v0.2.0

Added

• Added antigravitycli as a distinct Antigravity CLI integration with prompt
  context, instructions, and MCP support. CLI skills are intentionally deferred
  because Antigravity CLI uses flat .md skill files while this crate's
  SkillSurface models directory-scoped SKILL.md packages.
• Added OpenCode prompt rules and instruction-surface support through
  documented AGENTS.md files.

Fixed

• Antigravity global MCP installs now resolve the documented
  ~/.gemini/antigravity/mcp_config.json leaf symlink when it points to a
  shared config file under ~/.gemini, matching Antigravity's app-created
  layout without weakening the generic global symlink safety policy.
• Codex hook Matcher::Bash now emits Codex's documented Bash matcher.
• Codex streamable HTTP MCP entries now write http_headers and omit the
  unsupported type = "http" field. SSE transport is refused for Codex MCP.
• OpenCode's generated default plugin now uses the current
  tool.execute.before plugin callback signature.

Changed

• Antigravity app/IDE local prompt rules, instructions, and skills now write
  to .agents/* paths. Existing .agent/* installs are still detected for
  status and uninstall, but are not auto-migrated.
• OpenCode HookSpec::rules now injects into OpenCode's documented
  AGENTS.md prompt file instead of only writing the plugin file.

Changed (breaking, instruction surface)

• The instruction surface (InstructionPlacement::InlineBlock and
  InstructionPlacement::ReferencedFile) now writes its managed markdown
  fence with a distinct prefix:
  <!-- BEGIN AGENT-CONFIG-INSTR:<name> --> ... <!-- END AGENT-CONFIG-INSTR:<name> -->.
  Hook rules continue to use <!-- BEGIN AGENT-CONFIG:<tag> -->. This
  eliminates a silent collision where a hook with tag = T and an
  instruction with name = T installed into the same memory file (e.g.
  ~/.claude/CLAUDE.md, AGENTS.md, GEMINI.md) would overwrite each
  other.
• Status detection and uninstall on the instruction surface accept the
  legacy AGENT-CONFIG:<name> prefix as a fallback, so consumers upgrading
  from a pre-rename build will see existing installs detected and removed
  cleanly. Re-installing an instruction whose host carries the legacy
  fence prunes the legacy block before writing the new one. Pruning is
  gated on the instructions ledger already having an entry for the name,
  so a hook block sharing the same identifier is never erased.

Documentation

• README now notes that schema/agents.json is the Linux-canonical view of
  agent path layouts. macOS and Windows views differ for any agent whose
  config dir flows through paths::config_dir() (Cline, Roo). Regenerate
  on Linux for byte-stable output; the companion snapshot test
  (tests/schema_golden.rs) only enforces equality on Linux.
• CLAUDE.md records that the Crush hook integration relies on Crush's
  Go decoder ignoring unknown JSON fields. Crush's HookConfig schema
  declares additionalProperties: false; if upstream switches to strict
  decoding, the inline _agent_config_tag marker will need to migrate to
  a sidecar .agent-config-hooks.json ledger paralleling the existing
  MCP / skills / instructions ledgers. No hook ledger infrastructure
  exists today; the migration is deferred until needed.

Migration notes

• Outstanding installs from a pre-rename build do not require manual
  cleanup. Uninstall detects the legacy AGENT-CONFIG:<name> fence and
  removes it.
• Re-installing the same instruction (matching name) replaces the
  legacy block with the new-prefix block in a single operation.
• Consumers parsing memory files for installed agent-config blocks should
  match either prefix (AGENT-CONFIG: for hooks,
  AGENT-CONFIG-INSTR: for instructions).