Skip to content

phase_3_opencode_provider_runtime.md

Latest commit

 

History

History
95 lines (69 loc) · 5.99 KB

phase_3_opencode_provider_runtime.md

File metadata and controls

95 lines (69 loc) · 5.99 KB

Phase 3 — OpenCode Provider Runtime

Summary

Implemented the OpenCodeCliProvider class, registered it in the provider manager, and added it to the workspace-access guard in the launch command. 43 unit tests were written covering all acceptance criteria at 96% line coverage.

Files Created

File Purpose
src/cli_agent_orchestrator/providers/opencode_cli.py Full provider implementation (332 lines)
test/providers/test_opencode_cli_unit.py 43 unit tests (96% line coverage)

Files Modified

File Change
src/cli_agent_orchestrator/providers/manager.py Import + elif branch for opencode_cli
src/cli_agent_orchestrator/cli/commands/launch.py Added "opencode_cli" to PROVIDERS_REQUIRING_WORKSPACE_ACCESS

Acceptance Criteria Disposition

AC-1 — OpenCodeCliProvider subclasses BaseProvider with all required methods

Satisfied. The class implements initialize(), get_status(), extract_last_message_from_script(), get_idle_pattern_for_log(), exit_cli(), and cleanup(). The paste_enter_count property returns 1.

AC-2 — initialize() sends the correct inline-env launch command and waits for IDLE/COMPLETED with 120s timeout

Satisfied. _build_launch_command() emits all seven stability env vars (OPENCODE_CONFIG, OPENCODE_CONFIG_DIR, OPENCODE_DISABLE_AUTOUPDATE=1, OPENCODE_DISABLE_MOUSE=1, OPENCODE_DISABLE_TERMINAL_TITLE=1, OPENCODE_CLIENT=cao, TERM=xterm-256color) followed by opencode [--agent <name>] [--model <name>] built with shlex.join. The wait_until_status call uses a 120-second timeout to cover first-run npm install cold starts. Verified output:

OPENCODE_CONFIG=... OPENCODE_DISABLE_AUTOUPDATE=1 ... opencode --agent developer
OPENCODE_CONFIG=... OPENCODE_DISABLE_AUTOUPDATE=1 ... opencode --agent code-reviewer --model anthropic/claude-sonnet-4-6

AC-3 — get_status() correctly classifies all five states

Satisfied. Priority-order detection (WAITING_USER_ANSWER → PROCESSING → COMPLETED → IDLE → ERROR) with line-level position guard for stale esc interrupt alt-screen remnants.

Key design decisions:

  • Line-level position guard: During active processing, esc interrupt and ctrl+p commands appear on the same footer line. A stale alt-screen snapshot puts them on separate lines. The guard checks whether any idle-footer or completion-marker line appears after the line containing esc interrupt — if so, esc interrupt is stale.
  • esc_is_stale flag: When the position guard fires, this flag allows the IDLE branch to match even though esc interrupt text is still present in the buffer.
  • COMPLETED requires no trailing ▣: After completion, the new user input bar shows a partial ▣ Build · Big Pickle (no duration), which would make the IDLE-post-completion state look like a new incomplete turn. The COMPLETED check confirms no ▣ token follows the last full completion marker.

AC-4 — extract_last_message_from_script() correctly extracts agent text

Satisfied. Algorithm:

  1. Strip ANSI codes.
  2. Find last full COMPLETION_MARKER_PATTERN match (requires ·…·…Ns duration suffix).
  3. Search for last ┃ before the completion marker (unanchored — TUI lines have leading spaces).
  4. Extract between ┃ end and completion marker start.
  5. Skip user-message lines (still using ┃ indent).
  6. Strip Thinking: preamble lines.
  7. Dedent 5-space agent indent.
  8. Clean control characters and trailing whitespace.

Raises ValueError with descriptive messages for missing completion marker or missing user message.

AC-5 — Provider registered in manager.py and launch.py

Satisfied. ProviderManager.create_provider() routes "opencode_cli" to OpenCodeCliProvider passing model=model. PROVIDERS_REQUIRING_WORKSPACE_ACCESS includes "opencode_cli".

AC-6 — 90%+ test coverage

Satisfied. 96% line coverage on opencode_cli.py (5 uncovered lines are unreachable exception paths in string cleaning).

Smoke Test

Server-based launch was not used (port 9889 reserved per user instruction). Provider dispatch was verified directly:

from cli_agent_orchestrator.providers.manager import ProviderManager
pm = ProviderManager()
p = pm.create_provider('opencode_cli', 'test-terminal-1', 'cao-session', 'win-1', 'developer')
# → Provider type: OpenCodeCliProvider
# → launch command: OPENCODE_CONFIG=... opencode --agent developer
p2 = pm.create_provider('opencode_cli', 't2', 's', 'w', 'code-reviewer', model='anthropic/claude-sonnet-4-6')
# → launch command: ... opencode --agent code-reviewer --model anthropic/claude-sonnet-4-6

The cao install developer --provider opencode_cli CLI command was also verified in a prior session:

✓ Agent 'developer' installed successfully
✓ opencode_cli agent: /home/bajablast69/.aws/opencode_cli/agents/developer.md

Design Decisions and Trade-offs

  • --model only via launch flag, never frontmatter (§3.1 exception): OpenCode frontmatter does not support a model: key that overrides the active provider; passing it via --model at the CLI level is the correct path. The model parameter flows from ProviderManager.create_provider() → OpenCodeCliProvider.__init__() → _build_launch_command().
  • Unanchored ┃ search in extract_last_message_from_script: The module-level USER_MESSAGE_PATTERN = r"^┃\s{2}" uses ^ anchoring suitable for full-line matching, but TUI output lines have variable leading spaces (e.g. ┃ say hello). The extraction function uses r"┃\s{2}" without anchoring to handle this.
  • Completion-marker-first extraction order: Finding the completion marker first and searching backwards for ┃ avoids the bottom input-box trap — the last ┃ in a completed screen is the new input-box agent/model header which appears after the completion marker. Restricting the ┃ search to clean[:last_completion.start()] cleanly avoids it.
⚡