Merge pull request #166 from omkhar/agent/skill-improvement-plan

docs: add evidence-gated skill improvement plan

Author: omkhar

docs/authoritative-sources.md

@@ -2,6 +2,56 @@
 
 Latest research date: 2026-07-11.
 
+## 2026-07-11 — skill-improvement-plan startup/trust recheck
+
+Scope: the planning-only startup, precedence, trigger, and fail-closed support decisions in
+`docs/proposals/skill-improvement-plan.md`. Official documentation was re-fetched for every
+first-party supported agent during the plan's release review. This pass adopts no runtime or
+generated-envelope feature; it establishes only what current vendor documentation does and does
+not prove before Tranche 1 implementation.
+
+### Supported-agent documentation
+
+- **OpenAI Codex** — <https://learn.chatgpt.com/docs/customization/overview#skills> and
+  <https://learn.chatgpt.com/docs/customization/overview#agents-guidance>. Codex documents global
+  skills under `$HOME/.agents/skills`, repository skills under `.agents/skills`, implicit skill
+  selection, progressive disclosure, and closest-file precedence for `AGENTS.md`. It does not make
+  a loaded skill a signed pre-load trust anchor or document that a global skill neutralizes a
+  same-name repository skill before target read. Decision: keep the existing package and defer a
+  hardened external-startup claim to E1 loader/precedence evidence; unproved startup remains
+  no-target-read compatibility.
+- **Claude Code** — <https://code.claude.com/docs/en/slash-commands> and
+  <https://code.claude.com/docs/en/memory>. Claude documents personal and project skill locations,
+  automatic discovery, same-name priority, invocation controls, and on-demand loading. Those
+  controls do not verify package bytes or establish an external launcher before project discovery.
+  Decision: retain the generated package and read-only tool preapproval; defer hardened startup,
+  hooks, agents, and broader tool overrides pending E1 evidence.
+- **Gemini CLI** — <https://geminicli.com/docs/cli/using-agent-skills/> and
+  <https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/>.
+  Gemini documents user and workspace skill tiers and gives workspace skills higher same-name
+  precedence; Google's transition notice retains Agent Skills in Antigravity CLI while the
+  consumer-hosted Gemini tier has ended. Decision: a user-level install alone cannot be treated as
+  neutralizing a workspace package; keep Gemini as compatibility coverage and require a
+  no-target-read cap unless an external launcher proves otherwise.
+- **GitHub Copilot** — <https://docs.github.com/en/copilot/concepts/agents/about-agent-skills> and
+  <https://docs.github.com/en/copilot/reference/custom-instructions-support>. Copilot documents
+  project skills in `.github/skills`, `.claude/skills`, and `.agents/skills`, plus personal skills
+  in `~/.copilot/skills` or `~/.agents/skills`; it does not document signed pre-load verification or
+  project-skill neutralization as a security boundary. Decision: keep the existing shared packages
+  and minimal instructions; defer hardened-startup support and do not add a duplicate package.
+- **Google Antigravity** — <https://antigravity.google/docs/skills> and
+  <https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/>.
+  Antigravity documents global skills under `~/.gemini/config/skills`, workspace skills under
+  `.agents/skills`, backward support for `.agent/skills`, and progressive disclosure. The page does
+  not establish signed pre-load verification or a uniform neutralization/precedence mechanism.
+  Decision: retain canonical and legacy-compatibility packages; defer hardened-startup support to
+  E1 and otherwise cap the surface before target read.
+
+Net decision: current docs support the proposal's per-surface evidence matrix and fail-closed caps,
+not a universal trusted-global-install claim. Adopt no new Codex, Claude Code, Gemini CLI, GitHub
+Copilot, or Google Antigravity envelope feature in this planning release. Tranche 1 must perform
+isolated loader and precedence tests before changing a support tier.
+
 ## 2026-07-11 — v0.29.0 release-readiness and agent-support refresh
 
 Scope: release-readiness documentation only. The portable workflow, artifact contract, validator,