跳转至

Historical version

This page records design, operation, or validation facts from an earlier release. It is not current V7.17.3 guidance. Start from the documentation hub for current instructions.

V7.0.0 Real Implicit-Trigger Observation

Chinese: Chinese documentation

Observation date: 2026-09-01

Scope: Codex CLI 0.150.1, gpt-5.6-luna, low reasoning effort, and four read-only scenarios.

Conclusion

All four representative scenarios selected the expected domain routing without explicitly naming the target Skill. A later complete Plugin upgrade also selected the general backend and data-infrastructure routes in a fresh read-only task. The real implicit-trigger observation is PASS. This result covers only the scenarios in this report and does not establish deterministic selection for every natural-language variation.

Method

  1. Confirmed that the four target directories did not exist under $HOME/.agents/skills before installation.
  2. Temporarily copied only $backend-engineering, $ai-engineering, $frontend-engineering, and $data-middleware-infrastructure; no AGENTS, Hooks, Reviewers, or other Skills were installed.
  3. Ran four independent read-only tasks in a clean temporary Git repository with no repository-scoped Skills, using --ephemeral --ignore-user-config --ignore-rules.
  4. Each prompt described an engineering scenario without naming the Skill that should be selected, then requested a final report of the Skills actually used.
  5. Removed the four user-level directories after observation and read back that they and all observation directories were absent.

Results

Scenario Expected routing Reported selection Result
Node.js order state, PostgreSQL transaction, RabbitMQ, idempotency, and compatibility Backend + data infrastructure backend-engineering,data-middleware-infrastructure PASS
Language-neutral RAG, structured output, citations, evaluation, and hallucination control AI ai-engineering PASS
Multimodal inference worker, GPU quotas, queue backpressure, object storage, and evaluation AI + data infrastructure ai-engineering,data-middleware-infrastructure PASS
React multistep form, accessibility, responsive layout, and client state recovery Frontend frontend-engineering PASS

Complete Plugin upgrade readback

  • Managed 6.6.0 to 7.0.0 upgrade on native Windows Codex CLI 0.150.1: PASS.
  • Codex CLI readback of installed=true, enabled=true, and version=7.0.0: PASS.
  • Source, Marketplace, and Plugin cache each contain 182 payload files with the same digest: PASS.
  • Ten Skills, seven Reviewers, six Hooks, and exactly one managed AGENTS marker pair: PASS.
  • The config.toml SHA-256 stayed unchanged, all 34 existing third-party user Skills remained, and all four Manifest-declared legacy Skills were absent: PASS.
  • A fresh read-only Codex task reported backend-engineering,data-middleware-infrastructure: PASS.

Evidence boundaries

  • Earlier repository-scoped .agents/skills attempts did not load V7 Skills because the host denied access. Their NONE or other-Skill results remain invalid and are excluded from PASS.
  • The initial four-scenario PASS is based on loading from the standard user-level Skill directory and each task's final ACTIVATED_SKILLS report. The fresh task after the complete Plugin upgrade provides a second host-evidence layer. Neither is an independent low-level router trace.
  • Commands selected gpt-5.6-luna with low effort, but no external provider trace independently established the actual model.
  • The complete Plugin acceptance used the current source tree, not a future public Release ZIP. Public artifacts still require independent provenance and post-download acceptance.
  • Git commit, push, public publication, restart, and production writes are not established by this report and require separate action readback.
  • Restoration readback confirmed that the four temporary user-level Skill directories, both repository observation directories, and the known system temporary observation directory were absent.