diff --git a/local-e2e/README.md b/local-e2e/README.md
new file mode 100644
index 000000000..624a41831
--- /dev/null
+++ b/local-e2e/README.md
@@ -0,0 +1,104 @@
+# local-e2e — session-continuity canary harness
+
+Self-contained Docker-Compose harness that gates RFC#600-class template
+changes (session continuity, file-only messages, multimodal prompts,
+cross-session memory) **before** they reach customer canary.
+
+Per CTO standing directive "fully tested + separate CI": this is a
+dedicated, *fast* (target <3 min), *small-surface* harness that uses a
+Python tenant-CP simulator (not the full `workspace-server` Go service)
+to exercise the runtime image end-to-end against canonical canary turns.
+
+See [`feedback_no_single_source_of_truth`] — the harness IS the canonical
+session-continuity validator. Per-runtime unit tests still cover their
+own guard logic; the harness covers the live conversational behaviour
+that those unit tests cannot prove.
+
+See [`feedback_image_promote_is_not_user_live`] — every assertion reads
+state back from the *running container*, never from a publish-pipeline
+ack.
+
+## What it tests (the 4 canaries)
+
+| # | Scenario | Asserts |
+|---|----------|---------|
+| 1 | 2-turn name canary | turn 2 reply contains "Hongming" → SessionStore continuity |
+| 2 | File-only message (no caption) | NOT "(empty prompt — nothing to do)" + reply references filename or asks for clarification |
+| 3 | File + caption ("summarize this") | reply addresses attachment + caption |
+| 4 | Cross-session memory recall | new session pulls "blue" via memory tool |
+
+Each scenario re-uses the same A2A wire-shape that the production
+`workspace-server` POSTs to runtime `:8000` (canvas-thread-id semantics
+via `context_id`).
+
+## Architecture
+
+```
+local-e2e/
+  docker-compose.yml           # runtime under test + cp_sim
+  cp_sim/                      # ≈300 LoC Python A2A poster + file uploader
+    cp_sim.py
+    Dockerfile
+    requirements.txt
+  canary/
+    conftest.py
+    test_session_continuity.py # 4 canary scenarios
+    test_layer_diagnostics.py  # SessionStore state probe + key derivation
+  scripts/
+    run-canary.sh              # one-shot orchestration entrypoint
+```
+
+The CP simulator emits the **exact** JSON-RPC `message/send` envelope
+that `workspace-server` produces (verified against
+`tests/e2e/test_chat_attachments_e2e.sh`). No Go service is in the loop —
+this keeps the harness lean per the CTO directive.
+
+## Run locally
+
+```bash
+# from molecule-core repo root:
+export TEMPLATE_IMAGE=ghcr.io/molecule-ai/workspace-template-hermes:latest
+./local-e2e/scripts/run-canary.sh
+```
+
+Exit code 0 = all 4 canaries pass. Non-zero = at least one canary failed
+and the harness dumped SessionStore state + last 200 log lines from the
+runtime container into `./local-e2e/artifacts/`.
+
+## How it integrates into CI
+
+Each template repo's `.gitea/workflows/session-continuity-e2e.yml` calls
+`run-canary.sh` with its own freshly-built `TEMPLATE_IMAGE`. The
+template repo's Gitea branch-protection lists
+`session-continuity-e2e (pull_request)` as a required context.
+
+Rollout order (deliberate — per `feedback_image_promote_is_not_user_live`
+we bake before we cascade):
+
+1. `molecule-ai-workspace-template-hermes` — highest-traffic + most
+   recent RFC#600-class fixes — REQUIRED gate
+2. Bake for 5 business days
+3. Cascade to claude-code, langgraph, autogen, openclaw, smolagents,
+   google-adk (one PR per template — see `scripts/onboard-template.sh`)
+
+## Future extensions (out of scope for the initial PR)
+
+- Multi-session memory consistency (3+ sessions deep)
+- Tool-use canary (workspace seeded with skills/, agent must invoke)
+- Streaming-cancellation canary (mid-stream client disconnect)
+- Cross-runtime A2A peer call (currently covered by `e2e-peer-visibility`)
+
+## Why a thin Python simulator and not the real `workspace-server`?
+
+`workspace-server` is a 60+ MB Go binary that requires Postgres, Redis,
+admin-token wiring, registry plumbing, and a 30+ second cold-boot. None
+of that touches session-continuity behaviour, which is fully owned by
+the runtime container's `a2a_executor.py`. Per CTO directive "separate
+CI as possible" + the <3 min target, we excise the platform-tenant Go
+service from the loop and emit identical wire-shape envelopes from a
+single Python file.
+
+If the simulator diverges from `workspace-server` wire shape, the gate
+goes red — fix the simulator to match production. The wire shape is
+asserted in `tests/e2e/test_chat_attachments_e2e.sh` and the runtime's
+`workspace/a2a_executor.py:_core_execute`.
diff --git a/local-e2e/cp_sim/Dockerfile b/local-e2e/cp_sim/Dockerfile
new file mode 100644
index 000000000..09888e34f
--- /dev/null
+++ b/local-e2e/cp_sim/Dockerfile
@@ -0,0 +1,19 @@
+# Python tenant-CP simulator + canary test driver.
+# Single image — pytest + httpx + the canary tests baked in.
+FROM python:3.11-slim@sha256:e78299e55776ca065dcb769f80161f48465ad352014240eb5fe4712e22505e9b
+
+WORKDIR /harness
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Test files are bind-mounted by docker-compose at run time so a `pytest -x`
+# rerun loop doesn't require a rebuild. The COPY here is for the
+# self-contained image used by Gitea Actions (where bind mounts are awkward).
+COPY cp_sim.py /harness/cp_sim.py
+COPY canary /harness/canary
+
+ENV PYTHONUNBUFFERED=1
+
+# Default: run the 4 canaries with verbose output + JUnit XML for CI.
+CMD ["pytest", "-v", "--tb=short", "--junitxml=/harness/artifacts/junit.xml", "canary/"]
diff --git a/local-e2e/cp_sim/canary/__init__.py b/local-e2e/cp_sim/canary/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/local-e2e/cp_sim/canary/conftest.py b/local-e2e/cp_sim/canary/conftest.py
new file mode 100644
index 000000000..fc9562c6a
--- /dev/null
+++ b/local-e2e/cp_sim/canary/conftest.py
@@ -0,0 +1,31 @@
+"""Shared pytest fixtures for the canary suite."""
+
+from __future__ import annotations
+
+import os
+import sys
+import uuid
+
+# cp_sim.py lives one dir up — make it importable without packaging.
+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+import pytest  # noqa: E402
+
+from cp_sim import CPSim, CPSimConfig  # noqa: E402
+
+
+@pytest.fixture
+def sim() -> CPSim:
+    """Fresh CPSim per test — cheap, isolates connection state."""
+    return CPSim(
+        cfg=CPSimConfig(
+            runtime_url=os.environ.get("RUNTIME_URL", "http://localhost:18000"),
+        )
+    )
+
+
+@pytest.fixture
+def context_id() -> str:
+    """A unique canvas-thread-id per test — guarantees SessionStore isolation
+    between scenarios so a failing canary doesn't poison the next one."""
+    return f"canary-ctx-{uuid.uuid4().hex[:12]}"
diff --git a/local-e2e/cp_sim/canary/test_layer_diagnostics.py b/local-e2e/cp_sim/canary/test_layer_diagnostics.py
new file mode 100644
index 000000000..b6cc46d74
--- /dev/null
+++ b/local-e2e/cp_sim/canary/test_layer_diagnostics.py
@@ -0,0 +1,80 @@
+"""Layer-isolation diagnostics — runs alongside the 4 canaries.
+
+These probes are not strict pass/fail gates by themselves; they exist so
+when a canary fails, the artifacts include enough state to tell whether
+the regression is in the wire-shape layer, the SessionStore layer, or
+the memory layer. Each test always passes (returns early) when the
+underlying surface is unavailable on the runtime under test — different
+templates expose different debug endpoints.
+
+Cross-refs:
+  - feedback_verify_actual_endstate_not_ack_follow_sop — we read state
+    back, not the side-effect ack.
+  - feedback_image_promote_is_not_user_live — the verification is at
+    the running-container layer.
+"""
+
+from __future__ import annotations
+
+import os
+import uuid
+
+import httpx
+
+from cp_sim import CPSim
+
+
+def test_diag_agent_card_advertises_a2a(sim: CPSim) -> None:
+    """The runtime's /agent-card must advertise A2A capabilities.
+
+    If this fails, the canaries' transport assumption (POST /a2a) is
+    already broken — diagnose the runtime image, not the canary.
+    """
+    url = f"{sim.cfg.runtime_url}/agent-card"
+    r = httpx.get(url, timeout=10.0)
+    assert r.status_code == 200, (
+        f"/agent-card returned {r.status_code}: {r.text[:300]!r}"
+    )
+    body = r.json()
+    # AgentCard spec: capabilities object must exist, even if empty.
+    assert isinstance(body, dict), f"/agent-card body not an object: {body!r}"
+    # We don't require any specific capability flag — different templates
+    # advertise different sets. The point of this diag is "is the card
+    # there at all", which signals the runtime booted past entrypoint.
+
+
+def test_diag_context_id_required_for_continuity(sim: CPSim) -> None:
+    """Same context_id in two turns must not crash the runtime.
+
+    Pure smoke probe — proves the executor accepts a continuation
+    message without 5xx-ing. The substantive assertion is canary 1; this
+    one just guarantees the path is reachable.
+    """
+    ctx = f"diag-{uuid.uuid4().hex[:8]}"
+    r1 = sim.send_text("ping", context_id=ctx)
+    r2 = sim.send_text("ping again", context_id=ctx, task_id=r1.get("result", {}).get("id"))
+    # Both replies must parse — non-empty envelope, no JSON-RPC error.
+    for label, env in (("turn1", r1), ("turn2", r2)):
+        assert "error" not in env, f"{label} returned JSON-RPC error: {env['error']}"
+
+
+def test_diag_memory_root_writable_in_canary_mode(sim: CPSim) -> None:
+    """When MOLECULE_CANARY_MODE=1, the memory root must accept writes.
+
+    Probes via the recall_memory MCP tool — if /mcp is not exposed,
+    returns early (skip-style; we still pass because some templates
+    proxy MCP elsewhere).
+    """
+    # We can't write directly here — only confirm the read path doesn't
+    # 500 on a missing key. A real write happens in canary 4.
+    key = f"canary-probe-{uuid.uuid4().hex[:8]}"
+    try:
+        val = sim.probe_memory(key)
+    except Exception as e:
+        # /mcp may not be exposed on this template — canary 4 will
+        # surface the real defect if memory is actually broken.
+        if os.environ.get("CANARY_STRICT_MCP") == "1":
+            raise
+        return
+    # Unknown key → None is fine. The point is the call didn't crash.
+    assert val is None or isinstance(val, str)
diff --git a/local-e2e/cp_sim/canary/test_session_continuity.py b/local-e2e/cp_sim/canary/test_session_continuity.py
new file mode 100644
index 000000000..0c95a2971
--- /dev/null
+++ b/local-e2e/cp_sim/canary/test_session_continuity.py
@@ -0,0 +1,204 @@
+"""The 4 canonical session-continuity canaries (task #342, RFC#600 class).
+
+These tests speak A2A directly to the runtime under test. They are the
+authoritative gate that the runtime preserves conversation continuity,
+handles file-only messages without dropping to the empty-prompt error,
+addresses multimodal prompts, and persists memory across sessions.
+
+Wire-shape source of truth: see ../cp_sim.py docstring.
+"""
+
+from __future__ import annotations
+
+import re
+import uuid
+
+from cp_sim import CPSim
+
+
+# ---------- canary 1: 2-turn name continuity -------------------------------
+
+
+def test_canary_1_two_turn_name_continuity(sim: CPSim, context_id: str) -> None:
+    """SessionStore continuity — turn 2 must recall the name from turn 1.
+
+    Empirically tests:
+      - ``a2a_executor._core_execute`` injects prior-turn history via
+        ``_extract_history(context)`` (workspace/a2a_executor.py:313).
+      - The runtime's session store is keyed on ``context_id`` (canvas
+        thread id) NOT ``task_id`` — task_id is per-turn, context_id is
+        per-conversation. Regressions to that key derivation were the
+        root cause of the 2026-05 multi-turn-amnesia incidents
+        (#a60623344 diagnosis).
+    """
+    # Turn 1 — establish the fact.
+    r1 = sim.send_text(
+        "Hi, my name is Hongming.",
+        context_id=context_id,
+    )
+    reply1 = sim.extract_text_parts(r1)
+    assert reply1, f"Turn 1 produced empty reply. envelope={r1!r}"
+
+    # Turn 2 — ask back. Same context_id → same SessionStore key.
+    r2 = sim.send_text(
+        "What's my name?",
+        context_id=context_id,
+    )
+    reply2 = sim.extract_text_parts(r2)
+    assert reply2, f"Turn 2 produced empty reply. envelope={r2!r}"
+
+    # Substring match, case-insensitive — agents may reply
+    # "Your name is Hongming." or "It's Hongming!" or similar.
+    assert re.search(r"\bhongming\b", reply2, flags=re.IGNORECASE), (
+        f"Turn 2 reply does not contain 'Hongming' — SessionStore "
+        f"continuity regression suspected. context_id={context_id} "
+        f"turn1_reply={reply1[:200]!r} turn2_reply={reply2[:400]!r}"
+    )
+
+
+# ---------- canary 2: file-only message (no caption) -----------------------
+
+
+_DROPPED_TURN_MARKERS = (
+    "(empty prompt — nothing to do)",
+    "empty prompt",
+    "message contained no text content",
+    "no text content",
+)
+
+
+def test_canary_2_file_only_message(sim: CPSim, context_id: str) -> None:
+    """File-attached A2A message with NO text part must not be dropped.
+
+    Root cause this guards against: a long-standing executor bug where
+    ``extract_message_text`` returned "" for file-only messages and the
+    executor short-circuited with the "Error: message contained no text
+    content." reply, even though the attached file was the entire point
+    of the turn.
+
+    Hard assertions:
+      - Reply is non-empty AND not the dropped-turn marker.
+      - Reply references the file by name OR asks an actionable
+        clarifying question (NOT a flat error).
+    """
+    file_name = f"canary-{uuid.uuid4().hex[:8]}.txt"
+    file_body = b"Project status: nominal. Lighthouse score 98."
+
+    r = sim.send_with_file(
+        context_id=context_id,
+        text=None,  # ← THE CANARY: no caption.
+        file_name=file_name,
+        file_bytes=file_body,
+        mime_type="text/plain",
+    )
+    reply = sim.extract_text_parts(r)
+    assert reply, f"File-only message produced empty reply. envelope={r!r}"
+
+    low = reply.lower()
+    for marker in _DROPPED_TURN_MARKERS:
+        assert marker.lower() not in low, (
+            f"File-only message was dropped — reply contains "
+            f"{marker!r}. Full reply: {reply[:500]!r}"
+        )
+
+    # Soft assertion: reply must engage with the file (reference its
+    # name) OR ask an actionable clarification. We require ONE of those —
+    # a generic "Hello! How can I help?" reply is also a drop.
+    name_referenced = file_name.lower() in low or "file" in low or "attach" in low
+    asks_clarification = (
+        "what" in low or "would you like" in low or "?" in reply
+    )
+    assert name_referenced or asks_clarification, (
+        f"File-only reply neither references the file nor asks a "
+        f"clarifying question. Reply: {reply[:500]!r}"
+    )
+
+
+# ---------- canary 3: file + prompt (multimodal) ---------------------------
+
+
+def test_canary_3_file_with_prompt(sim: CPSim, context_id: str) -> None:
+    """File-attached A2A message WITH a caption — multimodal happy path.
+
+    Lower bar than canary 2: assert the agent acknowledges the file was
+    received and tries to address the caption. We deliberately don't
+    require a perfect summary because canary mode replies are canned —
+    the goal is to prove the executor's multimodal code path doesn't
+    drop EITHER the file OR the caption.
+    """
+    file_name = f"canary-doc-{uuid.uuid4().hex[:8]}.txt"
+    file_body = (
+        b"Quarterly review. Revenue up 14%. Churn down 3%. "
+        b"Team headcount steady. Action: ship RFC#600 by end of week."
+    )
+    r = sim.send_with_file(
+        context_id=context_id,
+        text="summarize this",
+        file_name=file_name,
+        file_bytes=file_body,
+        mime_type="text/plain",
+    )
+    reply = sim.extract_text_parts(r)
+    assert reply, f"File+prompt produced empty reply. envelope={r!r}"
+
+    low = reply.lower()
+    for marker in _DROPPED_TURN_MARKERS:
+        assert marker.lower() not in low, (
+            f"File+prompt was dropped — reply contains {marker!r}. "
+            f"Full reply: {reply[:500]!r}"
+        )
+
+    # At minimum: the reply must mention file/attach/summary semantics,
+    # demonstrating the executor accepted both parts.
+    engaged = any(
+        kw in low for kw in ("file", "attach", "summary", "summarize", "content", file_name.lower())
+    )
+    assert engaged, (
+        f"Multimodal reply doesn't engage with attached file or caption. "
+        f"Reply: {reply[:500]!r}"
+    )
+
+
+# ---------- canary 4: cross-session memory recall --------------------------
+
+
+def test_canary_4_cross_session_memory_recall(sim: CPSim) -> None:
+    """Memory persists across distinct context_ids → memory layer (NOT
+    SessionStore) is the storage.
+
+    Two distinct context_ids in this test — SessionStore CANNOT bridge
+    them. The bridge is the runtime's persistent memory (MOLECULE_MEMORY_ROOT
+    in canary mode). If the recall returns "blue" in session 2, the
+    memory layer is wired correctly.
+
+    Note: we ask the agent to commit the memory explicitly in session 1
+    so that the canary doesn't depend on memory auto-extraction
+    heuristics (which vary by runtime). The commit goes through the
+    same MCP tool the canvas would invoke.
+    """
+    ctx_a = f"canary-ctx-{uuid.uuid4().hex[:12]}"
+    ctx_b = f"canary-ctx-{uuid.uuid4().hex[:12]}"
+
+    # Session 1 — commit a fact via the memory tool. Use the explicit
+    # "remember" verb so canary-mode agents (which short-circuit to a
+    # deterministic tool-call) reliably invoke `commit_memory`.
+    r1 = sim.send_text(
+        "Please use the memory tool to remember: my favorite color is blue.",
+        context_id=ctx_a,
+    )
+    reply1 = sim.extract_text_parts(r1)
+    assert reply1, f"Session 1 produced empty reply. envelope={r1!r}"
+
+    # Session 2 — different context_id. Same workspace, same memory.
+    r2 = sim.send_text(
+        "Use the memory tool to recall my favorite color, then tell me what it is.",
+        context_id=ctx_b,
+    )
+    reply2 = sim.extract_text_parts(r2)
+    assert reply2, f"Session 2 produced empty reply. envelope={r2!r}"
+
+    assert re.search(r"\bblue\b", reply2, flags=re.IGNORECASE), (
+        f"Session 2 reply does not contain 'blue' — cross-session memory "
+        f"recall regression suspected. ctx_a={ctx_a} ctx_b={ctx_b} "
+        f"session1_reply={reply1[:200]!r} session2_reply={reply2[:400]!r}"
+    )
diff --git a/local-e2e/cp_sim/cp_sim.py b/local-e2e/cp_sim/cp_sim.py
new file mode 100644
index 000000000..48d735c97
--- /dev/null
+++ b/local-e2e/cp_sim/cp_sim.py
@@ -0,0 +1,214 @@
+"""Tenant control-plane simulator.
+
+Emits the byte-identical JSON-RPC `message/send` wire shape that the
+production `workspace-server` POSTs to the runtime's :8000 — see
+``workspace-server/internal/handlers/a2a.go`` and the canonical sample
+in ``tests/e2e/test_chat_attachments_e2e.sh``.
+
+This file is purposefully small (~250 LoC). It is NOT a re-implementation
+of `workspace-server`; it is just the minimum surface required to drive
+the 4 session-continuity canaries.
+
+If the runtime asserts on a header / envelope field that the production
+platform sets but this simulator omits, FIX THE SIMULATOR — never weaken
+the runtime to accept divergent wire shapes. The simulator is the
+canonical contract emitter for canary purposes
+(``feedback_no_single_source_of_truth``).
+"""
+
+from __future__ import annotations
+
+import base64
+import json
+import os
+import uuid
+from dataclasses import dataclass
+from typing import Any
+
+import httpx
+
+
+@dataclass
+class CPSimConfig:
+    runtime_url: str
+    """Base URL of the runtime under test (e.g. http://runtime:8000)."""
+    request_timeout_s: float = 60.0
+    """Per-A2A-call timeout. Generous — canary mode replies are fast,
+    but a real Provider-backed runtime under cold cache can take 30+s."""
+
+
+class CPSim:
+    """Thin client matching workspace-server's wire shape."""
+
+    def __init__(self, cfg: CPSimConfig | None = None) -> None:
+        self.cfg = cfg or CPSimConfig(
+            runtime_url=os.environ.get("RUNTIME_URL", "http://localhost:18000"),
+        )
+        self._client = httpx.Client(timeout=self.cfg.request_timeout_s)
+
+    # ------------------------------------------------------------------ A2A
+
+    def send_text(
+        self,
+        text: str,
+        *,
+        context_id: str,
+        task_id: str | None = None,
+    ) -> dict[str, Any]:
+        """POST a text-only A2A message. Returns the JSON-RPC envelope."""
+        msg_id = f"canary-{uuid.uuid4().hex[:12]}"
+        payload = {
+            "jsonrpc": "2.0",
+            "id": msg_id,
+            "method": "message/send",
+            "params": {
+                "message": {
+                    "role": "user",
+                    "messageId": msg_id,
+                    "kind": "message",
+                    "contextId": context_id,
+                    "taskId": task_id,
+                    "parts": [{"kind": "text", "text": text}],
+                },
+                "configuration": {
+                    "acceptedOutputModes": ["text/plain"],
+                    "blocking": True,
+                },
+            },
+        }
+        return self._post(payload)
+
+    def send_with_file(
+        self,
+        *,
+        context_id: str,
+        text: str | None,
+        file_name: str,
+        file_bytes: bytes,
+        mime_type: str = "text/plain",
+        task_id: str | None = None,
+    ) -> dict[str, Any]:
+        """POST an A2A message with an inline file part.
+
+        Uses the inline `bytes` form of A2A file parts (RFC#600 — the
+        no-URI variant added precisely so canary tests don't need a
+        `/chat/uploads` round-trip). Each runtime's executor calls
+        ``extract_attached_files`` which handles both forms — verified
+        in ``workspace/executor_helpers.py:903``.
+        """
+        msg_id = f"canary-{uuid.uuid4().hex[:12]}"
+        parts: list[dict[str, Any]] = []
+        if text:
+            parts.append({"kind": "text", "text": text})
+        parts.append(
+            {
+                "kind": "file",
+                "file": {
+                    "name": file_name,
+                    "mimeType": mime_type,
+                    "bytes": base64.b64encode(file_bytes).decode("ascii"),
+                },
+            }
+        )
+        payload = {
+            "jsonrpc": "2.0",
+            "id": msg_id,
+            "method": "message/send",
+            "params": {
+                "message": {
+                    "role": "user",
+                    "messageId": msg_id,
+                    "kind": "message",
+                    "contextId": context_id,
+                    "taskId": task_id,
+                    "parts": parts,
+                },
+                "configuration": {
+                    "acceptedOutputModes": ["text/plain"],
+                    "blocking": True,
+                },
+            },
+        }
+        return self._post(payload)
+
+    # ------------------------------------------------------------ helpers
+
+    def _post(self, payload: dict[str, Any]) -> dict[str, Any]:
+        url = f"{self.cfg.runtime_url}/a2a"
+        try:
+            r = self._client.post(url, json=payload)
+        except httpx.HTTPError as e:
+            raise CPSimError(f"A2A POST failed: {e}") from e
+        if r.status_code != 200:
+            raise CPSimError(
+                f"A2A non-200: status={r.status_code} body={r.text[:500]}"
+            )
+        try:
+            return r.json()
+        except json.JSONDecodeError as e:
+            raise CPSimError(f"A2A body not JSON: {r.text[:500]}") from e
+
+    @staticmethod
+    def extract_text_parts(envelope: dict[str, Any]) -> str:
+        """Return concatenated text from all text parts of a reply.
+
+        Handles both top-level `result.parts` (the canonical shape) and
+        `result.artifacts[*].parts` (which some runtimes emit when the
+        reply was streamed as artifact chunks). Matches the extractor in
+        ``tests/e2e/test_chat_attachments_e2e.sh``.
+        """
+        result = envelope.get("result") or {}
+        chunks: list[str] = []
+        for p in result.get("parts", []) or []:
+            if p.get("kind") == "text":
+                chunks.append(p.get("text", ""))
+        for art in result.get("artifacts", []) or []:
+            for p in art.get("parts", []) or []:
+                if p.get("kind") == "text":
+                    chunks.append(p.get("text", ""))
+        # Some runtimes return a status.message instead of/in addition to parts.
+        status = result.get("status") or {}
+        status_msg = status.get("message") or {}
+        for p in status_msg.get("parts", []) or []:
+            if p.get("kind") == "text":
+                chunks.append(p.get("text", ""))
+        return "\n".join(chunks).strip()
+
+    # ----------------------------------------------------- memory probe
+
+    def probe_memory(self, key: str) -> str | None:
+        """Read a memory value via the runtime's MCP memory tool.
+
+        Uses the same MCP transport the canvas uses
+        (``POST /workspaces/:id/mcp``-shaped JSON-RPC over /mcp).  Returns
+        the recalled string or None if the key is missing.
+        """
+        payload = {
+            "jsonrpc": "2.0",
+            "id": f"canary-mem-{uuid.uuid4().hex[:8]}",
+            "method": "tools/call",
+            "params": {"name": "recall_memory", "arguments": {"key": key}},
+        }
+        try:
+            r = self._client.post(f"{self.cfg.runtime_url}/mcp", json=payload)
+        except httpx.HTTPError as e:
+            raise CPSimError(f"MCP POST failed: {e}") from e
+        if r.status_code != 200:
+            return None
+        body = r.json()
+        result = body.get("result") or {}
+        # MCP responses wrap the tool output in result.content[*].text per
+        # the JSON-RPC tools/call contract.
+        for c in result.get("content", []) or []:
+            if c.get("type") == "text":
+                return c.get("text")
+        return None
+
+
+class CPSimError(RuntimeError):
+    """Raised on transport / envelope failures (NOT canary assertion failures).
+
+    Distinct from AssertionError so pytest reports them as ERROR not
+    FAILED — a transport-layer fault should be debugged differently from
+    a real session-continuity regression.
+    """
diff --git a/local-e2e/cp_sim/requirements.txt b/local-e2e/cp_sim/requirements.txt
new file mode 100644
index 000000000..b8bb5cd3e
--- /dev/null
+++ b/local-e2e/cp_sim/requirements.txt
@@ -0,0 +1,5 @@
+# Pinned (not floating) so the harness is reproducible across CI runs.
+# These versions match what tests/e2e/_lib.sh and tests/e2e/conftest.py use.
+httpx==0.27.2
+pytest==8.3.3
+pytest-asyncio==0.24.0
diff --git a/local-e2e/docker-compose.yml b/local-e2e/docker-compose.yml
new file mode 100644
index 000000000..e306b31ea
--- /dev/null
+++ b/local-e2e/docker-compose.yml
@@ -0,0 +1,58 @@
+# local-e2e/docker-compose.yml — minimal harness stack.
+#
+# Two services:
+#   runtime  — the template image under test (TEMPLATE_IMAGE env var).
+#              Exposes :8000 for A2A traffic. The simulator POSTs to it.
+#   cp_sim   — thin Python tenant-CP simulator. Drives the canary turns.
+#
+# Deliberately NO postgres, NO redis, NO platform Go service. SessionStore
+# continuity is a runtime-internal concern (a2a_executor + executor_helpers);
+# we test it without dragging the platform-tenant Go binary into the loop.
+# See README.md "Why a thin Python simulator" for rationale.
+
+services:
+  runtime:
+    image: ${TEMPLATE_IMAGE:?TEMPLATE_IMAGE env required, e.g. ghcr.io/molecule-ai/workspace-template-hermes:latest}
+    # The runtime entrypoint (workspace/entrypoint.sh) refuses to start when
+    # any operator-scope env var is present. We deliberately set no creds —
+    # the canary doesn't invoke a real LLM provider (see TEST_NO_PROVIDER below).
+    environment:
+      # Disable provider calls during canary — the runtime returns canned
+      # echo-style replies so the harness can assert continuity / file-handling
+      # behaviour without burning provider quota. The template image must
+      # honour MOLECULE_CANARY_MODE=1 (added in molecule-ai-workspace-runtime
+      # PR #46 — see molecule_runtime/a2a_executor.py canary short-circuit).
+      MOLECULE_CANARY_MODE: "1"
+      # Anonymous workspace identity so RBAC paths exercise the same code
+      # they would in tenant production.
+      WORKSPACE_ID: "canary-${CANARY_RUN_ID:-local}"
+      # Memory tool requires a writable scope; point at /tmp inside the
+      # container so cross-session canary (#4) works without bind mounts.
+      MOLECULE_MEMORY_ROOT: "/tmp/canary-memory"
+      # The provisioner's forbidden-env guard exits non-zero when any
+      # operator-scope literal is present; the canary intentionally sets
+      # zero of them. Leave guard ON (do NOT set MOLECULE_TENANT_GUARD_DISABLE)
+      # so we exercise the prod entrypoint code path verbatim.
+    ports:
+      - "${RUNTIME_PORT:-18000}:8000"
+    healthcheck:
+      # /agent-card is the universal A2A discovery endpoint — every template
+      # exposes it. /health varies per template.
+      test: ["CMD-SHELL", "wget -qO /dev/null --tries=1 http://localhost:8000/agent-card || exit 1"]
+      interval: 3s
+      timeout: 3s
+      retries: 20
+      start_period: 30s
+
+  cp_sim:
+    build:
+      context: ./cp_sim
+    depends_on:
+      runtime:
+        condition: service_healthy
+    environment:
+      RUNTIME_URL: "http://runtime:8000"
+      CANARY_RUN_ID: "${CANARY_RUN_ID:-local}"
+    # cp_sim doesn't expose a port — it's a one-shot driver invoked by
+    # run-canary.sh via `docker compose run cp_sim pytest ...`.
+    profiles: ["driver"]
diff --git a/local-e2e/scripts/onboard-template.sh b/local-e2e/scripts/onboard-template.sh
new file mode 100755
index 000000000..79b3c8dfc
--- /dev/null
+++ b/local-e2e/scripts/onboard-template.sh
@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+# onboard-template.sh — gitops helper to wire local-e2e into a new template.
+#
+# Drops .gitea/workflows/session-continuity-e2e.yml into the target template
+# repo (a thin shim that clones molecule-core's local-e2e harness, then runs
+# run-canary.sh against the locally-built template image). Opens a PR.
+#
+# Usage:
+#   ./local-e2e/scripts/onboard-template.sh molecule-ai-workspace-template-claude-code
+#
+# Per task #342 sequencing: do NOT run this for every template at once.
+# Bake the gate on hermes for ≥5 business days first; expand only after
+# the canary is empirically stable.
+#
+# Cross-refs:
+#   feedback_no_single_source_of_truth — the workflow content is identical
+#     across templates; this helper guarantees it.
+#   feedback_image_promote_is_not_user_live — we wire the gate at the
+#     CI layer; flipping it to REQUIRED in branch_protection is a
+#     separate step (see README.md).
+
+set -euo pipefail
+
+REPO="${1:?usage: onboard-template.sh <template-repo-name>}"
+HARNESS_ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}" )/.." && pwd )"
+
+# Sanity: ensure the template-side workflow file exists in this repo.
+TEMPLATE_WORKFLOW="$HARNESS_ROOT/templates/session-continuity-e2e.yml"
+[ -f "$TEMPLATE_WORKFLOW" ] || {
+    echo "ERROR: $TEMPLATE_WORKFLOW not found in this harness checkout"
+    exit 1
+}
+
+WORK_DIR=$(mktemp -d -t e2e-onboard-XXXXXX)
+trap 'rm -rf "$WORK_DIR"' EXIT
+
+cd "$WORK_DIR"
+
+# Use mol_clone — preserves the persona credential model.
+# shellcheck disable=SC1090
+source "$HOME/.molecule-ai/ops.sh"
+mol_clone "$REPO"
+cd "$REPO"
+
+git checkout -b "task342/session-continuity-e2e-gate"
+
+mkdir -p .gitea/workflows
+cp "$TEMPLATE_WORKFLOW" .gitea/workflows/session-continuity-e2e.yml
+
+git add .gitea/workflows/session-continuity-e2e.yml
+git commit -m "ci: add local-e2e session-continuity canary gate (task #342)
+
+Wires this template into the cross-template session-continuity harness
+in molecule-ai/molecule-core/local-e2e/. The gate boots THIS repo's
+locally-built image, drives 4 canonical canaries (2-turn name continuity,
+file-only message, file+prompt, cross-session memory recall), and fails
+PRs that regress any of them.
+
+Per CTO directive: required-context flip in branch_protection is a
+SEPARATE step after 5 business days of bake."
+
+# Push branch; do not auto-open PR — leave that to the operator so the
+# review-relay routing follows the same rules as a normal change.
+git push -u origin "task342/session-continuity-e2e-gate"
+
+echo
+echo "DONE. Branch pushed to $REPO. Open PR manually:"
+echo "  https://git.moleculesai.app/molecule-ai/$REPO/compare/main...task342/session-continuity-e2e-gate"
diff --git a/local-e2e/scripts/run-canary.sh b/local-e2e/scripts/run-canary.sh
new file mode 100755
index 000000000..ae9c98c9e
--- /dev/null
+++ b/local-e2e/scripts/run-canary.sh
@@ -0,0 +1,105 @@
+#!/usr/bin/env bash
+# run-canary.sh — one-shot orchestration for the local-e2e session-continuity
+# canary harness. Used by both interactive local runs and the per-template
+# .gitea/workflows/session-continuity-e2e.yml.
+#
+# Usage:
+#   TEMPLATE_IMAGE=ghcr.io/molecule-ai/workspace-template-hermes:latest \
+#       ./local-e2e/scripts/run-canary.sh
+#
+# Optional env:
+#   CANARY_RUN_ID   — disambiguator for parallel CI runs (default: random)
+#   RUNTIME_PORT    — host port for runtime :8000 (default: 18000)
+#   KEEP_RUNNING    — set =1 to leave containers up for post-mortem
+#
+# Exit codes:
+#   0   — all 4 canaries passed
+#   1   — at least one canary failed (artifacts/ has the dump)
+#   2   — harness infrastructure failure (image pull / compose / etc.)
+#
+# Cross-refs:
+#   feedback_image_promote_is_not_user_live — we verify at the running
+#     container layer, NOT at the pipeline-green layer.
+#   feedback_verify_actual_endstate_not_ack_follow_sop — every assert
+#     reads state back; no side-effect-ack claims success.
+
+set -euo pipefail
+
+: "${TEMPLATE_IMAGE:?TEMPLATE_IMAGE env required (the runtime image under test)}"
+
+# ----------------------------------------------------------------- paths
+HARNESS_ROOT="$( cd "$( dirname "${BASH_SOURCE[0]}" )/.." && pwd )"
+ARTIFACTS_DIR="$HARNESS_ROOT/artifacts"
+mkdir -p "$ARTIFACTS_DIR"
+
+export CANARY_RUN_ID="${CANARY_RUN_ID:-$(uuidgen 2>/dev/null | tr A-Z a-z | tr -d - | cut -c1-12 || date +%s)}"
+export RUNTIME_PORT="${RUNTIME_PORT:-18000}"
+export TEMPLATE_IMAGE
+COMPOSE_PROJECT="canary-${CANARY_RUN_ID}"
+COMPOSE_FILE="$HARNESS_ROOT/docker-compose.yml"
+
+log() { printf "\n=== [%s] %s ===\n" "$(date +%H:%M:%S)" "$*"; }
+
+# ----------------------------------------------------------- cleanup hook
+cleanup() {
+    local rc=$?
+    if [ "${KEEP_RUNNING:-0}" = "1" ]; then
+        log "KEEP_RUNNING=1 — leaving containers up (project=$COMPOSE_PROJECT)"
+        return $rc
+    fi
+    log "Tearing down compose project $COMPOSE_PROJECT"
+    # On non-zero exit, capture logs FIRST. Per feedback_image_promote_is_
+    # not_user_live: dump state from the actually-running container, not
+    # an inferred pipeline state.
+    if [ $rc -ne 0 ]; then
+        log "Canary FAILED — dumping artifacts to $ARTIFACTS_DIR"
+        docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" logs \
+            --no-color --tail=200 runtime \
+            > "$ARTIFACTS_DIR/runtime.log" 2>&1 || true
+        # SessionStore state probe — runtime exposes /admin/session-store
+        # in canary mode; if not present this 404s and the file is empty.
+        docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" exec -T runtime \
+            sh -c 'ls -la /tmp/canary-memory 2>/dev/null; find /tmp -name "session*.json" -exec cat {} \; 2>/dev/null' \
+            > "$ARTIFACTS_DIR/session-store.txt" 2>&1 || true
+    fi
+    docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" down --volumes --remove-orphans >/dev/null 2>&1 || true
+    return $rc
+}
+trap cleanup EXIT
+
+# ------------------------------------------------------ stack bring-up
+log "Building cp_sim image"
+docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" build cp_sim
+
+log "Pulling runtime image: $TEMPLATE_IMAGE"
+docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" pull runtime 2>&1 \
+    | tail -5 || true
+
+log "Starting runtime (host port $RUNTIME_PORT)"
+docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" up -d runtime
+
+# Wait for healthcheck — docker-compose `--wait` is the canonical mechanism
+# (introduced in v2.1.1 in 2021, available on every supported runner pool).
+log "Waiting for runtime healthcheck"
+if ! docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" up -d --wait runtime; then
+    log "Runtime never went healthy — dumping logs"
+    docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" logs --no-color --tail=200 runtime \
+        > "$ARTIFACTS_DIR/runtime-boot-failure.log" 2>&1 || true
+    exit 2
+fi
+
+# -------------------------------------------------------------- run tests
+log "Running canary suite"
+# Run cp_sim under the same compose project so DNS (runtime hostname)
+# resolves on the molecule-core-net bridge. --rm cleans the driver container
+# after pytest exits; volume bind mounts pytest's junit-xml back to host.
+if docker compose -p "$COMPOSE_PROJECT" -f "$COMPOSE_FILE" --profile driver run \
+        --rm \
+        -v "$ARTIFACTS_DIR:/harness/artifacts" \
+        cp_sim; then
+    log "All canaries PASSED"
+    exit 0
+else
+    log "At least one canary FAILED — see $ARTIFACTS_DIR/junit.xml"
+    exit 1
+fi
diff --git a/local-e2e/templates/session-continuity-e2e.yml b/local-e2e/templates/session-continuity-e2e.yml
new file mode 100644
index 000000000..901816f4d
--- /dev/null
+++ b/local-e2e/templates/session-continuity-e2e.yml
@@ -0,0 +1,85 @@
+name: session-continuity-e2e
+
+# Per-template wrapper for the molecule-core/local-e2e canary harness.
+# DO NOT EDIT THIS FILE IN A TEMPLATE REPO — the canonical copy lives at
+# molecule-ai/molecule-core:local-e2e/templates/session-continuity-e2e.yml
+# (feedback_no_single_source_of_truth). The onboard-template.sh script
+# copies it verbatim into each template; future fixes propagate via that
+# helper, not by editing the template-side copy.
+#
+# What this workflow does:
+#   1. Build THIS template's runtime image locally on the docker-host runner.
+#   2. Clone molecule-core (canonical harness source).
+#   3. Invoke local-e2e/scripts/run-canary.sh with TEMPLATE_IMAGE set to
+#      the just-built local image.
+#   4. Upload artifacts/ on failure for post-mortem.
+#
+# Required-context flip:
+#   This workflow posts a status under the literal context name
+#   "session-continuity-e2e (pull_request)" — Gitea's standard
+#   <workflow-name> (<event>) format. To make it REQUIRED, add that
+#   exact string to the template repo's branch_protection
+#   status_check_contexts list. See README.md for the bake-period rule.
+#
+# Gitea 1.22.6 / act_runner notes (cross-refs to known footguns):
+#   - No cross-repo `uses:` (feedback_gitea_cross_repo_uses_blocked) —
+#     we clone molecule-core via plain git instead.
+#   - Per-SHA concurrency (feedback_concurrency_group_per_sha).
+#   - Workflow-level GITHUB_SERVER_URL pinned to the Gitea host
+#     (feedback_act_runner_github_server_url).
+#   - Runs on docker-host pool — NOT the heavy CI pool — per CTO
+#     directive "separate CI as possible" and the <3 min target.
+
+on:
+  pull_request:
+    branches: [main]
+  push:
+    branches: [main]
+
+concurrency:
+  group: session-continuity-e2e-${{ github.workflow }}-${{ github.event_name }}-${{ github.event.pull_request.head.sha || github.sha }}
+  cancel-in-progress: true
+
+env:
+  GITHUB_SERVER_URL: https://git.moleculesai.app
+
+jobs:
+  session-continuity-e2e:
+    runs-on: docker-host
+    timeout-minutes: 8
+    steps:
+      - name: Checkout template
+        uses: actions/checkout@v4
+        with:
+          path: template
+
+      - name: Build template image
+        id: build
+        working-directory: template
+        run: |
+          IMAGE_TAG="local-e2e-${GITHUB_SHA::12}"
+          docker build -t "molecule-ai/template-under-test:${IMAGE_TAG}" .
+          echo "image=molecule-ai/template-under-test:${IMAGE_TAG}" >> "$GITHUB_OUTPUT"
+
+      - name: Clone harness from molecule-core
+        run: |
+          # Anonymous clone — molecule-core is internal-readable. NEVER bake
+          # an auth token into the URL (feedback_credentials_in_git_url).
+          git clone --depth 1 "${GITHUB_SERVER_URL}/molecule-ai/molecule-core.git" harness
+
+      - name: Run canary
+        env:
+          TEMPLATE_IMAGE: ${{ steps.build.outputs.image }}
+          CANARY_RUN_ID: ${{ github.run_id }}-${{ github.run_attempt }}
+        run: |
+          cd harness
+          ./local-e2e/scripts/run-canary.sh
+
+      - name: Upload artifacts on failure
+        if: failure()
+        uses: actions/upload-artifact@v4
+        with:
+          name: session-continuity-canary-${{ github.run_id }}
+          path: harness/local-e2e/artifacts/
+          if-no-files-found: warn
+          retention-days: 7