#AGENTS.md

This file provides guidance to AI coding agents when working with code in this repository.

#What this is

sigit ("siGit Code") is a single Rust binary: a local-first AI coding agent that runs LLM inference on-device (via the onde crate / GGUF models) or against a hosted/OpenAI-compatible endpoint. It exposes itself two ways from the same binary, chosen at startup by whether stdin is a TTY:

• ACP mode (stdin not a TTY): speaks the Agent Client Protocol over stdio for editor integration (Zed, VS Code ACP Client). Cross-platform.
• Interactive terminal mode (stdin is a TTY): a full-screen ratatui chat UI. Unix-only — it relies on fd redirection to keep logs out of the TUI, so Windows gets ACP mode only.

Before the TTY/ACP split, main also dispatches the account subcommands sigit login, sigit logout, sigit whoami (see src/main.rs main()).

#Working in this repo

IMPORTANT — branch naming: Prefix every working branch with feature/ (new functionality) or fix/ (bug fixes) — never a tool- or agent-name prefix like claude/. Name the branch after the changes it contains, as a short, descriptive, kebab-case slug that is self-explanatory from the name alone (e.g. feature/tool-permission-system, fix/glob-mtime-sort), never after a task, ticket, or session id (not feature/task-q003hm).

IMPORTANT — pull request target: Always open pull requests against the development branch, never main. main is release-only; development is where day-to-day work integrates, and the release merge (see the sigit-code-release skill) is the one thing that ever puts commits on main.

This rule needs help to hold, because the repository's default branch is main: both gh pr create and the GitHub web form pre-select it, so a PR lands on the wrong base unless the base is passed explicitly. Pass it every time, and check that it took:

gh pr create --base development --head <branch>
gh pr view <number> --json baseRefName

Retargeting a PR that was opened against main costs nothing before it merges: gh pr edit <number> --base development. After it merges it costs real work — the change sits on main while development, which the next release is cut from, does not have it, and someone has to carry it back by hand. PR #65 went in that way.

IMPORTANT — branch off development: Start every working branch from the latest origin/development. Exception: when new work depends on a feature branch that has not merged yet (e.g. it builds on tools or APIs that branch introduces), it may be stacked on top of that branch instead. When stacking: merge the base PR into development first, then rebase the stacked branch onto development before opening its pull request, so each PR shows only its own commits.

IMPORTANT — run CI before pushing: Run the full CI gate locally and confirm it is green before pushing a branch or opening a pull request — never push work that fails these:

cargo fmt -- --check                  # formatting
cargo clippy --tests -- -D warnings   # lint (warnings are errors)
cargo test --locked                   # tests

#Agent assets layout

.agents/ is the canonical home for agent assets; every other agent path in the repo is a symlink into it, never a copy:

• .agents/AGENTS.md — this file, the project instructions. The root AGENTS.md and CLAUDE.md are symlinks to it, so agents.md-standard tools, Claude Code, and sigit itself (via src/instructions.rs, which prefers AGENTS.md over CLAUDE.md in the same directory and therefore reads it once) all load the same content. Edit this file; never replace the root symlinks with real files.
• .agents/skills/ — all project skills. .claude/skills is a symlink to it, which is how both Claude Code and sigit itself (via src/skills.rs) discover them. Add new skills under .agents/skills/<name>/SKILL.md, never as real files under .claude/.

#Build / test / lint

cargo build                 # debug build
cargo build --release       # release binary at target/release/sigit
cargo run                   # launches interactive TUI (stdin is a TTY)
cargo test                  # CI runs: cargo test --locked --target <target>
cargo clippy --tests -- -D warnings   # CI gate: clippy is -D warnings on all 4 targets
cargo fmt -- --check        # CI gate (edition 2024)

CI (.github/workflows/ci.yml) runs fmt + clippy + test across four targets: aarch64-apple-darwin, x86_64-apple-darwin, x86_64-unknown-linux-gnu, x86_64-pc-windows-msvc. Clippy is -D warnings, so warnings fail the build.

Run a single test: cargo test <test_name>.

#Critical platform constraint: #[cfg(unix)] dead code

The interactive client is #[cfg(unix)]-only. The InferenceBackend seam (backend.rs) and provider resolution (provider.rs) are consumed by both the interactive client and the ACP server, but several of their items are reached only through the Unix-only interactive paths, so the dead-code lint is suppressed on non-Unix targets only.

Consequence: code can pass clippy on macOS/Linux but fail on the Windows target (or vice versa). When touching backend.rs, provider.rs, or the interactive path, keep the cfg gates intact — don't "fix" an unused-warning by deleting code that's live on Unix.

#Architecture

The agent loop is backend-agnostic. The flow: a turn (messages + tool specs) goes to an InferenceBackend, which returns assistant text and/or tool calls; the loop executes tools and feeds results back. Neither the loop nor ACP/TUI surfaces depend on a concrete backend.

• src/main.rs — entry point, mode dispatch, the full ACP Agent impl (session lifecycle: new/load/fork/prompt/cancel, config options, slash-command advertisement), and the SYSTEM_PROMPT (note: it bakes in smbCloud-specific context the agent should use when the repo is clearly smbCloud, and stay general otherwise).
• src/backend.rs — the InferenceBackend trait and neutral types (ToolSpec, ToolCall, ToolResult, TurnResult). Two impls: LocalBackend (on-device via onde::ChatEngine) and OpenAiBackend (any OpenAI-compatible HTTP endpoint).
• src/provider.rs — decides which backend serves inference. Resolution order, first match wins: (1) override via OPENAI_BASE_URL+OPENAI_API_KEY or active profile in ~/.config/sigit/providers.toml; (2) siGit Code Cloud when logged in; (3) on-device.
• src/tools.rs — agent tool schemas + execution: read_file, create_directory, list_directory, search_files, glob, read_website, create_file, edit_file, multi_edit, delete_file, run_command, write_todos, remember. Add a tool in both the spec list (all_tools) and the execute match (execute_tool). run_command also enforces commit attribution: when a command creates a new commit that lacks the Co-Authored-By: siGit Code trailer (COMMIT_CO_AUTHOR_TRAILER), it amends the trailer in — unless the commit already exists on a remote, which is never rewritten. Also owns the task tool: a nested agent loop in a fresh conversation, offered only when subagent_available() (a subagent factory is registered — see register_subagent_factory_for in main.rs; on-device registers a None-returning factory since onde has a single shared history). A subagent's toolset is a hard-gated read-only allow-list (SUBAGENT_TOOL_NAMES) that is never expanded by a configurable subagent type (see src/subagents.rs) — only ever narrowed — so a .sigit/agents/*.md file can't grant itself edit_file/run_command and bypass the permission system. Also owns web_search: a thin native wrapper around the official MCP server's Brave-Search-backed web_search tool (mcp__sigit__web_search, implemented server-side in sigit-si's WebSearchService + Mcp::Tools::WebSearch), offered only when that MCP tool was actually discovered (i.e. the user is signed in to siGit Code Cloud — see web_search_available). Wrapping it natively rather than leaving it as a raw mcp__* tool matters for two reasons: permissions::classify treats every mcp__* tool as mutating (an "ask" prompt on every call), while web_search is read-only like read_website; and the raw delegate name is filtered out of the assembled tool list (is_web_search_delegate) so the model sees one clean option, not two names for the same tool. Execution still forwards verbatim to mcp::call_tool — no second HTTP/JSON-RPC implementation.
• src/skills.rs — Agent Skills support. Discovers skill folders (each with a SKILL.md: YAML frontmatter name + description, then Markdown instructions) from .sigit/skills/ and .claude/skills/ in the cwd, $SIGIT_CONFIG_DIR/skills/, and ~/.claude/skills/. Progressive disclosure: the discovery list (name + description) is baked into the dynamically-built skill tool's description, and activating a skill (the model calls skill with a name) loads the full SKILL.md body. The skill tool is appended in the *_as_specs/build_tool_specs layer (not in all_tools()) so its description can be dynamic, and only when at least one skill exists.
• src/commands.rs — user-defined slash commands. Discovers Markdown files (each an optional YAML frontmatter block — description, argument-hint — followed by a prompt-template body) from .sigit/commands/ and .claude/commands/ in the cwd, $SIGIT_CONFIG_DIR/commands/, and ~/.claude/commands/. A subdirectory namespaces the command with : (.sigit/commands/git/commit.md → /git:commit). Unlike skills there's no tool-call indirection: invoking one works exactly like the built-in /init — commands::render substitutes $ARGUMENTS/$1..$9 in the body against whatever followed the command on the input line, and the result is fed to the model as a normal turn through the ordinary tools and permission checks. Resolution happens where each surface (main.rs/chat.rs) already special-cases /init: an unrecognized slash command is tried against commands::resolve_command before falling back to "unknown command". A custom command sharing a name with a built-in is unreachable (built-ins match first in parse_slash) and is skipped when advertised to ACP clients, with a warning logged.
• src/subagents.rs — configurable subagent types for the task tool. Discovers Markdown files (YAML frontmatter name + description, optional comma-separated tools: allow-list, then a Markdown body that becomes the subagent's system prompt) from .sigit/agents/ and .claude/agents/ in the cwd, $SIGIT_CONFIG_DIR/agents/, and ~/.claude/agents/. Passing a type's name as task's subagent_type argument swaps in that system prompt and, if tools: is set, narrows the offered toolset to its intersection with SUBAGENT_TOOL_NAMES — the security-relevant narrowing logic lives in tools.rs next to that constant, not here; this module only discovers and parses files. SubagentFactory (in tools.rs) takes the resolved system prompt per call rather than baking one in at registration, so a single registered factory serves both the default research subagent and every configured type.
• src/frontmatter.rs — shared "YAML frontmatter + Markdown body" parsing used by both src/skills.rs (SKILL.md) and src/commands.rs (.sigit/commands/*.md).
• src/mcp.rs — Model Context Protocol client. Two transports: Streamable HTTP (one JSON-RPC POST endpoint, url in mcp.toml; replies are application/json or SSE) and stdio (command + optional args/[server.env] in mcp.toml; sigit spawns the server and speaks newline-delimited JSON-RPC over its stdin/stdout, stderr inherited into sigit's log). url and command are mutually exclusive — both or neither is a config error, logged and skipped. Both transports run the same initialize/tools/list handshake and forward tools/call. Discovery is best-effort at startup (mcp::init, called from both branches of main()) and cached in a process-global so the synchronous spec builders (mcp::tool_specs) and the async dispatch (mcp::call_tool) can both read it; /reload does not re-run it, so config changes need a restart. stdio children live for the process; a dead child fails calls with an in-band error string (no auto-restart). Tools are namespaced mcp__<server>__<tool>, appended in the *_as_specs/build_tool_specs layer and routed in tools::execute_tool via mcp::is_mcp_tool. Two servers are baked in: the official server (<cloud>/mcp, default https://sigit.si/api/v1/mcp, always HTTP, authed with the cloud session token) and the smbCloud CLI server (smb --mcp, stdio, added only when the smb binary is on PATH; opt out with smbcloud = false in mcp.toml or SIGIT_MCP_SMBCLOUD=off). A user-defined entry named sigit or smbcloud overrides the corresponding baked-in one. Extra servers live in mcp.toml (global $SIGIT_CONFIG_DIR/mcp.toml and project-local .sigit/mcp.toml). The stdio path is covered by tests/mcp_stdio.rs, driven by the test-only src/bin/mcp_stdio_stub.rs helper binary (excluded from the published crate via exclude in Cargo.toml). The baked-in official server is also listed in the public MCP Registry as si.sigit/sigit — a remote Streamable-HTTP listing (server.json at the repo root, published by release-mcp-registry.yml). Because it's a remote server, the registry's URL-match rule forces a domain namespace (si.sigit ↔ sigit.si) verified by a DNS TXT record, not the GitHub-OIDC scheme smbcloud-cli uses for its package listing.
• src/permissions.rs — tool permission policy. Every tool call passes through decision_for before executing: read-only tools always run; mutating tools (and all mcp__*/unknown tools) are governed by, in order: per-session plan mode (/plan — deny all mutating tools with a present-a-plan message), session "always allow" grants, per-tool overrides and the default mode from [permissions] in settings.toml (allow/ask/deny, default ask; SIGIT_PERMISSIONS env overrides the default). On ask, the ACP path sends session/request_permission (allow once / allow for session / deny) and the TUI pauses the inference task on a y/a/n prompt. Note: ACP turn-affecting handlers run in cx.spawned tasks serialized by SiGitAgent::turn_lock so the dispatch loop can route the client's permission answer mid-turn — don't move them back inline, and don't await client requests from inline handlers (deadlock).
• src/instructions.rs — project instruction files, the always-on counterpart to skills. Reads AGENTS.md (the cross-tool agents.md standard) and CLAUDE.md, walking from the session cwd up to the repo root (nearest ancestor with .git, never above it), plus a global file under $SIGIT_CONFIG_DIR. Files are ordered outermost-first so the deepest (most specific) wins. The combined block is injected via session_context_message in main.rs — pushed as a system message at every ACP session entry point (new/load/fork + model switch) and appended to the system prompt on the cloud and TUI-startup paths.
• src/chat.rs — the Unix-only ratatui TUI. Loading-spinner phase then chat; uses tokio::select! to multiplex terminal events with streaming tokens.
• src/setup.rs — model cache location, local model discovery, selected-model persistence. Must run (setup_shared_model_cache) before anything touches ChatEngine/hf-hub, since those read env vars once at init.
• src/account.rs — siGit Code Cloud auth (/login, /logout, /whoami); authenticates against the account API and stores a session token. Performs no console I/O.
• src/browser_auth.rs — browser sign-in, the path the editor's "Sign in to siGit Code" button takes. sigit is a public OAuth client (client id sigit-code-cli, no secret, PKCE S256) against the authorization server at $SIGIT_API_URL/oauth. Two ways the code gets back: a loopback listener on 127.0.0.1:0 (the registration lists http://127.0.0.1/callback with no port, and RFC 8252 §7.3 matching ignores the port, so one registration covers every run), or the out-of-band URN, where the server renders the code and the user pastes it. The resulting token goes into credentials like any other, and is served by /api/v1/user rather than the deprecated /api/v1/me — OAuth tokens are rejected there. The scopes requested are user:read code:agent; code:agent is what reaches chat/completions and the official MCP server. Surfaces: ACP authenticate (loopback only — nowhere to show a code), a bare /login in either chat surface, and sigit login (which adds the paste fallback, plus --paste to force it and --password for the old email/password prompt).
• src/credentials.rs — local session-token store (TOML, 0600 on Unix).
• src/models.rs — model-picker types shared across platforms.

Slash commands (/help, /models, /skills, /agents, /commands, /mcp, /login, /logout, /whoami, /reload, /plan, /permissions, /init, /clear, /status) are advertised via advertise_commands in main.rs and handled in both the TUI and ACP sessions. /init is special: instead of replying directly it substitutes instructions::INIT_PROMPT for the user text and runs a normal agent turn that explores the repo and writes (or improves) AGENTS.md through the ordinary tools and permission checks. User-defined commands (see src/commands.rs) get the same treatment via the SlashCommand::Unknown fallback path, so anyone can add their own /name commands without touching the built-in command list.

#Model cache (macOS)

On macOS the HF model cache lives in an App Group container shared with the siGit desktop app: ~/Library/Group Containers/group.com.ondeinference.apps/models/. Other platforms fall back to ~/.cache/huggingface/. The CLI reuses a model the desktop app already downloaded. First run downloads a GGUF model (~1–2 GB) from Hugging Face.

#Logging

In TTY (interactive) mode, all output — log, tracing, stray println! — is redirected to $TMPDIR/sigit.log so the ratatui surface stays clean; the TUI holds a separate fd to the real terminal. In ACP mode, stdout is reserved for protocol JSON and logs go to stderr. Control verbosity with RUST_LOG.

#Relevant env vars

OPENAI_BASE_URL / OPENAI_API_KEY (provider override), SIGIT_API_URL (account API base, default https://sigit.si), SIGIT_CLOUD_URL, SIGIT_CONFIG_DIR (default ~/.config/sigit), SIGIT_MODEL, SIGIT_MCP (off disables MCP), SIGIT_MCP_SMBCLOUD (off drops the baked-in smbCloud CLI server), SIGIT_MCP_OFFICIAL (off drops the baked-in server), SIGIT_PERMISSIONS (allow/ask/deny — overrides the default permission mode for mutating tools; the escape hatch for clients without permission-request support), HF_HOME / HF_HUB_CACHE, RUST_LOG.

#Releasing

Version lives in Cargo.toml. Update CHANGELOG.md for releases.

Publishing splits into two groups. Language registries each get their own tag-triggered workflow: release-crates, release-npm, release-nuget, release-pypi. The npm/, nuget/, and pypi/ dirs hold their wrapper-package templates.

npm and NuGet authenticate with OIDC (Trusted Publishing), so neither needs a stored token. On npm that is configured per package, not per org: all seven @getsigit/* names carry a trusted publisher pointing at release-npm.yml, and a package added later needs the same setup or its publish step will fail. OIDC also requires npm >= 11.5.1 and Node >= 22.14.0, which is why the workflow upgrades npm and why .nvmrc cannot drop below 22.

OS package managers all publish somewhere outside this repo and all need checksums from the GitHub release, so release-github builds the binaries, attaches the assets, and then dispatches release-homebrew, release-scoop, release-winget, and release-aur. A dispatch failure in one is logged as a warning rather than failing the others, so an unconfigured channel does not block a release. Their inputs live in packaging/.

Every release asset now carries a .sha256 sidecar, not just the macOS Homebrew tarball. Scoop, winget, and the AUR PKGBUILD each need one, and they consume the raw binaries rather than the tarball. release-github also builds a .deb and .rpm per Linux target with nfpm (packaging/nfpm.yaml), packaging the already-built binary rather than re-invoking cargo.

Three of these need credentials or a one-time manual step before they work:

• Scoop needs a getsigit/scoop-bucket repo and a SCOOP_BUCKET_TOKEN secret, mirroring the Homebrew tap setup.
• winget needs a WINGET_TOKEN (PAT with public_repo) so wingetcreate can fork microsoft/winget-pkgs. It is passed as WINGET_CREATE_GITHUB_TOKEN rather than --token, which wingetcreate warns can leak the token into logs. wingetcreate update only works on a package that already exists in microsoft/winget-pkgs, so the workflow checks the manifests/g/getSigit/siGitCode path first and falls back to rendering packaging/winget/*.yaml.in and running wingetcreate submit for a first submission. That fallback runs once, then every later release takes the update path.
• The AUR needs AUR_USERNAME, AUR_EMAIL, and AUR_SSH_PRIVATE_KEY. It publishes sigit-bin (a prebuilt binary) so Arch users are not compiling the on-device inference stack to install a CLI.

[profile.release] sets strip = "symbols" because binary size is a distribution constraint, not just a nicety. See the NuGet note below.

The NuGet package (SiGit.Code, installed with dotnet tool install --global SiGit.Code) is the odd one out among the language registries: npm and PyPI publish one artifact per platform, but a .NET tool is a single package, so nuget/sigit/ bundles all six binaries under native/<os>-<arch>/ and a small managed shim (Program.cs) execs the right one. That shim leaves stdin/stdout/stderr unredirected on purpose, since siGit Code chooses TUI or ACP mode by testing whether stdin is a TTY. It also sets RollForward=Major: it targets net8.0, and without that a machine carrying only the .NET 10 runtime installs a tool that refuses to start.

Bundling every target means the package is large, so release-nuget.yml fails the pack job if the .nupkg crosses nuget.org's 250 MB limit. At v1.5.1 it lands around 160 MB. If a future release trips that check, the fix is not to drop targets but to split into RID-specific tool packages (.NET 10's DotnetToolRidPackage), which ship one binary per platform the way npm already does. That also needs a fallback package for pre-.NET-10 SDKs.