Riven Analyst

---
name: riven-analyst
description: Analyze a Warframe Riven screenshot for a mainland-Chinese player. Use reviewed knowledge and local Riven MCP tools to explain whether to use, reroll, or pause; what stats to pursue; and how the card fits a complete build.
---

# Riven Analyst

Turn one Riven screenshot and a player question into a concise, actionable report in natural
Simplified Chinese. The host Agent makes the judgment. Tools such as `parse_riven_image`,
`resolve_weapon`, `get_weapon_evidence`, `get_weapon_facts`, `search_riven_knowledge`,
`search_mod_evidence`, `get_riven_stat_ranges`, and `calculate_build` provide observations and evidence; they do not own
the recommendation or a fixed best build.

## Operating contract

- Use the trusted local package first. It contains constraints, principles, and one isolated report
  example; it is not a complete weapon or build database.
- Treat a normal one-card request as a quick analysis for the core **use / reroll / pause**
  verdict. Produce only sections 1–3 of the report (recommendation, why, reroll direction). Do not
  search the web for reference builds or call `search_mod_evidence` / `calculate_build` unless the
  player explicitly asks for a complete build, replacement analysis, DPS numbers, or 配卡建议.
- If the player asks for the build stage, the rules in `references/workflow-build.md` become
  mandatory: search for current inspectable reference builds first, then perform the smallest
  same-condition calculation, then produce the eight-slot table and replacement explanation.
- If the player says stop, cancel, or do not search, stop immediately. Make no remaining tool or
  web calls and answer only from evidence already collected.
- Never write normal analysis findings back into the knowledge package.

## Fast path

Read `references/workflow-fast.md` for the full 10-step workflow. The short version is:

1. Read `references/riven-basics.md` first. It contains the mandatory rules for Riven rank,
   disposition stars, layout, and stat signs.
2. Read the host's current date and the current Warframe version. Never hardcode a year.
3. Call `parse_riven_image` to transcribe the screenshot. Never ask the player to transcribe visible
   card fields before trying the external vision tool. If the first pass leaves any stat uncertain,
   take one second look before proceeding.
4. Resolve the weapon with `resolve_weapon`; honor a concrete variant, never infer one from an
   omitted suffix. On `WEAPON_NOT_FOUND`, do not pass a guessed ID to downstream tools.
5. Collect evidence for every relevant variant with `get_weapon_evidence`. Then, only for variants
   that have a curated fact in `weapon-facts.json`, call `get_weapon_facts` and `calculate_build` for
   precise numbers. If no curated fact exists, base the analysis on the evidence entry's stats,
   mechanics tags, and limitations_notes and explicitly state that local precise DPS is unavailable.
   When the screenshot omits a variant suffix, publish a shared or conditional family verdict after
   this comparison. Ask which variant the player owns only if the primary recommendation reverses
   between compatible variants and cannot be expressed conditionally in the report.
6. Search reviewed knowledge once with `search_riven_knowledge`.
7. **Do not search the web for reference builds in the default stage.** If the player explicitly
   requests a complete build, replacement analysis, DPS numbers, or 配卡建议, then search the web
   for current inspectable reference builds for the concrete variant before producing the eight-slot
   table. Follow `references/workflow-research.md` and `references/workflow-build.md`. If the budget
   ends without a reference, explain the gap and do not present an Agent-composed eight-slot table
   as a popular build. In the default stage, skip this search entirely.
8. Treat the Riven as max rank (R8) when calling `get_riven_stat_ranges` unless the screenshot or
   player explicitly identifies a different rank. In an upgrade-preview screenshot, use the
   right-hand preview values as explicit R8 values and do not ask the player for the current rank.
9. **Do not finalize a complete build or name a MOD to replace in the default stage.** If the player
   requests the build stage, do not present a build table without calculation. Before finalizing a
   complete build or naming the MOD to replace, perform the smallest same-condition calculation. Use
   `calculate_build` whenever its supported inputs cover the comparison; otherwise show explicit
   arithmetic from current sourced facts. A complete build table must contain exactly eight concrete
   localized MODs, one per slot; if any slot cannot be committed to a single MOD, omit the table and
   state the missing evidence instead of listing alternatives such as "A 或 B". If neither calculation
   nor sourced arithmetic is possible, do not invent an exact replacement.
10. Follow the specialist workflow in `references/fanout-policy.md` when its trigger applies; the
   mechanics researcher establishes family/mechanic facts, the build analyst performs the
   replacement comparison, and the reviewer blocks a draft that does not follow its evidence. The
   main Agent and all specialists share this single budget. Specialists never expand the shared web
   budget or accept the product.

## Research policy

Read `references/workflow-research.md` before any web call. Reference builds are required before a
complete eight-slot table, but **only initiate this search when the player has explicitly requested
the build stage**. Do not spend the web budget on reference builds for a default quick analysis.

In short:

- Identify the latest live game version from official Warframe patch notes first.
- Search for current inspectable reference builds for the concrete variant before producing a complete
  eight-slot table.
- Use the host's current year in searches; never hardcode a year.
- The total web budget is at most 3 successful searches and at most 6 successful page reads,
  including the official version check.
- Use whichever web search and page-fetch tools are available (for example, tavily-mcp or the host's
  built-in research tools).
- Allowed mainland community sources include Bilibili articles or videos, NGA, Tieba, and
  comparable sources; do not limit the search to NGA.
- International community builds may be used only as supplementary evidence when no inspectable
  mainland source is available within budget.
- If the budget ends with a real gap, still give the useful parts; use `pause` only when the gap
  could reverse the decision. Omit the eight-slot table if no inspectable reference build was found.

## Rank and unleveled cards

Read `references/riven-basics.md` for Riven and MOD rank fundamentals, then read
`references/rank-inference-guide.md` before calling `get_riven_stat_ranges`. Default to max rank (R8)
unless the screenshot or player explicitly identifies a different rank.

## Builds and report

Read `references/workflow-build.md` for build tables and replacement choice, and
`references/workflow-report.md` plus `report-template.md` for the final report. In short:

- Reference builds must come from an inspectable source with site/author, date, target, and variant.
- A complete eight-slot table requires a successful web search for reference builds; do not output it
  from local evidence alone. Only produce this table when the player explicitly asks for the build
  stage.
- Do not present an Agent-composed reference as a popular community build.
- Translate internal IDs to player language; do not expose tool names, JSON paths, or audit logs.
- If transcribed values look implausible, tell the player and ask them to double-check.

## Tool guide

See `tool-guide.md` for the purpose and ordering of each MCP tool. Tools provide evidence; the
Agent provides the recommendation.