flutter_agent_harness 0.1.294

Cross-platform AI agent harness for Dart and Flutter: streaming provider adapters, agent loop with tools, session persistence, context compaction.

flutter_agent_harness

Cross-platform AI agent harness for Dart and Flutter — streaming provider adapters, an agent loop with native tool calling, JSONL session persistence, and context compaction. Architecture ported from pi-mono (packages/ai + packages/agent), with a pure-Dart core that runs on the VM, Flutter desktop/mobile, and web.

Status: early development (Phase 0). See GOAL.md for the roadmap, quality gates, and design contract. The API is not yet stable.

Highlights (target design)

• Streaming-first: Stream<AgentEvent> from every provider, partial-first deltas — each event carries the current partial message.
• Errors-as-events: providers never throw; failures arrive as error events with a stopReason, so the agent loop never dies on a 429 or a dropped connection.
• Native tool calling (OpenAI tools, Anthropic tool_use, Google functionCalling) — no prompt-based JSON scraping.
• Token-based context management: inline usage accounting, overflow detection, LLM-powered compaction.
• Sessions as append-only JSONL trees behind a storage abstraction — portable to web and mobile.
• Cancellation everywhere via CancelToken.

Usage (current seed)

import 'package:flutter_agent_harness/flutter_agent_harness.dart';

void main() async {
  final source = CancelTokenSource();

  // Pass source.token into any long-running operation.
  final cancelled = source.token.onCancel.then((_) => print('aborted'));

  source.cancel('user pressed stop');
  await cancelled;
}

Provider adapters and the agent loop land in the next phases — see the roadmap in GOAL.md.

CLI (fah)

A pi-like terminal coding agent ships in bin/fah.dart: a REPL with the built-in read / write / ls / bash tools, JSONL session persistence under ~/.fah/sessions (cwd-encoded layout), automatic context compaction, and slash commands.

export OPENROUTER_API_KEY=sk-or-...   # or ANTHROPIC_API_KEY / GOOGLE_API_KEY
dart run bin/fah.dart                 # defaults: OpenRouter, claude-sonnet-4
dart run bin/fah.dart --provider anthropic --model claude-sonnet-4-5
dart run bin/fah.dart --model openai/gpt-4o-mini --cwd . --session-root /tmp/fah

Headless mode runs a single non-interactive prompt and exits — the response streams to stdout, tool indicators and notices go to stderr (stdout stays pipeable), nothing is ever prompted interactively, and the session persists like a REPL run. Exit codes: 0 ok, 1 provider error, 130 aborted (Ctrl-C). A first positional naming an existing file becomes the prompt source: text files (.md, .markdown, .txt) are inlined as the prompt; any other (binary) file is attached as a path reference for the agent's tools — in both cases trailing text appends as the instruction. A path that does not exist is treated as plain prompt text.

dart run bin/fah.dart "summarize the changelog"      # positional prompt
dart run bin/fah.dart -p "fix the typos in README.md"  # -p/--prompt alias
dart run bin/fah.dart CHANGELOG.md "summarize this"  # text file as prompt
dart run bin/fah.dart screenshot.png "describe it"   # binary → path reference
dart run bin/fah.dart "summarize the changelog" | pbcopy  # pipes cleanly

Flags: --model <id>, --provider openai-completions|anthropic|google|dial|minimax|zai, --base-url <url>, --cwd <dir>, --session-root <dir>, -p/--prompt <text>, --help, --version.

The chatgpt provider (Codex backend) is also available: sign in with a ChatGPT account via /provider chatgpt oauth in the REPL (OAuth-only — there is no headless --provider chatgpt flag). Several ChatGPT accounts can coexist: the flow offers the saved accounts first, each account keeps its own named entry and secure-store slot, and re-auth never touches a sibling account's credentials.

Env preconfig (Docker / headless)

FA_PROVIDER_TYPE + FA_PROVIDER_CONFIG boot a declared provider with no saved config, and the declaration becomes the session default for every model role (default/smol/slow/plan) — the same selection a /provider <name> switch makes:

FA_PROVIDER_TYPE=zai
FA_PROVIDER_CONFIG='{"baseUrl":"https://api.z.ai/api/coding/paas/v4","model":"glm-5.3","apiKeyEnvVar":"ZAI_API_KEY"}'
ZAI_API_KEY=sk-...

baseUrl and model are required — no catalog defaults fill gaps; a missing field fails loud at boot. apiKeyEnvVar is optional: declared, the named env var (or its _BASE64 twin) must hold the key; omitted, the provider boots keyless and the spec's usual env names are never probed. Every text value has a base64 twin for CI platforms that mangle special characters — FA_PROVIDER_CONFIG_BASE64, and <apiKeyEnvVar>_BASE64 for the key: the plain value wins when both carry the same value; mismatched or malformed twins fail loud.

# base64 twin form (identical boot):
FA_PROVIDER_CONFIG_BASE64=$(printf '%s' "$FA_PROVIDER_CONFIG" | base64)
ZAI_API_KEY_BASE64=$(printf '%s' "$ZAI_API_KEY" | base64)

GitHub Copilot is a first-class provider. /provider copilot connects a GitHub account via the device-code flow (open the shown verification_uri, enter the user_code) or by pasting an existing PAT — the flow works headless too — and the Flutter app offers the same connect as a sheet. Accounts save as named entries (copilot-<login> by default); the plan picks the host — individual api.githubcopilot.com, business api.business.githubcopilot.com, enterprise api.enterprise.githubcopilot.com, or a custom --base-url override — and several accounts can coexist side by side. Tokens live only in the OS secure store (Keychain / Secret Service); config.yaml carries name, plan, and baseUrl, never a token. CI runs store-less via the FA_KEY_COPILOT_<NAME> env (plus a _2… ring for more entries), and headless runs take --provider copilot --model <id> with COPILOT_GITHUB_TOKEN. Models come from a live GET /models (with capabilities and limits). The device-flow client id is overridable via FA_COPILOT_CLIENT_ID — that GitHub endpoint is undocumented, so a custom client id carries an account-ban risk; override only with cause.

Slash commands inside the REPL: /exit, /reset (new session), /compact (summarize history now), /stats (token/cost totals), /model <id> (show or switch model), /approval [always-ask|write|yolo] (tool approval mode), /allow [tool] (always-allow a tool), /help. While a run is streaming, typed input is steered into the agent; Ctrl-C aborts the run (Ctrl-C at the idle prompt exits).

Tool calls pass an approval gate (lib/src/approval/): every tool has a capability tier (read/write/exec; exec for undeclared custom tools), the session mode decides what runs unattended, and per-tool overrides plus a critical-pattern interceptor for bash (e.g. rm -rf /, fork bombs, curl … | sh) can force a prompt — even in yolo. The prompt UI is an injectable callback; piped (non-interactive) input denies prompt-policy calls with a reason. Mode and always-allowed tools persist in ~/.fah/config.yaml.

The CLI core (AgentCli + CliIO) is pure Dart and lives in lib/src/cli/agent_cli.dart; only bin/fah.dart and lib/io.dart touch dart:io.

Agent-to-agent messaging (DAP)

The CLI ships a default-on DAP/1 hub plugin (bin/fah_hub_plugin.dart + the hosted fah_hub_client pub package): agents connect to a hub over a signed WebSocket, exchange end-to-end encrypted channel messages and DMs (the hub only ever sees ciphertext), and see each other's presence. Inbound hub mail is drained into the agent loop as steering messages; /dap and the dap_* tools drive the connection. See docs/dap.md for the protocol, the hub server, and an end-to-end setup walkthrough.

Development

dart pub get
dart test --coverage=coverage --exclude-tags integration
dart run coverage:format_coverage --lcov -i coverage -o coverage/lcov.info
python3 scripts/check_coverage.py

Pre-commit hook (analyze + tests + coverage ≥ 80% + duplication < 1%):

cp scripts/pre-commit .git/hooks/pre-commit && chmod +x .git/hooks/pre-commit

License

MIT — see LICENSE.