⏩ thunderkit

Opinionated multi-model delegation for very large repos.

plannerplan

tk-plan

Use to turn an agreed big-repo change into dependency-layered lanes with file ownership, acceptance criteria and verification commands. Hands planning to one qualified native planner or a bound owned planner, preserves native artifacts and approvals, and prepares a lane summary for separate cross-family plan review before execution approval.

Delegates: omo:ulw-plan omh:ultrawork/ulw-plan

Contract: 1

Python 3.11+ standard library for the bundled resolver. Optional native handoffs require pinned oh-my-openagent on OpenCode/Codex or oh-my-hermes on Hermes, with proven loaded provenance and effective role bindings. No automatic installation or host reconfiguration.

Install: npx skills add thunderock/thunderkit -s tk-plan -g

---

tk-plan — decompose into parallel lanes

The heart of the thunderkit thesis. tk-plan takes a change and produces a plan whose unit is the lane: a disjoint, file-scoped slice of work that can run *in parallel* with its siblings without collision, ordered into dependency layers. Entangled work stays explicitly sequential; never manufacture parallelism or treat planning as permission to implement.

The planner is the user's one classes.planner, not a preferred model or the arbitrary current root. Preserve classes.executors and explicit classes.reviewers in their requested order, or retain the literal reviewers "all". Backend choice never changes those selections.

Inputs and paths

Skill root is the installed directory containing this file. Resolve references/dependencies.json, references/delegation.md, references/model-roster.md, references/models.json, references/config.schema.json and scripts/tk-resolve.py from that root, not the current directory, a checkout, or another installed skill.

Project root is the actual repository being planned. Read its required, explicit, project-contained .thunderkit/config.json, .thunderkit/BRIEF.md and .thunderkit/MAP.md inputs. Also read, validate and consume .thunderkit/SPEC.md, .thunderkit/CONTEXT.md and other upstream outputs whenever already produced or required by the approved scope/lifecycle. A full milestone requires the outputs of its preceding stages.

For input completeness on the minimum path, config/BRIEF/MAP suffice when spec/discuss were intentionally omitted and no additional upstream output is required or already produced. Record each intentional stage omission and its reason with the input record; a missing file alone does not establish omission.

Record input paths, content digests and source/base identity. Reject escaping paths and resolve aliases before checking containment. Supply the settled goal, scope, non-goals, constraints, accepted decisions, frozen_paths, max_layers, acceptance checks and verification requirements. A missing required or previously produced input, stale evidence (including optional inputs), contradictory artifacts or open brief unknown stops planning; do not silently ignore it, replace it with assumptions or reopen a settled decision.

Validate all three classes through the bundled config contract before model-bearing work, including owned work with delegation off. Missing choices are not defaults. Complete valid legacy configuration is a preview only; never save it or change a model without the user's approval. For reviewers "all", consider every catalog model, not just the planner and executors. Report unavailable optional candidates; every explicit selection must succeed and the responding review families must independently meet review_families_min. A native host's representable subset does not establish reachability or that later family gate. Preserve the controller's current preflight requirements; resolver admission cannot rescue missing or failed preflight evidence.

Before handing work to tk-router, tk-map, tk-spec, tk-discuss, tk-grill, tk-test, tk-review or tk-execute, check that the sibling is actually available. If absent, report the missing prerequisite and stop that transition; do not read a presumed sibling path or install it.

Delegation

The local manifest's tk-plan / plan entry is authoritative. Its targets are alternatives, both in handoff mode, not components or planners to launch for each lane:

These addresses are registry identities, not invented slash commands. Invoke the admitted selector through the host's real skill tool: OMO ulw-plan or OMH ultrawork/ulw-plan. OMH's catalog name ralplan is not a replacement selector. Its bundle root is neither skills_root nor HERMES_HOME. Compare package/version/source, root identity, loaded entrypoint and the real bytes of every manifest companion with the pinned fingerprints. A same-name skill, quarantined file, missing companion, null hash, self-reported checksum or ready flag does not qualify. Consume the local pins; do not qualify a different release on the fly.

Before handoff, verify every native_roles slot, including roles that may not run on this request. Use actual live host descriptors and effective agent/category or session mappings. Record each slot's class, selected catalog member, exact supported provider/model identity and supported effort. OMO requires all three binding classes; OMH planning requires only planner. Preserve the other selected classes for later stages without claiming they were exercised by OMH planning. For each required explicit plural selection retain every member's association and order; a slot may use only its own class. For "all", retain the request and the reported native subset separately from the later catalog-wide reviewer expansion. A run need not exercise every member, but an opaque, collapsed, reordered or unrepresentable selection is not admitted.

OMO task() has no per-call model parameter; load_skills supplies instructions, not a model binding. Inspect the actual root-session model as well as effective delegated role slots. Editing config does not prove the running root switched. If a selected model requires native configuration or restart, report operator guidance and wait for fresh binding evidence; never rewrite global/provider/auth configuration to make a route appear ready. OMH planning is one planner-bound session; its in-session critic is not an independent reviewer. OMO Momus/Oracle bindings also do not replace Thunderkit's separate cross-family plan review.

Gather the current capability snapshot without credentials, installation or doctor calls. With SKILL_ROOT, PROJECT_ROOT and RUN_ID set to the actual installed skill, repository and controller run, resolve the explicit operation:

python3 "$SKILL_ROOT/scripts/tk-resolve.py" --skill tk-plan --operation plan \
  --project-root "$PROJECT_ROOT" --config "$PROJECT_ROOT/.thunderkit/config.json" \
  --capabilities "$PROJECT_ROOT/.thunderkit/runs/$RUN_ID/capabilities.json" --json

The input paths must be explicit and inside the actual project root. For delegation off or no enabled ecosystem, omit --capabilities and do not discover native peers. Keep the complete resolver record unchanged, including decision, reason_code, target, requested/effective bindings and evidence paths. Exit 0 means routing was computed, not native completion. Only delegate / compatible admits the chosen native owner; blocked means stop.

Native ownership and return

Give the single admitted owner the settled inputs above and the required lane/output contract below. It owns its planning workflow, native approvals and native state until a known return. Thunderkit does not cut a competing plan, launch another planner per lane or start implementation while that owner is active. Preserve the native workflow rather than copying it into this skill.

• OMO writes only within its planning domains .omo/drafts/ and .omo/plans/. Keep drafts

distinct from the actual approved plan and preserve its native approval flow and evidence.

• OMH records under .omh/plans/ through its native omh hermes plan --record flow and

obtains native acceptance through omh hermes plan-accept <path>. Retain the actual acceptance evidence for the returned artifact; an in-session critique or a recorded draft is not acceptance.

• Neither planner may write into .thunderkit/, edit implementation files, dispatch execution,

push, open a PR, publish or merge. The controller alone performs later normalization, after the native owner returns. If the native workflow cannot preserve this boundary, do not invoke it.

• Do not activate conditional external-owner/ulw-maestro, durable-checkpoint/ulw-loop, or

no-plan execution paths. They remain unavailable at the pin; report capability_missing as the unmet capability separately from the unchanged resolver record. Do not launch them or add their sources to trust.

After a known terminal return, the controller reads the actual native artifact and acceptance evidence. Verify a regular, project-contained file under the selected .omo/plans/ or .omh/plans/ directory, with no traversal or symlink escape. Hash its unaltered bytes, record the actual repo-relative path, and bind native approval to that content identity. Missing, unapproved, conflicting or unusable output stops readiness; never infer success from returned Markdown, exit 0, done, a filename or an old approval. Request correction through the same native planning flow only after ownership is settled, then require approval for the corrected bytes.

Record genuine native session/resume identity and requested versus effective versus observed model identities alongside the returned artifacts. Missing facts remain null / unverified, including session_id, observed_model and observed_family; a config choice is not a runtime observation. Retain native output evidence even when incomplete, but missing required identity proof or an unapproved model change prevents acceptance.

An unknown, timed-out or still-in-flight native owner retains ownership. Inspect its real captured session and artifact state before any retry or fallback; do not invent a session ID or treat history metadata as proof that the session is resumable. Without an ID or known terminal state, stop as blocked/unknown and report the missing evidence. A known invocation/output failure is recorded separately; it never rewrites the earlier resolver reason into a different routing result.

Fallback

owned / disabled or owned_policy and a computed fallback may use the bounded lane procedure below only when no native owner remains active or uncertain. Keep the specific resolver reason and failure evidence. Delegation off performs no native invocation, discovery, doctor or routing helper call. No undeclared ecosystem substitutes for a failed peer.

Owned planning still requires a genuinely bound selected planner. Validating configuration or mentioning classes.planner in a prompt does not bind the current root. Use only an already supported channel proven to run that selected planner, with the same scope, limits and approval policy; otherwise stop as blocked and report the binding gap. A blocked resolver result never starts fallback. Once admitted, the owned planner produces the same lane contract and the controller writes the documented Thunderkit outputs; omit native_plan for owned work rather than fabricating native provenance or approval. A known failed handoff must be explicitly retired before an owned replacement is authorized; never hide an unusable native artifact behind a ready summary. The separate plan-review and execution-approval gates apply unchanged.

What a lane is

• Disjoint file scope — two lanes in the same layer must not write the same files. This is

what makes parallel execution safe. If two slices need the same file, they belong in different *layers*, not the same layer.

• A dependency layer — lanes in layer N may depend only on layers < N. Layer 0 lanes have no

intra-plan dependencies; they become eligible only after review and execution approval.

• Acceptance criteria — what "this lane is done" means, testably.
• A verification command — the exact command tk-review runs to gate the lane. No command

→ the lane is blocked, not plannable.

• A model hint — critical-path lane vs. breadth/cleanup lane, resolved within the selected

executor class. A hint cannot substitute a model or approve dispatch.

Asking the user

When this skill needs a decision from the user, ask through the host's structured choice tool as described in references/asking.md: one decision per question, two to four options with the recommended one first, free text always accepted. Use the numbered-list fallback only when the host has no such tool; in a non-interactive run record unknown and stop at the gate.

Output contract — .thunderkit/PLAN.md + .thunderkit/plan.json

Human-readable PLAN.md and a machine-readable plan.json that tk-execute consumes:

For a native handoff these are controller-derived lane summaries, not another executable plan. Normalize only after known return, approved artifact verification and lane validation; retain the native plan as the execution authority. Do not change its bytes to fit the summary. Preserve the existing goal/layers/lanes structure and every lane's fields:

{
  "goal": "one-line change description",
  "layers": [
    {
      "layer": 0,
      "lanes": [
        {
          "id": "L0-auth-token-refresh",
          "files": ["src/auth/token.rs", "src/auth/token_test.rs"],
          "depends_on": [],
          "acceptance": "token refresh retries 3x with backoff; expired token triggers refresh",
          "verify": "cargo test -p auth token::",
          "model_hint": "critical-path"
        }
      ]
    }
  ]
}

For native planning add native_plan: {ecosystem, package_version, skill_name, artifact, sha256, approval_status}. Use the actual verified native path and digest; approval_status comes from native acceptance evidence, not the controller's optimism. Attach a model-contract snapshot retaining requested classes/order/all, effective per-slot/member associations, observed identities or nulls, review_families_min, max_layers, frozen_paths and the supporting evidence paths.

Alongside existing harness output, retain the delegated record {lane_id, ecosystem, package_version, skill_name, requested_model, effective_model, observed_model, observed_family, artifact, artifact_sha256, session_id, status, evidence_paths}. Planning has no implementation lane yet: leave lane_id null unless an actual association exists. Keep the source-qualified selector and package/source snapshot in the unchanged routing record; the common bare skill_name alone cannot distinguish these two planners.

The controller records normalized output identities and the native/input identities they derive from. Any native byte change invalidates the dependent summary, plan-review readiness and execution approval. Changes to normalized lanes, source inputs or the model contract also require fresh validation and review. Do not update a stored digest simply to keep an old approval green.

Procedure

Use these bounded steps for an admitted owned planner. For a native handoff, supply their required outputs to the native owner and inspect the returned plan instead of running this procedure in parallel with it.

1. Refresh the map if stale against the current source/base identity — stop and route to

available tk-map; a timestamp alone is not freshness evidence.

1. Cut along seams, not arbitrarily. Use the boundaries in MAP.md so lanes fall on real

module edges and file scopes genuinely don't overlap.

1. Layer by dependency. Put independent slices in the same layer (they parallelize); put a

slice that needs another's output in a later layer.

1. Attach acceptance + verify to every lane from the map's per-area verification commands.

A lane with no runnable verify is blocked — record why and what's needed to unblock it.

1. Mark model hints. Flag the critical-path lane(s) so tk-router knows to ask the user

which selected executor implements them, without reopening settled class choices.

1. Check testability before finishing: can each lane's verify actually run in this repo? If

a command is aspirational (test doesn't exist yet), the lane's first task is to create it.

The parallelism check (do this before declaring the plan done)

• Every pair of lanes in the same layer has non-overlapping files. If not, re-layer.
• Enumerate concrete repo-relative files, including tests and generated outputs. Resolve aliases

and existing ancestors so directory scopes or symlinks cannot hide overlap or escape. No lane may write a frozen file or a file under a frozen directory.

• IDs are unique, every dependency names a real lane, and all edges point to earlier layers.

Self-dependencies, cycles and same-layer dependencies stop readiness, not just execution order.

• Stay within max_layers; do not silently increase it to repair an overlap or entanglement.
• Every lane has a verify or is explicitly blocked.
• Verify commands have an actual working directory, executable and known prerequisites. A test

to be created is an explicit owned file/task; its future result is not present verification. Unavailable prerequisites remain blockers with an owner and the evidence needed to unblock them.

• At least the critical-path lane has a model_hint for the user-choice step.
• Layer 0 is non-empty with no intra-plan dependencies. This is structural readiness, never

permission to start immediately; preserve genuinely serial work rather than inventing seams.

For a native plan, a failed check returns an unresolved finding to its owner after known return. Do not repair only the derived lanes while leaving the native execution authority contradictory.

Record unresolved tradeoffs

If a clean disjoint decomposition isn't possible (genuinely entangled code), say so explicitly: record the entanglement, propose the least-bad layering, and flag the lanes that must run serial. Don't flatten a real dependency into fake parallelism.

Record unresolved scope, approval, model, artifact and verification blockers alongside the lane summary and report the plan as not ready. A native approval does not erase an overlap, cycle, frozen-path conflict or exceeded layer budget. Get a corrected, newly approved native artifact before regenerating its summary; do not drop blocked lanes to manufacture a passing subset.

Plan review and execution approval

Planning ends with a readiness report and the next gate, not implementation. Check availability before routing to tk-review --plan; its plan operation is distinct from diff review. Require .thunderkit/PLAN-REVIEW.md from independent selected reviewers against the current native path/hash, normalized output hashes, source/input identity and model-contract snapshot. Count actual responding model families, not harness names: meet review_families_min (at least two), with at least one family different from the author. Explicit reviewers cannot disappear because of quota or host limitations; "all" retains its reported reachable expansion. Unresolved blocker or major findings and missing identity/verification evidence prevent execution readiness.

Native acceptance, in-session/native critique, Thunderkit plan-review approval and execution approval are separate gates. Even a currently passing cross-family plan review does not authorize execution. The controller must obtain separate execution approval for that exact reviewed artifact set and scope before an available tk-execute takes ownership. Immediately before that transition, compare identities again; stale hashes, missing or unapproved plans, changed selections or unresolved lane blockers stop dispatch. This skill never starts execution.