darman/ai-harness
1
0
Fork 0
Code Issues Pull Requests 5 Actions Packages Projects 1 Releases Wiki Activity

Labels Milestones

M1 — Headless core loop + Anthropic
0%
1 Open
0 Closed
Updated 2026-07-10 17:08:38 +02:00
No due date

Goal

A working headless engine loop against a real provider (Anthropic), with the tool trait,
config loading, and permission service it depends on.

Scope

  • Tool trait + config loading/schema
  • Engine loop: text → tool call → final text, with a doom-loop guard and retry scaffolding
  • Permission service (auto-approve stub frontend for headless use)
  • read, write, bash, glob, grep tools
  • edit tool: opencode's replacer chain ported verbatim, including its test table
  • Anthropic SSE codec + provider, registry wiring
  • harness-app composition root
  • harness run -p "<prompt>" debug command

Acceptance criteria

  • Integration test: a scripted MockProvider does text → tool call (read) → final text with
    correct parts persisted.
  • harness run -p "<prompt>" works against real Anthropic.
M2 — TUI
0%
1 Open
0 Closed
Updated 2026-07-10 17:10:00 +02:00
No due date

Goal

A ratatui-based terminal UI driving the M1 engine loop interactively.

Scope

  • EngineHandle bridging the async engine to the render loop
  • Event loop + chat viewport, with a markdown renderer and a render cache
  • Input handling
  • Permission modal wired to the real oneshot ask path
  • Session picker / resume
  • Abort (cancel an in-flight turn)
  • TestBackend snapshot tests: empty session, tool cards (running/completed/error),
    permission modal, session picker, user/assistant messages

Acceptance criteria

  • Manually drive a multi-turn session with permission-gated edits.
  • TestBackend snapshot tests pass.
M3 — All providers + auth
0%
0 Open
0 Closed
Updated 2026-07-10 17:03:30 +02:00
No due date

Goal

Round out provider support (OpenAI, Copilot) alongside Anthropic, with auth, cost tracking,
and retry/doom-loop hardening.

Scope

  • OpenAI chat + responses codecs, provider, registry wiring
  • auth.json credential storage
  • Copilot: device-flow login modal, token exchange with direct-Bearer fallback,
    /models routing across the three codecs
  • Model pricing + per-step cost accumulation
  • models.dev metadata catalog + cost display wiring in the TUI
  • Retry policy + doom-loop guard (landed in M1's engine loop, exercised fully once all
    three providers exist)

Acceptance criteria

  • The same session works on all three providers.
  • Device-flow login completes inside the TUI.
  • Costs shown.

Top risk carried into this milestone

Copilot token exchange variance — opencode ships direct-Bearer, but classic Copilot API wants
the copilot_internal/v2/token exchange. Mitigated with a dual-path token.rs; verify against
the live API early.

M4 — Multiagent
0%
1 Open
0 Closed
Updated 2026-07-10 17:11:31 +02:00
No due date

Goal

Let a session spawn and track subagents, foreground or background, with their own
permission scope.

Scope

  • Agent registry: bundled markdown agent definitions (designer, explorer, fixer,
    librarian, oracle, orchestrator) with override support
  • Permission intersection: a spawned subagent's effective rules are the parent's
    narrowed by its own
  • JobBoard: persistence + context injection into the parent's next turn
  • task tool + subagent spawner: aliases/reuse, depth limits
  • Subagent context-file reporting to the job board
  • Jobs pane, subtask drill-in, reminders for long-running background jobs
  • Dedicated opencode (Zen) provider

Acceptance criteria

  • An orchestrator session spawns a background explorer; the board appears in next-turn
    context (verify via request logging).
  • Reuse by alias continues the same child session.
M5 — MCP, LSP, commands/skills
0%
0 Open
0 Closed
Updated 2026-07-10 17:02:50 +02:00
No due date

Goal

Extend tool access beyond the built-ins: external MCP servers, language-server diagnostics,
and user-authored commands/skills.

Scope

  • LSP pool + diagnostics: harness-core::lsp seam trait, harness-lsp pool (lazily
    spawns one server per language, only when its binary is on PATH), surfaced in
    edit/write/bash/glob/grep output via harness-tools::diagnostics
  • rmcp stdio client + namespaced tool adapters, wired through harness-app
  • Skills: SKILL.md frontmatter/body parsing + skill tool, advertised in the system prompt
  • Slash commands: markdown command files under command/*.md, TUI input handling for
    /mycommand

Acceptance criteria

  • A config-declared MCP server's tool is callable with ask-permission.
  • Editing a broken .rs file surfaces rust-analyzer errors in the tool result.
  • /mycommand works.
M6 — Polish
0%
0 Open
0 Closed
Updated 2026-07-10 17:02:28 +02:00
No due date

Goal

Make long sessions durable and the TUI pleasant to actually live in.

Scope — done

  • Auto-compaction: Compactor + overflow trigger, threaded through the engine loop
  • Session-title generation via a configured small_model
  • Crash-resume hardening: an unfinished assistant message left by a killed process is
    marked aborted on load instead of replayed
  • Theming — theme: Option<String> exists in config::schema but isn't read anywhere in
    harness-tui
  • Syntax-highlighted code blocks in the markdown renderer
  • Streaming-render perf pass (incremental layout / tail-only re-render beyond the
    existing markdown render cache from M2)

Acceptance criteria

  • A session driven past the context limit compacts via Compactor and continues correctly,
    with prior turns summarized rather than dropped or truncated mid-message.
  • After a new session's first turn, its title reflects a small_model-generated summary
    instead of the default placeholder.
  • Resuming a session whose process was killed mid-turn shows the last assistant message as
    aborted, and does not replay or re-send it to the provider.
  • Setting theme in config produces a visibly different TUI color scheme; switching between
    at least two configured themes changes the rendered output.
  • Fenced code blocks in assistant messages render with syntax highlighting matching the
    fence's declared language.
  • Streaming a long assistant response renders incrementally without visible input lag or
    full-viewport re-layout on each token.