Merge pull request 'Wait out LLM outages with backoff (L5)' (#116) from llm-retry into main

A 502, 503 or 504 from the LiteLLM proxy, or a dropped connection, is
retried after 5, 15 and 45 seconds before the call fails. Twice on
2026-10-07 a short burst of 502s stopped work outright: the integration
tier at 16:39 and an e2e conversation at 20:43, where the item waited on
the operator after one refused call.

Not retried: a 500 (it may be the request's fault), a timeout (it already
had its request_timeout_s), or a budget refusal. Rate limits keep their
single 20-second retry. The time spent waiting is reported with the rate
limit's (last_rate_limit_wait_s), so it isn't counted as the model's.

Tests: an outage waited out, one that outlasts the backoff, a dropped
connection, a 500 and a timeout not retried, a budget refusal not waited
on. The offline agent in test_agent_build skips the backoff (its proxy
doesn't exist).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PPQ14L3LikyefCMFBqEbKE

.dockerignore

@@ -0,0 +1,9 @@
+.git
+.env
+**/.venv
+**/__pycache__
+**/.pytest_cache
+**/.ruff_cache
+_apalache-out
+tmp
+bench

.gitea/workflows/ci.yml

@@ -58,18 +58,15 @@ jobs:
   e2e:
     # The end-to-end suite (scripts/e2e.sh): the whole agent against real
     # services, every role on a local model, so it spends nothing but keeps
-    # the RTX 4090 busy for ~25 minutes. On PRs only, as its own job so the
-    # checks above report first; a newer push to the same PR cancels the
-    # run it supersedes. Locally, run only what you're writing (-k) or
-    # diagnosing (--lf). Only after the checks pass: a PR that's red already
-    # doesn't spend the GPU.
+    # the RTX 4090 busy for ~25 minutes. On PRs only, and only after the
+    # checks pass: a PR that's red already doesn't spend the GPU. There is
+    # one runner and Gitea doesn't honour cancel-in-progress, so a run a
+    # newer push supersedes holds the runner until it's cancelled by hand.
+    # Locally, run only what you're writing (-k) or diagnosing (--lf).
     if: github.event_name != 'push'
     needs: checks
     runs-on: ubuntu-latest
     timeout-minutes: 90
-    concurrency:
-      group: e2e-${{ github.head_ref || github.ref }}
-      cancel-in-progress: true
     steps:
       - uses: actions/checkout@v4
 

.gitea/workflows/release.yml

@@ -0,0 +1,63 @@
+name: Release
+# A published release builds the agent's image and pushes it to Gitea's
+# registry (Development Plan M1.5, X4): gitea.dongar.net/dev/nisse-dev
+# tagged with the version (the release tag without its v) and latest. Cut a
+# release in Gitea (Releases → New release, tag vX.Y.Z); a tag pushed for any
+# other reason builds nothing.
+#
+# The registry needs a personal access token with write:package
+# (REGISTRY_TOKEN, an organization secret of `dev`, belonging to the
+# registry-push bot), never the built-in job token: Gitea's registry
+# resolves an Actions job token to an empty package scope (go-gitea/gitea
+# #23642 and others). Packages belong to the repository's owner, which is
+# why this repository is the `dev` organization's: the bot writes packages
+# through the org's registry-push team. Never on pull_request: registry
+# credentials stay off that trigger.
+on:
+  release:
+    types: [published]
+
+env:
+  REGISTRY: gitea.dongar.net
+  REGISTRY_USER: registry-push  # the token's account
+  IMAGE: gitea.dongar.net/dev/nisse-dev
+
+jobs:
+  build-and-push:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Derive the image version from the release's tag
+        id: vars
+        env:
+          TAG: ${{ github.event.release.tag_name }}
+        run: echo "version=${TAG#v}" >> "$GITHUB_OUTPUT"
+
+      - name: Log in to Gitea's container registry
+        uses: docker/login-action@v3
+        with:
+          registry: ${{ env.REGISTRY }}
+          username: ${{ env.REGISTRY_USER }}
+          password: ${{ secrets.REGISTRY_TOKEN }}
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Build and push the agent
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          target: agent
+          push: true
+          build-args: |
+            VERSION=${{ steps.vars.outputs.version }}
+          labels: |
+            org.opencontainers.image.version=${{ steps.vars.outputs.version }}
+            org.opencontainers.image.revision=${{ github.sha }}
+          tags: |
+            ${{ env.IMAGE }}:${{ steps.vars.outputs.version }}
+            ${{ env.IMAGE }}:latest

Dockerfile

@@ -0,0 +1,68 @@
+# The nisse-dev agent as a container image (Development Plan M1.5, X4).
+#
+# Stateless: the state store and SSH known-hosts live in a volume
+# (/var/lib/nisse-dev); configuration is mounted read-only
+# (/etc/nisse-dev/nisse-dev.yaml); secrets come as environment variables
+# (docker --env-file); SSH and deploy keys are mounted read-only wherever the
+# configuration says. The agent runs no model-chosen command here (L13):
+# those go to build VMs over SSH.
+#
+#   docker build -t nisse-dev .                  the agent
+#   docker build --target test -t nisse-dev:test .   plus the tests, for
+#                                                   scripts/e2e.sh --image
+
+FROM python:3.12-slim AS build
+COPY --from=ghcr.io/astral-sh/uv:0.12.10 /uv /usr/local/bin/uv
+ENV UV_PROJECT_ENVIRONMENT=/opt/nisse UV_COMPILE_BYTECODE=1 UV_LINK_MODE=copy
+WORKDIR /src
+COPY pyproject.toml uv.lock ./
+COPY nisse-core nisse-core
+COPY nisse-dev nisse-dev
+COPY nisse-ops nisse-ops
+# The agent and what it needs, installed (not editable): the source tree
+# isn't part of the image.
+RUN uv sync --frozen --no-dev --package nisse-dev --no-editable
+
+FROM python:3.12-slim AS runtime
+ARG VERSION=dev
+LABEL org.opencontainers.image.title=nisse-dev org.opencontainers.image.version=$VERSION
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends openssh-client \
+    && rm -rf /var/lib/apt/lists/* \
+    && useradd --create-home --uid 10001 nisse \
+    && mkdir -p /var/lib/nisse-dev /etc/nisse-dev \
+    && chown nisse /var/lib/nisse-dev
+COPY --from=build /opt/nisse /opt/nisse
+ENV PATH=/opt/nisse/bin:$PATH \
+    NISSE_CONFIG=/etc/nisse-dev/nisse-dev.yaml \
+    NISSE_ENV=/run/nisse-dev/.env
+VOLUME /var/lib/nisse-dev
+USER nisse
+ENTRYPOINT ["nisse-dev"]
+
+# The image's own package plus the tests, so the end-to-end suite runs
+# against what was built, not the source tree.
+FROM runtime AS test
+USER root
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends git openssh-server \
+    && rm -rf /var/lib/apt/lists/* \
+    && mkdir -p /run/sshd
+COPY --from=build /usr/local/bin/uv /usr/local/bin/uv
+WORKDIR /src
+COPY pyproject.toml uv.lock ./
+COPY nisse-core nisse-core
+COPY nisse-dev nisse-dev
+COPY nisse-ops nisse-ops
+COPY tests tests
+COPY scripts scripts
+COPY conftest.py ./
+ENV UV_PROJECT_ENVIRONMENT=/opt/nisse
+# Test tools too, all installed (not editable), and `uv run` never re-syncs:
+# the tests import the package as built, not the source tree beside them.
+RUN uv sync --frozen --all-packages --no-editable --inexact
+ENV UV_NO_SYNC=1
+ENTRYPOINT []
+
+# The default target: the agent itself.
+FROM runtime AS agent

docs/Development-Plan.md

@@ -76,8 +76,9 @@ Only `nisse-dev` has real code; `nisse-core` and `nisse-ops` are placeholders.
 - **The end-to-end test runs on every PR** (M1.6): the real agent, every role on a
   local model (so it costs nothing per call), a repository of the run's own and a
   real build host, driven by a scripted operator. CI runs it as its own job on pull
-  requests, cancelling a run a newer push supersedes. Locally only the tests being
-  written (`-k`) or diagnosed (`--lf`); `scripts/e2e.sh` reports what a run spent. Manual
+  requests, after the checks pass. There is one runner, and a run a newer push
+  supersedes has to be cancelled by hand. Locally only the tests being written
+  (`-k`) or diagnosed (`--lf`); `scripts/e2e.sh` reports what a run spent. Manual
   live runs are for the new loop's first runs and phase ends. Findings are filed as
   issues and either fixed before moving on or deliberately deferred.
 - **The Quint models move with the loop:** `orchestrator.qnt` while `control.py`
@@ -240,8 +241,19 @@ deletes afterwards.
 - Evaluation capture: delegations, `ask_expert` and the reader's verdicts.
 - Deferred: the context-won't-fit trigger (L6); the action classifier (M3.2).
 
-**M2.7 Conversation.** *Pre-approved.* Explanations, progress in prose, asking
-mid-work (L5, #42, #77).
+**M2.7 Conversation.** *Pre-approved.* **Done** (this PR). Explanations, progress
+in prose, asking mid-work (L5, #42, #77).
+- Questions never change state: a message ending in `?` to a waiting project item
+  is answered in a question run (`reply` and look-only tools), and the item goes
+  back to the very wait it was in, so a restatement up stays up and a go binds to
+  it.
+- A restatement gives its reasons and the main alternative; a revision says what
+  changed (`restate(changed)`, sent back without it).
+- Progress: the driver's and a helper's own words for a turn are posted, trimmed,
+  at most once a minute per item; `details` shows the last tool calls. A message
+  arriving mid-run is acknowledged at once.
+- Deferred: streaming (#42); the question property in `specs/loop.qnt` (the walk
+  checks it against the code).
 
 **M1.4 Projects** (D8). *Pre-approved.* **Done** (this PR).
 - The project entry in configuration: repository, default branch, build VM, label
@@ -259,11 +271,20 @@ mid-work (L5, #42, #77).
 - Deferred: config reload without a restart (with M2.5's entry step if wanted);
   the trust dial's configuration (M3.3).
 
-**M1.5 The agent as a container image.** *Pre-approved.*
+**M1.5 The agent as a container image.** *Pre-approved.* **Built** (this PR);
+accepted once a tag's image passes the e2e suite.
 - A Dockerfile for `nisse-dev`, stateless with a volume for state; a Gitea Actions
   workflow that builds it and pushes to Gitea's registry on each release tag (X4).
 - `bringup.sh` runs the image instead of a venv.
 - Acceptance: a tagged release produces an image; the end-to-end test runs from it.
+- The image runs as its own user; state in `/var/lib/nisse-dev`, configuration
+  mounted at `/etc/nisse-dev/nisse-dev.yaml`, secrets as environment variables,
+  keys mounted read-only. Its `test` stage carries the tests over the installed
+  package, for `scripts/e2e.sh --image REF`. Images are built only on release:
+  `release.yml` pushes `gitea.dongar.net/dev/nisse-dev:<version>` and `:latest`
+  on a published release (`REGISTRY_TOKEN`, the `dev` organization's).
+- The live bring-up's switch to the image waits for the next live test (it needs
+  its own test repository too); `bringup.sh plan` shows what it will run.
 
 **M2.8 Switch over.** *Pre-approved, after acceptance.* `nisse-dev` runs the new
 loop; `control.py`, its tests and `orchestrator.qnt` are removed.

nisse-dev/src/nisse_dev/delegation.py

@@ -173,6 +173,7 @@ class Delegation:
         now: Callable[[], float] = time.monotonic,
         cancel: Cancel | None = None,
         reader: str = "",
+        on_turn: Callable[[str], None] | None = None,
     ) -> None:
         self._client = client
         self._store = store
@@ -180,6 +181,8 @@ class Delegation:
         # Who acts on the handback, from its capability profile (L9): never
         # a model's name.
         self._reader = reader or "a materially weaker model"
+        # The delegate's own text for each turn, for progress (#42).
+        self._on_turn = on_turn
         self._git = git_repo
         self._model = model
         self._backing_model = backing_model
@@ -400,6 +403,11 @@ class Delegation:
                 transcript.append(event)
                 if isinstance(event, ToolCall):
                     calls.append(event)
+            said = " ".join(
+                e.text.strip() for e in produced if isinstance(e, AssistantText)
+            ).strip()
+            if said and self._on_turn is not None:
+                self._on_turn(said)
 
             if not calls:
                 # A delegate that stops without handing back has not reported,

nisse-dev/src/nisse_dev/llm.py

@@ -39,6 +39,13 @@ class LLMError(RuntimeError):
 RATE_LIMIT_STATUS = 429
 RATE_LIMIT_RETRY_S = 20.0
 
+# The proxy, or what's behind it, briefly gone: a 502/503/504 or a dropped
+# connection. Outages delay, never fail (L5): waited out with backoff before
+# the work is stopped. Not a 500 (it may be the request) and not a timeout
+# (a hung model has had its request_timeout_s already).
+OUTAGE_STATUSES = frozenset({502, 503, 504})
+OUTAGE_BACKOFF_S = (5.0, 15.0, 45.0)
+
 # Generous by default. Output is billed by what is produced, not by the
 # ceiling, and a ceiling too low does not fail loudly — it truncates a tool
 # call mid-argument and hands back something that parses. Measured: Opus
@@ -381,20 +388,39 @@ class AnthropicPassthroughClient:
 def _post_with_rate_limit_retry(
     http: httpx.Client, path: str, body: dict[str, Any], sleep: Any
 ) -> tuple[httpx.Response, float]:
-    """The response, and how long was spent waiting out a rate limit."""
+    """The response, and how long was spent waiting: out a rate limit (once)
+    or an outage (with backoff), not on the model."""
     waited = 0.0
-    for attempt in (1, 2):
+    rate_limited = False
+    outages = iter(OUTAGE_BACKOFF_S)
+    while True:
         try:
             response = http.post(path, json=body)
+        except httpx.TimeoutException as exc:
+            raise UpstreamUnavailable(f"LLM request failed: {exc}") from exc
         except httpx.HTTPError as exc:
             # Could not reach the proxy at all: nothing about the request was
             # wrong, so callers holding work in progress should keep it.
+            pause = next(outages, None)
+            if pause is None:
                 raise UpstreamUnavailable(f"LLM request failed: {exc}") from exc
-        if response.status_code != RATE_LIMIT_STATUS or attempt == 2:
-            return response, waited
+            sleep(pause)
+            waited += pause
+            continue
+        if response.status_code == RATE_LIMIT_STATUS and not rate_limited:
+            rate_limited = True
             sleep(RATE_LIMIT_RETRY_S)
             waited += RATE_LIMIT_RETRY_S
-    raise AssertionError("unreachable")
+            continue
+        if response.status_code in OUTAGE_STATUSES and not _looks_like_budget_refusal(
+            response.text
+        ):
+            pause = next(outages, None)
+            if pause is not None:
+                sleep(pause)
+                waited += pause
+                continue
+        return response, waited
 
 
 # Model families we have a native renderer adapter for. L9: delegations take

nisse-dev/src/nisse_dev/loop/agent.py

@@ -29,6 +29,8 @@ from nisse_dev.canonical import (
     DelegationStarted,
     Fact,
     PlanSubmitted,
+    ToolCall,
+    ToolResult,
     UserMessage,
 )
 from nisse_dev.config import GENERAL
@@ -87,6 +89,9 @@ _NAMED = re.compile(r"(go|approve|abandon|stop) (?:([a-z0-9][a-z0-9-]*))?#(\d+)"
 _NEEDS_ITEM = frozenset({"go", "approve", "abandon"})
 BUG_GRACE_S = 600.0
 GENERAL_IDLE_S = 86400.0
+PROGRESS_EVERY_S = 60.0  # progress in prose at a readable rate (L5, #42)
+PROGRESS_CHARS = 400
+DETAILS = 10  # tool calls shown by `details`
 FILE_ISSUE = "file_issue"  # the intent of opening a record issue (L8)
 # Carried by a record issue the agent opens for work started in conversation:
 # intake never takes it for new work, even before its item exists (D1).
@@ -102,9 +107,14 @@ The issue and the conversation about it are below.
 
 Before any work, find out what you need, then use restate to put your
 understanding to the operator: what you take the request to mean, how it fits
-the project, and your first step. Wait for their go-ahead. If the request is
-unclear, ask_human instead of guessing: a restatement asks no questions, so
-the operator can agree to it as it stands.
+the project, your first step, and why that way rather than the main
+alternative. A later restatement says what changed. Wait for their go-ahead.
+If the request is unclear, ask_human instead of guessing: a restatement asks
+no questions, so the operator can agree to it as it stands. At any point,
+asking is better than guessing.
+
+Before each group of tool calls, say in a sentence what you're doing and why:
+the operator sees it as your progress.
 
 The project's documents are in its repository: read ARCHITECTURE.md and any
 build plans under docs/plans/ before you restate.
@@ -119,7 +129,8 @@ the same milestone (don't start another), run the tests, and call
 finish_milestone again to push the fix. When a review needs no change from
 you, or you are only waiting for the operator to merge (you can't merge it
 yourself), call wait_for_review. The next milestone starts only after the
-previous PR has merged. When the last PR has merged, call close_item with
+previous PR has merged. A milestone's summary says what was done, what was
+verified, and what is left open. When the last PR has merged, call close_item with
 what was done, what was verified, and what is left open. If a PR is closed
 without merging, ask the operator what next.
 
@@ -133,6 +144,12 @@ Keep ARCHITECTURE.md current: a design change updates it in the same PR.
 
 {workspace}"""
 
+QUESTION = """
+
+The operator has asked a question (the last message). Answer it with reply,
+from what you know and what you can look at; don't change anything. The item
+then goes back to waiting exactly as it was, so a restatement up stays up."""
+
 CONVERSATION = """You are nisse-dev, a software agent, in a conversation with
 the operator in a Matrix thread. It isn't about a project yet: it may be a
 question, thoughts on feasibility or hardware, or the start of some work.
@@ -257,6 +274,9 @@ class LoopAgent:
         # The project whose workspace was last worked on: switching to
         # another parks it (L1). None after a boot: unknown, so all are.
         self._last_project: str | None = None
+        # Per item: when progress was last posted (L5: a readable rate).
+        self._progressed: dict[str, float] = {}
+        self._active_name = ""  # the running item, for "noted" on arrival
         self._running = False
         self._stop = threading.Event()
         self._intake_thread: threading.Thread | None = None
@@ -743,6 +763,16 @@ class LoopAgent:
             # Straight to the flag: the run is what it has to reach (L2).
             self._cancel.set()
         about = self._shown.get(thread) if thread else None
+        if self._running and command not in ("stop", "status"):
+            # L5: say what happens to a message that can't be read yet.
+            here = thread and thread == self._active_thread
+            self._transport.post(
+                "Noted; I'll read it when this turn ends."
+                if here
+                else f"Noted; I'm in the middle of {self._active_name}, and I'll "
+                "read it when that turn ends.",
+                thread=thread or msg.event_id or None,
+            )
         self._worker.submit(
             self._settled("your message", lambda: self.handle(msg, about))
         )
@@ -814,6 +844,30 @@ class LoopAgent:
             self._close(item, loop, "abandoned", f"abandoned by {msg.sender}")
         elif command == "done" and loop.new_project:
             self._provisioned(item, loop)
+        elif command == "details":
+            self._transport.post(self._details(item), thread=loop.thread)
+        elif (
+            item.position == WAITING
+            and item.scope != GENERAL
+            and is_question(text)
+            and loop.wait in (GO_AHEAD, ANSWER, MERGE)
+        ):
+            # L5: questions never change state. Answered, then back to this
+            # very wait: a restatement up stays up, a go still binds to it.
+            save(
+                self._store,
+                item.id,
+                READY,
+                replace(
+                    loop,
+                    question=True,
+                    back_to=loop.wait,
+                    back_note=loop.note,
+                    wait="",
+                    note="",
+                ),
+                cause="the operator asked a question",
+            )
         elif command == "stop":
             # Stopping a run took effect on arrival; a stop is never an answer.
             if not loop.note.startswith("stopped ("):
@@ -1332,6 +1386,9 @@ class LoopAgent:
             caps=self._helpers.caps,
             cancel=self._cancel,
             reader=self._helpers.reader(roles),
+            on_turn=lambda said: self._progress(
+                f"{item.id}:helper", loop.thread, said, prefix="helper: "
+            ),
         )
         git = milestones.git
         before = git.head_sha()  # where the milestone is, uncommitted work aside
@@ -1432,6 +1489,54 @@ class LoopAgent:
         save(self._store, item.id, READY, loop_of(item), cause="a helper finished")
         return True
 
+    def _changed_problem(self, item_id: str, args: dict[str, object]) -> str | None:
+        """L5: a revision says what changed from the one before."""
+        item = self._store.work_item(item_id)
+        if item is None or loop_of(item).rev == 0:
+            return None
+        if str(args.get("changed", "")).strip():
+            return None
+        return (
+            f"not posted: this is a revision of rev-{loop_of(item).rev}, so say "
+            "in changed what's different and why"
+        )
+
+    def _progress(self, item_id: str, thread: str, said: str, prefix: str = "") -> None:
+        """A turn's own words, posted at a readable rate (L5, #42): at most
+        one a minute per item, the latest when the minute is up."""
+        now = self._clock()
+        if now - self._progressed.get(item_id, float("-inf")) < PROGRESS_EVERY_S:
+            return
+        self._progressed[item_id] = now
+        text = said if len(said) <= PROGRESS_CHARS else said[:PROGRESS_CHARS] + " …"
+        self._transport.post(f"{prefix}{text}", thread=thread)
+
+    def _details(self, item: WorkItem) -> str:
+        """L5: raw tool activity, on request: the item's last tool calls."""
+        events = self._store.transcript(item.id, "driver").read()
+        results = {
+            e.event.call_id: e.event for e in events if isinstance(e.event, ToolResult)
+        }
+        lines = []
+        for e in events:
+            if not isinstance(e.event, ToolCall):
+                continue
+            call, result = e.event, results.get(e.event.call_id)
+            target = call.arguments.get("path") or call.arguments.get("command") or ""
+            outcome = (
+                "no result yet"
+                if result is None
+                else ("ok" if result.success else "failed")
+                + (f": {result.result.strip()[:120]}" if result.result.strip() else "")
+            )
+            who = "" if e.actor == "driver" else f"{e.actor}: "
+            lines.append(f"- {who}{call.tool}({str(target)[:80]}) → {outcome}")
+        if not lines:
+            return f"No tool calls yet on {self._label(item)}."
+        return f"The last tool calls on {self._label(item)}:\n" + "\n".join(
+            lines[-DETAILS:]
+        )
+
     def _design_problem(self, item_id: str) -> str | None:
         """D1, D2: a `design` item agrees its design before any code. Until
         a PR of it has merged, a milestone holds documents only."""
@@ -1544,6 +1649,8 @@ class LoopAgent:
         save(self._store, item.id, RUNNING, loop, cause="driver run")
         # D1: a `discuss` item talks before it restates.
         hold = "discuss" in loop.labels and not loop.talked and not loop.started
+        if loop.question:
+            hold = True  # nothing but answering (below)
         tools = [
             *([] if hold else [RESTATE]),
             ASK_HUMAN,
@@ -1560,6 +1667,8 @@ class LoopAgent:
                 ]
             ),
         ]
+        if loop.question:
+            tools = [REPLY, *_LOOK]
         roles = self._roles_of(item)
         workspace = project.workspace
         system = SYSTEM.format(
@@ -1573,8 +1682,9 @@ class LoopAgent:
                 else ""
             ),
             workspace=workspace.describe() if workspace else "",
-        )
+        ) + (QUESTION if loop.question else "")
         self._active_thread = loop.thread
+        self._active_name = self._label(item)
         self._running = True
         try:
             outcome = self._driver.run(
@@ -1592,10 +1702,14 @@ class LoopAgent:
                     ),
                 },
                 checks={
-                    "restate": _restatement_problem,
+                    "restate": lambda args: (
+                        _restatement_problem(args)
+                        or self._changed_problem(item.id, args)
+                    ),
                     "finish_milestone": lambda args: self._design_problem(item.id),
                 },
                 model=roles.resolve("driver"),
+                on_turn=lambda said: self._progress(item.id, loop.thread, said),
             )
         finally:
             self._running = False
@@ -1607,6 +1721,31 @@ class LoopAgent:
         assert item is not None
         loop = loop_of(item)
         args = outcome.arguments or {}
+        if loop.question:
+            # Back to the wait the question interrupted, as it was (L5).
+            if outcome.kind == "wait" and outcome.tool == "reply":
+                self._transport.post(
+                    str(args.get("text", "")).strip(), thread=loop.thread
+                )
+            else:
+                self._transport.post(
+                    f"I couldn't answer that ({outcome.note or outcome.kind}).",
+                    thread=loop.thread,
+                )
+            self._wait(
+                item,
+                replace(
+                    loop,
+                    wait=loop.back_to,
+                    note=loop.back_note,
+                    question=False,
+                    back_to="",
+                    back_note="",
+                ),
+                None,
+            )
+            self._store.set_active(None, cause="question answered")
+            return
         if outcome.kind == "wait" and outcome.tool == "restate":
             text = str(args.get("text", "")).strip()
             rev = loop.rev + 1
@@ -1622,7 +1761,11 @@ class LoopAgent:
                     f"It's a bug, so I'll go ahead in {round(self._grace / 60)} min "
                     "unless you reply; say go to start now, or reply with changes."
                 )
-            self._transport.post(f"rev-{rev}: {text}\n\n{after}", thread=loop.thread)
+            changed = str(args.get("changed", "")).strip()
+            since = f"\n\nChanged from rev-{rev - 1}: {changed}" if rev > 1 else ""
+            self._transport.post(
+                f"rev-{rev}: {text}{since}\n\n{after}", thread=loop.thread
+            )
             self._wait(
                 item,
                 replace(loop, rev=rev, wait=GO_AHEAD, note="", grace_until=grace),

nisse-dev/src/nisse_dev/loop/driver.py

@@ -12,7 +12,7 @@ from dataclasses import dataclass
 from typing import Any, Literal
 
 from nisse_dev.cancel import Cancel, skip_calls
-from nisse_dev.canonical import ToolCall, ToolResult
+from nisse_dev.canonical import AssistantText, ToolCall, ToolResult
 from nisse_dev.llm import LLMClient, LLMError
 from nisse_dev.loop.tools import WAITING, names
 from nisse_dev.state import Transcript
@@ -63,6 +63,7 @@ class Driver:
         | None = None,
         checks: Mapping[str, Callable[[dict[str, Any]], str | None]] | None = None,
         model: str | None = None,
+        on_turn: Callable[[str], None] | None = None,
     ) -> RunOutcome:
         """``handlers`` answer control tools that don't end the run (e.g.
         start_milestone): each returns a result and whether it succeeded.
@@ -89,6 +90,12 @@ class Driver:
                 transcript.append(event)
                 if isinstance(event, ToolCall):
                     calls.append(event)
+            # The model's own words for the turn: its progress, in prose (#42).
+            said = " ".join(
+                e.text.strip() for e in produced if isinstance(e, AssistantText)
+            ).strip()
+            if said and calls and on_turn is not None:
+                on_turn(said)
             if not calls:
                 if nudged:
                     return RunOutcome("failed", note="kept replying in prose")

nisse-dev/src/nisse_dev/loop/items.py

@@ -58,6 +58,11 @@ class Loop:
     design_merged: bool = False  # a `design` item's design PR has merged (D1, D2)
     helps: int = 0  # delegations so far, for their branch names (L9)
     helped_at: int = 0  # the milestone a helper was last called in for, unasked
+    # A question from the operator is answered without changing the item
+    # (L5): the wait it interrupted, to go back to.
+    question: bool = False
+    back_to: str = ""
+    back_note: str = ""
     # Where the operator last spoke (D5: answer where asked): "" for the
     # Matrix thread, or the PR it came from, with the inline location and a
     # little of what they said, for quoting.

nisse-dev/src/nisse_dev/loop/tools.py

@@ -25,9 +25,12 @@ def _tool(tool: str, description: str, /, **params: str) -> dict[str, Any]:
 RESTATE = _tool(
     "restate",
     "Put your understanding of the request to the operator, in their item's "
-    "thread: what you take it to mean, how it fits the project, and your "
-    "first step. Ends your turn; work starts when they agree.",
+    "thread: what you take it to mean, how it fits the project, your first "
+    "step, and why that way (and the main alternative you considered). Ends "
+    "your turn; work starts when they agree.",
     text="Your restatement, in plain language.",
+    changed="What changed from your previous restatement, and why; leave it "
+    "empty for the first.",
 )
 ASK_HUMAN = _tool(
     "ask_human",

nisse-dev/tests/test_agent_build.py

@@ -143,6 +143,9 @@ class FakeTransport:
 def build_offline(tmp_path: Path, origin: Path, monkeypatch: Any) -> Any:
     """Assemble the real graph with only the network boundaries faked."""
     config = config_for(tmp_path)
+    # The proxy here doesn't exist: a call to it fails at once rather than
+    # waiting out an outage's backoff (L5).
+    monkeypatch.setattr("nisse_dev.llm.OUTAGE_BACKOFF_S", ())
 
     monkeypatch.setattr(
         "nisse_dev.agent.GiteaClient.whoami", lambda self: "nisse-dev-testing"

nisse-dev/tests/test_llm.py

@@ -14,6 +14,7 @@ from nisse_dev.canonical import (
 from nisse_dev.envfile import load_env_file
 from nisse_dev.llm import (
     DEFAULT_MAX_OUTPUT_TOKENS,
+    OUTAGE_BACKOFF_S,
     RATE_LIMIT_RETRY_S,
     AnthropicPassthroughClient,
     BudgetExhausted,
@@ -432,3 +433,81 @@ def test_the_answering_model_is_recorded() -> None:
     )
     client.complete(EVENTS, tools=None, model="local-fast")
     assert client.last_model == "gpt-test-1"
+
+
+# --- outages delay, never fail (L5) --------------------------------------------
+
+
+def test_a_gateway_error_is_waited_out_with_backoff() -> None:
+    responses = [
+        httpx.Response(502, text="bad gateway"),
+        httpx.Response(503, text="unavailable"),
+        completion({"content": "back"}),
+    ]
+    waited: list[float] = []
+    client = make_client(lambda request: responses.pop(0), sleep=waited.append)
+    [event] = client.complete(EVENTS, tools=None, model="local/qwen")
+    assert isinstance(event, AssistantText) and event.text == "back"
+    assert waited == list(OUTAGE_BACKOFF_S[:2])
+    # Waiting isn't the model's time.
+    assert client.last_rate_limit_wait_s == sum(OUTAGE_BACKOFF_S[:2])
+
+
+def test_an_outage_that_outlasts_the_backoff_is_surfaced() -> None:
+    attempts: list[int] = []
+    waited: list[float] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        attempts.append(1)
+        return httpx.Response(502, text="bad gateway")
+
+    client = make_client(handler, sleep=waited.append)
+    with pytest.raises(UpstreamUnavailable, match="502"):
+        client.complete(EVENTS, tools=None, model="local/qwen")
+    assert len(attempts) == len(OUTAGE_BACKOFF_S) + 1
+    assert waited == list(OUTAGE_BACKOFF_S)
+
+
+def test_a_dropped_connection_is_retried() -> None:
+    calls: list[int] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        calls.append(1)
+        if len(calls) == 1:
+            raise httpx.ConnectError("refused")
+        return completion({"content": "ok"})
+
+    [event] = make_client(handler).complete(EVENTS, tools=None, model="local/qwen")
+    assert isinstance(event, AssistantText) and len(calls) == 2
+
+
+def test_a_500_or_a_timeout_is_not_retried() -> None:
+    # A 500 may be the request's fault; a timeout already had its full
+    # request_timeout_s.
+    for failure in (
+        lambda request: httpx.Response(500, text="boom"),
+        lambda request: (_ for _ in ()).throw(httpx.ReadTimeout("slow")),
+    ):
+        calls: list[int] = []
+
+        def handler(
+            request: httpx.Request, failure=failure, calls=calls
+        ) -> httpx.Response:
+            calls.append(1)
+            return failure(request)
+
+        with pytest.raises(LLMError):
+            make_client(handler).complete(EVENTS, tools=None, model="local/qwen")
+        assert len(calls) == 1
+
+
+def test_a_budget_refusal_is_not_waited_on() -> None:
+    calls: list[int] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        calls.append(1)
+        return httpx.Response(503, text="Budget has been exceeded! Current cost: 5.1")
+
+    with pytest.raises(BudgetExhausted):
+        make_client(handler).complete(EVENTS, tools=None, model="local/qwen")
+    assert len(calls) == 1

nisse-dev/tests/test_loop.py

@@ -214,7 +214,9 @@ def test_go_starts_the_work_and_close_item_closes_the_issue(h: Harness) -> None:
 
 
 def test_a_go_for_an_earlier_revision_is_not_applied(h: Harness) -> None:
-    h.model.turns.append([call("restate", text="revised: in a new module")])
+    h.model.turns.append(
+        [call("restate", text="revised: in a new module", changed="a new module")]
+    )
     h.agent.intake()
     h.say("put it in its own module")  # changes, so the driver restates rev-2
     assert loop_of(h.item()).rev == 2

nisse-dev/tests/test_loop_help.py

@@ -301,3 +301,26 @@ def test_a_delegation_cut_off_by_a_crash_is_discarded_at_boot(
     ]
     assert any("cut off by a restart and discarded" in f for f in facts)
     assert git
+
+
+def test_a_helpers_progress_is_posted(world: dict[str, Any]) -> None:
+    helper = Spy(
+        [
+            AssistantText(text="Writing helped.py with the constant.", model="hlp"),
+            call("write_file", path="helped.py", content="H = 1\n"),
+        ],
+        [hand_back()],
+    )
+    model = ScriptedModel(
+        RESTATE,
+        [
+            call("start_milestone", name="words"),
+            call("delegate", task="write helped.py"),
+            call("ask_human", question="done?"),
+        ],
+    )
+    agent, transport = loop_agent(world, model, helpers=helpers(helper))
+    agent.intake()
+    say(transport, world, "go")
+    said = [m.text for m in transport.outbound]
+    assert "helper: Writing helped.py with the constant." in said

nisse-dev/tests/test_loop_labels.py

@@ -118,7 +118,7 @@ def test_a_plain_yes_is_a_go_ahead_said_so_first(tmp_path: Path) -> None:
 def test_anything_else_goes_to_the_driver(tmp_path: Path) -> None:
     h = Harness(
         tmp_path,
-        ScriptedModel(RESTATED, [call("restate", text="v2")]),
+        ScriptedModel(RESTATED, [call("restate", text="v2", changed="as you asked")]),
         FakeGitea(issue()),
         reader=Reader(),
     )
@@ -132,7 +132,7 @@ def test_a_question_never_reaches_the_reader(tmp_path: Path) -> None:
     reader = Reader("sounds good?")
     h = Harness(
         tmp_path,
-        ScriptedModel(RESTATED, [call("restate", text="v2")]),
+        ScriptedModel(RESTATED, [call("restate", text="v2", changed="as you asked")]),
         FakeGitea(issue()),
         reader=reader,
     )
@@ -145,7 +145,9 @@ def test_a_plain_yes_to_an_older_revision_is_not_a_go(tmp_path: Path) -> None:
     reader = Reader("yes")
     h = Harness(
         tmp_path,
-        ScriptedModel(RESTATED, [call("restate", text="v2")], STARTED),
+        ScriptedModel(
+            RESTATED, [call("restate", text="v2", changed="as you asked")], STARTED
+        ),
         FakeGitea(issue()),
         reader=reader,
     )
@@ -224,7 +226,12 @@ def test_a_reply_stops_the_bugs_countdown(tmp_path: Path) -> None:
 
 
 def test_a_new_revision_of_a_bug_gets_its_own_countdown(tmp_path: Path) -> None:
-    h, clock = bug(tmp_path, RESTATED, [call("restate", text="v2")], STARTED)
+    h, clock = bug(
+        tmp_path,
+        RESTATED,
+        [call("restate", text="v2", changed="as you asked")],
+        STARTED,
+    )
     h.agent.intake()
     clock.now += 300
     h.say("use the other file")

nisse-dev/tests/test_loop_multi.py

@@ -95,7 +95,7 @@ def test_the_next_milestone_waits_for_the_merge(world: dict[str, Any]) -> None:
     agent, transport = loop_agent(world, model)
     agent.intake()
     say(transport, world, "go")
-    say(transport, world, "how's it going?")
+    say(transport, world, "carry on")
     assert "hasn't merged yet" in result_of(world, "start_milestone")
     assert (item(world).position, loop_of(item(world)).wait) == (WAITING, MERGE)
     assert len(forge.prs) == 1

nisse-dev/tests/test_loop_properties.py

@@ -76,6 +76,8 @@ class PolicyModel:
             return [call("ask_human", question=f"question {self.n}?")]
         if roll < 0.12:
             return [ToolCall(call_id=f"p{self.n}", tool="nonsense", arguments={})]
+        if "reply" in offered and "propose_work" not in offered:
+            return [call("reply", text=f"because {self.n}")]  # a question run
         if "reply" in offered:  # a conversation: talk, or propose work
             if self.rng.random() < 0.1:
                 return [
@@ -99,7 +101,7 @@ class PolicyModel:
         if "restate" not in offered and "finish_milestone" not in offered:
             return [call("ask_human", question=f"talk first {self.n}?")]  # discuss
         if "finish_milestone" not in offered:
-            return [call("restate", text=f"restatement {self.n}")]
+            return [call("restate", text=f"restatement {self.n}", changed=f"c{self.n}")]
         if "delegate" in offered and self.rng.random() < 0.1:
             return [call("delegate", task=f"help {self.n}")]
         if "wait_for_review" in offered and self.rng.random() < 0.2:
@@ -248,7 +250,7 @@ class Walk:
         forge: Forge = self.worlds[project]["forge"]
         if item and loop:
             weights.update({"go": 3, "answer": 2, "abandon": 1, "go-stale": 1})
-            weights.update({"plain-yes": 2, "wait": 2})
+            weights.update({"plain-yes": 2, "wait": 2, "question": 2})
             if loop.new_project:
                 weights.update({"done": 2})
             if item.scope != "general":  # conversations aren't named by number
@@ -264,6 +266,28 @@ class Walk:
         elif action == "done" and loop:
             # The operator says a new project is ready; no entry was added.
             self.transport.receive(DON, "done", thread=loop.thread)
+        elif action == "question" and loop:
+            # L5: a question never changes the item it's asked about.
+            before = (item.position, loop.wait, loop.rev, loop.pr, loop.branch)
+            self.transport.receive(DON, f"why {len(self.steps)}?", thread=loop.thread)
+            after_item = self.world["store"].work_item(item.id)
+            after = loop_of(after_item)
+            if (
+                item.scope != "general"
+                and item.position == WAITING
+                and loop.wait in (GO_AHEAD, ANSWER, MERGE)
+            ):
+                now = (
+                    after_item.position,
+                    after.wait,
+                    after.rev,
+                    after.pr,
+                    after.branch,
+                )
+                assert now == before, (
+                    f"questions_change_nothing: {before} -> {now}: "
+                    f"seed {self.seed}, steps {self.steps}"
+                )
         elif action == "chat":
             # A top-level message: a conversation of its own (L1, L5).
             self.transport.receive(DON, f"thinking aloud {len(self.steps)}")

nisse-dev/tests/test_loop_talk.py

@@ -0,0 +1,163 @@
+"""Conversation on the new loop (M2.7, L5, #42, #77): questions answered
+without changing the item, revisions that say what changed, progress in
+prose at a readable rate, raw tool activity on request, and a note for a
+message that can't be read yet."""
+
+from pathlib import Path
+from typing import Any
+
+from test_loop import DON, RESTATED, FakeGitea, Harness, ScriptedModel, call, issue
+
+from nisse_dev.canonical import AssistantText
+from nisse_dev.loop.items import ANSWER, GO_AHEAD, WAITING, loop_of
+
+
+class Clock:
+    def __init__(self) -> None:
+        self.now = 1000.0
+
+    def __call__(self) -> float:
+        return self.now
+
+
+def harness(tmp_path: Path, *turns: list[Any], **kw: Any) -> Harness:
+    return Harness(tmp_path, ScriptedModel(*turns), FakeGitea(issue()), **kw)
+
+
+# --- questions never change state (#77) -------------------------------------
+
+
+def test_a_question_about_a_restatement_leaves_it_up(tmp_path: Path) -> None:
+    h = harness(
+        tmp_path,
+        RESTATED,
+        [call("reply", text="Because text.py already holds the string helpers.")],
+        [call("ask_human", question="started; anything else?")],
+    )
+    h.agent.intake()
+    h.say("Why text.py?")
+    loop = loop_of(h.item())
+    assert (h.item().position, loop.wait, loop.rev) == (WAITING, GO_AHEAD, 1)
+    assert not loop.question and not loop.started
+    assert h.posted()[-1] == "Because text.py already holds the string helpers."
+    h.say("go")  # still bound to rev-1
+    assert loop_of(h.item()).started
+
+
+def test_a_question_back_keeps_the_agents_own_question_open(tmp_path: Path) -> None:
+    h = harness(
+        tmp_path,
+        [call("ask_human", question="One function or two?")],
+        [call("reply", text="One handles words, two would split lines too.")],
+    )
+    h.agent.intake()
+    h.say("What would the second one do?")
+    loop = loop_of(h.item())
+    assert (loop.wait, loop.note) == (ANSWER, "One function or two?")
+
+
+def test_a_question_that_isnt_answered_still_goes_back(tmp_path: Path) -> None:
+    h = harness(
+        tmp_path,
+        RESTATED,
+        [call("restate", text="sneaky v2", changed="x")],  # not offered
+        [AssistantText(text="hmm", model="m")],
+        [AssistantText(text="hmm", model="m")],  # prose twice: the run fails
+    )
+    h.agent.intake()
+    h.say("why?")
+    loop = loop_of(h.item())
+    # restate isn't offered in a question run: no new revision.
+    assert (loop.wait, loop.rev) == (GO_AHEAD, 1)
+    assert h.posted()[-1].startswith("I couldn't answer that")
+
+
+# --- explanations ------------------------------------------------------------
+
+
+def test_a_revision_says_what_changed(tmp_path: Path) -> None:
+    h = harness(
+        tmp_path,
+        RESTATED,
+        [call("restate", text="count_words in words.py", changed="")],
+        [
+            call(
+                "restate",
+                text="count_words in words.py",
+                changed="a module of its own, as you asked",
+            )
+        ],
+    )
+    h.agent.intake()
+    h.say("put it in its own module")
+    loop = loop_of(h.item())
+    assert loop.rev == 2
+    assert "Changed from rev-1: a module of its own, as you asked" in h.posted()[-1]
+
+
+# --- progress and details (#42) ----------------------------------------------
+
+
+def test_progress_is_posted_at_a_readable_rate(tmp_path: Path) -> None:
+    clock = Clock()
+
+    class Ticking(ScriptedModel):
+        def complete(self, events, tools, model, system=None):  # type: ignore[no-untyped-def]
+            clock.now += 30  # each turn takes a while
+            return super().complete(events, tools, model, system)
+
+    h = Harness(
+        tmp_path,
+        Ticking(
+            [
+                AssistantText(text="Looking at text.py first.", model="m"),
+                call("list_dir", path="."),
+            ],
+            [
+                AssistantText(text="Reading the tests too.", model="m"),
+                call("list_dir", path="tests"),
+            ],
+            [
+                AssistantText(text="x" * 1000, model="m"),
+                call("list_dir", path="."),
+            ],
+            [call("restate", text="I'll add count_words.", changed="")],
+        ),
+        FakeGitea(issue()),
+        clock=clock,
+    )
+    h.agent.intake()
+    said = h.posted()
+    assert "Looking at text.py first." in said
+    assert "Reading the tests too." not in said  # 30 s later: folded away
+    long = [p for p in said if p.startswith("xxx")]
+    assert long and len(long[0]) < 420 and long[0].endswith(" …")  # 60 s on
+
+
+def test_details_shows_the_last_tool_calls(tmp_path: Path) -> None:
+    h = harness(
+        tmp_path,
+        [
+            call("list_dir", path="src"),
+            call("read_file", path="text.py"),
+            call("restate", text="I'll add count_words.", changed=""),
+        ],
+    )
+    h.agent.intake()
+    h.say("details")
+    shown = h.posted()[-1]
+    assert shown.startswith("The last tool calls on #14:")
+    assert "list_dir(src)" in shown and "read_file(text.py)" in shown
+    assert loop_of(h.item()).wait == GO_AHEAD  # nothing changed
+
+
+def test_a_message_mid_run_is_acknowledged(tmp_path: Path) -> None:
+    h = harness(tmp_path, RESTATED)
+    h.agent.intake()
+    h.agent._running = True  # as if a driver run were under way
+    h.agent._active_thread = "elsewhere"
+    h.agent._active_name = "#15"
+    h.transport.receive(DON, "one more thing", thread=h.thread())
+    assert "Noted; I'm in the middle of #15" in h.posted()[-1]
+    h.transport.receive(DON, "and another", thread="elsewhere")
+    assert h.posted()[-1] == "Noted; I'll read it when this turn ends."

scripts/bringup.sh

@@ -1,13 +1,19 @@
 #!/usr/bin/env bash
 # Bring-up runs of nisse-dev on this development VM (Development Plan M0.2).
 #
-# The agent runs as root, as it will on its own VM in production (D7): it may
-# install what it needs. It touches only test resources (the sandbox repo, the
-# test Matrix room, a budget-capped LiteLLM key), from the run's config. It
-# runs as a transient systemd unit, so it outlives the shell that started it.
+# The agent runs as its container image (M1.5), built from the code under
+# test, as a transient systemd unit, so it outlives the shell that started
+# it. It touches only test resources (an operator-made test repository, the
+# test Matrix room, a budget-capped LiteLLM key), from the run's config.
 #
-#   scripts/bringup.sh start [REF]    set the last run aside, export REF (default
-#                                     HEAD) and start the agent
+# $E2E is mounted at the same path inside the container, so the config's
+# paths (state_db, keys) mean the same there; for that it runs as root here,
+# since the keys are root's. In production the image runs as its own user
+# with a state volume (Dockerfile).
+#
+#   scripts/bringup.sh start [REF]    set the last run aside, build REF (default
+#                                     HEAD) into an image and start the agent
+#   scripts/bringup.sh plan [REF]     print what start would build and run
 #   scripts/bringup.sh stop           stop it cleanly (SIGTERM)
 #   scripts/bringup.sh status         the unit, and the agent's state
 #   scripts/bringup.sh logs           follow the agent's log
@@ -16,7 +22,7 @@
 #
 # Layout, under $E2E (default /root/nisse-e2e):
 #   nisse-dev.yaml, .env    the run config and its secrets (made by hand once)
-#   code/                   the exported code under test, with its own .venv
+#   code/                   the exported code under test, built into the image
 #   state.db                the agent's state (its workspace is on the build VM)
 #   ssh/                    the agent's key for the build VM, and push keys
 #   run-info, collected/    what was run, and the output of `collect`
@@ -27,11 +33,28 @@ E2E=${NISSE_E2E:-/root/nisse-e2e}
 SRC=${NISSE_SRC:-/root/nisse}            # repository to export code from
 CODE=$E2E/code
 UNIT=nisse-bringup
-UV=${UV:-$(command -v uv || echo /snap/bin/uv)}
+IMAGE=nisse-dev:bringup
 
 die() { echo "bringup: $*" >&2; exit 1; }
 running() { systemctl is-active --quiet "$UNIT"; }
 
+# The container the unit runs: host networking (Matrix, Gitea, LiteLLM as
+# here), $E2E at its own path, the config and secrets from it.
+docker_run() {
+    echo docker run --rm --name "$UNIT" --network host --user 0 \
+        -v "$E2E:$E2E" \
+        -e NISSE_CONFIG="$E2E/nisse-dev.yaml" -e NISSE_ENV="$E2E/.env" \
+        "$IMAGE"
+}
+
+cmd_plan() {
+    local ref=${1:-HEAD} sha
+    sha=$(git -C "$SRC" rev-parse --verify "$ref^{commit}") || die "unknown ref $ref"
+    echo "git -C $SRC archive $sha | tar -x -C $CODE"
+    echo "docker build -t $IMAGE $CODE"
+    echo "systemd-run --unit=$UNIT ... $(docker_run)"
+}
+
 cmd_start() {
     local ref=${1:-HEAD} sha
     [ -f "$E2E/nisse-dev.yaml" ] && [ -f "$E2E/.env" ] || die "no config in $E2E"
@@ -58,16 +81,16 @@ cmd_start() {
     rm -rf "$CODE"
     mkdir -p "$CODE"
     git -C "$SRC" archive "$sha" | tar -x -C "$CODE"
-    (cd "$CODE" && "$UV" sync --quiet --all-packages)
+    docker build --quiet -t "$IMAGE" "$CODE" > /dev/null
     printf 'ref: %s\ncommit: %s\nstarted: %s\n' "$ref" "$sha" "$(date -Is)" > "$E2E/run-info"
 
     systemctl reset-failed "$UNIT" 2>/dev/null || true
+    # SIGTERM reaches the agent through docker run's signal proxy; it shuts
+    # down cleanly (main.py).
+    # shellcheck disable=SC2046
     systemd-run --quiet --unit="$UNIT" --collect \
-        --property=WorkingDirectory="$CODE" \
         --property=KillSignal=SIGTERM --property=TimeoutStopSec=120 \
-        --setenv=NISSE_CONFIG="$E2E/nisse-dev.yaml" \
-        --setenv=NISSE_ENV="$E2E/.env" \
-        "$CODE/.venv/bin/nisse-dev"
+        $(docker_run)
     echo "started $UNIT at ${sha:0:8} ($ref)"
 }
 
@@ -78,15 +101,17 @@ remote_aside() {
     where=$(python3 - "$E2E/nisse-dev.yaml" <<'PY'
 import sys, yaml
 c = yaml.safe_load(open(sys.argv[1]))
-b = c["build_host"]
-print(b.get("user", "root"), b["address"], b.get("port", 22), b["key"], c["workspace"])
+for p in c["projects"].values():  # each project's workspace, on its build VM (D8)
+    b = p["build_host"]
+    print(b.get("user", "root"), b["address"], b.get("port", 22), b["key"], p["workspace"])
 PY
-) || die "cannot read build_host and workspace from the config"
-    read -r user address port key workspace <<< "$where"
-    ssh -i "$key" -p "$port" -o BatchMode=yes -o StrictHostKeyChecking=accept-new \
+) || die "cannot read the projects' build VMs and workspaces from the config"
+    while read -r user address port key workspace; do
+        ssh -n -i "$key" -p "$port" -o BatchMode=yes -o StrictHostKeyChecking=accept-new \
             -o UserKnownHostsFile="$E2E/ssh/known_hosts" "$user@$address" \
             "if [ -e '$workspace' ]; then mv '$workspace' '$workspace.$stamp' && echo 'build VM: set aside $workspace.$stamp'; fi" \
             || die "cannot reach the build VM $address"
+    done <<< "$where"
 }
 
 cmd_stop() {
@@ -98,28 +123,35 @@ cmd_stop() {
 cmd_status() {
     if running; then echo "unit: running"; else echo "unit: not running"; fi
     [ -f "$E2E/run-info" ] && sed 's/^/  /' "$E2E/run-info"
-    if [ -f "$E2E/state.db" ] && [ -x "$CODE/.venv/bin/python" ]; then
-        "$CODE/.venv/bin/python" -m nisse_dev.dump "$E2E/state.db" | head -n 3
+    if [ -f "$E2E/state.db" ]; then
+        dump "$E2E/state.db" | head -n 3
     fi
 }
 
 cmd_logs() { journalctl -u "$UNIT" -f -o cat; }
 
+# The state store, read with the image's own code.
+dump() {
+    docker run --rm --user 0 -v "$E2E:$E2E" --entrypoint python "$IMAGE" \
+        -m nisse_dev.dump "$@"
+}
+
 cmd_collect() {
     [ -f "$E2E/run-info" ] || die "no run to collect"
     local out=$E2E/collected since
     since=$(sed -n 's/^started: //p' "$E2E/run-info")
     mkdir -p "$out"
     journalctl -u "$UNIT" --since "$since" -o cat > "$out/agent.log"
-    "$CODE/.venv/bin/python" -m nisse_dev.dump --full "$E2E/state.db" > "$out/transcript.txt"
+    dump --full "$E2E/state.db" > "$out/transcript.txt"
     echo "collected into $out"
 }
 
 case "${1:-}" in
     start) shift; cmd_start "$@" ;;
+    plan) shift; cmd_plan "$@" ;;
     stop) cmd_stop ;;
     status) cmd_status ;;
     logs) cmd_logs ;;
     collect) cmd_collect ;;
-    *) sed -n '2,25p' "$0" | sed 's/^# \{0,1\}//'; exit 2 ;;
+    *) sed -n '2,30p' "$0" | sed 's/^# \{0,1\}//'; exit 2 ;;
 esac

scripts/e2e.sh

@@ -6,6 +6,8 @@
 #
 #   scripts/e2e.sh            on the bring-up build VM (configured below)
 #   scripts/e2e.sh --local    on a throwaway local sshd instead
+#   scripts/e2e.sh --image REF   inside an image's `test` stage (Dockerfile),
+#                                on its own sshd: checks a release as built
 #
 # Uses tests/integration/nisse-test.yaml and its secrets (.env at the repo
 # root), like the integration tests. Every role runs on a local model
@@ -14,6 +16,18 @@
 set -euo pipefail
 cd "$(dirname "$0")/.."
 
+if [ "${1:-}" = "--image" ]; then
+    image=${2:?--image needs an image reference}
+    shift 2
+    runs=${NISSE_E2E_RUNS:-/root/nisse-e2e/test-runs}
+    mkdir -p "$runs"
+    secrets=()
+    [ -f .env ] && secrets=(--env-file .env)
+    # Host networking: the suite reaches Matrix, Gitea and LiteLLM as here.
+    exec docker run --rm --network host "${secrets[@]}" \
+        -e NISSE_E2E_RUNS=/runs -v "$runs:/runs" \
+        "$image" scripts/e2e.sh --local "$@"
+fi
 if [ "${1:-}" = "--local" ]; then
     shift
 else

tests/integration/test_end_to_end_loop.py

@@ -634,3 +634,35 @@ def test_a_conversation_becomes_a_new_project(
         ]
         assert "ARCHITECTURE.md" in files, files
         assert any(f.startswith("docs/plans/") for f in files), files
+
+
+def test_a_question_leaves_the_restatement_up(
+    running: tuple[RunningLoop, InMemoryTransport, str],
+) -> None:
+    """M2.7: a question while a restatement is up is answered, and changes
+    nothing: the same revision stays up, and a go binds to it."""
+    agent, transport, run = running
+    operator = Operator(agent, transport)
+    repo = agent.config.gitea_repo
+    as_operator = operator_gitea(agent.config.gitea_url)
+    issues: list[int] = []
+    try:
+        issues.append(as_operator.create_issue(repo, f"vowels_{run}", task(run)))
+        as_operator._request(
+            "PATCH",
+            f"/repos/{repo}/issues/{issues[0]}",
+            json={"assignees": [agent.account]},
+        )
+        operator.tick()
+        item = operator.drive_until(GO_AHEAD, issues[0])
+        rev = loop_of(item).rev
+        said = len(transport.outbound)
+        operator.say("Why put it in that module rather than a new one?", issues[0])
+        answer = [m.text for m in transport.outbound[said:]]
+        assert answer and not answer[-1].startswith("rev-"), answer
+        loop = loop_of(operator.item(issues[0]))
+        assert (loop.wait, loop.rev) == (GO_AHEAD, rev), loop
+        operator.say("go", issues[0])
+        assert loop_of(operator.item(issues[0])).started
+    finally:
+        operator.tidy(issues, [])