Fix: Make main_pm + task_type=code impossible

agents/prompts/roles/prompter.md

@@ -89,6 +89,12 @@ When you are scoped to a **MegaTask**, the CEO wants several distinct tasks work
 
 Over-declaring a surface is safe (the worst case is a task waits a little); under-declaring is not. You do **not** compute the order yourself — declare each surface honestly and the analyzer derives the waves. Present all the tasks in prose first (a short paragraph each), then call `propose_batch` once. If the conversation changes the set, call it again with the full updated batch.
 
+**Each draft in a MegaTask becomes a Main-PM coordination root-subtask** — the Main PM coordinates it and delegates the actual code to the cells; the Main PM never writes the code itself. So draft each root-subtask as the coordination it is, not as code the Main PM will implement:
+
+- `task_type`: `"planning"` (the system coerces `code`→`planning` for a Main-PM root anyway, but draft it correctly — a Main PM task is never `code`).
+- `acceptance_criteria`: **coordination-level**, not code-level. Write criteria the Main PM can satisfy by delegating and assembling — e.g. *"the chart-first Metrics refactor is delegated to fe-pm and lands on a cell PR"*, *"the cell→root PR is assembled and passes the in-path review gate"*, *"all cell subtasks are terminal and the root→master PR is merged"*. Do **not** write code-level criteria on the root — specific file paths (`frontend/src/components/timeseries-chart.tsx`), "lint/build clean", exact APIs — those belong on the **cell/dev subtasks** the Main PM delegates to, not on the root. A root carrying code-level ACs is the structural mismatch behind the 2026-06-27 meltdown (the gate reviewed code the Main PM couldn't fix → an infinite re-submit loop).
+- `the_work` still names the per-cell breakdown (which cell does what) — that's the delegation plan the Main PM executes; it's correct here because it *is* the coordination spec.
+
 ## What happens after you call `propose_draft`
 
 A draft card appears for the human with three choices: **Keep chatting**, **Board review & Start**, or **Approve & Start**. **Choosing is the human's action, not yours** — you cannot create, start, or route the task. If they pick **Board review & Start**, it becomes a pending task owned by the Board (Product Owner + Head of Marketing) to review first; if they pick **Approve & Start**, it becomes a pending task that goes straight to the Main PM to delegate to the cells. Either way, your job ends the moment you call `propose_draft`. Do not say you'll "kick it off", "send it to the PM chain", or route it anywhere — you have no such ability, and which path it takes is the human's choice on the card.

roboco/foundation/policy/batch.py

@@ -15,6 +15,8 @@ between sites. Inputs are typed ``object | None`` because callers pass either OR
 
 from __future__ import annotations
 
+from roboco.models.base import TaskType, Team
+
 
 def is_batch_umbrella(
     *, batch_id: object | None, parent_task_id: object | None
@@ -97,3 +99,27 @@ def is_valid_batch_shape(
         return targets == 0
     # root-subtask: exactly one target
     return targets == 1
+
+
+def main_pm_cannot_own_code(*, team: object | None, task_type: object | None) -> bool:
+    """True when a Main-PM-owned task must NOT be ``code``.
+
+    A Main PM *coordinates* — it plans and delegates execution to the cells; it
+    has no verb to write code. A ``task_type=code`` task owned by ``main_pm`` is
+    the structural mismatch behind the 2026-06-27 MegaTask meltdown: the git/PR/
+    review layer treated the Main-PM root-subtask as a code root (branch + PR +
+    ``submit_root`` + the ``pr_review`` gate) while the ownership/dispatch layer
+    treated it as coordination, and a code-style ``pr_fail`` landed on a
+    coordinator with no code verb → an infinite re-submit loop. The root's code
+    ACs were never decomposed to children because the roll-up gate is inert
+    unless a child declares ``parent_ac_refs``.
+
+    This predicate is the single invariant every layer consults so ``main_pm`` +
+    ``code`` can never coexist: the intake coercion (``create_task_from_draft``
+    retypes ``code``→``planning``), the ``TaskService.create`` backstop reject,
+    the reassign/escalation diversion, and the claim guard. Accepts either ORM
+    enum members or their ``.value`` strings (callers pass both shapes).
+    """
+    team_value = str(getattr(team, "value", team))
+    type_value = str(getattr(task_type, "value", task_type))
+    return team_value == Team.MAIN_PM.value and type_value == TaskType.CODE.value

roboco/foundation/policy/content/models.py

@@ -143,10 +143,19 @@ class AcVerdict(_Base):
 
 
 class PrReviewContent(_Content):
-    """A PR-review comment / reviewer verdict."""
+    """A PR-review comment / reviewer verdict.
+
+    ``findings`` carries the *structured* review (file/line/severity/expected/
+    actual) used by the inbound external-PR path. The in-path gate instead
+    fails on free-text ``issues`` — those land in the additive ``issues`` slot
+    rather than being flattened into ``summary`` alone, so a reader of
+    ``notes_structured.pr_review`` (or the derived ``pr_reviewer_notes`` mirror)
+    gets the concrete change-requests either way.
+    """
 
     summary: str
     findings: list[Finding] = Field(default_factory=list)
+    issues: list[str] = Field(default_factory=list)
     verdict: Verdict
 
     @field_validator("findings", mode="before")
@@ -154,6 +163,11 @@ class PrReviewContent(_Content):
     def _coerce_findings(cls, v: Any) -> Any:
         return coerce_to_list(v)
 
+    @field_validator("issues", mode="before")
+    @classmethod
+    def _coerce_issues(cls, v: Any) -> Any:
+        return coerce_to_list(v)
+
     @field_validator("summary")
     @classmethod
     def _nontrivial_summary(cls, v: str) -> str:
@@ -174,6 +188,8 @@ class PrReviewContent(_Content):
                     f"| {f.expected} → {f.actual} |"
                 )
             parts.append("## Findings\n" + "\n".join(rows))
+        if self.issues:
+            parts.append(_section("Issues", _bullets(self.issues)))
         parts.append(_section("Verdict", self.verdict.value.replace("_", " ")))
         return _join(parts)
 

roboco/foundation/policy/lifecycle.py

@@ -30,7 +30,7 @@ from typing import TYPE_CHECKING, Any, Literal
 # Re-exported here so callers can import `Role` from this module alongside
 # the lifecycle tables that depend on it. New consumers may also import
 # from `roboco.foundation.identity` directly.
-from roboco.foundation.identity import Role
+from roboco.foundation.identity import Role, Team
 
 if TYPE_CHECKING:
     from collections.abc import Callable
@@ -779,6 +779,27 @@ def _next_hint_dev_revise(_t: Any) -> str:
     return "idle - dev will revise and re-submit"
 
 
+def _next_hint_pr_fail(t: Any) -> str:
+    # Steer a pr_fail verdict by owner shape. A Main-PM branch-bearing root is
+    # an assembled cell→root / root→master PR — coordination, not the Main PM's
+    # own code. The rejection is about the cells' merged code, which the Main PM
+    # cannot fix directly (no code verb); it must re-delegate the fixes to the
+    # owning cell PM(s) and wait for re-assembly. Re-submitting the unchanged
+    # root is the 2026-06-27 infinite pr_fail loop. A cell/dev task is revised
+    # in place by its dev, so keep the dev-revise hint there.
+    team = getattr(t, "team", None)
+    team_value = str(getattr(team, "value", team))
+    branch = bool(getattr(t, "branch_name", None))
+    if team_value == Team.MAIN_PM.value and branch:
+        return (
+            "assembled cell work failed review — re-delegate the fixes to the"
+            " owning cell PM(s) via delegate(...), then wait for the cell"
+            " subtasks to complete and the PR to be re-assembled; do NOT"
+            " re-submit the root until then"
+        )
+    return "idle - dev will revise and re-submit"
+
+
 def _next_hint_doc_after_claim(_t: Any) -> str:
     return (
         "write docs in your workspace, commit them, then call"
@@ -1152,7 +1173,7 @@ _INTENT_VERBS: dict[str, IntentSpec] = {
         composes=("pr_fail",),
         extra_preconditions=(),
         side_effects=(),
-        next_hint=_next_hint_dev_revise,
+        next_hint=_next_hint_pr_fail,
     ),
     # Phase 3: documenter verbs
     "claim_doc_task": IntentSpec(
@@ -1244,8 +1265,10 @@ _INTENT_VERBS: dict[str, IntentSpec] = {
             "Main PM opens the root→master PR and moves the root task to"
             " awaiting_pr_review for the main reviewer (the root analogue of the"
             " cell PM's submit_up). After pr_pass, call complete to escalate to"
-            " the CEO. Only for code roots; branchless coordination roots skip"
+            " the CEO. For branch-bearing roots (a Main-PM root-subtask assembles"
-            " the gate and complete directly."
+            " the cells' merged work); branchless coordination roots skip the"
+            " gate and complete directly. The gate is branch-keyed, not"
+            " task_type-keyed — a Main-PM root is planning-typed, never code."
         ),
         composes=("submit_for_review",),
         extra_preconditions=(),

roboco/services/gateway/choreographer/_impl.py

@@ -5780,12 +5780,15 @@ class Choreographer:
                     main_pm_agent_id, root_task_id
                 ),
             )
-        # A code root must pass the in-path PR-review gate first: submit_root
+        # A branch-bearing root must pass the in-path PR-review gate first:
-        # opens the root→master PR and moves it in_progress → awaiting_pr_review,
+        # submit_root opens the root→master PR and moves it in_progress →
-        # then the main reviewer pr_passes it to awaiting_pm_review. So complete
+        # awaiting_pr_review, then the main reviewer pr_passes it to
-        # accepts only awaiting_pm_review for a code root. A branchless
+        # awaiting_pm_review. So complete accepts only awaiting_pm_review for a
-        # coordination root (product fan-out, no repo/PR) skips the gate, so it
+        # branch-bearing root (a Main-PM root-subtask is planning-typed, never
-        # may still be walked from in_progress here.
+        # code, but it still assembles the cells' merged work into a real PR).
+        # A branchless coordination root (product fan-out, no repo/PR) skips the
+        # gate, so it may still be walked from in_progress here. The split is
+        # branch-keyed, not task_type-keyed.
         root_is_branchless = not bool(t.branch_name)
         allowed_statuses = (
             ("awaiting_pm_review", "in_progress")

roboco/services/gateway/choreographer/pr_gate.py

@@ -245,7 +245,7 @@ class PRGateMixin(_Base):
         # what keeps notes_structured.pr_review in lock-step with the decision —
         # a later pr_fail overwrites an earlier pr_pass verdict instead of
         # leaving a stale "passed" on a task that was just sent back.
-        self._record_gate_verdict(t, verb, notes)
+        self._record_gate_verdict(t, verb, notes, issues=issues)
         runner = self._verb_runner()
         try:
             t = await runner.run_intent(verb, t, agent, spec_ctx)
@@ -265,6 +265,49 @@ class PRGateMixin(_Base):
         # transition — a GitHub failure must not roll back the gate decision.
         reviewer_slug = getattr(agent, "slug", None) or role_str
         await self._post_gate_review_to_pr(t, verb, reviewer_slug, notes)
+        # Deliver the change-requests to the owner that now has to act on them
+        # — the cell PM the runner just re-assigned via _revision_pm_for_task.
+        # The reviewer posts the verdict on the PR itself but that never reaches
+        # any PM-readable channel (no a2a, and _briefing_for / build_task_handoff
+        # read neither pr_reviewer_notes nor notes_structured.pr_review). Without
+        # this the owning PM respawned into needs_revision saw a generic "needs
+        # revision" with zero concrete issues, concluded nothing to rework, and
+        # re-submitted the same PR — an infinite pr_fail loop (live on
+        # 9980d0a0 / PR #138). Mirrors QA's fail_review a2a to the dev (qa.py:671).
+        # Best-effort: the transition already committed, so a delivery failure
+        # must not roll the verdict back or 500 the reviewer.
+        if verb == "pr_fail" and t.assigned_to is not None:
+            # A Main-PM branch-bearing root is an assembled cell→root / root→master
+            # PR — coordination, not the Main PM's own code. The rejection is
+            # about the cells' merged code, which the Main PM cannot fix directly
+            # (no code verb). Steer the a2a body to re-delegate + wait for
+            # re-assembly so the PM doesn't re-submit the unchanged root (the
+            # 2026-06-27 infinite pr_fail loop). The Envelope ``next`` hint makes
+            # the same steer via _next_hint_pr_fail.
+            team = getattr(t, "team", None)
+            team_value = str(getattr(team, "value", team))
+            is_main_pm_root = team_value == spec_module.Team.MAIN_PM.value and bool(
+                getattr(t, "branch_name", None)
+            )
+            steer = (
+                " Assembled cell work failed — re-delegate the fixes to the"
+                " owning cell PM(s) and wait for re-assembly; do NOT re-submit"
+                " the root."
+                if is_main_pm_root
+                else ""
+            )
+            try:
+                await self.a2a.send(
+                    from_agent=reviewer_agent_id,
+                    to_agent=t.assigned_to,
+                    skill="code_review",
+                    task_id=task_id,
+                    body=f"PR review needs changes. {notes}{steer}",
+                )
+            except Exception:
+                logger.exception(
+                    "pr_fail a2a to owning PM failed", task_id=str(task_id)
+                )
         return Envelope.ok(
             status=str(t.status),
             task_id=str(task_id),
@@ -302,27 +345,50 @@ class PRGateMixin(_Base):
                 )
         return None
 
-    def _record_gate_verdict(self, t: Any, verb: str, notes: str) -> None:
+    def _record_gate_verdict(
+        self, t: Any, verb: str, notes: str, issues: tuple[str, ...] = ()
+    ) -> None:
         """Persist the gate verdict as the canonical ``pr_review`` note.
 
         The tracing gate only threads ``notes`` through a throwaway shim, so
         nothing wrote the task's structured PR-reviewer slot — a task passed
         once and later failed kept showing the stale ``verdict: passed``. This
         authors the slot on every decision (``pr_pass`` → passed, ``pr_fail`` →
-        failed) so it can never contradict the transition. Best-effort: content
+        failed) so it can never contradict the transition. For ``pr_fail`` the
-        validation (e.g. a too-short summary) must never roll back the gate, so a
+        free-text ``issues`` land in the structured ``issues`` slot (not the
-        malformed payload is logged and skipped rather than raised.
+        format-enforced ``findings`` list, which needs file/severity/expected/
+        actual) so a reader of ``notes_structured.pr_review`` — or the owning
+        PM's briefing that mirrors it — gets the concrete change-requests.
+        Best-effort: content validation (e.g. a too-short summary) must never
+        roll back the gate, so a malformed payload is logged and skipped.
         """
         from roboco.foundation.policy.content import ContentValidationError
         from roboco.services.content_notes import apply_structured_note
 
         verdict = "passed" if verb == "pr_pass" else "failed"
-        try:
+        if verb == "pr_fail" and issues:
-            apply_structured_note(
+            # The free-text issues render under their own ``## Issues`` section
-                t,
+            # (render_markdown). Baking them into ``summary`` too duplicated each
-                "pr_review",
+            # issue on the Task Details "PR Reviewer Notes" card (once under
-                {"summary": notes, "findings": [], "verdict": verdict},
+            # ## Summary, once under ## Issues). The summary is a substantive
+            # non-issues sentence; ``notes`` (with the issues) still drives the
+            # GitHub PR post and the a2a to the owning PM — those are raw text,
+            # not rendered through render_markdown, so no duplication there.
+            summary = (
+                f"In-path PR-review gate requested changes - "
+                f"{len(issues)} issue(s) listed below."
             )
+        else:
+            summary = notes
+        payload: dict[str, Any] = {
+            "summary": summary,
+            "findings": [],
+            "verdict": verdict,
+        }
+        if issues:
+            payload["issues"] = list(issues)
+        try:
+            apply_structured_note(t, "pr_review", payload)
         except ContentValidationError:
             logger.warning(
                 "gate verdict note skipped (invalid content)",

roboco/services/gateway/choreographer/pr_review.py

@@ -160,6 +160,23 @@ class PRReviewerMixin(_Base):
         apply_structured_note(t, "pr_review", structured)
         return structured.render_markdown()
 
+    @staticmethod
+    def _is_hand_formatted_verdict(body: str) -> bool:
+        """True when a free-text ``body`` carries verdict/section markdown headers
+        the system would otherwise generate — i.e. the reviewer hand-formatted a
+        verdict into ``body`` instead of passing structured ``findings``.
+
+        Matches the section headers the canonical renderer emits (``## Findings``)
+        plus the ones a hand-formatter reaches for (``## Summary`` / ``## Issues``
+        / ``## Verdict``). A real one-paragraph summary does not contain ``## ``
+        headers, so the prose word "summary" never trips this.
+        """
+        lowered = (body or "").lower()
+        return any(
+            header in lowered
+            for header in ("## summary", "## issues", "## verdict", "## findings")
+        )
+
     async def _post_review_side_effects(
         self,
         t: Any,
@@ -219,11 +236,15 @@ class PRReviewerMixin(_Base):
         if isinstance(pre, Envelope):
             return pre
         agent, role_str, briefing, spec_ctx = pre
-        # Refuse a verdict that contradicts the findings BEFORE anything is
+        # Content gates BEFORE anything is recorded or posted to the
-        # recorded or posted to the contributor's PR (e.g. a forgotten
+        # contributor's PR: (1) refuse a verdict that contradicts the findings
-        # event='APPROVE' that defaults to a blocking REQUEST_CHANGES with no
+        # (a forgotten event='APPROVE' defaulting to a blocking REQUEST_CHANGES
-        # findings cited).
+        # with no findings); (2) refuse a hand-formatted verdict body with no
-        conflict = await self._verdict_consistency_gate(
+        # findings (the tool contract is "body = a one-paragraph summary; the
+        # system GENERATES the comment from structured findings — do not
+        # hand-format"). Folded into one helper so neither slips through and the
+        # verb body stays under the return-count lint ceiling.
+        rejection = await self._post_pr_review_content_gates(
             t,
             reviewer_agent_id,
             task_id,
@@ -231,9 +252,10 @@ class PRReviewerMixin(_Base):
             briefing,
             event=event,
             findings=findings,
+            body=body,
         )
-        if conflict is not None:
+        if rejection is not None:
-            return conflict
+            return rejection
         slug = await self._project_slug_for(t)
         pr_number = t.pr_number
         post_body = self._resolve_post_body(t, body, findings, event)
@@ -351,6 +373,72 @@ class PRReviewerMixin(_Base):
             verb="post_pr_review",
         )
 
+    async def _post_pr_review_content_gates(
+        self,
+        t: Any,
+        reviewer_agent_id: UUID,
+        task_id: UUID,
+        role_str: str,
+        briefing: dict[str, Any],
+        *,
+        event: str,
+        findings: list[dict[str, Any]] | None,
+        body: str,
+    ) -> Envelope | None:
+        """Pre-side-effect content gates for ``post_pr_review``: verdict
+        consistency, then the no-hand-formatted-body guard. Returns the first
+        rejection ``Envelope`` or ``None`` to proceed.
+
+        The hand-format guard: the tool contract is "``body`` = a one-paragraph
+        summary; the system GENERATES the GitHub comment from structured
+        findings — do not hand-format it in ``body``". Nothing enforced that, so
+        a reviewer could pass ``findings=[]`` and dump a self-formatted
+        ``## Summary`` / ``## Issues`` / ``## Verdict`` blob into ``body``,
+        which ``_resolve_post_body`` posts verbatim (the renderer emits
+        ``## Findings``, never ``## Issues`` — so a ``## Issues`` section on the
+        PR is proof the body was hand-formatted). Observed live: a duplicated,
+        self-redundant hand-formatted verdict posted to a contributor's PR.
+        Refuse it and point the reviewer at the structured path. Scoped to empty
+        findings: with structured findings the system generates the comment, so
+        a header-shaped word in the summary is harmless; a genuine plain-note
+        ``COMMENT`` (no verdict headers) is still allowed.
+        """
+        conflict = await self._verdict_consistency_gate(
+            t,
+            reviewer_agent_id,
+            task_id,
+            role_str,
+            briefing,
+            event=event,
+            findings=findings,
+        )
+        if conflict is not None:
+            return conflict
+        if not findings and self._is_hand_formatted_verdict(body):
+            return await self._emit_rejection(
+                Envelope.invalid_state(
+                    message=(
+                        "post_pr_review body is hand-formatted as a verdict — "
+                        "pass structured findings instead"
+                    ),
+                    remediate=(
+                        "do not hand-format the review. Pass a one-paragraph "
+                        "summary in `body` plus structured "
+                        "`findings=[{file, line?, severity "
+                        "(blocker|major|minor|nit), expected, actual}, ...]`; the "
+                        "system generates the GitHub comment (summary + findings "
+                        "table + verdict). event='REQUEST_CHANGES' requires >=1 "
+                        "finding; a bare event='COMMENT' with no findings is for a "
+                        "plain note, not a verdict"
+                    ),
+                    context_briefing=briefing,
+                ).with_introspection(task=t, role=role_str),
+                agent_id=reviewer_agent_id,
+                task_id=task_id,
+                verb="post_pr_review",
+            )
+        return None
+
     async def _resolve_role(
         self,
         t: Any,

roboco/services/prompter.py

@@ -21,7 +21,7 @@ from sqlalchemy import select
 
 from roboco.db.tables import AgentTable, TaskTable
 from roboco.foundation.identity import CELL_TEAMS
-from roboco.foundation.policy.batch import is_batch_umbrella
+from roboco.foundation.policy.batch import is_batch_umbrella, main_pm_cannot_own_code
 from roboco.foundation.policy.content.validators import coerce_str_list
 from roboco.foundation.policy.sequencing.models import DraftSurface, SequencePlan
 from roboco.models.base import (
@@ -304,6 +304,23 @@ class PrompterService:
             default_lead=_lead,
         )
 
+        # A Main PM coordinates — it never owns a code task. A main_pm + code
+        # draft is the structural mismatch behind the 2026-06-27 MegaTask
+        # meltdown (the git/PR/review layer treated the root as code while the
+        # ownership layer treated it as coordination → pr_fail loop). Intake
+        # coerces code -> planning here so the combo can never persist; a
+        # root-subtask / umbrella / single-task main_pm route is a coordination
+        # root whose code ACs live on the delegated cell/dev leaves. The
+        # TaskService.create backstop rejects main_pm + code for non-intake
+        # create paths (the HTTP route).
+        if main_pm_cannot_own_code(team=team, task_type=task_type):
+            self.log.info(
+                "Main-PM intake task coerced code->planning",
+                team=str(getattr(team, "value", team)),
+                title=_text(draft_data.get("title")) or "",
+            )
+            task_type = TaskType.PLANNING
+
         req = TaskCreateRequest(
             title=draft_data["title"],
             description=draft_data["description"],
@@ -990,9 +1007,9 @@ def _compose_umbrella_draft(
     The umbrella targets neither project nor product (it is branchless) and
     carries no collision surface of its own — it exists to group the batch, hold
     the wave plan in its description for review, and be the one board-review /
-    CEO-approve / Main-PM-coordinate unit. ``task_type=code`` mirrors the existing
+    CEO-approve / Main-PM-coordinate unit. It is a Main-PM coordination root, so
-    product coordination roots created from intake (the branch/PR exemption comes
+    ``task_type=planning`` (a Main PM coordinates, it does not execute code); the
-    from the batch identity, not the type).
+    branch/PR exemption comes from the batch identity, not the type.
     """
     item_titles = [
         _text(d.get("title")) or f"Task {i + 1}" for i, d in enumerate(drafts)
@@ -1014,7 +1031,7 @@ def _compose_umbrella_draft(
         "acceptance_criteria": [
             "Every root-subtask in the MegaTask is completed and merged.",
         ],
-        "task_type": TaskType.CODE.value,
+        "task_type": TaskType.PLANNING.value,
         "nature": TaskNature.TECHNICAL.value,
         "estimated_complexity": Complexity.HIGH.value,
         "priority": 1,

roboco/services/task.py

@@ -39,6 +39,7 @@ from roboco.foundation.policy.batch import (
     is_batch_umbrella,
     is_branchless_coordination,
     is_valid_batch_shape,
+    main_pm_cannot_own_code,
 )
 from roboco.foundation.policy.content import markers
 from roboco.foundation.policy.content.validators import ContentValidationError
@@ -195,6 +196,20 @@ def _board_cannot_own(task: TaskTable) -> bool:
     )
 
 
+def _task_type_is_code(task_type: Any) -> bool:
+    """True when ``task_type`` is ``TaskType.CODE`` (enum member or its value).
+
+    Robust to the two shapes SQLAlchemy hands back: the ``TaskType`` enum or
+    its raw ``"code"`` string (the latter on detached/partially-hydrated rows).
+    Used by the Main-PM claim guard, which keys on the task's *type* (a Main PM
+    cannot execute code) rather than the team+type combo the create / reassign
+    guards use — claiming is owning, and a Main PM claiming a code task is a
+    mismatch regardless of the task's team.
+    """
+    value = task_type.value if isinstance(task_type, TaskType) else task_type
+    return str(value) == TaskType.CODE.value
+
+
 _PM_OWNED_CELL_TASK_TYPES: frozenset[str] = frozenset(
     {
         TaskType.PLANNING.value,
@@ -823,6 +838,23 @@ class TaskService(BaseService):
         if req.parent_task_id:
             await self._validate_parent_depth(req.parent_task_id)
 
+        # Impossibility backstop: a Main PM coordinates — it never owns a code
+        # task. ``main_pm`` + ``code`` on the same task is the structural
+        # mismatch behind the 2026-06-27 MegaTask meltdown (a root-subtask the
+        # git/PR/review layer treated as code while ownership treated it as
+        # coordination — never reconciled, pr_fail looped). Intake
+        # (create_task_from_draft) coerces code→planning, so this fires only on
+        # a non-intake create (the HTTP route / a direct internal create) that
+        # tries to persist the forbidden combo.
+        if main_pm_cannot_own_code(team=req.team, task_type=req.task_type):
+            raise ValidationError(
+                "MAIN_PM_NO_CODE: A Main PM task coordinates — it does not"
+                " execute code. Re-draft as `planning` with coordination-level"
+                " acceptance criteria, or target a cell so a developer owns the"
+                " code.",
+                field="task_type",
+            )
+
         # Stable per-criterion ids (1:1 with acceptance_criteria) so children can
         # reference specific parent criteria; generated here when not supplied.
         ac_ids = req.acceptance_criteria_ids or [
@@ -4728,6 +4760,24 @@ class TaskService(BaseService):
                 reason=reason,
             )
             return
+        # Impossibility backstop: a Main-PM target must never receive (back) a
+        # main_pm + code task — a coordinator with no code verb cannot fix the
+        # code, so escalating it to Main PM perpetuates the mismatch (the
+        # 2026-06-27 meltdown shape). Scoped to the team+type combo (NOT a broad
+        # code+main-pm-target rule) so a legacy main_pm+code task can still be
+        # escalated to a cell dev — the correct remediation. The combo is
+        # uncreatable going forward (create backstop + intake coercion), so this
+        # is a backstop for legacy / direct-ORM-write tasks.
+        if main_pm_cannot_own_code(
+            team=task.team, task_type=task.task_type
+        ) and await self._is_main_pm_agent(target_agent_id):
+            await self._release_code_task_to_pool(
+                task=task,
+                escalator_slug=escalator_slug,
+                blocked_target_slug=target_slug,
+                reason=reason,
+            )
+            return
         if task.assigned_to and not task.blocker_raised_by:
             task.blocker_raised_by = cast("Any", task.assigned_to)
         # Capture before mutating: the audit row must record the real prior
@@ -4983,6 +5033,20 @@ class TaskService(BaseService):
         # off the board — reflect the new owner. Team.MAIN_PM is a valid non-cell
         # team and does not affect dispatch (which routes by assignee, not team).
         task.team = cast("Any", Team.MAIN_PM)
+        # A board-approved task handed to Main PM must not stay `code` — a Main
+        # PM coordinates, it never owns a code task (the 2026-06-27 meltdown was
+        # a main_pm + code root). A board-routed PROJECT code task reaches here
+        # with team=cell and task_type=code (intake only coerces main_pm-team
+        # drafts); once team is flipped to MAIN_PM the combo would persist.
+        # Retype code→planning so it's a planning-typed coordination root the
+        # Main PM delegates to the cells. (Intake coerces this too; this is the
+        # board-review backstop.)
+        if main_pm_cannot_own_code(team=task.team, task_type=task.task_type):
+            self.log.info(
+                "approve_and_start retyped main-pm code task to planning",
+                task_id=str(task_id),
+            )
+            task.task_type = cast("Any", TaskType.PLANNING)
 
         if notes:
             # Coordination metadata, not a human handoff — store as a marker so
@@ -6128,6 +6192,28 @@ class TaskService(BaseService):
         if task.status == TaskStatus.AWAITING_PM_REVIEW:
             return await self._claim_review_state(task_id, claim_agent_id)
 
+        # Impossibility backstop (C8 — execution states only): a Main PM
+        # coordinates — it never claims a CODE task to execute (claiming here
+        # is owning through the lifecycle, NOT delegating; delegation uses the
+        # `delegate` verb, not claim). Scoped to run only AFTER the
+        # awaiting_pm_review review-claim above returned, so the legitimate
+        # Main-PM review/merge path is untouched. The effective claimant is the
+        # reassign target when ``allow_reassign`` (claim on behalf of), else the
+        # caller. The dispatcher never offers code tasks to main-pm (code→dev),
+        # so this is a backstop for a rogue / on-behalf-of-main-pm claim.
+        if allow_reassign:
+            claimant_is_main_pm = await self._is_main_pm_agent(claim_agent_id)
+        else:
+            claimant_is_main_pm = agent.role == AgentRole.MAIN_PM
+        if claimant_is_main_pm and _task_type_is_code(task.task_type):
+            raise UnauthorizedError(
+                action="claim",
+                reason=(
+                    "MAIN_PM_NO_CODE: A Main PM coordinates — it does not own a"
+                    " code task. Leave it for a developer; delegate instead."
+                ),
+            )
+
         claimed = await self.claim(
             task_id, claim_agent_id, allow_reassign=allow_reassign
         )
@@ -7079,6 +7165,46 @@ class TaskService(BaseService):
         )
         return task
 
+    async def _maybe_divert_main_pm_code_reassign(
+        self, task: TaskTable, task_id: UUID, new_assignee: UUID | None
+    ) -> TaskTable | None:
+        """Divert a main_pm+code → Main-PM reassign to the pool, or None.
+
+        Twin of ``_maybe_divert_board_advisory_reassign`` for the
+        impossibility invariant: a Main-PM target must never receive (back) a
+        ``main_pm`` + ``code`` task — a coordinator with no code verb cannot
+        fix the code, so re-handing it to Main PM perpetuates the mismatch
+        (the 2026-06-27 meltdown shape). Scoped to the team+type combo (NOT a
+        broad code+main-pm-target rule) so a legacy main_pm+code task can still
+        be reassigned to a cell dev — the correct remediation. The combo is
+        uncreatable going forward (create backstop + intake coercion +
+        approve_and_start retype), so this is a backstop for legacy /
+        direct-ORM-write tasks.
+
+        Returns the refreshed task when it diverted, or ``None`` when no
+        diversion applies and the caller should proceed with the normal handoff.
+        """
+        if not (
+            new_assignee is not None
+            and main_pm_cannot_own_code(team=task.team, task_type=task.task_type)
+            and await self._is_main_pm_agent(new_assignee)
+        ):
+            return None
+        await self._divert_owned_task_to_pool(
+            task,
+            note=(
+                "\n\n[REASSIGN REDIRECTED] attempted to hand this main_pm + code"
+                " task to a Main PM that cannot own code. Released to the pool"
+                " for a role-matched claim instead."
+            ),
+        )
+        self.log.info(
+            "main_pm + code task reassign to a Main PM diverted to pool",
+            task_id=str(task_id),
+            refused_assignee=str(new_assignee),
+        )
+        return task
+
     async def reassign(
         self, task_id: UUID, new_assignee: UUID | None
     ) -> TaskTable | None:
@@ -7101,6 +7227,14 @@ class TaskService(BaseService):
         diverted = await self._maybe_divert_board_advisory_reassign(
             task, task_id, new_assignee
         )
+        if diverted is not None:
+            return diverted
+        # main_pm + code → Main-PM hand-off is diverted to the pool (see
+        # `_maybe_divert_main_pm_code_reassign`); otherwise proceed with the
+        # normal handoff + cell-PM redirect.
+        diverted = await self._maybe_divert_main_pm_code_reassign(
+            task, task_id, new_assignee
+        )
         if diverted is not None:
             return diverted
         # Invariant backstop: cell-team planning/research/administrative children
@@ -7257,6 +7391,27 @@ class TaskService(BaseService):
                 refused_assignee=str(new_assignee),
             )
             return task
+        # Impossibility backstop (mirrors `reassign`): an active main_pm + code
+        # claim must not be handed to a Main PM — divert to the pool. Scoped to
+        # the team+type combo so a legacy main_pm+code task can still be
+        # reassign-claimed by a cell dev (the correct remediation).
+        if main_pm_cannot_own_code(
+            team=task.team, task_type=task.task_type
+        ) and await self._is_main_pm_agent(new_assignee):
+            await self._divert_owned_task_to_pool(
+                task,
+                note=(
+                    "\n\n[REASSIGN REDIRECTED] attempted to hand this active"
+                    " main_pm + code task to a Main PM that cannot own code."
+                    " Released to the pool for a role-matched claim instead."
+                ),
+            )
+            self.log.info(
+                "Active main_pm + code task reassign to a Main PM diverted",
+                task_id=str(task_id),
+                refused_assignee=str(new_assignee),
+            )
+            return task
         # Invariant backstop: an active claim on a cell-team planning/research/
         # administrative task must land on the cell PM, not main-pm.
         redirect = await self._resolve_cell_pm_redirect(task, new_assignee)
@@ -7697,6 +7852,19 @@ class TaskService(BaseService):
         role = result.scalar_one_or_none()
         return role in _BOARD_ADVISORY_ROLES
 
+    async def _is_main_pm_agent(self, agent_id: UUID) -> bool:
+        """True if ``agent_id`` is the Main PM role (the coordinator with no code verb).
+
+        Twin of ``_is_board_advisory_agent`` for the impossibility invariant —
+        consulted by the escalation / reassign / claim guards that divert or
+        refuse a ``main_pm`` + ``code`` hand-off to a Main-PM target.
+        """
+        result = await self.session.execute(
+            select(AgentTable.role).where(AgentTable.id == agent_id)
+        )
+        role = result.scalar_one_or_none()
+        return role == AgentRole.MAIN_PM
+
     async def _divert_owned_task_to_pool(self, task: TaskTable, *, note: str) -> None:
         """Clear ownership and return ``task`` to PENDING for a role-matched claim.
 

tests/foundation/test_lifecycle_spec.py

@@ -711,6 +711,34 @@ def test_run_all_validators_raises_on_unknown_intent_action(
         _validate.run_all_lifecycle_validators()
 
 
+def test_next_hint_pr_fail_main_pm_root_steers_to_redelegate() -> None:
+    """A ``pr_fail`` on a Main-PM branch-bearing root must steer the Main PM to
+    re-delegate the fixes, NOT re-submit the unchanged root. The root is an
+    assembled cell→root / root→master PR — coordination, not the Main PM's own
+    code — so re-submitting it is the 2026-06-27 infinite ``pr_fail`` loop."""
+    t = SimpleNamespace(team=spec.Team.MAIN_PM, branch_name="feature/main_pm/c80e19ff")
+    hint = _INTENT_VERBS["pr_fail"].next_hint(t)
+    assert "re-delegate" in hint
+    assert "do NOT re-submit" in hint
+
+
+def test_next_hint_pr_fail_cell_dev_keeps_dev_revise() -> None:
+    """A cell / dev task is revised in place by its dev, so ``pr_fail`` keeps the
+    dev-revise hint (the cell→root PR carries that dev's own code)."""
+    t = SimpleNamespace(team=spec.Team.BACKEND, branch_name="feature/backend/abc12345")
+    hint = _INTENT_VERBS["pr_fail"].next_hint(t)
+    assert hint == "idle - dev will revise and re-submit"
+
+
+def test_next_hint_pr_fail_branchless_main_pm_keeps_dev_revise() -> None:
+    """A branchless Main-PM umbrella (no ``branch_name``) assembles no PR of its
+    own, so the gate never lands a ``pr_fail`` on it — but defensively it keeps
+    the dev-revise hint rather than the re-delegate steer."""
+    t = SimpleNamespace(team=spec.Team.MAIN_PM, branch_name=None)
+    hint = _INTENT_VERBS["pr_fail"].next_hint(t)
+    assert hint == "idle - dev will revise and re-submit"
+
+
 def test_unmigrated_is_pinned() -> None:
     """The known-debt set; remove an entry once that consumer is migrated."""
     assert (

tests/unit/foundation/policy/content/test_content_models.py

@@ -206,6 +206,49 @@ def test_findings_single_dict_coerced_to_list() -> None:
     assert len(c.findings) == 1
 
 
+def test_pr_review_issues_carry_free_text_change_requests() -> None:
+    # The in-path gate fails on free-text issues (not structured Finding
+    # objects, which require file/severity/expected/actual). Those issues now
+    # land in the additive `issues` slot instead of being flattened into the
+    # summary string alone — so a reader of notes_structured.pr_review gets the
+    # concrete change-requests, and the rendered TEXT mirror gains an Issues
+    # section.
+    c = validate_content(
+        "pr_review",
+        {
+            "summary": "PR review needs changes before this can merge.",
+            "verdict": "changes_requested",
+            "issues": ["seam mismatch on the rebase path", "docs lag the diff"],
+        },
+    )
+    assert isinstance(c, PrReviewContent)
+    assert c.issues == ["seam mismatch on the rebase path", "docs lag the diff"]
+    rendered = c.render_markdown()
+    assert "## Issues" in rendered
+    assert "seam mismatch on the rebase path" in rendered
+    assert "docs lag the diff" in rendered
+
+
+def test_pr_review_issues_default_empty_and_single_scalar_coerced() -> None:
+    c = validate_content(
+        "pr_review",
+        {"summary": "Clean PR, no free-text issues to raise.", "verdict": "approved"},
+    )
+    assert isinstance(c, PrReviewContent)
+    assert c.issues == []
+    assert "## Issues" not in c.render_markdown()
+
+    coerced = validate_content(
+        "pr_review",
+        {
+            "summary": "One free-text issue passed as a bare string here.",
+            "verdict": "changes_requested",
+            "issues": "lone issue string",
+        },
+    )
+    assert coerced.issues == ["lone issue string"]
+
+
 def test_where_to_look_single_string_coerced() -> None:
     c = validate_content(
         "resumption",

tests/unit/foundation/policy/test_batch.py

@@ -10,7 +10,9 @@ from roboco.foundation.policy.batch import (
     is_batch_umbrella,
     is_branchless_coordination,
     is_valid_batch_shape,
+    main_pm_cannot_own_code,
 )
+from roboco.models.base import TaskType, Team
 
 
 def test_umbrella_is_batch_id_set_and_top_level() -> None:
@@ -163,3 +165,27 @@ def test_valid_batch_shape_denies_cell_map_alongside_another_target() -> None:
         product_id=uuid4(),
         has_cell_projects=True,
     )
+
+
+def test_main_pm_cannot_own_code_predicate() -> None:
+    """``main_pm`` + ``code`` must never coexist — the single invariant behind
+    the intake coercion, the create backstop, the reassign/escalation diversion,
+    and the claim guard. Accepts ORM enums or their .value strings."""
+    # The forbidden combo, in both enum and string form.
+    assert main_pm_cannot_own_code(team=Team.MAIN_PM, task_type=TaskType.CODE)
+    assert main_pm_cannot_own_code(
+        team=Team.MAIN_PM.value, task_type=TaskType.CODE.value
+    )
+    # A Main PM coordinating (planning / research / etc.) is fine.
+    assert not main_pm_cannot_own_code(team=Team.MAIN_PM, task_type=TaskType.PLANNING)
+    assert not main_pm_cannot_own_code(team=Team.MAIN_PM, task_type=TaskType.RESEARCH)
+    assert not main_pm_cannot_own_code(
+        team=Team.MAIN_PM, task_type=TaskType.DOCUMENTATION
+    )
+    # Code owned by any non-main_pm team (a cell dev, the board pre-approval) is fine.
+    assert not main_pm_cannot_own_code(team=Team.BACKEND, task_type=TaskType.CODE)
+    assert not main_pm_cannot_own_code(team=Team.FRONTEND, task_type=TaskType.CODE)
+    assert not main_pm_cannot_own_code(team=Team.BOARD, task_type=TaskType.CODE)
+    # Missing team or type cannot satisfy the invariant.
+    assert not main_pm_cannot_own_code(team=None, task_type=TaskType.CODE)
+    assert not main_pm_cannot_own_code(team=Team.MAIN_PM, task_type=None)

tests/unit/gateway/test_pr_gate_notifies_pm.py

@@ -0,0 +1,232 @@
+"""pr_fail delivers its change-requests to the owning PM, not just to GitHub.
+
+The in-path ``pr_fail`` gate persists the reviewer's verdict to
+``notes_structured.pr_review`` and posts it on the assembled PR — but historically
+never pushed the concrete issues to any channel the owning PM reads (no a2a,
+and ``_briefing_for`` / ``build_task_handoff`` read neither ``pr_reviewer_notes``
+nor ``notes_structured.pr_review``). So the cell PM respawned into
+``needs_revision`` saw a generic "needs revision" with zero actionable issues,
+concluded there was nothing to rework, and re-submitted the same PR — an
+infinite ``pr_fail`` loop (observed live on coordination root 9980d0a0 / PR #138).
+
+The fix mirrors QA's ``fail_review`` a2a to the dev (qa.py:671-678): on
+``pr_fail`` the gate now sends an a2a to the owner the runner just re-assigned
+(the cell PM, via ``_revision_pm_for_task``) carrying the issues, so the PM
+"knows" and can ``delegate`` a rework subtask or action the change-requests
+directly. ``pr_pass`` is unaffected.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock
+from uuid import uuid4
+
+import pytest
+from roboco.foundation.policy import lifecycle as spec_module
+from roboco.services.gateway.choreographer import Choreographer, ChoreographerDeps
+
+
+def _make_choreographer() -> Choreographer:
+    base: dict[str, Any] = {
+        "task": AsyncMock(),
+        "work_session": AsyncMock(),
+        "git": AsyncMock(),
+        "a2a": AsyncMock(),
+        "journal": AsyncMock(),
+        "audit": AsyncMock(),
+        "evidence_repo": AsyncMock(),
+    }
+    return Choreographer(ChoreographerDeps(**base))
+
+
+def _stub_gate_path(
+    c: Choreographer,
+    *,
+    reviewer_id: Any,
+    t_before: Any,
+    t_after: Any,
+) -> None:
+    """Drive ``_gate_decision`` past preflight/tracing/post and into the new
+    a2a step without exercising the heavy ownership/tracing logic (those have
+    their own tests). The runner rebinding to ``t_after`` is what simulates the
+    PM re-assignment the real ``_revision_pm_for_task`` performs.
+    """
+    agent = MagicMock(role="pr_reviewer", slug="be-pr-reviewer")
+    c._gate_preflight = AsyncMock(  # type: ignore[method-assign]
+        return_value=(
+            t_before,
+            agent,
+            "pr_reviewer",
+            {},
+            spec_module.Context(actor_id=reviewer_id),
+        )
+    )
+    c._gate_tracing = AsyncMock(return_value=None)  # type: ignore[method-assign]
+    c._record_gate_verdict = MagicMock()  # type: ignore[method-assign]
+    c._post_gate_review_to_pr = AsyncMock()  # type: ignore[method-assign]
+    runner = MagicMock()
+    runner.run_intent = AsyncMock(return_value=t_after)
+    c._verb_runner = MagicMock(return_value=runner)  # type: ignore[method-assign]
+
+
+@pytest.mark.asyncio
+async def test_pr_fail_notifies_reassigned_owning_pm() -> None:
+    reviewer_id = uuid4()
+    pm_id = uuid4()
+    task_id = uuid4()
+    parent_id = uuid4()
+    t_before = MagicMock(
+        id=task_id,
+        assigned_to=reviewer_id,  # owned by the reviewer until the runner runs
+        pr_number=138,
+        parent_task_id=parent_id,
+        status="awaiting_pr_review",
+    )
+    # The runner reassigns the task to the owning PM (needs_revision owner).
+    t_after = MagicMock(
+        id=task_id,
+        assigned_to=pm_id,
+        pr_number=138,
+        parent_task_id=parent_id,
+        status="needs_revision",
+    )
+
+    c = _make_choreographer()
+    _stub_gate_path(c, reviewer_id=reviewer_id, t_before=t_before, t_after=t_after)
+
+    await c.pr_fail(reviewer_id, task_id, ["seam mismatch", "docs lag the diff"])
+
+    c.a2a.send.assert_awaited_once()
+    kwargs = c.a2a.send.await_args.kwargs
+    assert kwargs["from_agent"] == reviewer_id
+    assert kwargs["to_agent"] == pm_id
+    assert kwargs["skill"] == "code_review"
+    assert kwargs["task_id"] == task_id
+    body = kwargs["body"]
+    assert "PR review needs changes." in body
+    assert "seam mismatch" in body
+    assert "docs lag the diff" in body
+
+
+@pytest.mark.asyncio
+async def test_pr_pass_does_not_notify_anyone() -> None:
+    reviewer_id = uuid4()
+    pm_id = uuid4()
+    task_id = uuid4()
+    t_before = MagicMock(
+        id=task_id,
+        assigned_to=reviewer_id,
+        pr_number=42,
+        parent_task_id=uuid4(),
+        status="awaiting_pr_review",
+    )
+    t_after = MagicMock(
+        id=task_id, assigned_to=pm_id, pr_number=42, status="awaiting_pm_review"
+    )
+
+    c = _make_choreographer()
+    _stub_gate_path(c, reviewer_id=reviewer_id, t_before=t_before, t_after=t_after)
+
+    await c.pr_pass(reviewer_id, task_id, "Assembled root scope is clean and covered.")
+
+    c.a2a.send.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_pr_fail_skips_a2a_when_no_assignee() -> None:
+    """If the runner left the task unassigned, there's nobody to notify — must
+    not raise and must not crash on ``a2a.send(None)``."""
+    reviewer_id = uuid4()
+    task_id = uuid4()
+    t_before = MagicMock(
+        id=task_id,
+        assigned_to=reviewer_id,
+        pr_number=9,
+        parent_task_id=uuid4(),
+        status="awaiting_pr_review",
+    )
+    t_after = MagicMock(
+        id=task_id, assigned_to=None, pr_number=9, status="needs_revision"
+    )
+
+    c = _make_choreographer()
+    _stub_gate_path(c, reviewer_id=reviewer_id, t_before=t_before, t_after=t_after)
+
+    env = await c.pr_fail(reviewer_id, task_id, ["one concrete issue here"])
+    c.a2a.send.assert_not_awaited()
+    assert env.status == "needs_revision"
+
+
+@pytest.mark.asyncio
+async def test_pr_fail_a2a_for_main_pm_root_steers_to_redelegate() -> None:
+    """A Main-PM branch-bearing root is an assembled cell→root / root→master PR —
+    coordination, not the Main PM's own code. The ``pr_fail`` a2a body must steer
+    the Main PM to re-delegate the fixes and NOT re-submit the unchanged root
+    (the 2026-06-27 infinite ``pr_fail`` loop), while still carrying the concrete
+    issues. ``_revision_pm_for_task`` returns main-pm for a non-cell team, so the
+    recipient stays the Main PM — correct, since the Main PM re-delegates."""
+    reviewer_id = uuid4()
+    main_pm_id = uuid4()
+    task_id = uuid4()
+    t_before = MagicMock(
+        id=task_id,
+        assigned_to=reviewer_id,
+        pr_number=139,
+        parent_task_id=uuid4(),
+        status="awaiting_pr_review",
+    )
+    t_after = MagicMock(
+        id=task_id,
+        assigned_to=main_pm_id,
+        pr_number=139,
+        parent_task_id=uuid4(),
+        status="needs_revision",
+    )
+    # Team is read off the runner-rebound task. ``getattr(team, "value", team)``
+    # must yield ``"main_pm"``; a MagicMock team would not, so set the attribute
+    # explicitly. branch_name set => an assembled root, not a branchless umbrella.
+    t_after.team = spec_module.Team.MAIN_PM
+    t_after.branch_name = "feature/main_pm/c80e19ff"
+
+    c = _make_choreographer()
+    _stub_gate_path(c, reviewer_id=reviewer_id, t_before=t_before, t_after=t_after)
+
+    await c.pr_fail(reviewer_id, task_id, ["duplicate TimeseriesChart export"])
+
+    c.a2a.send.assert_awaited_once()
+    kwargs = c.a2a.send.await_args.kwargs
+    assert kwargs["to_agent"] == main_pm_id
+    body = kwargs["body"]
+    assert "PR review needs changes." in body
+    assert "duplicate TimeseriesChart export" in body
+    assert "re-delegate" in body
+    assert "do NOT re-submit" in body
+
+
+@pytest.mark.asyncio
+async def test_pr_fail_a2a_failure_is_swallowed() -> None:
+    """The gate transition already committed; an a2a delivery failure must not
+    roll back the verdict or 500 the reviewer (same posture as the PR-post step,
+    and the inverse of the cell_pm_complete None-deref crash)."""
+    reviewer_id = uuid4()
+    pm_id = uuid4()
+    task_id = uuid4()
+    t_before = MagicMock(
+        id=task_id,
+        assigned_to=reviewer_id,
+        pr_number=7,
+        parent_task_id=uuid4(),
+        status="awaiting_pr_review",
+    )
+    t_after = MagicMock(
+        id=task_id, assigned_to=pm_id, pr_number=7, status="needs_revision"
+    )
+
+    c = _make_choreographer()
+    _stub_gate_path(c, reviewer_id=reviewer_id, t_before=t_before, t_after=t_after)
+    c.a2a.send = AsyncMock(side_effect=RuntimeError("db hiccup"))  # type: ignore[method-assign]
+
+    env = await c.pr_fail(reviewer_id, task_id, ["a concrete actionable issue"])
+    # Verdict still landed — the owning PM is in needs_revision.
+    assert env.status == "needs_revision"

tests/unit/gateway/test_pr_gate_records_verdict.py

@@ -89,3 +89,68 @@ def test_record_gate_verdict_swallows_invalid_note() -> None:
     # No exception, and the stale slot is left as-is rather than corrupted.
     assert t.notes_structured is not None
     assert t.notes_structured["pr_review"]["verdict"] == "passed"
+
+
+def test_pr_fail_stores_issues_structurally_not_summary_only() -> None:
+    """pr_fail's free-text issues must persist into the structured ``issues``
+    slot (so a reader of notes_structured.pr_review gets the concrete
+    change-requests), not be flattened into the summary string alone."""
+    c = _make_choreographer()
+    t = _TaskWithNoNotes()
+    c._record_gate_verdict(
+        t,
+        "pr_fail",
+        "Issues:\n- seam mismatch\n- docs lag the diff",
+        issues=("seam mismatch", "docs lag the diff"),
+    )
+    slot = t.notes_structured["pr_review"]
+    assert slot["verdict"] == "failed"
+    assert slot["issues"] == ["seam mismatch", "docs lag the diff"]
+    # The derived TEXT mirror surfaces them too, so pr_reviewer_notes carries
+    # the concrete change-requests for any future reader.
+    assert "seam mismatch" in t.pr_reviewer_notes
+    assert "docs lag the diff" in t.pr_reviewer_notes
+
+
+def test_pr_fail_summary_does_not_duplicate_issues() -> None:
+    """pr_fail's issues must render only under ## Issues, not also baked into
+    ## Summary — otherwise the Task Details "PR Reviewer Notes" card shows each
+    issue twice (once under Summary, once under Issues). The summary is a
+    substantive non-issues sentence; the structured ``issues`` slot carries the
+    change-requests. ``notes`` (with the issues) still drives the GitHub PR post
+    and the a2a to the owning PM — those are raw text, not rendered through
+    ``render_markdown``, so no duplication there."""
+    c = _make_choreographer()
+    t = _TaskWithNoNotes()
+    c._record_gate_verdict(
+        t,
+        "pr_fail",
+        "Issues:\n- seam mismatch\n- docs lag the diff",
+        issues=("seam mismatch", "docs lag the diff"),
+    )
+    slot = t.notes_structured["pr_review"]
+    assert slot["verdict"] == "failed"
+    # Issues live in the structured issues slot...
+    assert slot["issues"] == ["seam mismatch", "docs lag the diff"]
+    # ...NOT baked into the summary.
+    assert "seam mismatch" not in slot["summary"]
+    assert "docs lag the diff" not in slot["summary"]
+    # The rendered TEXT mirror surfaces each issue (under ## Issues)...
+    assert "seam mismatch" in t.pr_reviewer_notes
+    assert "docs lag the diff" in t.pr_reviewer_notes
+    # ...with both section headers present, and the summary is the substantive
+    # sentence (not the issues-joined string).
+    assert "## Summary" in t.pr_reviewer_notes
+    assert "## Issues" in t.pr_reviewer_notes
+    assert "requested changes" in t.pr_reviewer_notes
+
+
+def test_pr_pass_leaves_issues_slot_empty() -> None:
+    c = _make_choreographer()
+    t = _TaskWithNoNotes()
+    c._record_gate_verdict(
+        t, "pr_pass", "Assembled root scope is clean; every criterion is covered."
+    )
+    slot = t.notes_structured["pr_review"]
+    assert slot["verdict"] == "passed"
+    assert slot.get("issues", []) == []

tests/unit/gateway/test_pr_review_hand_format_guard.py

@@ -0,0 +1,183 @@
+"""post_pr_review refuses a hand-formatted verdict body with no findings.
+
+The tool's contract (flow_server.post_pr_review docstring) is explicit: ``body``
+is a one-paragraph summary; when ``findings`` are given the GitHub comment is
+GENERATED in the RoboCo format (summary + findings table + verdict) — "do not
+hand-format it in body". Nothing enforced that, so a reviewer could pass
+``findings=[]`` and dump a self-formatted ``## Summary`` / ``## Issues`` /
+``## Verdict`` markdown blob into ``body`` — which the system posts verbatim
+(``_resolve_post_body`` returns ``body`` as-is when there are no findings). The
+deployed renderer emits ``## Findings`` (never ``## Issues``), so a ``## Issues``
+section on the PR is proof the agent hand-formatted. Observed live: the reviewer
+posted a body that listed the issues under BOTH ``## Summary`` and ``## Issues``
+and then repeated the entire block twice — a duplicated, self-redundant
+hand-formatted blob the contributor sees.
+
+The guard: when ``findings`` is empty AND ``body`` carries verdict/section
+markdown headers, reject with ``invalid_state`` and a remediation that points the
+reviewer at the structured-findings path. The system never posts a hand-formatted
+verdict, so the duplication cannot recur. A clean plain-note ``COMMENT`` with no
+findings is still allowed (only verdict-shaped bodies are blocked), and a review
+with structured findings is unaffected (the system generates the comment).
+"""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock
+from uuid import uuid4
+
+import pytest
+from roboco.foundation.policy import lifecycle as spec_module
+from roboco.services.gateway.choreographer import Choreographer, ChoreographerDeps
+
+
+def _make_choreographer() -> Choreographer:
+    base: dict[str, Any] = {
+        "task": AsyncMock(),
+        "work_session": AsyncMock(),
+        "git": AsyncMock(),
+        "a2a": AsyncMock(),
+        "journal": AsyncMock(),
+        "audit": AsyncMock(),
+        "evidence_repo": AsyncMock(),
+    }
+    return Choreographer(ChoreographerDeps(**base))
+
+
+def _stub_post_path(c: Choreographer, *, reviewer_id: Any, t: Any) -> None:
+    """Drive ``post_pr_review`` past preflight + the verdict-consistency gate so
+    the hand-format guard is the thing under test. The runner / side-effects are
+    stubbed so a passing case does not hit GitHub or the DB transition."""
+    agent = MagicMock(role="pr_reviewer", slug="be-pr-reviewer")
+    c._post_pr_review_preflight = AsyncMock(  # type: ignore[method-assign]
+        return_value=(
+            agent,
+            "pr_reviewer",
+            {},
+            spec_module.Context(actor_id=reviewer_id),
+        )
+    )
+    c._verdict_consistency_gate = AsyncMock(return_value=None)  # type: ignore[method-assign]
+    c._project_slug_for = AsyncMock(return_value="proj")  # type: ignore[method-assign]
+    c._resolve_post_body = MagicMock(return_value="generated body")  # type: ignore[method-assign]
+    runner = MagicMock()
+    runner.run_intent = AsyncMock(return_value=t)
+    c._verb_runner = MagicMock(return_value=runner)  # type: ignore[method-assign]
+    c._post_review_side_effects = AsyncMock()  # type: ignore[method-assign]
+
+
+def _task() -> Any:
+    return MagicMock(
+        id=uuid4(),
+        pr_number=200,
+        status="in_progress",
+        notes_structured=None,
+        pr_reviewer_notes="",
+    )
+
+
+@pytest.mark.asyncio
+async def test_hand_formatted_verdict_body_with_no_findings_is_rejected() -> None:
+    reviewer_id = uuid4()
+    task_id = uuid4()
+    c = _make_choreographer()
+    _stub_post_path(c, reviewer_id=reviewer_id, t=_task())
+
+    hand_formatted = (
+        "## Summary\nIssues:\n- [BLOCKER] seam mismatch\n"
+        "## Issues\n- [BLOCKER] seam mismatch\n## Verdict\nfailed"
+    )
+    env = await c.post_pr_review(reviewer_id, task_id, hand_formatted, "COMMENT")
+
+    body = env.as_dict()
+    assert body["error"] == "invalid_state"
+    assert "hand-formatted" in body["message"].lower()
+    assert "findings" in body["remediate"].lower()
+    # Nothing posted / transitioned — the guard fired before any side effect.
+    c.git.post_pr_review.assert_not_awaited()
+    c._verb_runner().run_intent.assert_not_awaited()  # type: ignore[union-attr]
+
+
+@pytest.mark.asyncio
+async def test_hand_formatted_body_rejected_even_for_request_changes() -> None:
+    # pr_review_conflict already blocks REQUEST_CHANGES + no findings, but the
+    # guard is independent of event — a hand-formatted verdict must never post,
+    # whatever event the agent picked.
+    reviewer_id = uuid4()
+    task_id = uuid4()
+    c = _make_choreographer()
+    _stub_post_path(c, reviewer_id=reviewer_id, t=_task())
+
+    env = await c.post_pr_review(
+        reviewer_id, task_id, "## Verdict\nfailed\nbad", "REQUEST_CHANGES"
+    )
+    assert env.as_dict()["error"] == "invalid_state"
+
+
+@pytest.mark.asyncio
+async def test_clean_plain_note_with_no_findings_is_allowed() -> None:
+    # A genuine plain COMMENT note (no verdict headers, no findings) is a legit
+    # use of event=COMMENT — the guard must not block it.
+    reviewer_id = uuid4()
+    task_id = uuid4()
+    t = _task()
+    c = _make_choreographer()
+    _stub_post_path(c, reviewer_id=reviewer_id, t=t)
+
+    env = await c.post_pr_review(
+        reviewer_id,
+        task_id,
+        "Left a note for the contributor: the CI flake on job X is tracked separately.",
+        "COMMENT",
+    )
+    body = env.as_dict()
+    assert body.get("error") is None
+    assert body["status"] == "in_progress"
+
+
+@pytest.mark.asyncio
+async def test_structured_findings_with_summary_body_is_allowed() -> None:
+    # With structured findings the system generates the comment; the guard
+    # (scoped to empty findings) does not fire even if the summary body happens
+    # to contain a header-shaped word.
+    reviewer_id = uuid4()
+    task_id = uuid4()
+    t = _task()
+    c = _make_choreographer()
+    _stub_post_path(c, reviewer_id=reviewer_id, t=t)
+
+    env = await c.post_pr_review(
+        reviewer_id,
+        task_id,
+        "The 422 path is unguarded.",
+        "REQUEST_CHANGES",
+        findings=[
+            {
+                "file": "roboco/services/git.py",
+                "line": 42,
+                "severity": "blocker",
+                "expected": "retry as COMMENT",
+                "actual": "raises",
+            }
+        ],
+    )
+    body = env.as_dict()
+    assert body.get("error") is None
+
+
+@pytest.mark.asyncio
+async def test_envelope_invalid_state_has_introspection_role() -> None:
+    # The rejection must carry role introspection like the other gates.
+    reviewer_id = uuid4()
+    task_id = uuid4()
+    c = _make_choreographer()
+    _stub_post_path(c, reviewer_id=reviewer_id, t=_task())
+
+    env = await c.post_pr_review(
+        reviewer_id, task_id, "## Summary\nstuff\n## Verdict\nfailed", "COMMENT"
+    )
+    assert env.error == "invalid_state"
+    # with_introspection(role="pr_reviewer") populated the introspection fields.
+    assert env.current_state is not None
+    assert env.valid_next_verbs is not None

tests/unit/services/test_main_pm_code_guard.py

@@ -0,0 +1,406 @@
+"""``main_pm`` + ``code`` must never coexist — the impossibility guard.
+
+The 2026-06-27 MegaTask meltdown traced to a Main-PM-owned root-subtask that
+was ``task_type=code``: the git/PR/review layer treated it as a code root
+(branch + PR + ``submit_root`` + the in-path ``pr_review`` gate + complete),
+while the ownership/dispatch layer treated it as coordination (owned /
+claimed / submitted / completed by the Main PM). The two layers never
+reconciled — a ``pr_fail`` on the assembled code landed on a coordinator with
+no code verb, who re-submitted the unchanged root → infinite loop. The CEO's
+fix: *"It should be impossible to draft a task where main pm and task type
+code COEXIST."*
+
+This suite pins the layered backstop that closes the invariant at every site
+where the combo can be created or re-handed:
+
+* ``TaskService.create`` rejects ``main_pm`` + ``code`` (the HTTP-route /
+  direct-create backstop; intake coerces ``code``→``planning``).
+* ``approve_and_start`` retypes a board-routed code task to ``planning`` when
+  it hands it to Main PM (the board→main-pm handoff is where a project code
+  task would otherwise become ``main_pm`` + ``code``).
+* ``apply_escalation`` / ``reassign`` / ``reassign_active_claim`` divert a
+  ``main_pm`` + ``code`` task handed to a Main-PM target to the pool — but
+  STILL allow reassigning it to a cell dev (the correct remediation).
+* ``claim_task_for_agent`` rejects a Main-PM agent claiming a CODE task in an
+  execution state (claiming = owning through the lifecycle, not delegating),
+  while leaving the ``awaiting_pm_review`` review-claim path untouched (C8).
+
+The single predicate every layer consults is
+``roboco.foundation.policy.batch.main_pm_cannot_own_code`` (accepts ORM enums
+or their ``.value`` strings); these tests exercise the service-layer sites
+that call it.
+"""
+
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, MagicMock
+from uuid import uuid4
+
+import pytest
+from roboco.models.base import (
+    AgentRole,
+    Complexity,
+    TaskNature,
+    TaskStatus,
+    TaskType,
+    Team,
+)
+from roboco.models.task import TaskCreateRequest
+from roboco.services.base import UnauthorizedError, ValidationError
+from roboco.services.task import TaskService
+
+
+def _bind(svc: TaskService, name: str, value: object) -> None:
+    object.__setattr__(svc, name, value)
+
+
+def _service() -> TaskService:
+    session = MagicMock()
+    session.flush = AsyncMock()
+    session.commit = AsyncMock()
+    return TaskService(session)
+
+
+def _task(
+    *,
+    team: Team,
+    task_type: TaskType,
+    status: TaskStatus = TaskStatus.PENDING,
+    assigned_to: object = None,
+    board_review_complete: bool = True,
+) -> MagicMock:
+    return MagicMock(
+        id=uuid4(),
+        team=team,
+        task_type=task_type,
+        status=status,
+        assigned_to=assigned_to,
+        claimed_by=assigned_to,
+        active_claimant_id=assigned_to,
+        blocker_raised_by=None,
+        board_review_complete=board_review_complete,
+        dev_notes="",
+    )
+
+
+def _agent(role: AgentRole = AgentRole.DEVELOPER) -> MagicMock:
+    return MagicMock(role=role, agent_id=uuid4())
+
+
+def _perms() -> MagicMock:
+    p = MagicMock()
+    p.can_perform_task_action = MagicMock(return_value=True)
+    return p
+
+
+# ---------------------------------------------------------------------------
+# Site 1 — TaskService.create backstop (invariant #1: the combo on the task)
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_create_rejects_main_pm_plus_code() -> None:
+    svc = TaskService(
+        MagicMock(add=MagicMock(), flush=AsyncMock(), execute=AsyncMock())
+    )
+    req = TaskCreateRequest(
+        title="rogue main-pm code root",
+        description="should not persist",
+        acceptance_criteria=["ship it"],
+        team=Team.MAIN_PM,
+        created_by=uuid4(),
+        task_type=TaskType.CODE,
+        nature=TaskNature.TECHNICAL,
+        estimated_complexity=Complexity.MEDIUM,
+        project_id=uuid4(),
+    )
+    with pytest.raises(ValidationError, match="MAIN_PM"):
+        await svc.create(req)
+
+
+@pytest.mark.asyncio
+async def test_create_allows_main_pm_plus_planning() -> None:
+    # The backstop is the team+type combo, not main_pm alone — a Main-PM
+    # coordination root typed planning is the canonical shape, not a mismatch.
+    svc = TaskService(
+        MagicMock(add=MagicMock(), flush=AsyncMock(), execute=AsyncMock())
+    )
+    req = TaskCreateRequest(
+        title="main-pm coordination root",
+        description="fine",
+        acceptance_criteria=["coordinate the cells"],
+        team=Team.MAIN_PM,
+        created_by=uuid4(),
+        task_type=TaskType.PLANNING,
+        nature=TaskNature.TECHNICAL,
+        estimated_complexity=Complexity.MEDIUM,
+        project_id=uuid4(),
+    )
+    task = await svc.create(req)
+    assert task.task_type == TaskType.PLANNING
+
+
+# ---------------------------------------------------------------------------
+# Site 2 — approve_and_start retype (board→main-pm handoff)
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_approve_and_start_retypes_code_to_planning_for_main_pm(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    svc = _service()
+    main_pm_agent = MagicMock(id=uuid4(), slug="main-pm", role=AgentRole.MAIN_PM)
+    agent_svc = MagicMock()
+    agent_svc.get_by_slug = AsyncMock(return_value=main_pm_agent)
+    monkeypatch.setattr(
+        "roboco.services.agent.get_agent_service", lambda _session: agent_svc
+    )
+    task = _task(
+        team=Team.BOARD,
+        task_type=TaskType.CODE,
+        status=TaskStatus.PENDING,
+        board_review_complete=True,
+    )
+    _bind(svc, "get", AsyncMock(return_value=task))
+    _bind(svc, "_activate_batch_root_subtasks", AsyncMock())
+    _bind(svc, "_emit_task_event", AsyncMock())
+
+    await svc.approve_and_start(task.id)
+
+    # Handed to Main PM (team set) AND retyped off code in the same write.
+    assert task.team == Team.MAIN_PM
+    assert task.task_type == TaskType.PLANNING
+
+
+@pytest.mark.asyncio
+async def test_approve_and_start_leaves_planning_untouched(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    svc = _service()
+    main_pm_agent = MagicMock(id=uuid4(), slug="main-pm", role=AgentRole.MAIN_PM)
+    agent_svc = MagicMock()
+    agent_svc.get_by_slug = AsyncMock(return_value=main_pm_agent)
+    monkeypatch.setattr(
+        "roboco.services.agent.get_agent_service", lambda _session: agent_svc
+    )
+    task = _task(
+        team=Team.BOARD,
+        task_type=TaskType.PLANNING,
+        status=TaskStatus.PENDING,
+        board_review_complete=True,
+    )
+    _bind(svc, "get", AsyncMock(return_value=task))
+    _bind(svc, "_activate_batch_root_subtasks", AsyncMock())
+    _bind(svc, "_emit_task_event", AsyncMock())
+
+    await svc.approve_and_start(task.id)
+
+    assert task.team == Team.MAIN_PM
+    assert task.task_type == TaskType.PLANNING
+
+
+# ---------------------------------------------------------------------------
+# Site 3 — apply_escalation: divert a main_pm+code task → main-pm target to pool
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_apply_escalation_diverts_main_pm_code_to_main_pm_target() -> None:
+    svc = _service()
+    main_pm_target = uuid4()
+    task = _task(
+        team=Team.MAIN_PM,
+        task_type=TaskType.CODE,
+        status=TaskStatus.IN_PROGRESS,
+        assigned_to=uuid4(),
+    )
+    _bind(svc, "_is_board_advisory_agent", AsyncMock(return_value=False))
+    _bind(svc, "_is_main_pm_agent", AsyncMock(return_value=True))
+    release_mock = AsyncMock()
+    _bind(svc, "_release_code_task_to_pool", release_mock)
+
+    await svc.apply_escalation(
+        task=task,
+        target_agent_id=main_pm_target,
+        escalator_slug="cell-pm",
+        target_slug="main-pm",
+        reason="blocked",
+    )
+
+    release_mock.assert_awaited_once()
+    assert task.assigned_to != main_pm_target
+
+
+@pytest.mark.asyncio
+async def test_apply_escalation_does_not_divert_main_pm_code_to_cell_dev() -> None:
+    # The correct remediation for a legacy main_pm+code task is to reassign it
+    # to a cell dev who can actually fix the code — the guard must NOT block
+    # that, only block re-handing it BACK to a Main-PM target.
+    svc = _service()
+    cell_dev_target = uuid4()
+    task = _task(
+        team=Team.MAIN_PM,
+        task_type=TaskType.CODE,
+        status=TaskStatus.IN_PROGRESS,
+        assigned_to=uuid4(),
+    )
+    _bind(svc, "_is_board_advisory_agent", AsyncMock(return_value=False))
+    _bind(svc, "_is_main_pm_agent", AsyncMock(return_value=False))
+    release_mock = AsyncMock()
+    _bind(svc, "_release_code_task_to_pool", release_mock)
+    _bind(svc, "_emit_task_event", AsyncMock())
+
+    await svc.apply_escalation(
+        task=task,
+        target_agent_id=cell_dev_target,
+        escalator_slug="main-pm",
+        target_slug="be-dev-1",
+        reason="reassign to a dev to fix",
+    )
+
+    # Not diverted — proceeds to the normal escalation path.
+    release_mock.assert_not_awaited()
+
+
+# ---------------------------------------------------------------------------
+# Site 4 — reassign / reassign_active_claim diversion
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_reassign_diverts_main_pm_code_to_main_pm_target() -> None:
+    svc = _service()
+    main_pm_target = uuid4()
+    task = _task(team=Team.MAIN_PM, task_type=TaskType.CODE, assigned_to=uuid4())
+    _bind(svc, "get", AsyncMock(return_value=task))
+    _bind(svc, "_is_board_advisory_agent", AsyncMock(return_value=False))
+    _bind(svc, "_is_main_pm_agent", AsyncMock(return_value=True))
+    diverted = AsyncMock()
+    _bind(svc, "_divert_owned_task_to_pool", diverted)
+    # The cell-PM redirect must NOT be reached when the main-pm-code divert fires.
+    _bind(
+        svc,
+        "_resolve_cell_pm_redirect",
+        AsyncMock(
+            side_effect=AssertionError(
+                "cell-PM redirect must not run when main-pm-code divert fires"
+            )
+        ),
+    )
+
+    result = await svc.reassign(task.id, main_pm_target)
+
+    assert result is task
+    diverted.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_reassign_active_claim_diverts_main_pm_code_to_main_pm_target() -> None:
+    svc = _service()
+    main_pm_target = uuid4()
+    task = _task(
+        team=Team.MAIN_PM,
+        task_type=TaskType.CODE,
+        status=TaskStatus.IN_PROGRESS,
+        assigned_to=uuid4(),
+    )
+    _bind(svc, "get", AsyncMock(return_value=task))
+    _bind(svc, "_is_board_advisory_agent", AsyncMock(return_value=False))
+    _bind(svc, "_is_main_pm_agent", AsyncMock(return_value=True))
+    diverted = AsyncMock()
+    _bind(svc, "_divert_owned_task_to_pool", diverted)
+    _bind(
+        svc,
+        "_resolve_cell_pm_redirect",
+        AsyncMock(
+            side_effect=AssertionError(
+                "cell-PM redirect must not run when main-pm-code divert fires"
+            )
+        ),
+    )
+
+    result = await svc.reassign_active_claim(task.id, main_pm_target)
+
+    assert result is task
+    diverted.assert_awaited_once()
+
+
+# ---------------------------------------------------------------------------
+# Site 5 — claim_task_for_agent: main-pm claiming code (execution state only)
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_claim_rejects_main_pm_claiming_code_in_execution_state(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    svc = TaskService.__new__(TaskService)
+    svc.session = AsyncMock()
+    task = _task(team=Team.BACKEND, task_type=TaskType.CODE, status=TaskStatus.PENDING)
+    monkeypatch.setattr(svc, "_load_task_or_raise", AsyncMock(return_value=task))
+    plain_claim = AsyncMock(return_value=task)
+    monkeypatch.setattr(svc, "claim", plain_claim)
+
+    with pytest.raises(UnauthorizedError, match="MAIN_PM_NO_CODE"):
+        await svc.claim_task_for_agent(
+            task.id, _agent(role=AgentRole.MAIN_PM), _perms(), claim_target_slug=None
+        )
+
+    plain_claim.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_claim_allows_main_pm_claiming_planning_in_execution_state(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    # A Main PM coordinates planning roots — claiming a planning task is the
+    # legitimate coordination path, not a mismatch.
+    svc = TaskService.__new__(TaskService)
+    svc.session = AsyncMock()
+    task = MagicMock(
+        team=Team.MAIN_PM,
+        task_type=TaskType.PLANNING,
+        status=TaskStatus.PENDING,
+    )
+    monkeypatch.setattr(svc, "_load_task_or_raise", AsyncMock(return_value=task))
+    claimed = MagicMock()
+    plain_claim = AsyncMock(return_value=claimed)
+    monkeypatch.setattr(svc, "claim", plain_claim)
+
+    result = await svc.claim_task_for_agent(
+        task.id, _agent(role=AgentRole.MAIN_PM), _perms(), claim_target_slug=None
+    )
+
+    assert result is claimed
+    plain_claim.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_claim_does_not_reject_main_pm_code_in_awaiting_pm_review(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    # C8: awaiting_pm_review is a REVIEW state, not an execution state. A
+    # Main PM legitimately claims it to complete/merge — the code guard must
+    # not fire there (it runs only after the review-state early return).
+    svc = TaskService.__new__(TaskService)
+    svc.session = AsyncMock()
+    task = MagicMock(
+        team=Team.MAIN_PM,
+        task_type=TaskType.CODE,
+        status=TaskStatus.AWAITING_PM_REVIEW,
+    )
+    monkeypatch.setattr(svc, "_load_task_or_raise", AsyncMock(return_value=task))
+    review_claim = AsyncMock(return_value=task)
+    monkeypatch.setattr(svc, "_claim_review_state", review_claim)
+    plain_claim = AsyncMock(
+        side_effect=AssertionError("transitioning claim must not run for review state")
+    )
+    monkeypatch.setattr(svc, "claim", plain_claim)
+
+    result = await svc.claim_task_for_agent(
+        task.id, _agent(role=AgentRole.MAIN_PM), _perms(), claim_target_slug=None
+    )
+
+    assert result is task
+    review_claim.assert_awaited_once()

tests/unit/services/test_prompter.py

@@ -469,6 +469,9 @@ async def test_confirm_live_draft_product_routes_to_main_pm(db_session: Any) ->
     assert row.team == Team.MAIN_PM
     assert row.product_id == product_id
     assert row.project_id is None
+    # A Main-PM coordination root is never code — intake coerces code->planning
+    # so main_pm + code can never coexist (the 2026-06-27 meltdown shape).
+    assert row.task_type == TaskType.PLANNING
 
 
 # =============================================================================
@@ -556,6 +559,8 @@ async def test_confirm_live_batch_builds_umbrella_and_sequenced_subtasks(
     assert umbrella.team == Team.MAIN_PM
     assert umbrella.status == TaskStatus.PENDING
     assert umbrella.branch_name is None  # branchless
+    # A Main-PM coordination root is never code — the umbrella is planning-typed.
+    assert umbrella.task_type == TaskType.PLANNING
 
     a, b, c = [await db_session.get(TaskTable, UUID(sid)) for sid in ids]
     for sub in (a, b, c):
@@ -563,6 +568,11 @@ async def test_confirm_live_batch_builds_umbrella_and_sequenced_subtasks(
         assert sub.batch_id == umbrella.batch_id
         assert sub.team == Team.MAIN_PM
         assert sub.status == TaskStatus.PENDING
+        # Each root-subtask is a Main-PM coordination root: code->planning coerced
+        # at intake so main_pm + code can never coexist (the 2026-06-27 meltdown
+        # shape). It still gets its own branch + PR + submit_root + pr_review gate
+        # — the gate is branch-keyed, not task_type-keyed.
+        assert sub.task_type == TaskType.PLANNING
     assert a.project_id == project1
     assert b.project_id == project1
     assert c.project_id == project2