fix: clean up stale modes and fix auto-routing

factory/agents/prompts/ceo.md

@@ -44,11 +44,9 @@ You communicate directly with the user when running in foreground mode. You expl
 
 **You MUST complete ALL planned work before exiting.** This applies to every mode:
 
-- **Build mode:** All phases (B0–B6) must be attempted
-- **Improve mode:** Every approved hypothesis must have a Builder keep/revert verdict
-- **Discover mode:** The eval profile must be generated
+- **Design mode:** All phases must be attempted (build, discover, and improve are handled inline)
 - **Research mode:** Every approved hypothesis must have a verdict, or a termination condition must be met
-- **Meta mode:** Same as Improve, plus ACE playbook evolution
+- **Meta mode:** Same as Design, plus ACE playbook evolution
 
 **Self-judged early exits are FORBIDDEN.** Do not exit because:
 - "This is a good stopping point" — there are no stopping points, only completion
@@ -309,24 +307,20 @@ factory detect "$PROJECT_PATH"
 
 | State                  | Meaning                                       | Route to       |
 |------------------------|-----------------------------------------------|----------------|
-| `no_repo`              | No git repo at path                           | Build mode     |
-| `incomplete`           | Repo exists, open plan/implementation issues  | Build mode     |
-| `no_factory`           | Repo exists, no factory setup                 | Discover mode  |
-| `evals_pending_review` | Eval profile exists, not yet reviewed         | Review mode    |
-| `has_factory`          | Factory fully initialized, evals reviewed     | Improve mode   |
+| `no_repo`              | No git repo at path                           | Design mode    |
+| `incomplete`           | Repo exists, open plan/implementation issues  | Design mode    |
+| `no_factory`           | Repo exists, no factory setup                 | Design mode    |
+| `evals_pending_review` | Eval profile exists, not yet reviewed         | Design mode    |
+| `has_factory`          | Factory fully initialized, evals reviewed     | Design mode    |
 
 ### Step 2: Route to Mode via Skills
 
 Each mode's full instructions live in a workflow skill under `skills/workflow-<name>/SKILL.md`. After detecting project state, select and invoke the appropriate skill.
 
 **Default routing:**
-- `no_repo` or `incomplete` → read `skills/workflow-build/SKILL.md`
-- `no_factory` → read `skills/workflow-discover/SKILL.md`
-- `evals_pending_review` → read `skills/workflow-review/SKILL.md`
-- `has_factory` → read `skills/workflow-improve/SKILL.md`
+- All project states → read `skills/workflow-design/SKILL.md`
 
 **Mode overrides (from task directives):**
-- `--mode design` or `## Plan Loop (Interactive)` → read `skills/workflow-design/SKILL.md`
 - `--mode research` (with `research_target` configured) → read `skills/workflow-research/SKILL.md`
 - `--mode meta` → read `skills/workflow-meta/SKILL.md`
 - `--refine "<request>"` → read `skills/workflow-refine/SKILL.md`

factory/ceo_completion.py

@@ -292,66 +292,32 @@ def _detect_incomplete(
         cycle_started_at: If provided, only counts experiments created after this time.
             This prevents counting stale experiments from previous cycles.
     """
-    if mode in ("improve", "meta", "research"):
+    if mode in ("design", "meta", "research"):
         planned = _count_hypotheses(project_path)
         completed = _count_verdicts(project_path, since_ts=cycle_started_at)
 
         if planned == 0:
-            # No strategy yet — not an incomplete cycle, probably discover mode
-            return None
-
-        if completed >= planned:
-            return None
-
-        next_h = completed + 1
-        reason_prefix = "research" if mode == "research" else "improve"
-        return IncompleteGap(
-            mode=mode,
-            planned=planned,
-            completed=completed,
-            next_item=f"H{next_h}",
-            reason=f"{reason_prefix}.incomplete: {completed}/{planned} hypotheses have verdicts",
-        )
-
-    elif mode == "discover":
-        if _has_eval_profile(project_path):
-            return None
-        return IncompleteGap(
-            mode=mode,
-            planned=1,
-            completed=0,
-            next_item="eval_profile",
-            reason="discover.incomplete: no eval_profile.json",
-        )
-
-    elif mode == "build":
-        # For build mode, check if Builder completed at least one hypothesis
-        # In build mode, the strategy file should have phases marked as hypotheses
-        planned = _count_hypotheses(project_path)
-        completed = _count_verdicts(project_path, since_ts=cycle_started_at)
-
-        if planned == 0:
-            # No strategy means we're in scaffold phase — check for eval profile
-            if not _has_eval_profile(project_path):
+            if mode == "design" and not _has_eval_profile(project_path):
                 return IncompleteGap(
                     mode=mode,
                     planned=1,
                     completed=0,
                     next_item="discovery",
-                    reason="build.incomplete: no eval profile yet",
+                    reason="design.incomplete: no eval profile yet",
                 )
             return None
 
         if completed >= planned:
             return None
 
         next_h = completed + 1
+        reason_prefix = "research" if mode == "research" else "design"
         return IncompleteGap(
             mode=mode,
             planned=planned,
             completed=completed,
-            next_item=f"Phase{next_h}",
-            reason=f"build.incomplete: {completed}/{planned} phases have verdicts",
+            next_item=f"H{next_h}",
+            reason=f"{reason_prefix}.incomplete: {completed}/{planned} hypotheses have verdicts",
         )
 
     # Unknown mode — assume complete
@@ -386,26 +352,14 @@ def _build_continuation_task(gap: IncompleteGap, cycle_state: CycleState | None
             f"research cycle (R3–R5) for each remaining hypothesis. "
             f"Progress so far: {gap.completed}/{gap.planned} hypotheses have verdicts."
         )
-    elif gap.mode in ("improve", "meta"):
+    elif gap.mode in ("design", "meta"):
         body = (
-            f"Resume execution from hypothesis {gap.next_item}. "
+            f"Resume execution from {gap.next_item}. "
             f"Strategy is already approved at .factory/strategy/current.md — "
             f"do not re-plan, do not re-run Researcher or Strategist. "
             f"Spawn Builder for {gap.next_item} immediately. "
             f"Progress so far: {gap.completed}/{gap.planned} hypotheses have verdicts."
         )
-    elif gap.mode == "build":
-        body = (
-            f"Resume Build pipeline from {gap.next_item}. "
-            f"Plan is already approved at .factory/strategy/current.md. "
-            f"Progress so far: {gap.completed}/{gap.planned} phases complete. "
-            f"Continue with the next phase immediately."
-        )
-    elif gap.mode == "discover":
-        body = (
-            "Resume Discovery. The eval profile has not been generated yet. "
-            "Complete the Discover mode workflow to produce .factory/eval_profile.json."
-        )
     else:
         body = f"Resume from {gap.next_item}."
 

factory/checkpoint.py

@@ -36,6 +36,7 @@ class CheckpointState(BaseModel):
 def save_checkpoint(project_path: Path, state: CheckpointState) -> None:
     """Serialize checkpoint state to .factory/checkpoint.json."""
     from factory.store import ensure_factory_dir
+
     factory_dir = project_path / ".factory"
     ensure_factory_dir(factory_dir)
     checkpoint_path = factory_dir / _CHECKPOINT_FILE
@@ -45,6 +46,8 @@ def save_checkpoint(project_path: Path, state: CheckpointState) -> None:
 
 def load_checkpoint(project_path: Path) -> CheckpointState | None:
     """Load checkpoint from .factory/checkpoint.json, or None if absent/corrupt."""
+    from factory.cli._helpers import DEAD_MODES
+
     checkpoint_path = project_path / ".factory" / _CHECKPOINT_FILE
     if not checkpoint_path.exists():
         log.debug("checkpoint.not_found", path=str(checkpoint_path))
@@ -55,6 +58,10 @@ def load_checkpoint(project_path: Path) -> CheckpointState | None:
     except (json.JSONDecodeError, Exception) as exc:
         log.warning("checkpoint.corrupt", path=str(checkpoint_path), error=str(exc))
         return None
+    if state.mode in DEAD_MODES:
+        old_mode = state.mode
+        state.mode = DEAD_MODES[old_mode]
+        log.warning("checkpoint.mode_migrated", old=old_mode, new=state.mode)
     log.info("checkpoint.loaded", path=str(checkpoint_path))
     return state
 

factory/cli/_ceo_helpers.py

@@ -57,6 +57,7 @@ def _tool_exec_protocol(wt_path: Path) -> str:
     overview = ""
     try:
         from factory.workflow.tool import tool_overview
+
         overview = tool_overview(p, fmt="linear")
     except Exception:
         pass
@@ -69,11 +70,7 @@ def _tool_exec_protocol(wt_path: Path) -> str:
     )
 
     if overview:
-        protocol += (
-            "\n## Workflow Map\n"
-            "\n"
-            f"{overview}\n"
-        )
+        protocol += f"\n## Workflow Map\n\n{overview}\n"
 
     protocol += (
         "\n## Commands\n"
@@ -90,20 +87,20 @@ def _tool_exec_protocol(wt_path: Path) -> str:
         '1. Run "next" to see your current task — it tells you the node type, '
         "role, and what to do\n"
         "2. Execute the task:\n"
-        "   - Agent nodes: run factory agent <role> --task \"...\" --project <path>\n"
+        '   - Agent nodes: run factory agent <role> --task "..." --project <path>\n'
         "   - Study nodes: run the study command shown\n"
         "   - Function nodes: run the command shown\n"
         '3. Run "next" again — the tool auto-detects that the previous node completed\n'
         "   (by checking for output files) and advances to the next task\n"
         "4. Repeat until GATE or DONE\n"
         "5. For GATE nodes: the tool asks you to evaluate — read the artifacts, then\n"
         '   call "submit" with your verdict (PROCEED, RETRY, or HALT)\n'
-        "6. If RETRY: the tool rewinds — run \"next\" to get the retry task\n"
+        '6. If RETRY: the tool rewinds — run "next" to get the retry task\n'
         "7. If DONE: report completion\n"
         "\n"
         "## Important\n"
         "\n"
-        "- For most nodes, just run the command and call \"next\" — the tool handles tracking\n"
+        '- For most nodes, just run the command and call "next" — the tool handles tracking\n'
         '- Only call "submit" for gate verdicts (PROCEED/RETRY/HALT)\n'
         "- The tool auto-detects agent completion via .factory/reviews/ files\n"
         "- The tool auto-evaluates fn gates (precheck, guard) on your behalf\n"
@@ -144,16 +141,32 @@ def _tool_exec_protocol(wt_path: Path) -> str:
 
 def _validate_ceo_flags(
     args: argparse.Namespace,
-) -> tuple[str, bool, bool, bool, str | None, str | None, str | None, str | None, bool, str | None, bool] | int:
+) -> (
+    tuple[
+        str,
+        bool,
+        bool,
+        bool,
+        str | None,
+        str | None,
+        str | None,
+        str | None,
+        bool,
+        str | None,
+        bool,
+    ]
+    | int
+):
     """Validate and resolve top-level CLI flags. Returns parsed values or an error code."""
     mode: str = getattr(args, "mode", "auto")
     if mode == "interactive":
         mode = "design"
     if mode.startswith("project:"):
-        mode = mode[len("project:"):]
+        mode = mode[len("project:") :]
     all_modes = get_all_ceo_modes()
     if mode not in all_modes and mode != "auto":
         from factory.workflow.registry import WorkflowRegistry
+
         raw_path = getattr(args, "path", None)
         project_path = Path(raw_path).resolve() if raw_path else Path.cwd()
         entries = WorkflowRegistry.discover(project_path)
@@ -209,6 +222,7 @@ def _validate_ceo_flags(
     raw_path = getattr(args, "path", None)
     if not raw_path:
         from factory.plugins import get_registry
+
         plugin_registry = get_registry()
         has_pre_hooks = bool(plugin_registry.ceo_pre_hooks)
         if not has_pre_hooks:
@@ -294,7 +308,19 @@ def _validate_ceo_flags(
         )
         return 1
 
-    return (mode, headless, bg, bg_agents, prompt_file, focus, dir_name, refine_request, auto_approve, from_plan, just_plan)
+    return (
+        mode,
+        headless,
+        bg,
+        bg_agents,
+        prompt_file,
+        focus,
+        dir_name,
+        refine_request,
+        auto_approve,
+        from_plan,
+        just_plan,
+    )
 
 
 # ── project resolution ────────────────────────────────────────
@@ -454,9 +480,23 @@ def _validate_late_flags(
         )
         return 1
 
-    if focus and mode not in ("improve", "research", "create", "evolve", "study", "frontend-design", "frontend-design-discover") and not design_existing and not just_plan:
+    if (
+        focus
+        and mode
+        not in (
+            "design",
+            "research",
+            "create",
+            "evolve",
+            "study",
+            "frontend-design",
+            "frontend-design-discover",
+        )
+        and not design_existing
+        and not just_plan
+    ):
         print(
-            f"Error: --focus (targeted mode) only works in improve, research, create, evolve, study, frontend-design, "
+            f"Error: --focus (targeted mode) only works in design, research, create, evolve, study, frontend-design, "
             f"frontend-design-discover, or design (with --just-plan) mode, "
             f"got '{mode}'. The project must already be built before targeting specific items.",
             file=sys.stderr,
@@ -602,7 +642,7 @@ def _execute_ceo(
     elif mode == "design":
         ceo_mode = "design"
     elif interactive:
-        ceo_mode = "build"
+        ceo_mode = "design"
     else:
         ceo_mode = mode
 
@@ -746,12 +786,18 @@ def _execute_ceo(
 
         if engine == "tool":
             base_prompt = resolve_prompt(
-                "ceo", wt_path, use_profile=use_profile, workflow_mode=None,
+                "ceo",
+                wt_path,
+                use_profile=use_profile,
+                workflow_mode=None,
             )
             prompt = base_prompt + _tool_exec_protocol(wt_path)
         else:
             prompt = resolve_prompt(
-                "ceo", wt_path, use_profile=use_profile, workflow_mode=ceo_mode,
+                "ceo",
+                wt_path,
+                use_profile=use_profile,
+                workflow_mode=ceo_mode,
             )
         runner = get_runner(runner_name)
         extras: dict[str, object] = {}
@@ -776,6 +822,7 @@ def _execute_ceo(
         if engine == "tool":
             try:
                 from factory.workflow.tool import tool_finalize
+
                 finalize_result = tool_finalize(wt_path)
                 log.info("tool_exec.finalized", result=finalize_result)
             except Exception:
@@ -847,13 +894,18 @@ def _run_headless(
         executor = WorkflowExecutor(wf, wt_path, agent_pool=DEFAULT_AGENT_POOL)
         try:
             exec_result = asyncio.run(executor.execute())
-            print(json.dumps({
-                "workflow": ceo_mode,
-                "engine": "deterministic",
-                "success": exec_result.success,
-                "nodes_executed": exec_result.nodes_executed,
-                "duration_ms": round(exec_result.duration_ms, 1),
-            }, indent=2))
+            print(
+                json.dumps(
+                    {
+                        "workflow": ceo_mode,
+                        "engine": "deterministic",
+                        "success": exec_result.success,
+                        "nodes_executed": exec_result.nodes_executed,
+                        "duration_ms": round(exec_result.duration_ms, 1),
+                    },
+                    indent=2,
+                )
+            )
             code = 0 if exec_result.success else 1
             if code != 0:
                 return code
@@ -863,7 +915,7 @@ def _run_headless(
                 min_growth=min_growth,
                 max_new=max_new,
                 branch=branch,
-                already_improved=mode in ("improve", "meta") or discover_only,
+                already_improved=mode in ("design", "meta") or discover_only,
                 model=model,
                 no_github=no_github,
                 use_profile=use_profile,
@@ -916,7 +968,7 @@ def _run_headless(
             min_growth=min_growth,
             max_new=max_new,
             branch=branch,
-            already_improved=mode in ("improve", "meta") or discover_only,
+            already_improved=mode in ("design", "meta") or discover_only,
             model=model,
             no_github=no_github,
             use_profile=use_profile,
@@ -929,6 +981,7 @@ def _run_headless(
         if engine == "tool":
             try:
                 from factory.workflow.tool import tool_finalize
+
                 tool_finalize(wt_path)
             except Exception:
                 pass