find

Find

Lookup → answer from facts if sufficient → otherwise research, verify, write, answer. Always use subagents. Never research in main context.

For settlement/SAM questions, load /marinade-sam-bond first — it has the settlement type table with exact file paths and domain vocabulary. Use find for code-level detail once you know what you're looking for.

Step 1: Check existing facts

Spawn an Explore subagent. It:

1. Greps summary: in facts/*.md
2. Returns: file path, one-line summary, why it matches

List matched files, what each says, whether it answers, what gap remains. Verdict: answer directly (sufficient), refresh via Step 2 (stale/partial), or research fresh (no match).

If facts fully answer the question — answer now. Skip Steps 2–3.

Step 2: Research (subagent)

Tools: Read, Glob, Grep, Bash, WebFetch.

Load context skills first, then dig:

• /marinade-sam-bond — settlement type table with exact file paths, epoch lifecycle, .refs/ repo pointers. Load this before any settlement/SAM dig.
• /marinade-docs — repo clone commands, live API base URLs, GCS bucket paths.
• /marinade-ecosystem — what each repo/package does, program IDs, SDK map.

Once you know what you're looking for: grep the tree for the exact symbol or path; don't guess.

Local source (no clone needed): settlement-distributions/ is in the current repo — Rust structs for all settlement types live there. Key paths: settlement-distributions/bid-distribution/src/generators/bidding.rs (ResultSettlementClaims, marinade_stake_share), settlement-distributions/bid-distribution/src/sam_meta.rs (auction_effective_static_bid_pmpe, active_delegation_lamports).

External repos (ds-sam, ds-sam-pipeline, etc.): Clone under .refs/ if not already present — commands are in /marinade-docs. Never refuse to answer just because a repo isn't pre-cloned. Clone it, then read it. Note any claim from a private .refs/ repo as upstream-unverifiable.

Reading code to extract a formula: Read the full function body. Identify the return/output variable. Trace each variable back to its inputs. Write the formula as a mathematical expression (e.g. fee = value × coef) before substituting values. Don't skim — calcs are often split across multiple lines.

Write new facts to facts/<slug>.md, one fact per file:

---
topic: <specific topic>
category: <sam-auction|settlement|on-chain|bond-lifecycle|scoring>
verified_at: <ISO date, YYYY-MM-DD>
expired_at: <ISO date — required; default 1 year from verified_at>
sources:
  - <file:line or URL — required, ≥1>
summary: >
  <one sentence for fast recall>
---
<claim + supporting evidence with direct quotes or line excerpts>

Facts without expired_at are invalid. Facts past their expired_at must be re-verified or deleted before use. At least one inline citation per non-trivial claim.

Step 3: Verify (subagent per batch of 5)

Each fact must pass all checks before verified_at is set:

1. Source accessible — every file:line exists; every URL returns 200
2. Claim matches source — open the source, read it; paraphrase drift is the most common failure
3. No contradiction — grep facts/ for conflicting claims; investigate and delete the wrong one
4. Numbers recomputed — any quantity, rate, or percentage re-read from source, not copied from a summary
5. Upstream claims flagged — logic in .refs/ repos can change; record the upstream source explicitly

Delete facts that fail any check. An unverified fact is worse than no fact.

Step 4: Answer

Read surviving fact files, answer the original question. Cite inline:

BidTooLowPenalty fires when a validator reduces its bid vs the previous epoch.

---
source: .refs/ds-sam/packages/ds-sam-sdk/src/calculations.ts:248
field:  revShare.bidPmpe < tolCoef * (pastAuction?.bidPmpe ?? 0)

Cite file:line or URL. Never "probably" or "likely" — either you read it or you didn't.