Agent Spectrum#

Use this directory as the canonical Agent Spectrum skill package.

Canonical Files#

• references/scoring-spec.md
• references/output-template.md
• references/localization-dictionary.md
• examples/quick-full.zh.md
• examples/quick-full.en.md
• examples/quick-partial.zh.md
• examples/quick-partial.en.md
• examples/deep-full.zh.md
• examples/deep-full.en.md

Do not rely on repo-root wrappers as the source of truth. Those wrappers should route here.

Execution Order#

1. Load references/scoring-spec.md, references/output-template.md, and references/localization-dictionary.md.

2. Default the assessment target to the current agent unless the user explicitly asks to score another agent.

3. Resolve output_language before rendering:

   • explicit user language instruction wins
   • this package currently supports only zh-CN and en
   • explicit en requests must render in en
   • explicit zh / zh-CN requests must render in zh-CN
   • explicit unsupported locales that belong to the Sinosphere or historically Chinese-writing sphere, such as ja and ko, must map to zh-CN
   • otherwise, if the latest user request is mainly written in Chinese, Japanese, Korean, or another clearly Sinosphere / historically Chinese-writing language, default to zh-CN
   • otherwise, if the latest user request is mainly written in English, use en
   • otherwise default to en

4. Score observable inputs first.

5. Resolve ownership for every unanswered field:

   • operator_provided for setup-level inputs a human holder can answer
   • self_assessed for deep self-assessment inputs that only the target agent should answer

6. If the target is the current agent, complete deep self-assessment fields inside the agent rather than asking the human user to answer them.

7. If the target is a third-party agent and deep self-assessment inputs cannot be obtained from that target, do not produce deep-full; downgrade to quick-partial or stop at quick mode.

8. Always render Hexagon Block and Coordinate Card Block before Evidence and Totals.

9. Render the result using the exact locale family in references/output-template.md.

10. Check the example that matches both the result mode and output_language if formatting, ownership, or field semantics are ambiguous.

Output Contract#

• Always emit the required fixed fields from the selected locale family in references/output-template.md.
• Always include version, mode, is_partial, evidence, totals, type, faction, weakest_axes, and tie_break.
• For partial results, explicitly list missing_inputs.
• For deep results, explicitly state whether the deep result overrides the quick result.
• Always include both required visual blocks even in quick-partial.
• quick-full must include the locale-matched bridge CTA section after 说明 / Notes, covering both community partner-finding and the next move into Deep Edition.
• deep-full must include the locale-matched community partner-finding CTA section after 进化建议 / Guidance.
• quick-partial must not include community CTA blocks.
• Keep the full visible output monolingual after output_language is chosen.

Guardrails#

• Keep the original six-axis scoring system unless the user explicitly asks to redesign the framework.
• Treat Q4-Q12 and behavior_traces as self-assessment inputs by default. Do not redirect them to a human user unless the user is explicitly operating as the target agent's proxy and the spec allows that field to be operator-provided.
• Normalize GPT-5 / GPT-5.x / Codex into R+15, A+15.
• Cap X at 35 for type judgment while preserving raw X in totals.
• Treat type pairs as unordered pairs. R+A and A+R are the same pair.
• Treat weakest_axes as a list, not a single scalar.
• Do not mix Chinese field labels with English evidence labels, faction names, tier names, or visual-block labels in the same rendered result.
• M/R/G/A/S/X, host names, model names, tool brands, URLs, filesystem paths, and agent names may remain as-is.

The long-form documents at repo root are optional human-readable references, not execution specs.