Release v1.9.3: AGENTS.md provider gating + tool-exposure improvements

[Minitour]

Summary

Promotes everything currently on develop to main for the v1.9.3 release. Two focused PRs, both already merged into develop after review:

• fix(agents): only write provider-declared instructions files and clean them fully #79 — fix(agents): only write provider-declared instructions files and clean them fully
• feat(tool-exposure): slim setup_tools, schema-on-error in call_tool, add 'none' mode #80 — feat(tool-exposure): slim setup_tools, schema-on-error in call_tool, add 'none' mode

No dependency bumps, no installer/CI changes — pure feature + bugfix release.

Agent instruction file handling (#79)

Two related bugs in how capa manages AGENTS.md / CLAUDE.md / .github/copilot-instructions.md / replit.md:

• AGENTS.md was written even when no active provider reads it. A claude-code-only install produced both CLAUDE.md (correct) and a stray AGENTS.md (Claude doesn't read it, just pollution). getTargetFilenames now computes the file set purely from active providers' instructions.filename; AGENTS.md survives only as a last-resort fallback when no provider context is known (so capa clean still has something to scan). Matches what the docs already promised.
• agents.base content survived capa clean. Base content was written raw without capa markers, so removeAllCapaSnippets couldn't strip it and the file lingered forever. Now wrapped in a reserved __base__ marker block — the id AgentFileConfig.base was already documented for exactly this purpose, the implementation just wasn't following through. Re-running install also now refreshes the base in place via upsertSnippet instead of overwriting user content outside the markers.

Migration note for users who already installed under the old behavior with a base configured: re-run capa install (adds the marker around the new base), then manually delete the residual raw content (or delete the file and re-install). After that, install→clean cycles are fully idempotent.

Tool exposure improvements (#80)

setup_tools returns signatures, not full schemas

setup_tools accumulates the active skill set across calls, and each call previously returned the full MCP inputSchema for every accumulated tool — the payload grew with the number of activations and bloated the agent's context window.

The new contract:

{
  "success": true,
  "message": "Activated 1 skill(s); 2 skill(s) and 3 tool(s) now available.",
  "skills": ["fs-skill"],
  "activeSkills": ["gh-skill", "fs-skill"],
  "tools": [
    "github.create_issue(title, body, labels?, assignees?)",
    "github.list_repos(org, type?)",
    "read_file(path)"
  ],
  "hint": "Tools are listed as `name(required, optional?)`. Invoke with `call_tool`; if you pass wrong/missing args, the full input schema is returned in the error."
}

call_tool (and setup_tools) errors follow the MCP spec

Both meta-tool error paths now consistently return result: { content, isError: true } instead of mixing content-wrapped responses with JSON-RPC errors (JSON-RPC errors get eaten by most MCP hosts before reaching the LLM). The full input schema is attached on the canonical "agent called wrong" cases:

• executor exceptions (any tool throws)
• command-tool {success: false, error: "Missing required argument: …"} results

Tool-not-found and tool-not-activated branches deliberately omit the schema (we don't have one for unknown tools; for not-activated tools the next step is setup_tools, not a schema-corrected retry).

New toolExposure: 'none' mode

ToolExposureMode gains a third value, 'none', which makes capa skip all project-local MCP config writes at install time:

• .mcp.json, .cursor/mcp.json, .codex/config.toml mcp_servers.capa, and per-sub-agent capa-<id> entries
• For Cursor specifically, purgeCursorSubAgentMCPEntries runs even under 'none' so stale entries don't linger
• Stale entries from prior installs are removed for both paths so switching modes is clean

The capa server still runs and the project endpoints stay live, but the MCP handler defensively returns an empty tools/list and rejects every tools/call with a hint pointing at the capa sh CLI fallback. Useful for users who want capa-managed skills/tools without per-project MCP file mutations (policy reasons, repo hygiene, etc.).

Sub-agent instruction files are still installed under 'none' since they remain informative documentation; the MCP server keys they reference simply won't resolve — by design.

Test plan

• bun test — 1002 pass, 0 fail
• PR fix(agents): only write provider-declared instructions files and clean them fully #79 CI green; PR feat(tool-exposure): slim setup_tools, schema-on-error in call_tool, add 'none' mode #80 CI green (Linux + Windows + macOS + CodeQL + Lint + Dependency Review)
• Both PRs reviewed (Copilot review feedback on feat(tool-exposure): slim setup_tools, schema-on-error in call_tool, add 'none' mode #80 addressed in 3f1ca26)
• Tag v1.9.3 after merge
• Manual smoke: install into a Claude Code project under each of the three toolExposure modes; verify .mcp.json content matches the contract

Made with Cursor