Feat: Agent model settings with three quality presets

.gitignore

@@ -47,3 +47,4 @@ temp/
 docs/maestro/
 docs/superpowers/
 hooks/permissions.json
+.worktrees

claude/src/platforms/gemini/commands/maestro/orchestrate.toml

@@ -18,10 +18,12 @@ This preamble maps generic step references to Gemini CLI tool syntax.
 | Load template/reference/protocol | `get_skill_content(resources: ["<name>"])` |
 | Delegate to agent | Call agent tool by name: `coder(query: "...")`, `tester(query: "...")`, `design_system_engineer(query: "...")` |
 | MCP tools | `mcp_maestro_<tool_name>` (Gemini CLI also accepts bare names like `resolve_settings`) |
-| Enter Plan Mode | `enter_plan_mode` — if unavailable, tell user: "Run `gemini --settings` and set `experimental.plan` to `true`, then restart." Offer to continue without Plan Mode using `ask_user` for approvals. |
+| Enter Plan Mode | `enter_plan_mode` — if unavailable, tell user: "Run `gemini --settings` and set `experimental.plan` to `true`, then Request Restart." Offer to continue without Plan Mode using `ask_user` for approvals. |
 | Exit Plan Mode | `exit_plan_mode` with `plan_filename` |
 | User prompt (choose) | `ask_user` with `type: 'choice'` |
 | User prompt (approve) | `ask_user` with `type: 'yesno'` |
+| Define Agent Models | `ask_user` for valid mode (`quality`, `balanced`, `economic`, `skip`) and call `setup_models` with the selected mode. Then Request Restart. |
+| Request Restart | STOP here and wait for the user to restart and signal to continue. Do NOT proceed. |
 
 ## Execute
 

claude/src/references/orchestration-steps.md

@@ -8,13 +8,14 @@ STARTUP (Turn 1 — tool calls only, no text output)
  6. STOP. Turn 1 is ONLY steps 1-5. No text, no design questions, no file reads.
 
 CLASSIFICATION (Turn 2)
- 7. Load the architecture reference: ["architecture"]. Do NOT load templates yet — they are loaded at their consumption points (steps 13, 15, 20).
+ 7. Load the architecture reference: ["architecture"]. Do NOT load templates yet — they are loaded at their consumption points (steps 14, 16, 21).
  8. Classify task as simple/medium/complex. Present classification with rationale.
- 9. Route: simple → Express (step 31). Medium/complex → continue to step 10.
+ 9. Route: simple → Express (step 32). Medium/complex → continue to step 10.
 
 DESIGN (Phase 1)
 10. Enter Plan Mode. If unavailable, follow the runtime preamble's Plan Mode fallback instructions.
-11. Call `get_skill_content` with resources: ["design-dialogue"]. Follow the loaded protocol for:
+11. Define Agent Models.
+12. Call `get_skill_content` with resources: ["design-dialogue"]. Follow the loaded protocol for:
     - Design depth selector (first design question)
     - Repository grounding (for existing codebases, skip for greenfield)
     - One question at a time via user prompt
@@ -31,51 +32,51 @@ DESIGN (Phase 1)
     WRONG: user requests "fan site" → options include React, Next.js, Astro
     CORRECT: user requests "fan site" → recommended option is vanilla HTML/CSS/JS
     </ANTI-PATTERN>
-12. Present design sections one at a time, per the design-dialogue skill's convergence protocol.
+13. Present design sections one at a time, per the design-dialogue skill's convergence protocol.
     <HARD-GATE>
     Each section must be presented individually and approved via user prompt before
     proceeding to the next. Do NOT present the full design as a single block.
     Quick depth may combine sections. Standard/Deep MUST validate individually.
     </HARD-GATE>
-13. Call `get_skill_content` with resources: ["design-document"]. Write approved design document to <state_dir>/plans/ (or Plan Mode tmp path).
-14. If Plan Mode is active, exit Plan Mode with the plan path. Copy approved document to <state_dir>/plans/.
+14. Call `get_skill_content` with resources: ["design-document"]. Write approved design document to <state_dir>/plans/ (or Plan Mode tmp path).
+15. If Plan Mode is active, exit Plan Mode with the plan path. Copy approved document to <state_dir>/plans/.
 
 PLANNING (Phase 2)
-15. Call `get_skill_content` with resources: ["implementation-planning", "implementation-plan"]. Follow the loaded skill protocol.
-16. Call validate_plan with the generated plan and task_complexity.
+16. Call `get_skill_content` with resources: ["implementation-planning", "implementation-plan"]. Follow the loaded skill protocol.
+17. Call validate_plan with the generated plan and task_complexity.
     <HARD-GATE>
     You MUST call validate_plan BEFORE presenting the plan for approval. Do NOT
-    present the plan, write it to state_dir, or proceed to step 17 without first
+    present the plan, write it to state_dir, or proceed to step 18 without first
     calling validate_plan and resolving any error-severity violations.
     validate_plan enforces server-side: phase count limits, dependency cycles,
     unknown agents, file ownership conflicts, and agent-deliverable compatibility
     (read-only agents cannot be assigned to file-creating phases). If it returns
     violations with severity "error", fix them in the plan and re-validate.
     </HARD-GATE>
-17. Present plan for user approval (Approve / Revise / Abort via user prompt).
-18. Write approved implementation plan to <state_dir>/plans/.
+18. Present plan for user approval (Approve / Revise / Abort via user prompt).
+19. Write approved implementation plan to <state_dir>/plans/.
 
 EXECUTION SETUP (Phase 3 — pre-delegation)
-19. Call `get_skill_content` with resources: ["execution"]. Follow its Execution Mode Gate.
+20. Call `get_skill_content` with resources: ["execution"]. Follow its Execution Mode Gate.
     <HARD-GATE>
     Present ONLY "Parallel" and "Sequential" as execution mode options.
     Do NOT present "Ask" as a user-facing choice — "ask" is a setting value
     that means "prompt the user", not an execution mode the user selects.
     </HARD-GATE>
-20. Call `get_skill_content` with resources: ["session-management", "session-state"].
-21. Create session via create_session with resolved execution_mode. Do NOT create before mode is resolved.
-22. Call `get_skill_content` with resources: ["delegation", "validation", "agent-base-protocol", "filesystem-safety-protocol"].
+21. Call `get_skill_content` with resources: ["session-management", "session-state"].
+22. Create session via create_session with resolved execution_mode. Do NOT create before mode is resolved.
+23. Call `get_skill_content` with resources: ["delegation", "validation", "agent-base-protocol", "filesystem-safety-protocol"].
 
 EXECUTION (Phase 3 — delegation loop)
-23. For each phase (or parallel batch): call `get_agent` for the assigned agent, then delegate using the returned methodology and tool restrictions.
+24. For each phase (or parallel batch): call `get_agent` for the assigned agent, then delegate using the returned methodology and tool restrictions.
     <HARD-GATE>
     Dispatch by calling the agent's registered tool directly.
     Do NOT use the built-in generalist tool or invoke agents by bare name.
     Each Maestro agent carries specialized methodology, tool restrictions, temperature,
     and turn limits from its frontmatter that the generalist ignores.
     </HARD-GATE>
-24. After each agent returns, parse Task Report + Downstream Context from response.
-25. Call transition_phase to persist results.
+25. After each agent returns, parse Task Report + Downstream Context from response.
+26. Call transition_phase to persist results.
     <HARD-GATE>
     For parallel batches: call transition_phase INDIVIDUALLY for EVERY completed
     phase in the batch. The MCP tool writes files_created, files_modified,
@@ -85,73 +86,74 @@ EXECUTION (Phase 3 — delegation loop)
     merge all agents' files into one call — the archive attributes files per
     phase, so empty payloads mean lost traceability.
     </HARD-GATE>
-26. Repeat steps 23-25 until all phases complete.
+27. Repeat steps 24-26 until all phases complete.
 
 COMPLETION (Phase 4)
-27. Call `get_skill_content` with resources: ["code-review"].
-28. If execution changed non-documentation files, delegate to the code reviewer agent. Block on Critical/Major findings.
+28. Call `get_skill_content` with resources: ["code-review"].
+29. If execution changed non-documentation files, delegate to the code reviewer agent. Block on Critical/Major findings.
     <HARD-GATE>
     If Critical/Major findings: re-delegate to the implementing agent to fix.
     The orchestrator MUST NOT write code directly.
     </HARD-GATE>
-29. If MAESTRO_AUTO_ARCHIVE is true (or unset), call archive_session. If false, inform user session is complete but not archived.
-30. Present final summary with files changed, phase outcomes, and next steps.
+30. If MAESTRO_AUTO_ARCHIVE is true (or unset), call archive_session. If false, inform user session is complete but not archived.
+31. Present final summary with files changed, phase outcomes, and next steps.
 
 RECOVERY (referenced from any step on user request)
 If the user says the flow moved too fast: return to the most recent unanswered approval gate.
 If the user asks for implementation before approval: remind them Maestro requires approval first.
 If the user asks to skip execution-mode: remind them parallel/sequential is required unless MAESTRO_EXECUTION_MODE pins it.
 If an answer invalidates a prior choice: restate the updated assumption and re-run the relevant gate.
-If delegation collapses to parent session without fallback approval: return to step 19 or re-scope the child-agent work packages.
+If delegation collapses to parent session without fallback approval: return to step 20 or re-scope the child-agent work packages.
 
 EXPRESS WORKFLOW (simple tasks only — jumped to from step 9)
 
 EXPRESS MODE GATE BYPASS: Express bypasses the execution-mode gate entirely. Express always dispatches sequentially. Do NOT prompt for parallel/sequential.
 
 EXPRESS MCP FALLBACK: If MCP state tools (create_session, transition_phase, archive_session) are unavailable, fall back to direct file writes on <state_dir>/state/active-session.md.
 
-31. Verify classification is simple. If task requires multiple phases or agents, override to medium → step 10.
+32. Verify classification is simple. If task requires multiple phases or agents, override to medium → step 10.
     <HARD-GATE>
     Express sessions MUST have exactly one implementation phase with exactly one agent.
     </HARD-GATE>
-32. Ask 1-2 clarifying questions from Area 1 only.
+33. Define Agent Models.
+34. Ask 1-2 clarifying questions from Area 1 only.
     <HARD-GATE>
     Each question MUST use the user prompt tool (not plain text). Use the choose
     variant with 2-4 options where possible. Do NOT ask questions as plain text
     in the model response — the user prompt tool is the only input mechanism.
     </HARD-GATE>
-33. Present structured Express brief as plain text, then ask for approval.
+35. Present structured Express brief as plain text, then ask for approval.
     <HARD-GATE>
     The brief MUST be plain text output in the model response.
     The approval MUST be a SEPARATE user prompt tool call — not embedded in the
     brief text. The prompt contains only: "Approve this Express brief to proceed?"
     These are two distinct actions: first emit the brief as text, then call the
     user prompt tool for approval. Do NOT combine them into one text block.
     </HARD-GATE>
-34. On approval, create session with workflow_mode: "express", exactly 1 phase.
+36. On approval, create session with workflow_mode: "express", exactly 1 phase.
     On rejection, revise. On second rejection, escalate to Standard → step 10.
-35. Call `get_skill_content` with resources: ["agent-base-protocol", "filesystem-safety-protocol"] and prepend them to the delegation prompt.
-36. Delegate to the assigned agent.
+37. Call `get_skill_content` with resources: ["agent-base-protocol", "filesystem-safety-protocol"] and prepend them to the delegation prompt.
+38. Delegate to the assigned agent.
     <HARD-GATE>
-    Same dispatch rule as step 23: call agent by registered tool name, not generalist.
+    Same dispatch rule as step 24: call agent by registered tool name, not generalist.
     </HARD-GATE>
-37. Parse Task Report from the agent's response. Call transition_phase to persist results.
+39. Parse Task Report from the agent's response. Call transition_phase to persist results.
     <HARD-GATE>
     You MUST call transition_phase after the implementing agent returns. Extract
     files_created, files_modified, files_deleted, and downstream_context from the
     Task Report and pass them to transition_phase. Without this call, the session
     state has no record of what was delivered. Do NOT skip to code review or archive
     without calling transition_phase first.
     </HARD-GATE>
-38. Delegate to the code reviewer agent.
+40. Delegate to the code reviewer agent.
     <HARD-GATE>
     If Critical/Major findings: re-delegate to implementing agent (1 retry).
     Orchestrator MUST NOT write code directly. If retry fails, escalate to user.
     </HARD-GATE>
-39. Call archive_session.
-40. Present summary.
+41. Call archive_session.
+42. Present summary.
 
 EXPRESS RESUME (when resuming an Express session from get_session_status)
-If phase is pending: re-generate and present brief (step 33). On approval, proceed to delegation (step 36).
-If phase is in_progress: re-delegate with same scope (step 36).
-If phase is completed but session is in_progress: run code review (step 38), then archive (step 39).
+If phase is pending: re-generate and present brief (step 35). On approval, proceed to delegation (step 38).
+If phase is in_progress: re-delegate with same scope (step 38).
+If phase is completed but session is in_progress: run code review (step 40), then archive (step 41).

commands/maestro/orchestrate.toml

@@ -18,10 +18,12 @@ This preamble maps generic step references to Gemini CLI tool syntax.
 | Load template/reference/protocol | `get_skill_content(resources: ["<name>"])` |
 | Delegate to agent | Call agent tool by name: `coder(query: "...")`, `tester(query: "...")`, `design_system_engineer(query: "...")` |
 | MCP tools | `mcp_maestro_<tool_name>` (Gemini CLI also accepts bare names like `resolve_settings`) |
-| Enter Plan Mode | `enter_plan_mode` — if unavailable, tell user: "Run `gemini --settings` and set `experimental.plan` to `true`, then restart." Offer to continue without Plan Mode using `ask_user` for approvals. |
+| Enter Plan Mode | `enter_plan_mode` — if unavailable, tell user: "Run `gemini --settings` and set `experimental.plan` to `true`, then Request Restart." Offer to continue without Plan Mode using `ask_user` for approvals. |
 | Exit Plan Mode | `exit_plan_mode` with `plan_filename` |
 | User prompt (choose) | `ask_user` with `type: 'choice'` |
 | User prompt (approve) | `ask_user` with `type: 'yesno'` |
+| Define Agent Models | `ask_user` for valid mode (`quality`, `balanced`, `economic`, `skip`) and call `setup_models` with the selected mode. Then Request Restart. |
+| Request Restart | STOP here and wait for the user to restart and signal to continue. Do NOT proceed. |
 
 ## Execute
 

plugins/maestro/src/platforms/gemini/commands/maestro/orchestrate.toml

@@ -18,10 +18,12 @@ This preamble maps generic step references to Gemini CLI tool syntax.
 | Load template/reference/protocol | `get_skill_content(resources: ["<name>"])` |
 | Delegate to agent | Call agent tool by name: `coder(query: "...")`, `tester(query: "...")`, `design_system_engineer(query: "...")` |
 | MCP tools | `mcp_maestro_<tool_name>` (Gemini CLI also accepts bare names like `resolve_settings`) |
-| Enter Plan Mode | `enter_plan_mode` — if unavailable, tell user: "Run `gemini --settings` and set `experimental.plan` to `true`, then restart." Offer to continue without Plan Mode using `ask_user` for approvals. |
+| Enter Plan Mode | `enter_plan_mode` — if unavailable, tell user: "Run `gemini --settings` and set `experimental.plan` to `true`, then Request Restart." Offer to continue without Plan Mode using `ask_user` for approvals. |
 | Exit Plan Mode | `exit_plan_mode` with `plan_filename` |
 | User prompt (choose) | `ask_user` with `type: 'choice'` |
 | User prompt (approve) | `ask_user` with `type: 'yesno'` |
+| Define Agent Models | `ask_user` for valid mode (`quality`, `balanced`, `economic`, `skip`) and call `setup_models` with the selected mode. Then Request Restart. |
+| Request Restart | STOP here and wait for the user to restart and signal to continue. Do NOT proceed. |
 
 ## Execute