@polderlabs/bizar 10.15.0 → 10.16.0

package/AGENTS.md

@@ -31,6 +31,21 @@ make session-end              # lifecycle compatibility target
 - **MUST NOT** use a persistent Claude daemon. Claude Code and the Agent SDK run in-process; background work uses Claude Code's Agent tool.
 - **MUST NOT** rebase or force-push under the default project policy.
 
+### Worktree discipline
+
+Every code-writing Agent call dispatched by an orchestrator MUST pass
+`isolation: "worktree"`. Read-only agents stay foreground. Dispatched
+worktrees branch from HEAD as `wt/<agent_type>-<short-task-id>`. The
+SubagentStart hook (`config/claude/hooks/worktree-bootstrap.mjs`) bootstraps
+the worktree; the SubagentStop hook (`config/claude/hooks/worktree-archive.mjs`)
+records the branch into `~/.config/bizar/worktree-queue.json`. The
+orchestrator merges the queue with `bizar worktree-merge --all` (or one
+branch at a time with `bizar worktree-merge <branch>`); the sequencer
+tags each source tip as `merge-archive/<branch>-<sha>` before
+`git merge --no-ff` and surfaces conflicts instead of silently dropping
+work. Conflicts MUST be reported to the user for resolution; never silently
+force a resolution you do not understand.
+
 ## Autonomy and parallelism
 
 Agents execute clear, local, reversible work autonomously — they inspect,
@@ -43,14 +58,18 @@ still deny prohibited actions and escalate externally visible or irreversible
 actions with `permissionDecision: "ask"`; that escalation list is the
 authoritative floor, not a starting point.
 
-Subagent dispatch through the Agent tool is the default for focused disjoint
-work. Native dynamic workflows under `~/.claude/workflows/` are used when a
-large or repeatable task benefits from deterministic scripted fan-out. Native
-agent teams are used only when independent workers need direct communication or
-shared task state. When two or more subtasks have non-overlapping file scopes
-and no data dependency on each other's intermediate output, the orchestrator
-MUST dispatch them concurrently. Sequential dispatch is reserved for dependent
-phases and integration.
+Native dynamic workflows under `config/workflows/` and `~/.claude/workflows/`
+are the primary dispatch mechanism for non-trivial tasks. Mike dispatches a
+named workflow when the request maps to a research / implement / debug /
+review shape. For work that needs 3+ long-lived workers with bounded cross-
+talk, Mike invokes a workflow that fans out as a native agent team; the team
+is host-side state under `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`, and per
+Anthropic's docs `team_name` is deprecated and ignored. Plain `Agent` calls
+are reserved for trivial, single-shot, or fully isolated work. When two or
+more subtasks within a workflow have non-overlapping file scopes and no data
+dependency on each other's intermediate output, the orchestrator MUST dispatch
+them concurrently through `parallel([...])`. Sequential dispatch is reserved
+for dependent phases and integration.
 
 Agent roles are model-agnostic. Mike selects the cheapest sufficient tier for
 each dispatch. It passes a concrete model only when live discovery proves a
@@ -58,11 +77,51 @@ configured tier candidate; otherwise it omits `model` and inherits the active
 session. Bizar never retries a failed dispatch by cycling aliases, providers, or
 tiers.
 
-The authoritative hard approval list (cannot be auto-approved) is: commits, pushes, pull-request
+The authoritative hard approval list (cannot be auto-approved) is: pushes, pull-request
 mutations, releases, package publication, deployments, production/shared-
 infrastructure writes, credential changes, public exposure, irreversible
 destruction. Everything else proceeds.
 
+Under F-176 the seven-category floor above is enforced by `permission-request.mjs`
+(the destructive subset: force-push, rebase, root deletion, system-destructive
+commands) and surfaced for the rest via `git-workflow-guard.mjs` as advisory
+reminders in `hookSpecificOutput.additionalContext`. `permissions.deny` and
+`permissions.ask` are emptied by design — the floor is enforced by hook output,
+not by Claude Code prompts.
+
+> Note: `config/claude/settings.json` ships `permissions.ask` moved into
+> `permissions.allow` so subagents do not prompt for commits, pushes, PRs,
+> or deploys. Operators who want HITL back can move the following exact
+> patterns from `permissions.allow` into `permissions.ask` in their local
+> `~/.claude/settings.json` override:
+>
+> - `Bash(git push *)`
+> - `Bash(git -C * push *)`
+> - `Bash(git --git-dir=* push *)`
+> - `Bash(git push --force *)`
+> - `Bash(git push -f *)`
+> - `Bash(git rebase *)`
+> - `Bash(gh pr create *)`
+> - `Bash(gh pr merge *)`
+> - `Bash(gh release create *)`
+> - `Bash(npm publish *)`
+> - `Bash(bun publish *)`
+> - `Bash(pnpm publish *)`
+> - `Bash(vercel deploy *)`
+> - `Bash(wrangler deploy *)`
+> - `Bash(flyctl deploy *)`
+
+> Note: `disableAutoCompact: true` is shipped by default. Sessions rely on
+> manual `/compact` instead. Hook `precompact-priorities.sh` preserves
+> evidence and decisions on compaction.
+Local `git commit` is always allowed silently via `permissions.allow`; the seven
+HITL categories above are gated by `permission-request.mjs` plus the advisory
+hook chain.
+
+Agents always fetch current official documentation via WebSearch + WebFetch at
+task start and whenever uncertainty appears during work. Guess-and-try is
+prohibited.
+
 ## Execution model
 
 The autonomy and approval policy above governs this execution model. The project defaults to `acceptEdits`; eligible operators may opt into Claude Code Auto mode.

package/cli/bin.mjs

@@ -416,14 +416,37 @@ async function main() {
       break;
     }
 
+    case 'models': {
+      const mod = await importCommand('models');
+      if (!mod) {
+        console.error(chalk.red(`  ✗ Could not load models command module`));
+        process.exit(EXIT_ERROR);
+        return;
+      }
+      dbg('loaded command module:', 'models');
+      const found = await mod.run(cmd, cmdArgs, isHelpRequest);
+      if (found === false) {
+        console.error(chalk.red(`  ✗ Usage: bizar models [--list|--set|--clear|--json] — run 'bizar models --help'`));
+        process.exit(EXIT_USAGE);
+      }
+      break;
+    }
+
     case 'model': {
+      // Deprecated alias. Routes to the original `model.mjs` so the
+      // legacy JSON shape (`{ providers: { ... }, total: N }`) and table
+      // output keep working for existing scripts and tests. New code
+      // should use `bizar models` (see `case 'models'`).
       const mod = await importCommand('model');
       if (!mod) {
         console.error(chalk.red(`  ✗ Could not load model command module`));
         process.exit(EXIT_ERROR);
         return;
       }
-      dbg('loaded command module:', 'model');
+      dbg('loaded command module (deprecated alias):', 'model');
+      if (!isHelpRequest) {
+        console.error(chalk.yellow('  ! `bizar model` is deprecated; use `bizar models` for the new picker surface.'));
+      }
       const found = await mod.run(cmd, cmdArgs, isHelpRequest);
       if (found === false) {
         console.error(chalk.red(`  ✗ Usage: bizar model <subcommand> — run 'bizar model --help'`));

package/cli/commands/hook.mjs

@@ -47,6 +47,7 @@ export const HOOK_PROGRAMS = Object.freeze({
   'team-lifecycle': 'team-lifecycle.mjs',
   'verify-deliverables': 'verify-deliverables.mjs',
   'worker-suggest': 'worker-suggest.mjs',
+  'worktree-archive': 'worktree-archive.mjs',
   'worktree-bootstrap': 'worktree-bootstrap.mjs',
 });
 
@@ -81,7 +82,7 @@ export const EVENT_CHAINS = Object.freeze({
     'advisor-context',
     'worktree-bootstrap',
   ]),
-  'subagent-stop': Object.freeze(['verify-deliverables']),
+  'subagent-stop': Object.freeze(['verify-deliverables', 'worktree-archive']),
   'task-created': Object.freeze(['team-lifecycle']),
   'task-completed': Object.freeze(['team-lifecycle']),
   'teammate-idle': Object.freeze(['team-lifecycle']),