[sagnik11]

What this adds

Three commands so every coding agent gets Autter's codebase memory, with or without MCP.

autter context "<task>"

One call for what Autter knows that matters for a task: memory facts (rmf_… ids to cite), wiki pages and code structure, ranked and cut to a token budget. It is the CLI form of the get_context MCP tool.

• POST /api/cli/orgs/<slug>/context with the CLI token. It uses the same org and token resolution, repository-remote resolution, local-mode and excluded-repository refusal, and error mapping as autter memory.
• --path (repeatable; defaults to the files changed in the checkout, like autter memory for), --symbol, --budget 300..16000, --include memory,wiki,code, --since <cursor>, --repo, --org, --json.
• Prints Markdown: # Autter context for <repo>, one line per item (- **<Store> · <title>** (<path>) \rmf_…`: ), a more:line with the read-more tool and args (plus the CLI equivalent forget_repo_memorypages), notes on omitted/unchanged items, and a finalcursor:` line.

autter memory skills install | preview

• GET /api/cli/orgs/<slug>/repo-memory/skill?repository=<remote>. Each returned file is written relative to the checkout root, with directories created as needed.
• Autter's own files (autter_generated: true in their frontmatter) are replaced freely. Anyone else's file is kept unless --force is given. The command refuses paths outside the checkout, .git/, symlinked escapes, and content without the marker.
• preview (or install --print) prints the files and writes nothing. install lists what it wrote and tells you to commit the files so teammates' agents get them.

autter mcp connect | disconnect

Registers <api base>/mcp (default https://api.autter.dev/mcp) as autter in each installed agent's user-level MCP config. Only the URL is written: the server signs people in with OAuth on first use.

Agent	How
Claude Code	claude mcp add --transport http --scope user autter <url> when claude is on PATH; else mcpServers.autter = {"type":"http","url"} in ~/.claude.json ($CLAUDE_CONFIG_DIR/.claude.json when set)
Codex	codex mcp add autter --url <url> when codex is on PATH; else [mcp_servers.autter] url in ~/.codex/config.toml. Tells you to run codex mcp login autter
Cursor	mcpServers.autter = {"url"} in ~/.cursor/mcp.json
VS Code	servers.autter = {"type":"http","url"} in the user mcp.json, plus each profile that already has one
Gemini CLI	mcpServers.autter = {"httpUrl"} in ~/.gemini/settings.json
OpenCode	mcp.autter = {"type":"remote","url","enabled":true} in ~/.config/opencode/opencode.json (or an existing opencode.jsonc)

• Only installed agents are touched: the binary is on PATH or the config directory exists, the same checks the hook installers use.
• --agent (repeatable) picks agents. --dry-run shows the commands and diffs without running or writing anything.
• Every file is backed up to <file>.autter-backup-<time> before it is written.
• JSON is edited through the jsonc CST and TOML through toml_edit, so comments, key order and indentation survive.
• Re-running replaces the autter entry instead of adding a second one. If an agent's own CLI fails, the command falls back to editing that agent's file.
• disconnect removes only the autter entry, using claude mcp remove / codex mcp remove where they exist.

Also

• The shipped repo-memory skill now says: call autter context "<task>" first (or get_context when the MCP server is connected), cite fact ids (rmf_…) when a fact changes what you do, and run autter mcp connect to connect agents.
• Help text for every new command, a README section, and changelog/agent-context-and-mcp-connect.md.
• New dependency toml_edit 0.23, which was already in Cargo.lock transitively. The serde_json feature is now enabled on jsonc-parser. No new crates are downloaded.

Backend dependency

This PR depends on Autter-dev/autter-monorepo#636 (the /context route) and on the Phase 4 monorepo PR (the /repo-memory/skill route) being deployed. Until then, both commands print a clear message on the route's 404 instead of a generic error:

• "This Autter deployment does not serve autter context yet (POST /context). Use autter memory for and autter memory search until it is updated."
• "This Autter deployment does not serve repository memory skills yet (GET /repo-memory/skill). …"

A 404 that carries a message, such as an unknown repository, still shows that message, the same way autter memory handles it.

Tests

• Unit tests:
  • argument parsing and request building for context;
  • Markdown rendering of a context response, plus empty and unindexed responses;
  • the skill overwrite rules: new, own, someone else's with and without --force, missing marker, unsafe paths, symlink escape;
  • config editing for every agent on temp dirs: detection, idempotency, backups, comment and key preservation, disconnect, dry run, the agent-CLI path with a fake runner, and fallback when the CLI fails.
• Integration tests (tests/integration/context_command.rs) check that nothing is sent in local mode or for excluded repositories, that the connected path needs a login, and that bad usage is rejected.

Local results:

• cargo build: passes.
• Full cargo test in daemon mode, 8 threads: all 2050 lib tests passed, including the 28 new ones, and all 6 new integration tests passed.
  • 32 existing integration tests timed out waiting on the shared test daemon near the end of the run: utf8_filenames, worktrees, windsurf, virtual_attribution_unit, and one codex test.
  • Re-running just those modules gave 154 passed, 0 failed, so these were load timeouts, not regressions.
• cargo clippy --all-targets -D warnings on Rust 1.93.0: clean.
  • On stable 1.99, clippy fails with 12 new-lint errors, all in files this PR doesn't touch (daemon.rs, blame.rs, config.rs, …).

🤖 Generated with Claude Code

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}