@polderlabs/bizar-omp 0.6.1 → 0.6.3

package/docs/releases/0.6.3.md

@@ -0,0 +1,27 @@
+# 0.6.3
+
+## Highlights
+
+- `omb stop` is the single teardown path for everything Bizar owns: Bizar-named tmux sessions, the daemon, and the tray. Each target can be named on its own (`--sessions`, `--session NAME`, `--daemon`/`--dashboard`, `--tray`) or stopped together (`--all`, the default); `--all` with `--session NAME` narrows the session selection to the sessions named while the daemon and the tray are still stopped.
+- Every `omb stop` row is reported as `stopped`, `skipped` with a reason, `planned` for a dry run, or `failed`, both in the human summary and in `--json`, and the exit code is nonzero only when a resource failed. `omb restart` rows also use `started` for a resource it stopped and started again. Stopping something that is not running stays a success.
+- `omb restart` stops the whole runtime and starts it again: the daemon, the tray only when hub settings enable it, and a fresh single-use dashboard link. It reports one row per resource — each session, the daemon, the tray, and the dashboard — refuses every `omb stop` target flag including `--idle` with exit code 2, and `--dry-run` reports that same one-row shape with action `planned`. `--json` always carries `url`: the new dashboard link, or `null` when none was minted.
+- `omb help` lists the commands and `omb help <command>` prints one command's own usage, and `--help` works on every command. Usage text now lives beside the command that owns it, so the help table cannot drift from the implementation.
+- The OMP pass-through contract is unchanged and now stated in the help itself: anything `omb` does not recognise is handed to OMP, so `omb --resume`, `omb --session`, and a bare prompt keep their meaning. `help`, `--help`, and `-h` are omb's own tokens: `omb --help` and `omb -h` print omb's table, and `omp --help` prints OMP's.
+
+## Safety
+
+- Only Bizar-owned sessions are ever signalled, and the daemon and the tray are signalled only while the recorded lock's PID is verified to still be that process, so a lock that cannot be verified is skipped instead. An address that is not a Bizar-owned name, not a session id this host knows, and not a digest of a discovered session is refused with exit code 2 and nothing is killed.
+- An address that matches more than one session is refused with exit code 2 before anything is stopped, so a short id shared by two sessions can never end the wrong work.
+- `omb stop` never stops the session it is running inside, even with `--force`, so the client asking for the stop is not disconnected from its own report.
+- `--idle` stops only sessions with no attached client and no work in flight, so a detached session that is still working survives.
+- `--force` is required to stop a session a client is attached to, and `--dry-run` performs no kills, no daemon stop, and no tray stop.
+- A stop is reported as `stopped` only after the session is confirmed absent from tmux. A `kill-session` the multiplexer refuses is reported as `failed` with the reason, instead of as a stop that did not happen.
+- An unhosted session that still has a live journal is reported rather than signalled, because there is no tmux session behind it to kill.
+- A Bizar-owned name that is not running is reported `not found` and is a success, so running the same `omb stop` twice exits 0 both times.
+
+## Verification
+
+- 38 Vitest cases cover `omb help`, `omb stop`, and `omb restart` against a stateful fake multiplexer and injected host services: the help table against the command registry, the per-command usage, the README's own invocations against that usage, the unknown-command path, `--json` payloads, name/id/digest addressing, the ambiguous-address refusal, attached/`--idle`/`--force`/`--dry-run` selection, self-protection, unhosted journals, a refused kill, a refused daemon stop, restart ordering, tray settings, and the refusal of `stop` target flags on `restart`.
+- `tests/tray.test.ts` covers the tray half of the ownership rule: a tray lock whose PID is not the tray process, and a tray whose recorded lock is verified.
+- 215 Vitest tests across 28 files, TypeScript typecheck, build, package verification, generated-doc check, and secret scan.
+- Live smoke on Linux against real tmux servers and a scratch state home: `omb help`, `omb stop --dry-run`, `omb stop` twice (the second run reports nothing left to stop), `omb stop --session <unowned>` and `omb stop --session <shared id>` (both exit 2, nothing killed), `omb stop --session <digest>` (stops exactly one, leaves a non-Bizar session alone), an attached session skipped with its reason and then stopped with `--force`, `omb restart` (fresh access code exchanged for a token, single-use refused on re-exchange, authenticated session list returned), and `omb stop --daemon` followed immediately by `omb daemon status` reporting a stopped daemon with no stale lock.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@polderlabs/bizar-omp",
-  "version": "0.6.1",
+  "version": "0.6.3",
   "description": "Native oh-my-pi engineering workflow extension with typed evidence and serialized integration.",
   "type": "module",
   "engines": {

package/skills/bizar-omp/SKILL.md

@@ -36,7 +36,8 @@ Escalate when the scope grows, a check fails, the cause is uncertain, generated
 
 - `/bizar run <objective>` starts a bounded Bizar workflow.
 - `/bizar status`, `/bizar next`, `/bizar evidence`, and `/bizar cancel` inspect or advance that workflow.
-- `/bizar-models` scans OMP's live authenticated model catalog locally, assigns the best available models to every `bizar_*` role, validates native resolution, preserves unrelated roles, and reports a compact mapping. It does not start a planning/research agent turn. Never invent a provider/model selector.
+- `/bizar-models` scans OMP's live authenticated model catalog locally, reports what changed since the last observation, and hands the catalog, the current mapping, and a deterministic baseline to the active agent, which decides the distribution and persists it with the native `bizar_apply_model_roles` tool. That tool validates each selector against the live catalog and writes the whole mapping or nothing, preserving unrelated roles. A host without agent messaging applies the deterministic baseline itself. Never invent a provider/model selector.
+- `/bizar-models preview` (alias `dry-run`) scores the live catalog and prints the deterministic baseline plus catalog changes without messaging an agent and without writing settings.
 - `/bizar-models health` validates configured Bizar selectors without changing settings.
 
 The semantic roles include `bizar_orchestrator` for the main conversation and `bizar_advisor` for independent passive review, alongside coordination, common work, hard work, research, planning, architecture, implementation, review, security, verification, and documentation. OMP's native Advisor is enabled by default and resolves through the native `advisor` role; Bizar exposes `bizar_advisor` as a visible, assignable alias. Use the role that matches the work; do not assume a role is available until OMP resolves its selector.