bot-coordinator: Run Opus 5.5 at high effort

bot/SETUP.md

@@ -312,6 +312,10 @@ homegit's skills rendered for this operator, with the goals in
 `~/.config/bot-harness/goals.md`:
 
     install -m 644 bot/coordinator/operator-goals.md ~/.config/bot-harness/goals.md
+    install -m 644 bot/coordinator/operator-notes.md ~/.config/bot-harness/notes.md
+
+The notes also tell it when to hand reviews and well-defined work to
+opencode (`bot-opencode`, with the Copilot login from §5).
 
     podman build -t localhost/bot-coordinator bot/coordinator/
     mkdir -p ~/.config/containers/systemd

bot/coordinator/Containerfile

@@ -6,7 +6,8 @@ FROM localhost/bootc-bot-agent:latest
 
 COPY --chmod=755 entrypoint.sh /usr/local/bin/bot-coordinator-entrypoint
 COPY --chmod=755 render-skills.py /usr/local/bin/render-skills
-COPY operator-goals.md /usr/local/share/bot-coordinator/operator-goals.md
+COPY operator-goals.md operator-notes.md /usr/local/share/bot-coordinator/
+COPY --chmod=755 bot-opencode /usr/local/bin/bot-opencode
 ENV PATH=/home/agent/.local/bin:/home/agent/.cargo/bin:/usr/local/bin:/usr/bin:/bin \
     BOT_LOGIN=jmarrero-bot \
     HOMEGIT_URL=https://github.com/jmarrero-forge/homegit.git

bot/coordinator/README.md

@@ -45,6 +45,19 @@ writes it, so the fork is easy to update. If a homegit change moves a
 section the renderer replaces, it stops with an error rather than render
 his goals as yours; `bot/test_render.py` checks the result.
 
+## opencode for reviews and well-defined work
+
+The coordinator also has `bot-opencode`, which runs one task with opencode
+on a GitHub Copilot model (`BOT_OPENCODE_MODEL`, with the
+`jmarrero-bot-copilot-auth` secret; usage counts against the license
+holder's premium requests). The operator's notes
+(`~/.config/bot-harness/notes.md`, default
+[`operator-notes.md`](operator-notes.md), appended to the rendered
+coordinator skill) say when: a second, different-model review of every
+forge PR, and items well-defined enough to need no design choices. Edit the
+notes and restart the coordinator to change that. Runs are logged in
+`~/.local/state/bot-opencode/` in the container.
+
 ## Use
 
     podman build -t localhost/bot-coordinator bot/coordinator/
@@ -59,7 +72,8 @@ the bot."); comment it out to have it wait for you. If the Claude session
 ends (`/exit`, or by accident), systemd restarts it after 60 seconds; stop it
 for good with `systemctl --user stop bot-coordinator`. Attach to watch or
 talk to it, and detach with Ctrl-b d. Its usage counts against the Claude
-Max subscription.
+Max subscription. It runs `claude-opus-5-5` at `high` effort
+(`COORDINATOR_MODEL`, `COORDINATOR_EFFORT` in the `.container` file).
 
 The coordinator only hears of news while `bot-poll` runs in its background,
 and it sometimes stops it and doesn't start it again. The entrypoint's

bot/coordinator/bot-coordinator.container

@@ -12,6 +12,9 @@ Image=localhost/bot-coordinator:latest
 # Credentials as podman secrets (see bot/SETUP.md), never in the image.
 Secret=jmarrero-bot-gh-token,type=env,target=GH_TOKEN
 Secret=jmarrero-bot-claude-token,type=env,target=CLAUDE_CODE_OAUTH_TOKEN
+# The Copilot login for bot-opencode (bot/copilot-login.sh), and its model.
+Secret=jmarrero-bot-copilot-auth,type=env,target=OPENCODE_AUTH_CONTENT
+Environment=BOT_OPENCODE_MODEL=github-copilot/gpt-6.1-sol
 # The persistent home: clones, Claude's sessions, caches, devspace keys.
 # No keep-id: the container's users map to subordinate ids, not to you.
 Volume=bot-coordinator-home:/home/agent
@@ -21,6 +24,10 @@ Volume=%h/.config/bot-harness:/home/agent/.config/bot-harness:ro,z
 # botd answers "@BOT <command> ..." at the start of a line; bot-notify leaves
 # those to it. Keep in step with botd's agents ([agents.*]) plus help.
 Environment="BOT_NOTIFY_SKIP_COMMANDS=claude opencode help"
+# The coordinator's model, pinned by full name (an alias like "opus" moves
+# with new releases), and its effort (low, medium, high, xhigh, max).
+Environment=COORDINATOR_MODEL=claude-opus-5-5
+Environment=COORDINATOR_EFFORT=high
 # Network through the host (pasta), which is on the tailnet.
 PodmanArgs=--memory=16g --cpus=4 --pids-limit=8192
 NoNewPrivileges=true

bot/coordinator/bot-opencode

@@ -0,0 +1,43 @@
+#!/bin/bash
+# Run one task with opencode on a GitHub Copilot model, for the coordinator:
+# a second model family for reviews, and well-defined work. See the operator
+# notes in the coordinator skill for when to use it.
+#
+#   bot-opencode [--model PROVIDER/MODEL] [--dir DIR] [--name NAME] PROMPT
+#
+# Runs `opencode run --auto` in DIR (default: the current directory) and
+# prints its final output; the whole run is kept in
+# ~/.local/state/bot-opencode/NAME-TIMESTAMP.log. Start it with the Bash
+# tool's run_in_background, so its exit wakes the session. The login comes
+# from OPENCODE_AUTH_CONTENT (a podman secret); the model from --model or
+# BOT_OPENCODE_MODEL.
+set -euo pipefail
+
+model=${BOT_OPENCODE_MODEL:-github-copilot/gpt-6.1-sol}
+dir=$PWD
+name=task
+while [ $# -gt 1 ]; do
+    case $1 in
+        --model) model=$2; shift 2 ;;
+        --dir) dir=$2; shift 2 ;;
+        --name) name=$2; shift 2 ;;
+        --) shift; break ;;
+        -*) echo "bot-opencode: unknown option $1" >&2; exit 2 ;;
+        *) break ;;
+    esac
+done
+[ $# -eq 1 ] || { sed -n '6,7p' "$0" | sed 's/^# //' >&2; exit 2; }
+: "${OPENCODE_AUTH_CONTENT:?no Copilot login (OPENCODE_AUTH_CONTENT)}"
+[[ "$name" =~ ^[A-Za-z0-9._-]+$ ]] || { echo "bot-opencode: bad --name" >&2; exit 2; }
+
+logdir=$HOME/.local/state/bot-opencode
+mkdir -p "$logdir"
+log=$logdir/$name-$(date -u +%Y%m%dT%H%M%SZ).log
+echo "bot-opencode: $model in $dir; log $log" >&2
+cd "$dir"
+set +e
+opencode run --auto -m "$model" "$1" > >(tee "$log") 2>>"$log"
+rc=$?
+set -e
+echo "bot-opencode: exit $rc (log $log)" >&2
+exit $rc

bot/coordinator/entrypoint.sh

@@ -29,8 +29,10 @@ fi
 # the model sometimes copies). Rendered again on every start, after the pull.
 goals=$HOME/.config/bot-harness/goals.md
 test -s "$goals" || goals=/usr/local/share/bot-coordinator/operator-goals.md
+notes=$HOME/.config/bot-harness/notes.md
+test -s "$notes" || notes=/usr/local/share/bot-coordinator/operator-notes.md
 rendered=$HOME/.local/share/bot-skills
-render-skills "$homegit" "$rendered" "$goals"
+render-skills "$homegit" "$rendered" "$goals" "$notes"
 ln -sfn "$rendered/skills" "$HOME/.agents/skills"
 rm -f "$HOME"/.claude/skills/*
 for skill in "$rendered"/skills/*/; do
@@ -71,6 +73,10 @@ fi
 # are skipped (it runs unattended, in this container); COORDINATOR_PROMPT,
 # if set, is the first message, e.g. "Load the coordinator skill and run".
 cmd=(claude --dangerously-skip-permissions)
+# Its model and effort (COORDINATOR_MODEL, COORDINATOR_EFFORT); unset means
+# Claude Code's defaults. Subagents inherit the model unless they set one.
+[ -n "${COORDINATOR_MODEL:-}" ] && cmd+=(--model "$COORDINATOR_MODEL")
+[ -n "${COORDINATOR_EFFORT:-}" ] && cmd+=(--effort "$COORDINATOR_EFFORT")
 [ -n "${COORDINATOR_PROMPT:-}" ] && cmd+=("$COORDINATOR_PROMPT")
 cd "$clones"
 tmux new-session -d -s coordinator -x 200 -y 50 "$(printf '%q ' "${cmd[@]}")"

bot/coordinator/operator-notes.md

@@ -0,0 +1,46 @@
+## Operator's notes
+
+These come from jmarrero (`~/.config/bot-harness/notes.md`) and add to the
+rest of this skill.
+
+### After a restart
+
+A session that ends takes its workers with it: subagents and `bot-opencode`
+runs die mid-task, while their items stay In Progress on the board. So at
+the start of every session, right after the first `bot-poll --once`, run
+`bot-board list --status "In Progress"` (and check the board for items
+without a priority too). For each item, look at its PR, branch and Why to
+see how far it got, then resume it with a new worker (or `bot-opencode`)
+briefed with that state, or set it to Needs human with the reason. Say
+which in your start-up summary. Never leave an item In Progress without a
+worker on it.
+
+### Second opinions and well-defined work with opencode
+
+`bot-opencode` runs one task with opencode on a GitHub Copilot model (a
+different model family from yours; usage counts against jmarrero's Copilot
+premium requests). Always start it with the Bash tool's
+`run_in_background`, so its exit wakes you, and give it `--name` (e.g. the
+item or PR) and `--dir` (a worktree of its own). Run at most two at once.
+
+- **Reviews.** Every forge PR also gets a review from opencode, besides the
+  reviewer subagent: brief it like a reviewer, starting with "Read
+  ~/.agents/skills/coordinator/reviewer-preamble.md and follow it", plus the
+  PR URL and head SHA, and have it post its verdict as a review comment on
+  the PR (it has the bot's `gh` login). Weigh both verdicts; when they
+  disagree, say so on the PR and in the item's Why, and treat the stricter
+  one as standing until resolved.
+- **Well-defined work.** An item is well-defined when its issue says what to
+  change and how to check it, with no design choices left: dependency
+  bumps, mechanical refactors and renames, a test for a described case, a
+  clear compiler, clippy or CI fix, docs that follow a stated change. Send
+  those to opencode instead of a Claude worker: brief it starting with
+  "Read ~/.agents/skills/coordinator/worker-preamble.md and follow it",
+  plus the item URL, the repository and the worktree. Everything else
+  (design, ambiguous asks, debugging an unknown cause, anything touching
+  sign-off, credentials or the sandbox) stays with Claude workers.
+- Its result gets the same review as any worker's: a reviewer subagent, and
+  you before it reaches jmarrero. If it fails or wanders off, hand the item
+  to a Claude worker and note it in the Why field.
+- The default model is `BOT_OPENCODE_MODEL`; `--model` overrides it
+  (`opencode models github-copilot` lists them).

bot/coordinator/render-skills.py

@@ -8,11 +8,12 @@
 link), so the coordinator gets copies with this operator's values filled in.
 homegit stays as cgwalters writes it, which keeps the fork easy to update.
 
-  render-skills.py HOMEGIT DEST GOALS
+  render-skills.py HOMEGIT DEST GOALS [NOTES]
 
 writes DEST/skills/<name>/... (from dotfiles/.agents/skills), DEST/AGENTS.md
 and DEST/agents/builder.md, with values from `bot-operator --json`. GOALS is
-a Markdown file replacing cgwalters' goals and priority rules. What still
+a Markdown file replacing cgwalters' goals and priority rules; NOTES, the
+operator's own notes, is appended to the coordinator skill. What still
 names cgwalters afterwards (links to his own repositories, issues and the
 coordination channel, kept on purpose) is listed on stderr.
 """
@@ -137,6 +138,7 @@ def stash(m):
 
 def main():
     homegit, dest, goals_file = sys.argv[1:4]
+    notes = open(sys.argv[4]).read() if len(sys.argv) > 4 else ""
     cfg = json.loads(subprocess.run(["bot-operator", "--json"], check=True,
                                     capture_output=True, text=True).stdout)
     goals = open(goals_file).read()
@@ -160,6 +162,9 @@ def main():
             sys.exit(f"render-skills: expected one match of {pattern[:60]!r} in {rel}, found {n}")
         open(path, "w").write(new)
 
+    if notes.strip():
+        # After rendering, so the operator's text is taken as written.
+        notes_path = os.path.join(tmp, "skills/coordinator/SKILL.md")
     leftovers = []
     for root, _, files in os.walk(tmp):
         for name in files:
@@ -172,6 +177,9 @@ def main():
                 if re.search(r"(?i)cgwalters|colin|verbum", line):
                     leftovers.append(f"{os.path.relpath(path, tmp)}:{i}: {line.strip()[:120]}")
 
+    if notes.strip():
+        with open(notes_path, "a") as f:
+            f.write("\n" + notes.strip() + "\n")
     shutil.rmtree(dest, ignore_errors=True)
     os.rename(tmp, dest)
     print(f"render-skills: rendered for {cfg['operator']['login']}/{cfg['bot']['login']}; "

bot/test_render.py

@@ -11,6 +11,7 @@
 HERE = os.path.dirname(os.path.abspath(__file__))
 RENDER = os.path.join(HERE, "coordinator", "render-skills.py")
 GOALS = os.path.join(HERE, "coordinator", "operator-goals.md")
+NOTES = os.path.join(HERE, "coordinator", "operator-notes.md")
 HOMEGIT = os.environ.get("HOMEGIT", os.path.expanduser("~/development/github/jmarrero-forge/homegit"))
 CONFIG = json.load(open(os.path.join(HERE, "operator.json")))
 
@@ -30,7 +31,7 @@ def setUpClass(cls):
         os.chmod(stub, stat.S_IRWXU)
         cls.out = os.path.join(cls.tmp, "rendered")
         env = dict(os.environ, PATH=f"{cls.tmp}:{os.environ['PATH']}")
-        cls.proc = subprocess.run([sys.executable, RENDER, HOMEGIT, cls.out, GOALS],
+        cls.proc = subprocess.run([sys.executable, RENDER, HOMEGIT, cls.out, GOALS, NOTES],
                                   env=env, capture_output=True, text=True)
 
     def read(self, rel):
@@ -69,6 +70,11 @@ def test_peers_are_cgwalters_harness(self):
         self.assertIn("coordination questions from cgwalters-bot and cgwalters", self.read("skills/bot-notify/SKILL.md"))
         self.assertIn("the channel with cgwalters' harness", self.read("skills/coordinator/worker-preamble.md"))
 
+    def test_operator_notes_are_appended(self):
+        coordinator = self.read("skills/coordinator/SKILL.md")
+        self.assertTrue(coordinator.rstrip().endswith(open(NOTES).read().strip()))
+        self.assertIn("bot-opencode", coordinator)
+
     def test_paths_point_at_the_rendered_skills(self):
         self.assertIn("~/.agents/skills/coordinator/", self.read("skills/coordinator/SKILL.md"))
         self.assertNotIn("homegit/dotfiles/.agents/skills", self.read("skills/coordinator/SKILL.md"))