Miwarang Senpi Delegate v0.2.0

Miwarang Senpi Delegate v0.2.0

Background delegations are now visible live: a session status line, a /miwarang view, and an automatic dag watch run rendered in /dag and any dag pane.

Included

• Session-visible progress for every background delegation: a TUI footer status line (ctx.ui.setStatus) showing active runs with their runtime, plus the /miwarang command listing run id, state, elapsed time, target, and backend.
• Automatic dag watch bridge: each background delegation starts a dag run whose single category-routed watcher node polls the delegated run (the command-backend run record, or opencode api POST /api/experimental/session/<id>/wait) until it reaches a terminal state. The run renders in /dag, the TUI status widget, and dag panes such as omo-tmux-dag.
• New top-level dagWatch config key: false disables the bridge; { category, pollSeconds, timeoutMinutes } overrides the watcher's category route (default quick) and its bounds (default 30 s poll, 60 min watch).
• The miwarang_delegate background result reports the watch run id (Dag watch: dag_…) or the reason the watch could not start; a bridge failure never fails the delegation.
• The status line polls command run records every 15 seconds only while a command run is alive, then stops itself; opencode session state flips when miwarang_wait or miwarang_cancel reports it.

Verification

• bun test: 135 → 162 passing tests, 0 failures across 6 files (test/watch.test.js added, plus new dag watch bridge and progress surfaces suites).
• On-disk proof (.omo/evidence/2026-10-20-watch-bridge/proof-output.txt): real modules + real RunStore + the real 15 s poller — the status line showed miwarang: 1 active · run_22721626 0s on start, cleared itself after the run record flipped to completed, and /miwarang reported the completed run with target and backend.
• Dag watch trigger verified live against the OMO dag engine: a watch run starts from inside the tool handler through the eval relay (ctx.executeTool("eval", …) → kernel tool.workflow({ action: "start" }), returning a dag_<uuid> run id). A direct ctx.executeTool("workflow", …) is refused by the host ("Tool workflow not found"), which is why the bridge relays through eval.
• Bridge degradation verified by tests: a host without ctx.executeTool, a throwing relay, and dagWatch: false each keep the delegation working (details.dagWatch carries { error } or is absent).
• config.example.json validates against validateConfig with the new dagWatch key present.

Miwarang Senpi Delegate v0.1.1

Miwarang Senpi Delegate v0.1.1

Fixes result reporting for OpenCode v2 sessions and stops masking CLI failures.

Fixed

• wait and foreground delegation now return the newest assistant step that has text. The OpenCode message API returns messages newest-first, so the previous picker returned the run's opening line, or failed with "returned no assistant text", while the session held a complete report.
• A timed-out opencode api invocation is now reported as exit 124 with its stderr preserved, instead of exit 1 with the cause dropped.

Included

• MIT license, contributing guide, and license metadata in package.json.

Verification

• bun test: 29 passing tests, including regressions for descending message order, assistant steps without text, and timeout-shaped CLI failures driven through a real child process.
• Live end-to-end against local OpenCode v2 sessions: wait returned the full final reports (5,147 and 6,765 characters) where the previous code returned a 48-character opening line or an error.
• The new tests run against the pre-fix source: 6 failing expectations, confirming they catch the defect.

Miwarang Senpi Delegate v0.1.0

Miwarang Senpi Delegate v0.1.0

Initial release of a Senpi extension for delegating OMO work to a local OpenCode v2 service or a configured command-line agent.

Included

• Senpi tools for foreground delegation, background start, wait, and cancellation.
• OpenCode v2 integration through the authenticated opencode api CLI.
• Live OpenCode model discovery and free-only route validation.
• Default OMO routes for quick, unspecified-low, deep-low, explore, and librarian.
• A generic executable-plus-argv backend with stdin or argument prompt delivery and no shell interpolation.
• Configurable model and OpenCode agent overrides.
• Background OpenCode session-ID tracking and local command-run metadata.

Verification

• bun test: 23 passing tests.
• OpenCode v2.0.26 catalog: 74 models; configured free routes resolve to available zero-cost models.
• Live foreground, background/wait, cancellation, and OMO category:quick delegation were exercised with OpenCode free-tier models.
• Pi, Claude Code, and Codex CLI implementations were not individually installed or tested; use the generic command backend with each CLI's supported flags.