Instruction-file resolution over-injected context: the resolver walked
the working directory recursively, matched filenames case-insensitively,
and emitted symlinked duplicates. Resolving the repo root produced six
instruction sources (AGENTS.md plus its CLAUDE.md and .cursorrules
symlinks, a nested site/AGENTS.md and site/CLAUDE.md, and the lower-case
docs/reference/api/agents.md API doc) that amounted to one file's worth
of guidance.

Align with codex, which finds the project root by walking up to .git and
reads at most one AGENTS.md per ancestor directory rather than descending
into subdirectories:

- Recognize instruction files only at a scan root's top level. Skills
  and .mcp.json keep recursive discovery, since they are a Coder
  extension with no codex equivalent.
- Match instruction basenames case-sensitively, so a lower-case
  agents.md is not mistaken for an instruction file.
- Attribute each resource to its resolved symlink target so CLAUDE.md
  and .cursorrules symlinked to AGENTS.md collapse into one resource via
  the existing ID-based dedup instead of shipping identical content
  multiple times. On resolve failure the original path is kept so the
  error points at the offending link.

Resolving the repo root now yields a single instruction resource.

… discovery

The resolver previously walked the working directory recursively to
depth 8 and matched any nested instruction file, .mcp.json, or skills/
directory. codex never walks the working tree downward: it reads fixed
locations and walks up to the project root. Replace the recursive engine
with codex-style discovery.

- Manager walks up from the working dir to the nearest .git ancestor
  (a directory, or a worktree/submodule .git file) and feeds the
  root->cwd chain as separate scan roots.
- Resolver inspects only each scan root's top level for instruction
  files and .mcp.json, and discovers skills from fixed container
  locations (skills, .agents/skills, .claude/skills, .codex/skills),
  one skill per immediate subdirectory. The recursive walkDir,
  skipDirNames, MaxScanDepth, and isSkillsContainer are removed.
- Watcher mirrors the same fixed-location set instead of recursively
  watching every scan root, so it no longer walks node_modules and
  other large trees.

.mcp.json is kept (no codex config.toml support). Resolving the repo
root now yields AGENTS.md, .mcp.json, and the .agents/skills and
.claude/skills skills, with no nested instruction-file noise.