feedback: add run-feedback supervisor (stable docker env + auto-resume)

tools/feedback/discord-feedback

@@ -3213,6 +3213,145 @@ def latest_run_id(artifact_root_dir: pathlib.Path) -> Optional[str]:
     return runs[-1].name
 
 
+def scrape_feedback_sources(
+    repo,
+    artifact_dir,
+    token,
+    token_env,
+    guild_id,
+    channel_ids,
+    include_threads,
+    include_authors,
+    max_messages,
+    hours,
+    pr_feedback,
+    pr_feedback_hours,
+    directions,
+    watch_pr,
+    watch_pr_interval,
+    max_watch_fix_attempts,
+):
+    """Scrape Discord + recent maintainer PR feedback into the run's artifact dir.
+
+    Returns (transcript_path, jsonl_path, pr_feedback_path). Only the analysis agent
+    consumes these, so a resume whose analysis is already checkpointed skips this entirely
+    (see the caller) -- no Discord token, no minutes-long re-scrape on every resume/retry.
+    """
+    env_channel_ids = os.environ.get("PWN_FEEDBACK_DISCORD_CHANNEL_IDS")
+    selected_channel_ids = parse_channel_ids(channel_ids, env_channel_ids)
+    until = utc_now()
+    since = resolve_scrape_since(repo, until, hours, DEFAULT_FALLBACK_HOURS)
+
+    if not token:
+        raise click.ClickException(
+            f"{token_env} is not set. Create a Discord bot token and export it first."
+        )
+    api = DiscordAPI(token)
+    click.echo(f"Scraping Discord messages since {since.isoformat()} into {artifact_dir}")
+    channels, channels_by_id = discover_channels(
+        api, guild_id, selected_channel_ids, since, include_threads
+    )
+    click.echo(f"Discovered {len(channels)} readable channel(s)/thread(s)")
+    messages_by_channel: list[tuple[dict[str, Any], list[dict[str, Any]]]] = []
+    remaining = max_messages
+    for channel in channels:
+        if remaining <= 0:
+            break
+        messages = fetch_messages_since(api, channel, since, remaining)
+        remaining -= len(messages)
+        if messages:
+            messages_by_channel.append((channel, messages))
+            click.echo(
+                "Fetched "
+                f"{len(messages)} message(s) from "
+                f"{channel_display_name(channel, channels_by_id)}"
+            )
+            logger.info(
+                "fetched %d message(s) from %s",
+                len(messages),
+                channel_display_name(channel, channels_by_id),
+            )
+
+    messages = sanitize_messages(
+        messages_by_channel,
+        channels_by_id,
+        guild_id,
+        include_authors,
+        attachment_dir=artifact_dir / "attachments",
+    )
+    image_count = sum(
+        1
+        for message in messages
+        for attachment in message["attachments"]
+        if attachment.get("path")
+    )
+    if image_count:
+        click.echo(
+            f"Downloaded {image_count} image attachment(s) into "
+            f"{artifact_dir / 'attachments'}"
+        )
+    jsonl_path, transcript_path = write_transcript(artifact_dir, messages, since, until)
+
+    pr_feedback_path: Optional[pathlib.Path] = None
+    pr_feedback_prs = 0
+    pr_feedback_events = 0
+    if pr_feedback:
+        pr_since = until - datetime.timedelta(hours=pr_feedback_hours)
+        try:
+            api = GitHubAPI.from_environment(repo)
+            owner, repo_name = github_repo_slug(repo)
+            recent_pulls = fetch_recent_pull_requests(
+                api, owner, repo_name, pr_since.isoformat()
+            )
+            maintainer_feedback = collect_maintainer_pr_feedback(
+                api, owner, repo_name, recent_pulls, {}
+            )
+            pr_feedback_path = write_pr_feedback(
+                artifact_dir, maintainer_feedback, pr_since, until
+            )
+            pr_feedback_prs = len(maintainer_feedback)
+            pr_feedback_events = sum(
+                len(pull["events"]) for pull in maintainer_feedback
+            )
+            click.echo(
+                "Collected maintainer PR feedback: "
+                f"{pr_feedback_events} comment(s)/review(s) across "
+                f"{pr_feedback_prs} recent PR(s)"
+            )
+        except (click.ClickException, GitHubAPIError, OSError) as error:
+            click.echo(f"Skipping PR feedback collection: {error}")
+            pr_feedback_path = None
+
+    (artifact_dir / "run.json").write_text(
+        json.dumps(
+            {
+                "guild_id": guild_id,
+                "hours": hours,
+                "since": since.isoformat(),
+                "until": until.isoformat(),
+                "messages": len(messages),
+                "image_attachments": image_count,
+                "channels": len(channels),
+                "include_threads": include_threads,
+                "include_authors": include_authors,
+                "directions": directions,
+                "pr_feedback": pr_feedback,
+                "pr_feedback_hours": pr_feedback_hours,
+                "pr_feedback_prs": pr_feedback_prs,
+                "pr_feedback_events": pr_feedback_events,
+                "watch_pr": watch_pr,
+                "watch_pr_interval": watch_pr_interval,
+                "max_watch_fix_attempts": max_watch_fix_attempts,
+            },
+            indent=2,
+            sort_keys=True,
+        )
+        + "\n"
+    )
+    click.echo(f"Wrote {len(messages)} sanitized message(s) to {transcript_path}")
+    return transcript_path, jsonl_path, pr_feedback_path
+
+
 @click.command("discord-feedback")
 @click.option(
     "--guild-id",
@@ -3480,118 +3619,40 @@ def feedback_command(
             "falling through to the normal flow."
         )
 
-    env_channel_ids = os.environ.get("PWN_FEEDBACK_DISCORD_CHANNEL_IDS")
-    selected_channel_ids = parse_channel_ids(channel_ids, env_channel_ids)
-    until = utc_now()
-    since = resolve_scrape_since(repo, until, hours, DEFAULT_FALLBACK_HOURS)
-
-    if not token:
-        raise click.ClickException(
-            f"{token_env} is not set. Create a Discord bot token and export it first."
+    # The scrape only feeds the analysis agent. On resume with analysis already
+    # checkpointed, skip it entirely -- no Discord token, no minutes-long re-scrape on every
+    # resume/retry, no misleading "scraping from PR #N" line -- and reuse the PR feedback
+    # already recorded for this run.
+    if phase_done("analysis"):
+        click.echo("Resume: analysis already complete; skipping Discord scrape.")
+        recorded_pr_feedback = artifact_dir / "recent-pr-feedback.md"
+        pr_feedback_path = (
+            recorded_pr_feedback
+            if pr_feedback and recorded_pr_feedback.exists()
+            else None
         )
-    api = DiscordAPI(token)
-    click.echo(f"Scraping Discord messages since {since.isoformat()} into {artifact_dir}")
-    channels, channels_by_id = discover_channels(
-        api, guild_id, selected_channel_ids, since, include_threads
-    )
-    click.echo(f"Discovered {len(channels)} readable channel(s)/thread(s)")
-    messages_by_channel: list[tuple[dict[str, Any], list[dict[str, Any]]]] = []
-    remaining = max_messages
-    for channel in channels:
-        if remaining <= 0:
-            break
-        messages = fetch_messages_since(api, channel, since, remaining)
-        remaining -= len(messages)
-        if messages:
-            messages_by_channel.append((channel, messages))
-            click.echo(
-                "Fetched "
-                f"{len(messages)} message(s) from "
-                f"{channel_display_name(channel, channels_by_id)}"
-            )
-            logger.info(
-                "fetched %d message(s) from %s",
-                len(messages),
-                channel_display_name(channel, channels_by_id),
-            )
-
-    messages = sanitize_messages(
-        messages_by_channel,
-        channels_by_id,
-        guild_id,
-        include_authors,
-        attachment_dir=artifact_dir / "attachments",
-    )
-    image_count = sum(
-        1
-        for message in messages
-        for attachment in message["attachments"]
-        if attachment.get("path")
-    )
-    if image_count:
-        click.echo(
-            f"Downloaded {image_count} image attachment(s) into "
-            f"{artifact_dir / 'attachments'}"
+        transcript_path = artifact_dir / "transcript.md"
+        jsonl_path = artifact_dir / "messages.jsonl"
+    else:
+        transcript_path, jsonl_path, pr_feedback_path = scrape_feedback_sources(
+            repo,
+            artifact_dir,
+            token,
+            token_env,
+            guild_id,
+            channel_ids,
+            include_threads,
+            include_authors,
+            max_messages,
+            hours,
+            pr_feedback,
+            pr_feedback_hours,
+            directions,
+            watch_pr,
+            watch_pr_interval,
+            max_watch_fix_attempts,
         )
-    jsonl_path, transcript_path = write_transcript(artifact_dir, messages, since, until)
-
-    pr_feedback_path: Optional[pathlib.Path] = None
-    pr_feedback_prs = 0
-    pr_feedback_events = 0
-    if pr_feedback:
-        pr_since = until - datetime.timedelta(hours=pr_feedback_hours)
-        try:
-            api = GitHubAPI.from_environment(repo)
-            owner, repo_name = github_repo_slug(repo)
-            recent_pulls = fetch_recent_pull_requests(
-                api, owner, repo_name, pr_since.isoformat()
-            )
-            maintainer_feedback = collect_maintainer_pr_feedback(
-                api, owner, repo_name, recent_pulls, {}
-            )
-            pr_feedback_path = write_pr_feedback(
-                artifact_dir, maintainer_feedback, pr_since, until
-            )
-            pr_feedback_prs = len(maintainer_feedback)
-            pr_feedback_events = sum(
-                len(pull["events"]) for pull in maintainer_feedback
-            )
-            click.echo(
-                "Collected maintainer PR feedback: "
-                f"{pr_feedback_events} comment(s)/review(s) across "
-                f"{pr_feedback_prs} recent PR(s)"
-            )
-        except (click.ClickException, GitHubAPIError, OSError) as error:
-            click.echo(f"Skipping PR feedback collection: {error}")
-            pr_feedback_path = None
 
-    (artifact_dir / "run.json").write_text(
-        json.dumps(
-            {
-                "guild_id": guild_id,
-                "hours": hours,
-                "since": since.isoformat(),
-                "until": until.isoformat(),
-                "messages": len(messages),
-                "image_attachments": image_count,
-                "channels": len(channels),
-                "include_threads": include_threads,
-                "include_authors": include_authors,
-                "directions": directions,
-                "pr_feedback": pr_feedback,
-                "pr_feedback_hours": pr_feedback_hours,
-                "pr_feedback_prs": pr_feedback_prs,
-                "pr_feedback_events": pr_feedback_events,
-                "watch_pr": watch_pr,
-                "watch_pr_interval": watch_pr_interval,
-                "max_watch_fix_attempts": max_watch_fix_attempts,
-            },
-            indent=2,
-            sort_keys=True,
-        )
-        + "\n"
-    )
-    click.echo(f"Wrote {len(messages)} sanitized message(s) to {transcript_path}")
     if fetch_only:
         return
 

tools/feedback/run-feedback

@@ -0,0 +1,102 @@
+#!/usr/bin/env bash
+#
+# Supervisor for the discord-feedback bot.
+#
+# Two failure modes have repeatedly left the bot's screen window dead at a shell prompt:
+#
+#   1. It was launched from a plain shell (not `nix develop`), so DOCKER_HOST was unset and
+#      pwnshop fell through to the host docker -- whose docker0 bridge is broken -- making
+#      every challenge "fail" to build with "adding interface veth... to bridge docker0
+#      failed: Device does not exist". Nothing is actually wrong with the challenges.
+#   2. A transient hiccup (docker, a GitHub blip, an OOM) killed the bot mid-run, and it
+#      stayed dead until someone noticed and hand-typed the resume command.
+#
+# This wrapper makes both impossible:
+#
+#   * It runs the bot inside ONE persistent `nix develop` session. The dev shell's shellHook
+#     starts the project-local dockerd and exports DOCKER_HOST at it exactly once, and it
+#     stays up for the whole run -- so the bot always has a correct, stable docker and can
+#     never fall through to the broken host daemon. (It re-execs itself into `nix develop`
+#     if it isn't already inside one.)
+#   * It runs the bot in a loop: on any nonzero exit it waits (capped backoff) and resumes
+#     from the last checkpoint with --resume-latest. A transient failure self-heals on the
+#     next iteration instead of leaving the window dead. It stops only on a clean exit
+#     (the PR merged / work done) or Ctrl-C.
+#
+# Usage:
+#   tools/feedback/run-feedback [bot args...]
+#
+# The bot args are passed verbatim to the FIRST iteration (use --resume <id> to pick up a
+# specific run, or pass fresh-run flags to start one). Every retry afterwards forces
+# --resume-latest so it continues whatever run the first iteration advanced. With no args it
+# defaults to the standard apply/create-pr/watch invocation and resumes the latest run.
+
+set -uo pipefail
+
+REPO="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
+cd "$REPO"
+
+# Re-exec inside a single persistent dev shell if we aren't already in one. `nix develop
+# --command` runs the shellHook first (starting dockerd + exporting DOCKER_HOST), then this
+# same script with IN_NIX_SHELL set, so the branch below runs the supervised loop with a
+# stable docker for its entire lifetime.
+if [ -z "${IN_NIX_SHELL:-}" ]; then
+    echo "run-feedback: entering nix develop (starts project dockerd, sets DOCKER_HOST)..."
+    exec nix develop --command "$0" "$@"
+fi
+
+BOT="${FEEDBACK_BOT:-$REPO/.discord-feedback/discord-feedback}"
+if [ ! -x "$BOT" ]; then
+    BOT="$REPO/tools/feedback/discord-feedback"
+fi
+
+# Default first-iteration args if the caller passed none.
+DEFAULT_ARGS=(--apply --create-pr --resume-latest)
+if [ "$#" -gt 0 ]; then
+    first_args=("$@")
+else
+    first_args=("${DEFAULT_ARGS[@]}")
+fi
+
+# Retry args: strip any caller-supplied --resume <id> / --resume-latest, then force
+# --resume-latest so every retry continues the run the first iteration advanced.
+retry_args=()
+skip_next=0
+for arg in "${first_args[@]}"; do
+    if [ "$skip_next" = 1 ]; then skip_next=0; continue; fi
+    case "$arg" in
+        --resume) skip_next=1; continue ;;          # drop "--resume <id>"
+        --resume=*) continue ;;
+        --resume-latest) continue ;;
+        *) retry_args+=("$arg") ;;
+    esac
+done
+retry_args+=(--resume-latest)
+
+echo "run-feedback: DOCKER_HOST=${DOCKER_HOST:-<unset>}"
+echo "run-feedback: bot=$BOT"
+
+attempt=0
+args=("${first_args[@]}")
+while true; do
+    attempt=$((attempt + 1))
+    echo "==================== run-feedback iteration $attempt ===================="
+    echo "run-feedback: $BOT ${args[*]}"
+    "$BOT" "${args[@]}"
+    rc=$?
+    if [ "$rc" -eq 0 ]; then
+        echo "run-feedback: bot exited cleanly (work done); stopping."
+        break
+    fi
+    if [ "$rc" -eq 130 ]; then
+        echo "run-feedback: interrupted (Ctrl-C); stopping."
+        break
+    fi
+    # Capped linear backoff: 30s, 60s, ... up to 300s. A transient docker/API failure clears
+    # within this window; a persistent one keeps retrying (loudly) rather than dying.
+    delay=$((attempt * 30))
+    [ "$delay" -gt 300 ] && delay=300
+    echo "run-feedback: bot exited $rc; resuming from checkpoint in ${delay}s (next: --resume-latest)."
+    sleep "$delay"
+    args=("${retry_args[@]}")
+done