[Epic] Harness compatibility roadmap: core adapters, ACP tap, and community adapters

[Airine]

Objective

Track the next USL compatibility phase: adapter infrastructure, project-maintained head harnesses, experimental ACP live capture, and import-first community coverage.

Delivery principles

• Project-maintained: Claude Code, Codex, Gemini CLI, OpenCode, and Pi; retain dimagent support.
• Community adapters are import-first. An exporter is allowed only after upstream exposes a stable public restore/resume entry point and native resume is verified.
• Adapters are compile-time repository modules registered in USL; runtime third-party plugin loading is out of scope.
• Native evidence, unknown records, and opaque payloads must survive; all semantic loss is explicit.
• SQLite readers use database + WAL/SHM snapshots and never mutate native storage.
• A harness enters the roadmap only when it has a publicly obtainable native session/export format.

Phase 1 — Adapter Foundation (v0.2 Adapter Foundation)

• refactor(usl-convert): add a harness adapter registry and conformance kit #1

Phase 2 — First-party harnesses (v0.3 Core Harnesses)

• feat(adapter): complete Claude Code round-trip and verified resume export #2
• feat(adapter): add Gemini CLI session compatibility #3
• feat(adapter): add OpenCode session compatibility #4
• test(adapter): audit Pi compatibility against the current upstream format #5

Codex does not need a duplicate feature issue: #1 must preserve its current import/export behavior and encrypted-reasoning regression coverage.

Phase 3 — ACP tap

• agent-session-protocol/asp#1
• feat(capture): ingest live sessions through an ACP tap #6

The first USL implementation is experimental and stdio-only. It must be validated against at least one real ACP harness.

Community adapter backlog

All v1 contributions are import-only and depend on #1.

• feat(adapter): import GitHub Copilot CLI sessions #7 — GitHub Copilot CLI
• feat(adapter): import Kimi Code sessions #8 — Kimi Code
• feat(adapter): import Qwen Code sessions #9 — Qwen Code
• feat(adapter): import Goose sessions #10 — Goose
• feat(adapter): import Cline tasks #11 — Cline
• feat(adapter): import Continue CLI sessions #12 — Continue CLI
• feat(adapter): import Aider chat history #13 — Aider
• feat(adapter): import Crush sessions #14 — Crush
• feat(adapter): import OpenHands conversations #15 — OpenHands
• feat(adapter): import Mistral Vibe sessions #16 — Mistral Vibe
• feat(adapter): import Zed Agent threads #17 — Zed Agent
• feat(adapter): import OpenClaw sessions #18 — OpenClaw

Deferred candidates

Do not open implementation issues yet for:

• Cursor
• Windsurf
• Amp
• Factory Droid
• Amazon Q
• archived Roo Code

Reconsider a candidate only when an active upstream publishes a stable, publicly obtainable native session or export format with enough structure to build sanitized fixtures and deterministic import tests. Archived projects additionally need an active maintained successor or a clear compatibility owner.

Definition of done

• Existing 25 conversion tests remain green after registry migration.
• Every adapter covers determinism, source read-only behavior, malformed input, fidelity/loss, and a realistic native-structure fixture.
• SQLite adapters test WAL/SHM copying and live-database immutability.
• Every exporter has same-harness native round-trip and verified resume coverage.
• ACP tests cover patch merge, duplicates, opaque payloads, thinking degradation, disconnect, recovery, and identity mapping.
• npm run check, cargo test --workspace, and npm run lint:architecture pass.
• Issue labels, assignees, dependencies, fixtures, and acceptance criteria have been audited.

Explicitly out of scope

SessionQL, an OTEL exporter, and management-console expansion are not part of this roadmap.