duet-cli 0.2.8__tar.gz → 0.2.10__tar.gz

{duet_cli-0.2.8/duet_cli.egg-info → duet_cli-0.2.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: duet-cli
-Version: 0.2.8
+Version: 0.2.10
 Summary: Two CLI agents in conversation. One Python file. Stdlib only.
 Author: Volkan Altan
 License-Expression: MIT
@@ -47,7 +47,9 @@ Plain `/duet` runs Claude Code's real `/review`, then loops Codex and Claude in
 a worktree until they converge. Pass any upstream command as the kickoff:
 `/duet 'npm test 2>&1' --turns 4`. The plugin shells out to the `duet` CLI, so
 install that first (see below) and make sure `command -v duet` passes in Claude
-Code's shell. Full guide:
+Code's shell. Custom topology flags such as `--no-worktree` and
+`--allow-worktree-fallback` replace the plugin defaults without emitting
+conflicting CLI flags. Full guide:
 [docs/CLAUDE_CODE_PLUGIN.md](https://github.com/volkan/duet/blob/main/docs/CLAUDE_CODE_PLUGIN.md).
 If Claude Code disambiguates plugin commands with namespaces, `/duet:duet` is
 the same command.
@@ -108,8 +110,10 @@ Then invoke `/duet` in the OpenCode TUI (or `opencode run --command duet "..."`
 non-interactively). Like the other plugins it shells out to the `duet` CLI, so
 install that first and make sure `command -v duet` passes in OpenCode's shell.
 The command runs on OpenCode's `build` agent; plain `/duet` runs the same
-`claude -p /review` kickoff, and `/duet 'npm test 2>&1' --turns 4` seeds from
-any command. Full guide:
+`claude -p /review --model sonnet` kickoff, and
+`/duet 'npm test 2>&1' --turns 4` seeds from any command. Custom worktree
+overrides are constructed conditionally, matching the Claude and Codex entry
+points. Full guide:
 [docs/OPENCODE_PLUGIN.md](https://github.com/volkan/duet/blob/main/docs/OPENCODE_PLUGIN.md).
 (duet can also drive OpenCode as a backend — `--partner opencode:coder` — so
 OpenCode can be one of the two looped agents too.)
@@ -155,9 +159,28 @@ duet --task "Review the latest commit; fix only what the reviewer requests." \
 `/review`, a test run, or any command:
 
 ```bash
-duet --task-from-cmd 'claude -p /review' \
-    --lead claude:reviewer --partner codex:coder \
-    --worktree --recap --cwd ~/workspace/project --turns 6
+duet --recipe review --cwd ~/workspace/project
+```
+
+The review recipe uses recap mode, `claude:reviewer` + `codex:coder`, six
+turns, `claude -p /review --model sonnet`, and strict worktree isolation.
+Claude agents use the stable `sonnet` alias unless an explicit slot model
+overrides it. The recipe also sets
+`--on-turn-timeout continue`, so when the coder turn times out the reviewer
+still gets the timeout block plus the worktree diff to review instead of the
+run dying reviewless. Explicit flags override recipe values.
+
+**Deterministic automation** — discover the run without parsing terminal prose,
+then poll the curated status schema:
+
+```bash
+control_dir=$(mktemp -d)
+duet --recipe review --run-info-file "$control_dir/run.json" &
+duet_pid=$!
+while [ ! -s "$control_dir/run.json" ]; do sleep 0.1; done
+run_dir=$(python3 -c 'import json,sys; print(json.load(open(sys.argv[1]))["run_dir"])' "$control_dir/run.json")
+duet --status "$run_dir" --json
+wait "$duet_pid"
 ```
 
 **Deep planner, fast coder** — Claude plans at high effort while Codex coder
@@ -202,18 +225,31 @@ other.
 To converge, an agent must include an `LGTM rationale:` explaining why the work
 is done, followed by the sentinel `<<<LGTM>>>` on its own line — a bare
 sentinel is ignored, and **both** agents must agree in back-to-back turns. The
-loop also stops on `--turns`, a per-turn timeout, or Ctrl-C. After a normal
+loop also stops on `--turns`, a per-turn timeout, or Ctrl-C. With
+`--on-turn-timeout continue` (the review recipe's default) a single turn
+timeout is handed to the partner instead of ending the run; two consecutive
+timeouts still stop. Without `--timeout`, the per-turn budget scales with the
+reasoning level (900s → 1800s at `xhigh`/`max`). After a normal
 stop, duet opens a `force>` prompt so you can push another round.
 
 Every run writes a directory with `transcript.md`, `state.json`, per-turn
 stderr logs, and the `wt/` worktree when `--worktree` is on. Inspect a run with
-`duet --status <run-id>`, list runs with `duet --list`, and start a fresh run
+`duet --status <run-id> --json` (or omit `--json` for the human view), list runs
+with `duet --list`, and start a fresh run
 from saved state with `duet --continue <run> --task "next thing"`.
 
 - **Backends:** `claude`, `codex`, `gemini`, `copilot`, `opencode`
 - **Roles:** `planner`, `coder`, `reviewer`, `triage-reviewer`, or a custom one
 - **Reasoning:** `--reasoning minimal|low|medium|high|xhigh|max`
 
+## Limits / future
+
+Timeout continuation applies only to automatic loop turns. Seed extraction
+and forced turns still stop on timeout. A final-turn timeout also stops because
+no partner turn remains, and two consecutive timeouts stop the run. Poll
+`duet --status <run-id> --json` to watch the current budget instead of assuming
+a deep-reasoning turn is stuck.
+
 ## Documentation
 
 [docs/USAGE.md](https://github.com/volkan/duet/blob/main/docs/USAGE.md) is the

{duet_cli-0.2.8 → duet_cli-0.2.10}/README.md

@@ -26,7 +26,9 @@ Plain `/duet` runs Claude Code's real `/review`, then loops Codex and Claude in
 a worktree until they converge. Pass any upstream command as the kickoff:
 `/duet 'npm test 2>&1' --turns 4`. The plugin shells out to the `duet` CLI, so
 install that first (see below) and make sure `command -v duet` passes in Claude
-Code's shell. Full guide:
+Code's shell. Custom topology flags such as `--no-worktree` and
+`--allow-worktree-fallback` replace the plugin defaults without emitting
+conflicting CLI flags. Full guide:
 [docs/CLAUDE_CODE_PLUGIN.md](https://github.com/volkan/duet/blob/main/docs/CLAUDE_CODE_PLUGIN.md).
 If Claude Code disambiguates plugin commands with namespaces, `/duet:duet` is
 the same command.
@@ -87,8 +89,10 @@ Then invoke `/duet` in the OpenCode TUI (or `opencode run --command duet "..."`
 non-interactively). Like the other plugins it shells out to the `duet` CLI, so
 install that first and make sure `command -v duet` passes in OpenCode's shell.
 The command runs on OpenCode's `build` agent; plain `/duet` runs the same
-`claude -p /review` kickoff, and `/duet 'npm test 2>&1' --turns 4` seeds from
-any command. Full guide:
+`claude -p /review --model sonnet` kickoff, and
+`/duet 'npm test 2>&1' --turns 4` seeds from any command. Custom worktree
+overrides are constructed conditionally, matching the Claude and Codex entry
+points. Full guide:
 [docs/OPENCODE_PLUGIN.md](https://github.com/volkan/duet/blob/main/docs/OPENCODE_PLUGIN.md).
 (duet can also drive OpenCode as a backend — `--partner opencode:coder` — so
 OpenCode can be one of the two looped agents too.)
@@ -134,9 +138,28 @@ duet --task "Review the latest commit; fix only what the reviewer requests." \
 `/review`, a test run, or any command:
 
 ```bash
-duet --task-from-cmd 'claude -p /review' \
-    --lead claude:reviewer --partner codex:coder \
-    --worktree --recap --cwd ~/workspace/project --turns 6
+duet --recipe review --cwd ~/workspace/project
+```
+
+The review recipe uses recap mode, `claude:reviewer` + `codex:coder`, six
+turns, `claude -p /review --model sonnet`, and strict worktree isolation.
+Claude agents use the stable `sonnet` alias unless an explicit slot model
+overrides it. The recipe also sets
+`--on-turn-timeout continue`, so when the coder turn times out the reviewer
+still gets the timeout block plus the worktree diff to review instead of the
+run dying reviewless. Explicit flags override recipe values.
+
+**Deterministic automation** — discover the run without parsing terminal prose,
+then poll the curated status schema:
+
+```bash
+control_dir=$(mktemp -d)
+duet --recipe review --run-info-file "$control_dir/run.json" &
+duet_pid=$!
+while [ ! -s "$control_dir/run.json" ]; do sleep 0.1; done
+run_dir=$(python3 -c 'import json,sys; print(json.load(open(sys.argv[1]))["run_dir"])' "$control_dir/run.json")
+duet --status "$run_dir" --json
+wait "$duet_pid"
 ```
 
 **Deep planner, fast coder** — Claude plans at high effort while Codex coder
@@ -181,18 +204,31 @@ other.
 To converge, an agent must include an `LGTM rationale:` explaining why the work
 is done, followed by the sentinel `<<<LGTM>>>` on its own line — a bare
 sentinel is ignored, and **both** agents must agree in back-to-back turns. The
-loop also stops on `--turns`, a per-turn timeout, or Ctrl-C. After a normal
+loop also stops on `--turns`, a per-turn timeout, or Ctrl-C. With
+`--on-turn-timeout continue` (the review recipe's default) a single turn
+timeout is handed to the partner instead of ending the run; two consecutive
+timeouts still stop. Without `--timeout`, the per-turn budget scales with the
+reasoning level (900s → 1800s at `xhigh`/`max`). After a normal
 stop, duet opens a `force>` prompt so you can push another round.
 
 Every run writes a directory with `transcript.md`, `state.json`, per-turn
 stderr logs, and the `wt/` worktree when `--worktree` is on. Inspect a run with
-`duet --status <run-id>`, list runs with `duet --list`, and start a fresh run
+`duet --status <run-id> --json` (or omit `--json` for the human view), list runs
+with `duet --list`, and start a fresh run
 from saved state with `duet --continue <run> --task "next thing"`.
 
 - **Backends:** `claude`, `codex`, `gemini`, `copilot`, `opencode`
 - **Roles:** `planner`, `coder`, `reviewer`, `triage-reviewer`, or a custom one
 - **Reasoning:** `--reasoning minimal|low|medium|high|xhigh|max`
 
+## Limits / future
+
+Timeout continuation applies only to automatic loop turns. Seed extraction
+and forced turns still stop on timeout. A final-turn timeout also stops because
+no partner turn remains, and two consecutive timeouts stop the run. Poll
+`duet --status <run-id> --json` to watch the current budget instead of assuming
+a deep-reasoning turn is stuck.
+
 ## Documentation
 
 [docs/USAGE.md](https://github.com/volkan/duet/blob/main/docs/USAGE.md) is the