Add per-role reasoning effort, default to medium, move Claude roles to Opus 5.5

.ai/allowlist.json

@@ -25,22 +25,33 @@
   "codex_reasoning_effort": "medium",
   "models": {
     "planner": "gpt-6-astra",
-    "plan-reviewer-a": "claude-opus-5",
+    "plan-reviewer-a": "claude-opus-5-5",
     "plan-reviewer-b": "gpt-6-astra",
-    "arbiter": "claude-opus-5",
+    "arbiter": "claude-opus-5-5",
     "builder": "gpt-6-astra",
-    "code-reviewer-a": "claude-opus-5",
+    "code-reviewer-a": "claude-opus-5-5",
     "code-reviewer-b": "gpt-6-astra",
-    "review-arbiter": "claude-opus-5",
+    "review-arbiter": "claude-opus-5-5",
     "fixer": "gpt-6-astra"
   },
+  "effort": {
+    "planner": "high",
+    "plan-reviewer-a": "medium",
+    "plan-reviewer-b": "medium",
+    "arbiter": "high",
+    "builder": "high",
+    "code-reviewer-a": "high",
+    "code-reviewer-b": "high",
+    "review-arbiter": "high",
+    "fixer": "medium"
+  },
   "quest_startup": {
     "branch_mode": "branch",
     "branch_prefix": "quest/",
     "worktree_root": ".worktrees/quest"
   },
   "quest_id_format": "slug-first",
-  "review_mode": "full",
+  "review_mode": "auto",
   "fast_review_thresholds": {
     "max_files": 5,
     "max_loc": 300
@@ -253,7 +264,7 @@
     "require_approval_before_commit": true,
     "require_approval_before_push": true,
     "require_approval_before_delete": true,
-    "max_plan_iterations": 4,
+    "max_plan_iterations": 3,
     "max_fix_iterations": 3
   }
 }

.ai/quest.md

@@ -146,6 +146,7 @@ When a role invocation fails (missing/unparsable handoff), Quest uses a three-ti
 The Creator controls quest permissions via `.ai/allowlist.json`:
 - `auto_approve_phases` — which phases need human approval
 - `models` — default role model map (`planner`, `plan-reviewer-a`, `plan-reviewer-b`, `builder`, `code-reviewer-a`, `code-reviewer-b`, `arbiter`, `fixer`)
+- `effort` — reasoning effort per role, applied on whichever runtime the role's model selects. Run `scripts/quest_sync_model_defaults.py` after editing `models` or `effort`; the generator is what keeps the four dispatch surfaces in step.
 - `review_mode` — `auto` (default), `fast`, or `full` for Codex reviews
 - `fast_review_thresholds` — file/LOC thresholds for auto fast mode
 

.ai/schemas/allowlist.schema.json

@@ -45,7 +45,18 @@
     "codex_reasoning_effort": {
       "type": "string",
       "enum": ["low", "medium", "high", "xhigh", "max", "ultra"],
-      "description": "Reasoning effort pinned for Codex roles in new quests. Must be supported by the selected model and dispatch surface."
+      "description": "Legacy fallback: reasoning effort for Codex roles absent from the per-role `effort` map. New quests should use `effort`."
+    },
+    "effort": {
+      "type": "object",
+      "propertyNames": {
+        "enum": ["planner", "plan-reviewer-a", "plan-reviewer-b", "arbiter", "builder", "code-reviewer-a", "code-reviewer-b", "review-arbiter", "fixer"]
+      },
+      "additionalProperties": {
+        "type": "string",
+        "enum": ["low", "medium", "high", "xhigh", "max", "ultra"]
+      },
+      "description": "Reasoning effort per role, applied on whichever runtime the role's model selects. `ultra` is Codex-only and blocks a Claude-backed role at dispatch."
     },
     "models": {
       "type": "object",

.claude/agents/arbiter.md

@@ -3,6 +3,7 @@ name: arbiter
 description: Gatekeeper for plan quality. Synthesizes reviews, filters noise, decides if plan is ready for implementation.
 tools: Read, Glob, Grep, Write
 model: inherit
+effort: high
 ---
 
 You are the Arbiter Agent in a quest orchestration system.

.claude/agents/builder.md

@@ -3,6 +3,7 @@ name: builder
 description: Implements approved plans. Writes code, runs tests, produces PR description.
 tools: Read, Glob, Grep, Write, Edit, Bash
 model: inherit
+effort: high
 ---
 
 You are the Builder Agent in a quest orchestration system.

.claude/agents/code-reviewer.md

@@ -3,6 +3,7 @@ name: code-reviewer
 description: Reviews code changes for correctness, quality, security, and adherence to project patterns.
 tools: Read, Glob, Grep, Write
 model: inherit
+effort: high
 ---
 
 You are the Code Review Agent in a quest orchestration system.

.claude/agents/fixer.md

@@ -3,6 +3,7 @@ name: fixer
 description: Fixes issues identified by code review. Applies targeted fixes and re-runs tests.
 tools: Read, Glob, Grep, Write, Edit, Bash
 model: inherit
+effort: medium
 ---
 
 You are the Fixer Agent in a quest orchestration system.

.claude/agents/plan-reviewer.md

@@ -3,6 +3,7 @@ name: plan-reviewer
 description: Reviews implementation plans for feasibility, completeness, and alignment with acceptance criteria.
 tools: Read, Glob, Grep, Write
 model: inherit
+effort: medium
 ---
 
 You are a Plan Review Agent in a quest orchestration system.

.claude/agents/planner.md

@@ -3,6 +3,7 @@ name: planner
 description: Creates and refines implementation plans from quest briefs. Writes plans to .quest/ folder.
 tools: Read, Glob, Grep, Write, Edit, Bash
 model: inherit
+effort: high
 ---
 
 You are the Planner Agent in a quest orchestration system.

.claude/agents/review-arbiter.md

@@ -3,6 +3,7 @@ name: review-arbiter
 description: Impartial judge for the code-review phase. Adjudicates the two code-reviewer slot findings against the diff and emits the canonical review_findings.json.
 tools: Read, Glob, Grep, Write
 model: inherit
+effort: high
 ---
 
 You are the Review Arbiter Agent in a quest orchestration system.

.opencode/opencode.json

@@ -41,7 +41,7 @@
     "plan-reviewer-a": {
       "description": "Reviews implementation plans (Reviewer A - local subagent)",
       "mode": "subagent",
-      "model": "opencode/claude-opus-5",
+      "model": "opencode/claude-opus-5-5",
       "prompt": "{file:agents/plan-reviewer.md}",
       "permission": {
         "edit": {
@@ -67,7 +67,7 @@
     "arbiter": {
       "description": "Synthesizes dual reviews, renders APPROVE or ITERATE verdict",
       "mode": "subagent",
-      "model": "opencode/claude-opus-5",
+      "model": "opencode/claude-opus-5-5",
       "prompt": "{file:agents/arbiter.md}",
       "permission": {
         "edit": {
@@ -105,7 +105,7 @@
     "code-reviewer-a": {
       "description": "Reviews code changes (Reviewer A - local subagent)",
       "mode": "subagent",
-      "model": "opencode/claude-opus-5",
+      "model": "opencode/claude-opus-5-5",
       "prompt": "{file:agents/code-reviewer.md}",
       "permission": {
         "edit": {
@@ -141,7 +141,7 @@
     "review-arbiter": {
       "description": "Impartial judge for the code-review phase; adjudicates A/B findings against the diff and emits canonical review_findings.json",
       "mode": "subagent",
-      "model": "opencode/claude-opus-5",
+      "model": "opencode/claude-opus-5-5",
       "prompt": "{file:agents/review-arbiter.md}",
       "permission": {
         "edit": {

.skills/quest/SKILL.md

@@ -236,7 +236,8 @@ Before creating the quest folder, present the routing classification to the user
    - `claude_transport_resolved`: the `transport` field from the Step 2b preflight result (Codex-led sessions; `null` otherwise)
    - `claude_transport_downgraded`: compatibility field; write `false` for new runs (Codex-led preflight also emits `false`)
    - `codex_auth_mode`: persist the resolved Step 2b `checks.codex_auth_mode`, not a later allowlist re-read; pass it as `codex_auth_mode` to `write_default_from_allowlist`. In Codex-led startup retain the explicitly resolved mode for any later Claude-led resume; native Codex dispatch continues using the host identity.
-   - `codex_reasoning_effort`: copy the optional allowlist value verbatim and pass it to `write_default_from_allowlist`; validate with `validate_codex_reasoning_effort`. Omit it for older allowlists with no setting. Show the selected effort beside the model table before accepting defaults.
+   - `effort`: `.ai/allowlist.json` `.effort` passed verbatim to `write_default_from_allowlist`, which fills omitted roles from the shipped `DEFAULT_EFFORT`. Show the per-role effort beside the model table before accepting defaults.
+   - `codex_reasoning_effort`: legacy fallback for allowlists with no `.effort` map — copy the optional value verbatim and pass it through; validate with `validate_codex_reasoning_effort`. Omit it when unset.
    - `source: "default"`
    - `overridden_roles: []`
    - `preflight_validated_at: <ISO8601 now>`
@@ -271,7 +272,7 @@ Before creating the quest folder, present the routing classification to the user
    - `models`: merged block (all 9 keys present; unused-in-mode roles still carry the default value)
    - `claude_role_transport` / `claude_transport_resolved`: same sourcing as the Y path above; `claude_transport_downgraded: false` for compatibility
    - `codex_auth_mode`: same resolved Step 2b mode as the Y path; pass it to `write_orchestration_json`.
-   - `codex_reasoning_effort`: same allowlist sourcing as the Y path; pass it to `write_orchestration_json`. Model-only overrides do not change effort; validate model support before dispatch.
+   - `effort` / `codex_reasoning_effort`: same allowlist sourcing as the Y path; pass both to `write_orchestration_json`. Model-only overrides do not change effort; validate model support before dispatch.
    - `source: "overridden"`
    - `overridden_roles`: list of role names that were actually overridden
    - `preflight_validated_at: <ISO8601 now>`

.skills/quest/delegation/workflow.md

@@ -38,14 +38,14 @@ Quest dispatch separates **runtime** from **entrypoint**:
 | Orchestrator | Selected role runtime | Entrypoint | Rule |
 |--------------|-----------------------|------------|------|
 | Codex-led | Codex | local Codex subagent (the `spawn_agent` tool family — versioned namespace such as `multi_agent_v2` varies by Codex CLI release — or repo-supported equivalent) | Use the saved model and Codex effort according to the explicit-selection contract below. Do not use Codex MCP. |
-| Codex-led | Claude | `python3 scripts/quest_claude_runner.py` when `claude_transport_available` is true | The runner owns the transport underneath: background-agent (`scripts/quest_claude_bg_run.py`, `claude --bg`, subscription billing) when preflight proved it, or the bridge (`scripts/quest_claude_bridge.py`, `claude --print`) only when bridge was explicitly configured/selected. Pass `--model <models.<role> from .quest/<id>/orchestration.json>` and `--transport <claude_transport_resolved from orchestration.json>`. The exact `claude` model sentinel means use the Claude CLI/account default and must not be sent to the CLI as `--model claude`; concrete configured model strings pass through unchanged. Block with transport guidance if unavailable and no explicit Codex fallback exists. |
+| Codex-led | Claude | `python3 scripts/quest_claude_runner.py` when `claude_transport_available` is true | The runner owns the transport underneath: background-agent (`scripts/quest_claude_bg_run.py`, `claude --bg`, subscription billing) when preflight proved it, or the bridge (`scripts/quest_claude_bridge.py`, `claude --print`) only when bridge was explicitly configured/selected. Pass `--model <models.<role> from .quest/<id>/orchestration.json>` and `--transport <claude_transport_resolved from orchestration.json>`. The runner resolves `effort.<role>` from the same saved file and passes `claude --effort`; do not pass `--effort` yourself unless deliberately overriding the saved config. The exact `claude` model sentinel means use the Claude CLI/account default and must not be sent to the CLI as `--model claude`; concrete configured model strings pass through unchanged. Block with transport guidance if unavailable and no explicit Codex fallback exists. |
 | Claude-led | Codex | `python3 <installation-root>/scripts/quest_codex_runner.py role` when `codex_available` is true | The runner launches installed `codex exec`, reads saved model, effort and auth, and validates current-attempt artifacts. |
-| Claude-led | Claude | native `Task(...)` | Use the orchestrator's native Claude task path. |
+| Claude-led | Claude | native `Task(...)` | Use the orchestrator's native Claude task path. Native tasks cannot read `orchestration.json`, so do not try to pass effort here; the only available control is the subagent's `effort:` frontmatter, generated from the allowlist by `scripts/quest_sync_model_defaults.py`. **Unverified:** on Claude Code 2.1.280 that key could not be shown to change behavior (an invalid value raises no warning, and low-vs-max showed no token separation), unlike `claude --effort`, which measurably does. Treat effort on this path as best-effort; route through the runner when it must be guaranteed. |
 | Either orchestrator | Antigravity | `python3 scripts/quest_antigravity_runner.py` when `antigravity_available` is true | Selected by Gemini-family model IDs. Antigravity is never an orchestrator, only ever a dispatched runtime, so one runner serves both session types — there is no MCP path and no transport choice. Pass `--model <models.<role> from .quest/<id>/orchestration.json>`; the exact `gemini` sentinel means use the agy default model and must not be sent to the CLI as `--model gemini`. **Always pass `--add-dir` covering the quest directory** — see the Antigravity containment rule below. Block with the preflight `warning` lines if unavailable. |
 
-**Explicit model and effort selection:** Before every Codex role dispatch, read `models.<role>` and optional `codex_reasoning_effort` from the active quest's `orchestration.json`. Do not substitute defaults from this skill, the GPT skill, or the current allowlist. For local subagents, pass the exact `model` and, when set, `reasoning_effort` through the tool's exposed controls. This repository configuration authorizes explicit selection. When the tool requires a fresh or bounded context fork for overrides, use that mode and include the role instructions and artifact paths in the prompt. If the controls are unavailable, inherit only after verifying the parent matches the saved model and any pinned effort; otherwise stop and report the mismatch. Never substitute MCP or nested `codex exec` for Codex-led dispatch.
+**Explicit model and effort selection:** Before every Codex role dispatch, read `models.<role>` and `effort.<role>` from the active quest's `orchestration.json` (quests written before the per-role map fall back to the legacy scalar `codex_reasoning_effort`). Do not substitute defaults from this skill, the GPT skill, or the current allowlist. For local subagents, pass the exact `model` and, when set, `reasoning_effort` through the tool's exposed controls. This repository configuration authorizes explicit selection. When the tool requires a fresh or bounded context fork for overrides, use that mode and include the role instructions and artifact paths in the prompt. If the controls are unavailable, inherit only after verifying the parent matches the saved model and any pinned effort; otherwise stop and report the mismatch. Never substitute MCP or nested `codex exec` for Codex-led dispatch.
 
-For Claude-led Codex calls, use the installed runner's `role` mode. It reads `models.<role>`, optional `codex_reasoning_effort` and `codex_auth_mode` (legacy default `cached`) from the saved orchestration file. Do not pass model, effort or auth overrides in role mode. Unsupported settings block dispatch. Log requested settings separately from effective settings; effective values require runtime metadata, otherwise record `unknown`.
+For Claude-led Codex calls, use the installed runner's `role` mode. It reads `models.<role>`, `effort.<role>` (falling back to the legacy `codex_reasoning_effort`) and `codex_auth_mode` (legacy default `cached`) from the saved orchestration file. Do not pass model, effort or auth overrides in role mode. Unsupported settings block dispatch. Log requested settings separately from effective settings; effective values require runtime metadata, otherwise record `unknown`.
 
 **Orchestration violation:** A Codex-led attempt to dispatch another Codex role through MCP or nested `codex exec` is an entrypoint violation. Use local Codex subagents with the saved model and effort. Missing native controls or mismatched parent settings block, with no CLI/MCP substitution or saved-setting rewrite.
 

scripts/quest_claude_bridge.py

@@ -77,6 +77,12 @@ def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
         default="",
         help="Optional Claude model override (passed through if provided)",
     )
+    parser.add_argument(
+        "--effort",
+        default="",
+        choices=["", "low", "medium", "high", "xhigh", "max"],
+        help="Optional reasoning effort; empty uses the CLI/session default",
+    )
     parser.add_argument(
         "--system-prompt",
         default="",
@@ -160,13 +166,16 @@ def run_claude(
     add_dirs: list[str],
     allowed_tools: str,
     disallowed_tools: str,
+    effort: str = "",
 ) -> dict[str, Any]:
     cmd = ["claude", "--print", prompt, "--output-format", output_format]
     # `claude` is the runtime-family sentinel meaning account-default model —
     # never a CLI model name (defense-in-depth mirroring the bg runner; the
     # quest layer normalizes it away, but direct callers reach here unfiltered).
     if model and model != "claude":
         cmd.extend(["--model", model])
+    if effort:
+        cmd.extend(["--effort", effort])
     if system_prompt:
         cmd.extend(["--system-prompt", system_prompt])
     if append_system_prompt:
@@ -240,6 +249,7 @@ def main(argv: list[str] | None = None) -> int:
         output_format=args.output_format,
         timeout=args.timeout,
         model=args.model,
+        effort=args.effort,
         system_prompt=args.system_prompt,
         append_system_prompt=args.append_system_prompt,
         permission_mode=args.permission_mode,

scripts/quest_claude_runner.py

@@ -20,6 +20,7 @@
 from pathlib import Path
 
 from quest_runtime.artifacts import expected_artifacts_for_role
+from quest_runtime.orchestration import effort_for_role
 from quest_runtime.claude_runner import (
     DEFAULT_BG_RUNNER_SCRIPT,
     DEFAULT_BRIDGE_SCRIPT,
@@ -79,6 +80,14 @@ def parse_args() -> argparse.Namespace:
         default=1800.0,
         help="Command timeout seconds (default: 1800)",
     )
+    parser.add_argument(
+        "--effort",
+        choices=["low", "medium", "high", "xhigh", "max"],
+        help=(
+            "Reasoning effort. Omit to resolve effort.<agent> from the quest's "
+            "orchestration.json (the normal path); pass only to override it."
+        ),
+    )
     parser.add_argument("--permission-mode", default="bypassPermissions")
     parser.add_argument(
         "--transport",
@@ -127,6 +136,27 @@ def parse_args() -> argparse.Namespace:
     return args
 
 
+def _resolve_effort(quest_dir: str, agent: str, override: str | None) -> str | None:
+    """Effort for this dispatch: explicit flag, else the saved per-role map.
+
+    Resolving from orchestration.json here (rather than trusting the
+    orchestrator to pass it) mirrors Codex role mode and keeps the saved quest
+    config authoritative. A quest with no orchestration.json yet resolves to
+    None, which leaves the CLI default in place.
+    """
+    if override:
+        return override
+    try:
+        saved = json.loads(
+            (Path(quest_dir) / "orchestration.json").read_text(encoding="utf-8")
+        )
+    except (OSError, json.JSONDecodeError):
+        return None
+    if not isinstance(saved, dict):
+        return None
+    return effort_for_role(saved, agent)
+
+
 def main() -> int:
     args = parse_args()
     transport = resolve_claude_transport(args.transport)
@@ -149,6 +179,9 @@ def main() -> int:
             agent=args.agent,
             artifact_subset=args.artifact_subset,
         )
+        effort = _resolve_effort(
+            args.quest_dir, args.agent, getattr(args, "effort", None)
+        )
     except ValueError as exc:
         payload = {
             "exit_code": 1,
@@ -184,6 +217,7 @@ def main() -> int:
         handoff_file=args.handoff_file,
         bridge_script=resolve_path(args.cwd, args.bridge_script),
         model=args.model,
+        effort=effort,
         timeout=args.timeout,
         permission_mode=args.permission_mode,
         artifact_paths=artifact_paths,