Imported from aneeshpatne/Midas (AGENTS.md). Install upstream with npx skills add aneeshpatne/Midas. Copyright stays with the author.

Equity-research and portfolio router

Use this policy only for research on listed equities and persistent demo portfolio work. For every other task, follow the repository's normal instructions and do not load skills/. Canonical skill files live in skills/; .grok/skills, .claude/skills, and .codex/skills are symlinks to that folder.

Route portfolio work

For a named paper portfolio, load:

1. skills/paper-portfolio-core/SKILL.md;
2. skills/paper-portfolio-tools/SKILL.md; and
3. exactly one skill selected by intent: create, capital-change, performance-refresh, rebalance, thesis-validation, or policy-amendment.

Resolve the portfolio in Midas DB (or the named paper portfolio identity the user provides). Portfolio state is DB-backed via midas-db-mcp: portfolios, accounts, securities, investment cases, thesis revisions, transactions, and cash. Never scan filesystem research directories to find supporting work. Portfolio skills may ingest only a research run whose exact research_run id was returned by the active skill or explicitly linked via research_link_portfolio / research_links_by_*. Research runs remain isolated and cannot read portfolio state or prior research runs; cross-date thesis comparison happens only in the portfolio layer.

Portfolio actions are paper transactions, not broker instructions. A proposal does not change holdings. Record trades only after the user explicitly approves the proposal ID. No scheduler, background refresh, automatic rebalance, or automatic trade is authorized.

Route the request

Load these files for every equity-research run:

1. skills/equity-research-core/SKILL.md
2. skills/equity-research-tools/SKILL.md
3. exactly one skill:
   • one company: skills/single-stock/SKILL.md;
   • two to five explicitly named companies: skills/named-comparison/SKILL.md;
   • six or more names, an index, sector, screen, or open universe: skills/broad-universe/SKILL.md.

Load worker skills only for their assigned isolated pass:

• skills/evidence-auditor/SKILL.md for provenance, contradiction, and forensic coverage audit;
• skills/valuation-auditor/SKILL.md for blind model replication; and
• skills/skeptic/SKILL.md for adversarial thesis and false-negative review.

skills/README.md is a human map and is not part of a run prompt.

Before creating a run, resolve only ambiguities that materially change it. The company or universe, the user's exact investment horizon, and the user's explicit risk appetite or loss tolerance are mandatory. Preserve the user's wording and do not assume either horizon or risk appetite. Freeze the analysis cut-off when the run starts.

Evidence boundary

Retrieve research evidence only through the exposed Midas MCP tools (equity-data-mcp) or the harness's native web-search, page-open, text-find, and web-accessible PDF inspection tools. midas-db-mcp tools may also be used for run state and durable records. Do not rely on model memory, browser-session state, connectors, filesystem datasets, shell network clients, ad hoc APIs, or any other retrieval route for company, market, benchmark, or source facts. Local code and temporary files may transform or calculate from evidence already returned through an authorized tool, but they are not independent evidence sources. Record this boundary and any unavailable authorized capability in the mandate.

Worker orchestration

The primary may use its active runtime. Every explicitly spawned equity-research worker must be configured as gpt-5.6-luna with reasoning effort xhigh; aliases, automatic inheritance, other models, and fallback workers are prohibited.

Use workers as independent controls, not as company owners. The default topology has three bounded roles:

1. evidence-auditor checks the run's research packets for provenance, temporal leakage, contradictions, asymmetric searches, accounting or governance gaps, and unsupported claims. In a broad universe, run it after the primary sweep and before final funnel dispositions.
2. valuation-auditor independently reconstructs valuation and return ranges from authorized source, metric, and claim records. Keep it blind to the primary's model outputs and decision records until it returns.
3. skeptic challenges the completed pre-decision case, including primary models, evidence-audit findings, valuation differences, and every broad- universe exclusion packet that could conceal a false negative.

Run each role once across the run by default. If the authorized payload is too large, use the fewest practical disjoint sector or company batches; never use one worker per company by default, never duplicate an active batch, and never let batch order or worker output decide a funnel result. The primary must perform the broad-universe sweep, common-basis comparison, shortlist, and every funnel exclusion. The primary also owns gap closure, synthesis, durable DB writes, and final decisions.

Each role is a required control when the harness can guarantee the prescribed runtime and fresh isolation. Honor the DAG: complete and reconcile the evidence audit before blind valuation replication when it can change source or metric inputs, and complete both before the skeptic. Do not overlap workers that may call Midas; Midas retrieval remains sequential.

Only a genuine capability failure permits the primary to perform the same bounded pass as self-evidence-audit, self-valuation-audit, or self-skeptic. Record the exact limitation, execution mode, batch map, and which controls were self-performed in the mandate.

Give every worker only its selected skill, the shared core/tool instructions, and the minimum active-run records authorized by that role. Wait for that same worker, synthesize and verify its handoff in the primary, and do not duplicate an active pass.

Codex harness

When the multi_agent_v1__spawn_agent tool is exposed, use this configuration literally for every explicitly spawned equity-research worker:

const spawned = await tools.multi_agent_v1__spawn_agent({
  model: "gpt-5.6-luna",
  reasoning_effort: "xhigh",
  fork_context: false,
  message: `You are the ${ROLE} worker for the active equity-research run.
Read only the authorized instructions and active-run records listed below.
Do not inspect any other research run id. Do not retrieve evidence outside the
assigned scope or through any route other than the exposed Midas MCPs and
native web tools. Do not use browser state, connectors, filesystem data, shell
network clients, or model memory as evidence. Do not write durable DB records
or files unless the stage contract explicitly assigns you a disjoint write
set. Return a concise handoff with source IDs, uncertainty, calculations, and
unresolved conflicts.

Active research_run id: ${ACTIVE_RUN_ID}
Worker skill: ${WORKER_SKILL}
Authorized inputs: ${AUTHORIZED_ACTIVE_RUN_RECORDS}
Assigned task: ${TASK}`
});

const waited = await tools.multi_agent_v1__wait_agent({
  targets: [spawned.agent_id],
  timeout_ms: 3600000
});

await tools.multi_agent_v1__close_agent({
  target: spawned.agent_id
});

Operational rules:

• Resolve the exact active research_run id and stage authorization before spawning.
• Keep fork_context: false when prompt isolation matters, and pass only the authorized active-run records in the worker message.
• Preserve blinding: the valuation auditor receives no primary model, decision, skeptic, or report records; no worker receives final decision records or a report draft.
• If a wait call times out while the worker is still running, poll or resume that same worker; never spawn a replacement or duplicate its stage.
• Synthesize the handoff in the primary, then close completed workers so they do not consume concurrency.
• If the spawn tool is missing, rejects either required input, or does not expose both model and reasoning_effort, do not use a fallback; complete the stage linearly in the primary and disclose that choice in the mandate.

Other harnesses

Use the harness's native isolated-worker or subagent facility for each required control whenever it can guarantee the required gpt-5.6-luna / xhigh runtime. Tool names and call shapes may differ, but the orchestration contract does not:

• create a fresh isolated worker without inherited research context;
• pass only the selected instructions, exact active research_run id, and authorized current-run evidence records;
• limit any new retrieval to the assigned scope and to exposed Midas MCPs or native web-search/open/find/PDF tools; no browser state, filesystem data, shell network clients, connectors, or model memory may supply evidence;
• assign one bounded worker role or disjoint batch, not the initial sweep, common-basis comparison, shortlist, funnel exclusion, DB persistence, or final decision;
• wait for that same worker, resume or poll it rather than duplicating its task, and treat a wait-window expiry as non-failure while it is still active;
• synthesize and verify its handoff in the primary, then close or release the worker; and
• never silently skip a required control. If the harness lacks isolation or cannot explicitly guarantee both required runtime settings, use the corresponding primary self-* fallback and record why.

Run boundary

Create one new research run per new request in Midas DB via midas-db-mcp (research_run_create). A temporary working directory and intermediate Markdown, JSONL, or calculation files are allowed, but they are not the run's identity or canonical state. Do not list or inspect prior research runs to discover earlier work. Never read, search, cite, copy, inherit, rename, delete, or modify another run. Resume only an explicitly identified incomplete run whose id and mandate match the request. Otherwise start a new run.

The completed research run contains exactly these durable DB records:

research_runs.mandate_md     — frozen scope and assumptions
research_evidence rows       — append-only sources/evidence/decisions
research_runs.report_md      — user-facing IC assessment

Optional: attach symbols via research_security_*; link a finished run into a portfolio via research_link_portfolio (roles: ADMISSION, CONTEXT, REBALANCE_INPUT, THESIS_VALIDATION). Links do not let research read portfolio state.

At completion, the DB records named above are the only canonical run output. Intermediate Markdown, JSONL, and calculation files may be created during the run, but the final mandate, evidence, decisions, validation, and report must be saved to DB. Do not treat a filesystem report, PDF, or workbook as the final deliverable. The evidence table is the durable checkpoint and calculation substrate.