its agent, but cannot create or edit indicators.

Trading reads current Analysis context and owns synthesis and execution. It should write a trading_decision for each evaluated instrument when possible, including no-trade and position-management outcomes, linked to the evidence it considered.

Scheduled sub-agents receive their prompt as input but cannot modify it. See Sub-agents and Chat.

When granted its indicator-write capability, Review can inspect indicators and create an evidence-backed definition or immutable version. Indicator source must not encode trading policy. Trading has no indicator access and continues to consume Analysis research memories.