v0.7.0 feat: add zero-config agent discovery

[minglong51]

Summary

• Run bare clilane on a clean setup and immediately choose any installed Codex, Kimi, or Claude CLI without creating hub.json first.
• Resolve agents from the shell PATH captured when the switcher opens, then pin the exact executable path so launching cannot drift after changing projects.
• Start new sessions in the directory where clilane was invoked; returning from a managed job uses that job's project directory.
• Keep present default and explicit configs authoritative, including intentionally empty configs, with state-specific guidance when discovery is disabled.
• Extend the native PTY suite with deterministic zero-config, source-project, config-precedence, and helper edge-case coverage.

Test Coverage

AI-assessed persistent regression coverage: 83% (20/24 paths), above the 80% target.

Discovery
  [tested] missing and empty PATH
  [tested] found subset, unavailable agents omitted, exact executable retained
  [tested] relative PATH anchored to launch cwd
  [tested] Codex → Kimi → Claude ordering
  [tested] non-absolute discovery directory rejected
Config
  [tested] absent default config enables discovery
  [tested] present default or explicit config is authoritative
  [tested] intentionally empty config suppresses discovery
  [tested] explicit missing config errors
Native flows
  [tested] zero-config launch in invocation cwd with prompt delivery and detach
  [tested] launch from a managed job uses source-job cwd
  [tested] configured, empty-config, multi-client, privacy, and oversized-env flows

Coverage: 20/24 (83%)
Remaining gaps: legacy-context fallback/validation, an empty PATH component,
and the no-installed-agent journey after Tab/Enter.

Pre-Landing Review

Approved with no findings. Independent specialist quality score: 9.6/10.

The review verified terminal safety, exact executable resolution, old-context compatibility, cleanup behavior, PTY reliability, version consistency, and macOS/Linux portability. It also confirmed that the earlier coverage and empty-config guidance findings are resolved.

Design Review

No frontend files changed; design review was not applicable.

Eval Results

No prompt-related files changed; evals were not applicable.

Scope Drift

Scope Check: CLEAN. The diff is limited to zero-config agent discovery, project-aware launch context, user guidance, regression coverage, and v0.7.0 assertions.

Plan Completion

• Zero-config discovery and authoritative config precedence
• Exact executable paths and captured PATH behavior
• Invocation and source-job project directories
• State-specific no-agent guidance and README documentation
• Durable native PTY regression coverage
• macOS/Linux Python 3.10 matrix

Verification Results

• PYTHONDONTWRITEBYTECODE=1 python3 tests/hub_pty.py
• PYTHONDONTWRITEBYTECODE=1 python3 tests/test_provider_evidence.py (23 tests)
• Python 3.10 AST grammar for the CLI and PTY test
• Workflow YAML parsing
• Version assertions and git diff --check
• GitHub macOS/Linux smoke matrix on the final head

Refactor opportunities

• Split the 4,000-line single-file executable into lifecycle, hub UI, and config/discovery modules when the Homebrew packaging strategy can preserve the current one-file install.
• Break the growing PTY smoke file into scenario-focused runners if more hub journeys are added, so failures localize faster without losing real-terminal coverage.
• Persist structured lane provenance instead of inferring BOT origin from environment markers.

Test plan

• Native hub PTY suite passes locally
• Provider-evidence suite passes locally
• GitHub macOS/Linux smoke matrix passes

Generated with OpenAI Codex.