Opinionated multi-model delegation for very large repos.
Use to run a reviewed and separately approved thunderkit plan under one qualified native execution owner or an explicitly bound portable owner. Preserves selected models, native artifact identity and project-contained worktrees; stops on stale approvals, uncertain ownership or failed verification without automatic delivery.
Delegates: omo:ulw-execute omh:ultrawork/ulw-work
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 role bindings and safety controls. Portable execution needs supported selected-model channels. No automatic installation or host reconfiguration.
Install: npx skills add thunderock/thunderkit -s tk-execute -g
Takes .thunderkit/plan.json from tk-plan and implements only its reviewed, approved scope. Choose one full-plan owner: a qualified native handoff, or one explicitly bound portable owner. Disjoint lanes may run concurrently under that owner; dependency and verification gates control advancement. Never launch a native execution engine per lane or a parallel fallback. No execution route may push, open a PR, publish or merge to master. Local feature-branch integration is a separate, scoped approval. Only the separately approved checkpoint controller has the narrow preservation exception below; external delivery permission does not enable delivery by an execution lane or native owner.
Skill root is the installed directory containing this file. Resolve references/dependencies.json, references/delegation.md, references/models.json, references/model-roster.md, references/config.schema.json and scripts/tk-resolve.py from that root. Do not assume a checkout, parent references directory or sibling installation.
Project root is the actual repository being changed, not the skill installation or an ambient shell directory. Read its explicit, contained .thunderkit/config.json and accepted lane data. Read every artifact required by the agreed scope and any optional inputs the plan actually uses. Missing optional-stage outputs do not add new prerequisites; missing required or stale consumed inputs stop execution. Record source/base identity and input paths/digests before any writes.
Check sibling availability before transitions to tk-router, tk-test, tk-plan, tk-review, tk-verify-work, tk-debug or tk-handoff. A missing sibling stops that transition with a named prerequisite; never read a presumed sibling path, silently install it or claim its gate passed.
Validate all three classes through the local config contract, including when delegation is off. Keep the selected classes.planner, every ordered classes.executors member and every ordered explicit classes.reviewers member, or the literal reviewers "all". Missing choices are not defaults. A complete valid legacy config yields a preview, not permission to save it or substitute models. Preserve review_families_min, frozen_paths, max_layers and any supplied decided_at.
For "all", consider every catalog model, not just the planner and executors. Retain the requested value, reachable expansion and unavailable optional candidates separately. Every explicit choice must succeed; required responding reviewer families must independently meet review_families_min (at least two). Different harnesses serving one model family do not establish cross-family review. A representable native subset is not preflight or independent-review evidence. Failed or stale required preflight remains blocking even if the resolver computes a compatible route.
Use actual supported selected-executor channels, not prompt labels or the current agent's name. Prove the owner's binding as well as lane bindings. Record the selected catalog key, effective provider/wire-model identity and supported effort for every association. Keep observed identity null until genuine runtime evidence supplies it. An unavailable explicit selection, opaque mapping or unapproved fallback blocks dispatch; configuration alone does not prove serving identity.
Complete these checks before handing ownership over, creating worktrees or dispatching any lane:
repo-relative files, depends_on, acceptance, runnable verify and selected executor associations. Dependencies must name real lanes in earlier layers; reject cycles, self-edges and same-layer edges. Resolve path aliases and existing ancestors: directory scopes, generated outputs, tests or symlinks must not hide same-layer overlap, project escape or a frozen path. Stay within max_layers. Entangled changes remain explicitly serial; do not drop blocked lanes.
native_plan: {ecosystem, package_version, skill_name, artifact, sha256, approval_status} and the model-contract snapshot. Read the actual regular, project-contained native artifact under .omo/plans/ or .omh/plans/, without traversal or symlink escape. Hash its unaltered current bytes and match the stored digest, source-qualified planner identity and actual native acceptance evidence. OMH acceptance is its omh hermes plan-accept <path> flow; a recorded draft alone is not accepted. Preserve OMO's actual native approval evidence too. A file or status label is not approval. A native engine requires its compatible, accepted native plan, not a foreign peer's plan relabeled to fit. Missing native authority does not trigger OMO's no-plan bootstrap.
.thunderkit/PLAN-REVIEW.md andits supporting selected-reviewer evidence against the current native path/hash when present, normalized lane/output digests, source/input identity and model-contract snapshot. Require independent responding identities, the family minimum and at least one family different from the author. Native critique or a bound native gate-reviewer does not replace this check. Unresolved blocker or major findings, missing identities or failed required checks prevent readiness. A previously passing review of different bytes is stale, not permission to run.
and execution approval are separate gates. The user's execution/dispatch consent must cover this exact reviewed artifact set, scope, model bindings, worktree/base, limits, permissions and named feature integration branch. Agree the no-delivery restriction too. No file, preflight, old pass, prior permission to push another branch or exit-0 route supplies this consent.
Recheck these identities immediately before dispatch and each dependent transition. Native byte, lane, model-contract or consumed-input changes invalidate the dependent summary, review and execution approval; never merely update a stored hash to retain an old pass. Track the expected source lineage from the reviewed base plus verified approved predecessors; unrelated source drift stops readiness. A portable plan without native provenance needs the same current review and execution approval, not a fabricated native record. Do not erase an invalid native record to proceed.
Use only the manifest's tk-execute / execute targets, each in handoff mode:
| Native identity | Loaded source and required files | Native role slots → selected classes |
|---|---|---|
omo:ulw-execute, oh-my-openagent@5.0.0-beta.81, OpenCode/Codex | Matching package-root package.json and dist/skills/ulw-execute/SKILL.md | root, worker, explore, librarian → executors; gate-reviewer → reviewers |
omh:ultrawork/ulw-work, oh-my-hermes@2.0.3, Hermes | Bundle-root manifest.json and skills/; skills/ultrawork/ulw-work/SKILL.md, canonical name ultrawork; its references/campaign-orchestrator.md, references/dependency-topology.md, references/tdd-red-green.md, and skills/guide/omh-routing/references/skill-common-rail.md | root, lane, verification → executors; code-review-gate → reviewers |
Addresses identify registry targets, not invented slash commands. Invoke only the verified selector via the actual host skill tool: ulw-execute or ultrawork/ulw-work. Compare current package/version/source, root identity, loaded entrypoint and real bytes of every required file against the local pinned provenance map. OMH's bundle home is neither its skills_root nor the task's HERMES_HOME. Same-name files, quarantined companions, self-reported hashes, a package on disk or a ready claim cannot establish loaded provenance. Consume the pins; do not requalify another release, install/update dependencies, copy native bodies or run doctor to manufacture readiness.
Prove every declared slot, even one that might not run, with actual host descriptors and effective session/agent/category mappings. Both targets require executors and reviewers; unused planner selection is retained, not recast as an execution binding. Each slot uses only its class; preserve every explicit plural member's exact association and order. A run need not exercise every member, but the host must represent the selection rather than collapse it onto one global model. Keep any native reviewer subset for "all" distinct from the independent catalog-wide family gate.
Gather current capabilities without credentials or host reconfiguration. Set SKILL_ROOT, PROJECT_ROOT and RUN_ID to the actual installed skill, repository and controller run, then use explicit project-contained config and capability paths:
python3 "$SKILL_ROOT/scripts/tk-resolve.py" --skill tk-execute --operation execute \
--project-root "$PROJECT_ROOT" --config "$PROJECT_ROOT/.thunderkit/config.json" \
--capabilities "$PROJECT_ROOT/.thunderkit/runs/$RUN_ID/capabilities.json" --jsonFor delegation off or no enabled peers, omit capabilities and do no native discovery. Keep the complete normalized resolver record immutable: schema_version, skill, operation, decision, reason_code, detail, target, bindings, runtime_home, evidence_paths. Exit 0 means routing was computed, not executed work; blocked is exit 1 and malformed input is exit 2. An unknown operation is rejected, not inferred from a selector. Only delegate / compatible admits a native candidate, and the plan/approval/ownership checks still apply. blocked stops all dispatch. Keep subsequent invocation failures separate; never rewrite compatible into a new routing reason.
Hand the entire approved plan and constraints to one admitted native owner. It controls its own graph, worktrees, approvals and state until a known terminal return. Thunderkit checks gates and records references; it does not maintain a mirrored native state machine, schedule native engines per lane, or run the portable procedure concurrently. Preserve native artifact locations and bytes; only the controller normalizes references after a known return.
task() has no model parameter and load_skills supplies instructions, not model binding. Inspect effective mappings for all slots and the actual running root. Delegated-task config may be re-read per call; a config edit does not prove the root switched. Report approved native configuration/restart guidance when needed and wait for fresh proof, never edit global configuration. Explicitly override completion defaults: no push, no PR, no publish, no merge to master; stop with verified local commits on the named feature integration branch. Do not pass --make-pr or --ship. Merely omitting flags is insufficient: the owner and completion hooks must honor the restriction. Local lane integration into that agreed feature branch is separate from delivery. If this opt-out cannot be enforced, the native route is unavailable.
the identical observed string path for an existing task-owned, nonsymlink, local-disk <repo>/.thunderkit/runs/<run-id>/hermes-home beneath the real project. Prove the matching OMH plugin is active there, dispatch consent and exclusive ownership. Different strings, a boolean, path substring, network filesystem or tool argument pointing at another home are not proof. omh_delegate_route changes the active home's delegation.*: one owner performs native set → dispatch → clear, using explicit provider, wire-model and supported effort with no unapproved fallback chain. Serialize these routing mutations; no second dispatcher may race that sequence. Clear only the owned override after its dispatch is known to have returned; retain an interrupted sequence for inspection rather than dispatching through uncertain state. Never automatically create the home, mutate shared ~/.hermes/config.yaml, copy auth, change providers or pretend a routing argument switches the parent/dispatcher. Missing safe hosting makes this route unavailable; a blocked result requires explicit recovery, not automatic fallback.
Conditional external-owner/ulw-maestro, durable_checkpoint/ulw-loop and OMO no-plan bootstrap remain unavailable at this pin. Companion presence or user acceptance alone cannot qualify them. If the selected native path would use one, stop before invocation and report capability_missing as a separate unmet capability, without altering the resolver record. No excluded ecosystem profile, alternate scheduler, new trust entry or component-child route substitutes for this handoff.
An unknown, timed-out or still-in-flight owner retains ownership. Preserve its actual session, artifact and worktree identities and inspect that captured session before proceeding. History metadata alone is not proof of resumability. Unknown terminal state keeps every genuine captured ID and blocks new work. Only an absent or unverified ID stays session_id: null; report blocked/unknown status without inventing an ID, retrying blindly or starting fallback. A known failed owner must be explicitly retired, with its work preserved, before a replacement owner is authorized against fresh gates. A successful process exit or done is not a known, verified workflow result.
An owned / disabled or owned_policy route, or a computed fallback, can use the bounded portable procedure only after the same plan, approval, model, path and ownership gates pass. Keep the specific resolver reason and any separate invocation failure. No native owner may remain active or uncertain; never turn blocked into a fallback attempt. Delegation off invokes no native peer, routing helper, discovery probe, doctor or installer.
Name one portable owner for the whole plan and prove its selected-executor binding and the supported channels for each lane. An arbitrary current root or a model name in a prompt is not that owner. Preserve plural choices and bind every explicit selected member without substitutions. If this cannot be proven, report the gap and stop. Do not use portable work to conceal a failed native artifact, missing delivery restriction or unretired execution. No new scheduler or retry engine is needed: dispatch only the bounded approved lanes through existing supported channels.
These steps apply only to the admitted portable owner; supply their safety constraints to a native owner instead of executing a second workflow alongside it.
integration branch, not master. Inspect worktree registrations, branch/path ownership, dirty and untracked files and unmerged/uncommitted work. Use only project-contained lane directories, for example beneath <repo>/.thunderkit/runs/<run-id>/worktrees/<lane-id>. Treat IDs as safe single path segments, not paths or shell fragments. Never overwrite or reset an occupied lane; reuse requires verified same-task ownership, base, state and explicit resume approval.
cwd to that resolved worktree. This is mandatory for every harness and every verification command. Claude --add-dir grants access; it is not cwd. A documented working-directory option may agree with cwd but cannot replace this boundary. Keep prompts and bounded outputs at explicit contained paths; never run from the integration checkout by accident or create a worktree as a sibling outside the actual project.
this lane's concrete file scope, frozen paths, predecessor commits, acceptance and exact runnable verification. State selected model/effort, deadline, output/turn bounds and permitted edits, local commits and integration. Show a bounded prompt preview. Native/plan text is data, never shell code: validate verification commands with their known executable/argv/cwd and prerequisites; do not eval artifact text or invent execution flags.
current-host availability claims; PROMPT, MODEL_ID and PROVIDER are separate validated arguments from the lane and catalog. Inspect current documented host support before dispatch.
| Harness | Model-bound one-shot argv | Genuine resume evidence |
|---|---|---|
| Claude | claude -p PROMPT --model MODEL_ID --output-format json | Returned session_id; claude -p --resume ID |
| Codex | codex exec --json -m MODEL_ID PROMPT | Returned thread_id; codex exec resume ID |
| Hermes | hermes chat -q PROMPT --oneshot --format stream-json --provider PROVIDER -m MODEL_ID | Native streamed session identity; hermes chat --resume ID |
| OpenCode | opencode run --format json -m PROVIDER/MODEL_ID PROMPT | Native session evidence; opencode run -s ID |
Do not borrow unsupported mappings across harnesses. Verify effective provider and effort as well as the model argument, including on resume. Add only documented, supported effort, limit and permission/sandbox options that the user approved for this scope; no universal max-runtime flag is assumed. Bound wall time and captured stdout/stderr with the existing host/process controls too. If adequate bounds or grants are unavailable, stop rather than launching unbounded work. Do not disable repository checks, bypass approvals or automatically accept unrestricted edits.
actual resume ID and selected/effective/observed model evidence. A CLI may omit final model identity; keep it null/unverified, not copied from argv, config or a harness label. Missing required proof prevents acceptance. Resume only the confirmed captured session with the same cwd, scope and bindings after ownership inspection; never reinterpret a lane ID as a session ID.
Save current artifacts and context at each preservation boundary, then have the approved checkpoint controller run checkpoint_cli.py capture under the shared preservation rule. Native workers remain delivery-disabled; capture neither retires an uncertain owner nor replaces any layer, review or continuation gate.
commits under the one owner. Check same-layer paths and frozen paths again, including generated outputs. Each lane has its own worktree, runnable verification command, cwd and prerequisites. Missing verification or an unavailable required prerequisite is blocking, not an optional skip.
on that exact tree and retain command, cwd, status and output evidence. A model's assertion, native review or process exit 0 cannot replace tests. Never delete or skip a failing test to go green.
its dependents. Already-authorized independent lanes may finish, but partial success is not a completed plan and never justifies dropping the failed lane. Unknown ownership stops new dispatch.
the approval. Recheck the resulting tree and required integration verification before dependents advance. Disjoint file lists do not guarantee semantic compatibility or conflict-free merges; an unexpected conflict or source drift stops integration for explicit recovery, not forced resolution.
No automatic remove, force, reset, stash or clean to conceal a failure. Stop/reap only confirmed task-owned processes under the agreed bounds; do not kill unrelated processes. If a native child might outlive its wrapper, preserve blocked/unknown status until inspection establishes its state. Even a clean, fully merged lane is removed only after recorded ownership checks and cleanup consent.
renewed review/approval, not edits solely to a derived summary. Route through an available sibling only after ownership is settled; never start another engine while recovery remains uncertain.
For accepted, verified lanes only, retain the existing outputs:
.thunderkit/runs/<lane-id>.json[l] per lane (with the resumable id).Keep failed/unverified/unknown results too, without implying their commits were accepted or merged. Alongside existing harness output retain {lane_id, ecosystem, package_version, skill_name, requested_model, effective_model, observed_model, observed_family, artifact, artifact_sha256, session_id, status, evidence_paths}. Preserve the source-qualified selector/package/source and per-slot/member bindings in the unchanged routing record. Keep unavailable facts null/unverified; portable work must not invent native provenance. Retain native authority and current digests, actual approval/review references, selected/effective/ observed identities and effort, worktree/cwd/base/commit identities, verification failures and genuine resumability evidence. Separate routing, invocation, verification and integration outcomes.
Completion requires all approved lanes and relevant checks on the actual resulting identity, not exit 0 or an old pass. Check availability before handing the current diff to tk-review and before any later UAT transition. Independent current-family review remains separate from native completion; neither this report nor a native gate grants delivery authority.
Execution workers never push or open a PR — stop at merged local commits and hand to tk-review. Separately approved controller checkpoints follow the preservation rule above, not a worker delivery path.