kanban-orchestrator

name	kanban-orchestrator
description	Decomposition playbook + anti-temptation rules for an orchestrator profile routing work through Kanban. The "don't do the work yourself" rule and the basic lifecycle are auto-injected into every kanban worker's system prompt; this skill is the deeper playbook when you're specifically playing the orchestrator role.
version	3.0.0
platforms	["linux","macos","windows"]
metadata	{"hermes":{"tags":["kanban","multi-agent","orchestration","routing"],"related_skills":["kanban-worker"]}}

Kanban Orchestrator — Decomposition Playbook

The core worker lifecycle (including the kanban_create fan-out pattern and the "decompose, don't execute" rule) is auto-injected into every kanban process via the KANBAN_GUIDANCE system-prompt block. This skill is the deeper playbook when you're an orchestrator profile whose whole job is routing.

Profiles are user-configured — not a fixed roster

Hermes setups vary widely. Some users run a single profile that does everything; some run a small fleet (docker-worker, cron-worker); some run a curated specialist team they've named themselves. There is no default specialist roster — the orchestrator skill does not know what profiles exist on this machine.

Before fanning out, you must ground the decomposition in the profiles that actually exist. The dispatcher silently fails to spawn unknown assignee names — it doesn't autocorrect, doesn't suggest, doesn't fall back. So a card assigned to researcher on a setup that only has docker-worker just sits in ready forever.

Step 0: discover available profiles before planning.

Use one of these:

• hermes profile list — prints the table of profiles configured on this machine. Run it through your terminal tool if you have one; otherwise ask the user.
• kanban_list(assignee="<some-name>") — sanity-check a single name. Returns an empty list (rather than an error) for an unknown assignee, so this only confirms a name you're already considering.
• Just ask the user. "What profiles do you have set up?" is a fine first turn when the goal needs more than one specialist.

Cache the result in your working memory for the rest of the conversation. Re-asking every turn wastes a tool call.

When to use the board (vs. just doing the work)

Create Kanban tasks when any of these are true:

1. Multiple specialists are needed. Research + analysis + writing is three profiles.
2. The work should survive a crash or restart. Long-running, recurring, or important.
3. The user might want to interject. Human-in-the-loop at any step.
4. Multiple subtasks can run in parallel. Fan-out for speed.
5. Review / iteration is expected. A reviewer profile loops on drafter output.
6. The audit trail matters. Board rows persist in SQLite forever.

If none of those apply — it's a small one-shot reasoning task — use delegate_task instead or answer the user directly.

The anti-temptation rules

Your job description says "route, don't execute." The rules that enforce that:

• Do not execute the work yourself. Your restricted toolset usually doesn't even include terminal/file/code/web for implementation. If you find yourself "just fixing this quickly" — stop and create a task for the right specialist.
• For any concrete task, create a Kanban task and assign it. Every single time.
• Split multi-lane requests before creating cards. A user prompt can contain several independent workstreams. Extract those lanes first, then create one card per lane instead of bundling unrelated work into a single implementer card.
• Run independent lanes in parallel. If two cards do not need each other's output, leave them unlinked so the dispatcher can fan them out. Link only true data dependencies.
• Never create dependent work as independent ready cards. If a card must wait for another card, pass parents=[...] in the original kanban_create call. Do not create it first and link it later, and do not rely on prose like "wait for T1" inside the body.
• If no specialist fits the available profiles, ask the user which profile to create or which existing profile to use. Do not invent profile names; the dispatcher will silently drop unknown assignees.
• Decompose, route, and summarize — that's the whole job.

Decomposition playbook

Step 1 — Understand the goal

Ask clarifying questions if the goal is ambiguous. Cheap to ask; expensive to spawn the wrong fleet.

Step 2 — Sketch the task graph

Before creating anything, draft the graph out loud (in your response to the user). Treat every concrete workstream as a candidate card:

1. Extract the lanes from the request.
2. Map each lane to one of the profiles you discovered in Step 0. If a lane doesn't fit any existing profile, ask the user which to use or create.
3. Decide whether each lane is independent or gated by another lane.
4. Create independent lanes as parallel cards with no parent links.
5. Create synthesis/review/integration cards with parent links to the lanes they depend on. A child created with unfinished parents starts in todo; the dispatcher promotes it to ready only after every parent is done.

Examples of prompts that should fan out (using placeholder profile names — substitute whatever exists on the user's setup):

• "Build an app" → one card to a design-oriented profile for product/UI direction, one or two cards to engineering profiles for implementation, plus a later integration/review card if the user has a reviewer profile.
• "Fix blockers and check model variants" → one implementation card for the blocker fixes plus one discovery/research card for config/source verification. A final reviewer card can depend on both.
• "Research docs and implement" → a docs-research card can run in parallel with a codebase-discovery card; implementation waits only if it truly needs those findings.
• "Analyze this screenshot and find the related code" → one card to a vision-capable profile for the visual analysis while another searches the codebase.

Words like "also," "finally," or "and" do not automatically imply a dependency. They often mean "make sure this is covered before reporting back." Only link tasks when one card cannot start until another card's output exists.

Show the graph to the user before creating cards. Let them correct it — including which actual profile name should own each lane.

Step 3 — Create tasks and link

Use the profile names from Step 0. The example below uses placeholders <profile-A>, <profile-B>, <profile-C> — replace them with what the user actually has.

t1 = kanban_create(
    title="research: Postgres cost vs current",
    assignee="<profile-A>",  # whichever profile handles research on this setup
    body="Compare estimated infrastructure costs, migration costs, and ongoing ops costs over a 3-year window. Sources: AWS/GCP pricing, team time estimates, current Postgres bills from peers.",
    tenant=os.environ.get("HERMES_TENANT"),
)["task_id"]

t2 = kanban_create(
    title="research: Postgres performance vs current",
    assignee="<profile-A>",  # same profile, run in parallel
    body="Compare query latency, throughput, and scaling characteristics at our expected data volume (~500GB, 10k QPS peak). Sources: benchmark papers, public case studies, pgbench results if easy.",
)["task_id"]

t3 = kanban_create(
    title="synthesize migration recommendation",
    assignee="<profile-B>",  # whichever profile does synthesis/analysis
    body="Read the findings from T1 (cost) and T2 (performance). Produce a 1-page recommendation with explicit trade-offs and a go/no-go call.",
    parents=[t1, t2],
)["task_id"]

t4 = kanban_create(
    title="draft decision memo",
    assignee="<profile-C>",  # whichever profile drafts user-facing prose
    body="Turn the analyst's recommendation into a 2-page memo for the CTO. Match the tone of previous decision memos in the team's knowledge base.",
    parents=[t3],
)["task_id"]

parents=[...] gates promotion — children stay in todo until every parent reaches done, then auto-promote to ready. No manual coordination needed; the dispatcher and dependency engine handle it.

If the task graph has dependencies, create the parent cards first, capture their returned ids, and include those ids in the child card's parents list during the child kanban_create call. Avoid creating all cards in parallel and linking them afterward; that creates a window where the dispatcher can claim a child before its inputs exist.

Step 4 — Complete your own task

If you were spawned as a task yourself (e.g. a planner profile was assigned T0: "investigate Postgres migration"), mark it done with a summary of what you created:

kanban_complete(
    summary="decomposed into T1-T4: 2 research lanes in parallel, 1 synthesis on their outputs, 1 prose draft on the recommendation",
    metadata={
        "task_graph": {
            "T1": {"assignee": "<profile-A>", "parents": []},
            "T2": {"assignee": "<profile-A>", "parents": []},
            "T3": {"assignee": "<profile-B>", "parents": ["T1", "T2"]},
            "T4": {"assignee": "<profile-C>", "parents": ["T3"]},
        },
    },
)

Step 5 — Report back to the user

Tell them what you created in plain prose, naming the actual profiles you used:

I've queued 4 tasks:

• T1 (<profile-A>): cost comparison
• T2 (<profile-A>): performance comparison, in parallel with T1
• T3 (<profile-B>): synthesizes T1 + T2 into a recommendation
• T4 (<profile-C>): turns T3 into a CTO memo

The dispatcher will pick up T1 and T2 now. T3 starts when both finish. You'll get a gateway ping when T4 completes. Use the dashboard or hermes kanban tail <id> to follow along.

Common patterns

Fan-out + fan-in (research → synthesize): N research-style cards with no parents, one synthesis card with all of them as parents.

Parallel implementation + validation: one implementer card makes the change while one explorer/researcher card verifies config, docs, or source mapping. A reviewer card can depend on both. Do not make the implementer own unrelated verification just because the user mentioned both in one sentence.

Pipeline with gates: planner → implementer → reviewer. Each stage's parents=[previous_task]. Reviewer blocks or completes; if reviewer blocks, the operator unblocks with feedback and respawns.

Same-profile queue: N tasks, all assigned to the same profile, no dependencies between them. Dispatcher serializes — that profile processes them in priority order, accumulating experience in its own memory.

Human-in-the-loop: Any task can kanban_block() to wait for input. Dispatcher respawns after /unblock. The comment thread carries the full context.

Pitfalls

Inventing profile names that don't exist. The dispatcher silently fails to spawn unknown assignees — the card just sits in ready forever. Always assign to a profile from your Step 0 discovery; ask the user if you're unsure.

Bundling independent lanes into one card. If the user asks for two independent outcomes, create two cards. Example: "fix blockers and check model variants" is not one fixer task; create a fixer/engineer card for the fixes and an explorer/researcher card for the variant check, then optionally gate review on both.

Over-linking because of wording. "Finally check X" may still be parallel with implementation if X is static config, docs, or source discovery. Link it after implementation only when the check depends on the implementation result.

Forgetting dependency links. If the task graph says research -> implement -> review, do not create all tasks as independent ready cards. Use parent links so implement/review cannot run before their inputs exist.

Reassignment vs. new task. If a reviewer blocks with "needs changes," create a NEW task linked from the reviewer's task — don't re-run the same task with a stern look. The new task is assigned to the original implementer profile.

Argument order for links. kanban_link(parent_id=..., child_id=...) — parent first. Mixing them up demotes the wrong task to todo.

Don't pre-create the whole graph if the shape depends on intermediate findings. If T3's structure depends on what T1 and T2 find, let T3 exist as a "synthesize findings" task whose own first step is to read parent handoffs and plan the rest. Orchestrators can spawn orchestrators.

Tenant inheritance. If HERMES_TENANT is set in your env, pass tenant=os.environ.get("HERMES_TENANT") on every kanban_create call so child tasks stay in the same namespace.

Goal-mode cards (persistent workers)

By default a dispatched worker gets one shot at its card: it does its work, calls kanban_complete/kanban_block, and exits. For open-ended cards where one turn rarely finishes the job, pass goal_mode=True to wrap that worker in a Ralph-style goal loop — the same engine behind the /goal slash command:

kanban_create(
    title="Translate the full docs site to French",
    body="Acceptance: every page translated, no English left, links intact.",
    assignee="<translator-profile>",
    goal_mode=True,        # judge re-checks the card after each turn
    goal_max_turns=15,     # optional budget (default 20)
)["task_id"]

How it behaves:

• After each worker turn, an auxiliary judge evaluates the worker's response against the card's title + body (treated as the acceptance criteria).
• Not done + budget remains → the worker keeps going in the same session (full context retained — not a fresh respawn).
• Worker calls kanban_complete/kanban_block itself → loop stops, normal lifecycle.
• Budget exhausted without completion → the card is blocked for human review (sticky), never a silent exit.

When to use it: long, multi-step, or "keep going until X is true" cards. When NOT to: cheap one-shot cards (translation of a single string, a quick lookup) — the judge overhead isn't worth it, and the dispatcher's existing retry/circuit-breaker already handles transient worker failures.

Write the body as explicit acceptance criteria — the judge is only as good as the goal text. "Translate the README" is weaker than "Translate every section of the README to French; no English sentences remain."

Worker Lifecycle & Pitfalls

The core lifecycle (6 steps: orient → work → heartbeat → block/complete) is auto-injected via KANBAN_GUIDANCE. This section covers deeper detail: good handoff shapes, retry diagnostics, edge cases.

Workspace handling

Kind	What it is	How to work
scratch	Fresh tmp dir, yours alone	Read/write freely; it gets GC'd when the task is archived.
dir:<path>	Shared persistent directory	Other runs will read what you write. Treat it like long-lived state.
worktree	Git worktree at the resolved path	If .git doesn't exist, run git worktree add <path> ${HERMES_KANBAN_BRANCH:-wt/$HERMES_KANBAN_TASK} from the main repo first.

Good summary + metadata shapes

Coding task:

kanban_complete(
    summary="shipped rate limiter — token bucket, keys on user_id with IP fallback, 14 tests pass",
    metadata={
        "changed_files": ["rate_limiter.py", "tests/test_rate_limiter.py"],
        "tests_run": 14, "tests_passed": 14,
        "decisions": ["user_id primary, IP fallback for unauthenticated requests"],
    },
)

Review-required coding task: Block instead of complete, with reason prefixed review-required:. Drop structured metadata into a comment first.

Research task:

kanban_complete(
    summary="3 competing libraries reviewed; vLLM wins on throughput",
    metadata={"sources_read": 12, "recommendation": "vLLM"},
)

Claiming cards you actually created

If your run produced new kanban tasks, pass the ids in created_cards on kanban_complete. The kernel verifies each id exists and was created by your profile; any phantom id blocks the completion. Only list ids you captured from a successful kanban_create return value — never invent ids.

Block reasons that get answered fast

Bad: "stuck". Good: one sentence naming the specific decision you need. Leave longer context as a comment.

Retry scenarios

If kanban_show returns runs: [...] with closed runs, you're a retry:

• outcome: "timed_out" — hit max_runtime_seconds, chunk the work
• outcome: "crashed" — OOM or segfault, reduce memory
• outcome: "spawn_failed" — profile config issue, ask human via kanban_block
• outcome: "blocked" — previous attempt blocked; unblock comment should be in thread

Do NOT

• Call delegate_task as substitute for kanban_create
• Call clarify — you're headless. Use kanban_comment + kanban_block instead
• Modify files outside $HERMES_KANBAN_WORKSPACE unless the task body says to
• Complete a task you didn't actually finish — block it instead

Codex Lane (Hermes + Codex dual-lane pattern)

Archived standalone skill: The more detailed kanban-codex-lane skill has been consolidated into this orchestrator umbrella. The full original content is at references/kanban-codex-lane.md with the prompt template at templates/pmb-codex-lane-prompt.md.

Use when a Kanban worker wants to run Codex CLI as an isolated implementation lane while Hermes keeps ownership of task lifecycle, reconciliation, testing, and handoff.

When to use the Codex lane

Use when ALL of these are true:

• Coding/refactor/docs/test task with clear acceptance criteria
• A bounded diff can be evaluated by Hermes in one run
• Repo can be checked out in an isolated git worktree/branch
• Hermes can run the relevant tests itself after Codex exits

Do NOT use when:

• Task requires human judgment not captured in Kanban body
• Change touches secrets, credentials, or production systems
• A small direct edit is faster than spawning another agent

Ownership rules

1. Hermes owns the Kanban lifecycle. Codex must never call kanban_complete, kanban_block, etc.
2. Hermes owns final acceptance. Treat Codex commits/diffs as untrusted patches.
3. Hermes owns test execution. Codex test runs are advisory; repeat from Hermes.
4. Hermes owns safety. Reject lane if Codex changes safety boundaries.
5. Hermes owns cleanup. Kill stuck Codex processes, remove temporary worktrees.

Quick Reference: Worktree, Mode Selection, Reconciliation

The full expanded guidance (prompt construction, monitoring, timeout, kill behavior, kanban_complete metadata schema, PMB safety constraints, and common pitfalls) is in references/kanban-codex-lane.md. The Codex prompt template for prediction-market-bot is at templates/pmb-codex-lane-prompt.md.

Worktree pattern:

TASK_ID="${HERMES_KANBAN_TASK:-t_manual}"
REPO="/path/to/repo"
BASE="$(git -C "$REPO" rev-parse --abbrev-ref HEAD)"
SAFE_TASK="$(printf '%s' "$TASK_ID" | tr -cd '[:alnum:]_-')"
BRANCH="codex/${SAFE_TASK}/$(date -u +%Y%m%d%H%M%S)"
WORKTREE="/tmp/${SAFE_TASK}-codex-lane"
git -C "$REPO" worktree add -b "$BRANCH" "$WORKTREE" "$BASE"

Mode selection: codex exec for bounded one-shot edits; Codex /goal for broader multi-step work.

Reconciliation checklist:

• git status --short shows only expected files
• git diff reviewed by Hermes
• No secrets or unrelated changes included
• Hermes ran canonical tests independently
• Accepted commits applied to Hermes-owned workspace
• Rejected work has concrete reason and artifact path

Acceptance outcomes: accepted, partial, rejected, timed_out.

Archived Sub-Skills

Archived Skill	Absorbed As	Content
kanban-codex-lane	references/kanban-codex-lane.md + templates/pmb-codex-lane-prompt.md	Full Codex lane guidance: worktree setup, prompt construction, monitoring/kill, reconciliation, PMB safety constraints, metadata schema, common pitfalls
kanban-worker	Keep separate	Worker pitfalls and examples — kept as a separate skill because it serves a different role (worker vs orchestrator) and is auto-loaded by the dispatcher

Recovering stuck workers

When a worker profile keeps crashing, hallucinating, or getting blocked by its own mistakes (usually: wrong model, missing skill, broken credential), the kanban dashboard flags the task with a ⚠ badge and opens a Recovery section in the drawer. Three primary actions:

1. Reclaim (or hermes kanban reclaim <task_id>) — abort the running worker immediately and reset the task to ready. The existing claim TTL is ~15 min; this is the fast path out.
2. Reassign (or hermes kanban reassign <task_id> <new-profile> --reclaim) — switch the task to a different profile (one that exists on this setup) and let the dispatcher pick it up with a fresh worker.
3. Change profile model — the dashboard prints a copy-paste hint for hermes -p <profile> model since profile config lives on disk; edit it in a terminal, then Reclaim to retry with the new model.

Hallucination warnings appear on tasks where a worker's kanban_complete(created_cards=[...]) claim included card ids that don't exist or weren't created by the worker's profile (the gate blocks the completion), or where the free-form summary references t_<hex> ids that don't resolve (advisory prose scan, non-blocking). Both produce audit events that persist even after recovery actions — the trail stays for debugging.