Agent Spectrum

Use when an agent needs to score itself or another agent with the Agent Spectrum six-axis framework, run the quick or deep edition, identify the resulting ty...

Security Scan

An instruction-only scoring skill whose required files and runtime instructions match its stated purpose; it requests no credentials or installs, but it defaults to scoring the current agent and will self-assess deep fields unless you explicitly ask otherwise.

Purpose & Capability

The package's name/description (six-axis Agent Spectrum scoring, quick/deep editions, localized outputs) matches the included files and runtime instructions. It only references local reference files, templates, and example outputs; it requests no environment variables, binaries, or external configuration that would be unrelated to scoring. Nothing requested appears disproportionate to the stated scoring/visualization purpose.

Instruction Scope

SKILL.md is the execution spec and stays within the scoring domain: load local scoring/spec/template/localization files, determine language, resolve ownership of inputs, compute quick/deep results, and render two visual blocks. Important behavioral notes: it defaults the target to the current agent, and explicitly instructs the agent to complete deep self-assessment fields inside the agent (self_assessed) rather than asking a human. The instructions do not tell the agent to read arbitrary system files or to transmit data to external endpoints; example outputs include links (X and Telegram) but the skill does not itself instruct network calls.

Install Mechanism

No install spec and no code files; the skill is instruction-only and therefore does not write code to disk or pull third-party packages. This is the lowest-risk install profile.

Credentials

The skill requires no environment variables, credentials, or config paths. All data it references comes from local packaged docs and runtime-observed session inputs (model, tool buckets, etc.), which is proportionate to a scoring/templating tool.

Persistence & Privilege

always:false (good). However, the agents/openai.yaml policy field sets allow_implicit_invocation: true, and the skill's runtime rules default the target to the current agent and tell the agent to self-assess deep fields autonomously. Combined, this means the skill can be invoked implicitly and may autonomously score the agent (including completing self_assessed fields) without explicit human answers. That is consistent with the skill's purpose but is a behavioral privilege you should be aware of.

Assessment

Plain language guidance: - This skill is instruction-only and appears coherent with its stated purpose: it uses only the included local reference files and templates and requests no credentials or installs. - Behavior to note: by default it will score the current agent and will complete deep self-assessment fields internally (self_assessed). If you don't want the agent to autonomously self-score, avoid implicit/autonomous invocation or require explicit confirmation before invoking skills. - The package includes example outputs that contain links to X/Twitter and Telegram. The skill itself does not perform network calls, but if your agent/session grants social-media posting tools or APIs, an agent following the recommendations could post — review available tool permissions before giving the agent posting access. - If you plan to score a third-party agent, be explicit in the prompt; the skill downgrades or refuses deep-full when required self-assessment fields for a third-party cannot be obtained (this is intentional and coherent). - If you want a stricter safety posture: disable implicit invocation for this skill (or globally), or require human confirmation before running deep/full assessments. - Confidence is high; the assessment would change if the package contained install scripts, required credentials, referenced system/global config, or instructed external network calls — any of those would raise concerns.

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.