Uni-CLI Roadmap

Current: v0.225.2 — Apollo · Gordon. 317 sites, 1772 commands, 103 pipeline steps.

This file tracks current engineering direction for the agent control plane for real software: a universal agent-to-computer control platform. Historical release notes live in CHANGELOG.md; they do not belong in the roadmap.

Shipped

Area	Status
Operation catalog	Web, browser, desktop, macOS, bridge, local-tool, and protocol operations are discoverable through list, search, and do.
v2 output envelope	Normal command surfaces return structured success and error envelopes.
Operation policy	open, confirm, and locked profiles expose effect/risk/capability/resource scope, plus private remembered approvals.
Run recording	--record and UNICLI_RECORD_RUN=1 write append-only run traces that agents can list, show, probe, replay, and compare.
Browser evidence	Browser operator actions can emit pre/post evidence, movement dimensions, stale-ref details, and watchdog results.
Local computer control	compute exposes app discovery, snapshots, refs, clicks, typing, keys, scrolling, app launch, screenshots, and assertions.
Runtime exposure	Native CLI, JSON stream, MCP, ACP, HTTP API, OpenAI-compatible, and bridge routes are modeled explicitly.
Self-repair and delivery	Errors carry adapter path, step, retryability, suggestion, alternatives; delivery commands assess objectives and trajectories.
Docs site	VitePress landing page, guide/reference split, local search, and GitHub Pages deployment workflow are available.

Next Priorities

1. Computer-control root model

   • Keep public docs, architecture tree, command descriptions, and agent surfaces aligned around agent-to-computer control over real software.
   • Treat browser automation, computer-use sandboxes, MCP, natural-language local execution, and per-site wrappers as substrates below Uni-CLI, not as competing product identities.
   • Make architecture tree and architecture audit the executable check that catches regressions back into catalog/lifecycle-first framing.
   • Use capability_matrix and workflow_readiness from architecture audit to separate cataloged coverage from behavior that still needs live evidence.

2. Operation-contract parity

   • Project adapter commands and core Commander commands into the same operation-contract shape.
   • Keep argument schemas identical across search, describe, --dry-run, direct CLI, MCP, ACP, generated agent configs, and docs.
   • Add regression tests when a generated command family gets new arguments or a core command becomes externally callable.

3. Control-kernel hardening

   • Keep run traces append-only, local, and private by default.
   • Let agents probe replayability before repeating a command, then compare the replay trace against the original.
   • Extend evidence coverage across transport classes without making opaque browser screenshots the only proof.
   • Keep result envelopes, permission evaluations, and browser action evidence queryable enough for reviews and repair tasks.

4. Operation policy coverage

   • Keep adapters open by default.
   • Expand effect/risk/capability-scope inference where commands still lack enough metadata.
   • Use --yes --remember-approval when a team wants repeat approval for the same command capability scope without storing raw args.

5. Substrate bus

   • Make HTTP, CDP, accessibility, subprocess, service, and Visual dispatch share one invocation kernel and one evidence model.
   • Keep ACP/MCP/HTTP as wrappers over the same operation contracts rather than separate behavior definitions.
   • Surface unavailable transports as structured errors with install/setup suggestions.

6. Desktop and Visual stack

   • Build repeatable control paths for WeChat, WeCom, DingTalk, Lark, Mail, Notes, Word, PowerPoint, Excel, and common Electron apps.
   • Prefer app APIs, CDP, and accessibility before Visual.
   • For partial accessibility shells, add screenshot planning, background action primitives, and post-action verification before marking commands live.

7. Delivery loop alignment

   • Support parallel/background agent workflows with isolated worktrees, compact command discovery, and reviewable evidence.
   • Keep Uni-CLI command execution independent from any single agent loop or editor protocol.
   • Feed adapter failures back into repair tasks that can be run by coding agents.

8. Continuous trend intake

   • Periodically review agent-loop, computer-use, editor-agent, and desktop automation trends through a private research process.
   • Convert durable insights into architecture or roadmap updates, not prompt lore.
   • Keep source-specific attribution internal and keep public docs at the capability level.
   • Keep code decisions grounded in local tests, diffs, and runtime evidence.

9. Adapter authoring loop

   • Keep browser analyze, init, verify, fixtures, field maps, and site memory as first-class authoring artifacts.
   • Store reusable site notes under ~/.unicli/sites/SITE/.
   • Make repair output directly reusable by coding agents without extra prose.

10. Browser network detail

• Preserve request/response detail, cache hits, filters, and timing in browser capture commands.
• Make captured network evidence usable as adapter fixtures.

11. Backend honesty

    • Keep ACP as compatibility.
    • Prefer native CLI, JSON stream, and MCP when a backend exposes them.
    • Do not mark Visual as live unless a configured backend performs real actions.

12. Docs as product surface

    • README stays install-first and capability-first.
    • Public docs should explain the computer-control loop, operation contracts, action substrates, repair paths, and integration routes.
    • Keep the public entry path short, current, and directly useful.

13. Workflow evidence closure

    • Start from vehicle-assistant workflows: media, video search, browser tabs, installed apps, productivity state, and open/navigate destinations.
    • Promote a workflow from cataloged to claimed only after command execution, post-state evidence, and auth/policy posture have been recorded.
    • Do not add commands for coverage optics; new commands need a runner, fixture, live smoke, or platform doctor evidence.

Non-Goals

• No winner-take-all backend policy.
• No hidden success when an adapter failed.
• No theory-first README and no catalog-first identity.
• No new protocol shim unless it reduces latency, preserves session semantics, or unlocks a real client.
• No positioning as only a browser library, MCP wrapper, computer-use sandbox, natural-language shell, scraper, or per-site wrapper collection.

Verify

For public positioning and docs changes:

npm run docs:build
npm run docs:check-public

For release readiness:

npm run build
npm run release:check
npm run verify