Skip to main content

Task model assignments

KBD planning recommends a concrete model for each task, including integration and review work. It chooses demonstrated task suitability first, within explicit user selections, project policy and budget. Cost and latency break close ties. A change-level complexity label does not choose the model for all its tasks.

This is an authored planning contract and dispatch procedure. It does not introduce a worker runtime, service or task-state schema; inspect the installed skill and native tool capabilities before using a route.

Plan before dispatch​

  1. Draft tasks and resolve their backend identities.
  2. Assess reasoning difficulty, uncertainty, scope, tools, context size, modalities and review independence for each task.
  3. Discover the installed harness version, configured providers, exposed models, supported reasoning controls and actual worker-launch mechanism.
  4. Compare task fit using dated local results or authoritative capability documentation. Label unknown or stale evidence; a model name is not evidence of quality.
  5. Write one Task model assignments table in the active phase's plan.md and reconcile it against every emitted task before handoff.

Planning records prerequisites. It does not launch workers, change providers, start services or copy credentials into artifacts.

The assignment key and fields​

Each row is keyed by full phase path + change ID + exact backend task ID. For example, task 1 in changes api and docs has two distinct keys even if both display 1.1. Use the ID returned by the backend and canonical identity resolver; a displayed checkbox ordinal is not a replacement ID.

Required fieldWhat to record
Task requirementsReasoning, uncertainty, scope, tools, context, modalities and independence
Provider/modelConcrete configured identity, not only small/medium/frontier
Reasoning effortA supported setting, or explicitly unsupported/unknown
Rationale and evidenceWhy it fits this task, evidence date/source and limitations
Harness and routeInstalled harness/version and native or liter-llm inference route
Worker launch and handoffDocumented invocation, scope, working directory, tools, skills and result destination
Native alternativeBest viable native option if different; not permission to substitute
AvailabilityConfigured, verified or unresolved, with evidence
PrerequisitesMissing model access, protocol support, gateway or tool-enabled launcher

OpenSpec tasks.md and native KBD artifacts reference the matching rows in ordinary prose. Keep checkbox syntax and task titles unchanged. Table rows do not become extra tasks, and assignment keys do not replace canonical runtime IDs.

Harness discovery​

HarnessCheck before selecting a route
CodexExposed model choices, supported fresh-agent overrides and fork inheritance
Claude CodeInstalled subagent model controls and the protocol required by its worker
OpenCodeConcrete provider/model identifiers and configured per-agent or per-invocation controls
DeepSeek HarnessInstalled provider/agent controls; global settings do not prove per-member overrides
Kimi CodeSecondary-model pools and invocation controls; avoid ignored model frontmatter

There is no permanent best-model ranking. Re-evaluate when tasks, harnesses, model access or capability evidence change.

Native execution and liter-llm​

Use the selected model natively when the harness supports it. Otherwise record two separate parts: liter-llm supplies inference, and an existing documented tool-enabled worker executes the task in the workspace. Check protocol compatibility as well as model availability.

Discover the installed tool schema or CLI documentation. Do not invent a launch_agent interface or a liter-llm complete command. An inference response alone does not prove filesystem or tool execution.

If access or a worker is missing, leave the route unresolved and record its native alternative separately. A different selection requires an explicit recorded rationale. Other eligible tasks may continue; silently substituting a cheaper or available model is not allowed.

Execute, revise and review​

Before dispatch, kbd-execute and kbd-apply recheck the exact scoped assignment and copy its reference and actual route into execution.md and the worker handoff. Workers return artifacts, results and blockers to the KBD driver. The driver owns begin/end task boundaries and canonical completion updates.

Legacy plans remain usable: perform explicit task selection at execution time before running the task. Changed, split, removed or renumbered tasks require reconciliation so each current task has exactly one current assignment.

Independent plan/artifact review and available sycophancy screening run at the completed-production boundary. Assigning an integration or review task does not authorize early verification: final integration and review follow the complete production delivery boundary. Readiness, implementation, evidence, certification and publication remain distinct.

Separate planned selection from actual execution​

Keep evidence states explicit: operator-declared catalog, configured discovery, exposed native control, successful inference, tool-enabled worker execution and final acceptance. Record the actual provider/model/effort and serving route in execution.md, with a dated result or an unresolved limitation. A requested alias can resolve to a different backend; its spelling is not identity evidence. Never record a plan recommendation as an observed inference or worker result.

The portable team helper supports only model, tier, capabilities, maxInputPerMillion and maxOutputPerMillion. It accumulates capabilities across team → role → explicitly ordered skill → task layers, while scalar values override. low/medium/hard are declared tiers; reasoning effort, context size, whole-task budget and fallbacks live in plan/native evidence, not invented portable fields. Selection filters constraints, then sorts by known price sum; KBD's task-fit-first choice remains the planning authority.

Use the canonical model handbook for discovery, effective policy, native controls and budget limits. The full-pack request example provides actual models-discover, models-select and revisioned team-update requests. Both packs need Node.js 22+ for local comparison. Live discovery needs configured endpoints; execution needs the selected native/tool-enabled worker. Full-only card/intake and Python/Cortex controls are not mini model prerequisites.

Budget and fallback decisions​

Unknown capability/price/availability cannot satisfy its requirement. Catalog pricing is an estimate with provenance/freshness; per-million ceilings do not bound total tokens, reasoning, retries, multimodal usage or provider extras. No automatic fallback chain is implemented by the team selector. An unavailable route stays unresolved until an explicit recorded alternative respects owner policy. A native alternative in a table is not permission to substitute silently.

For a changed task/harness/model, reconcile the exact assignment key, update its dated rationale and record the actual dispatch. Fresh Codex controls and fork inheritance depend on the exposed tool schema; Claude model aliases and provider settings can affect the actual model; OpenCode variants are provider/model specific; Kimi model frontmatter and DeepSeek static per-member settings cannot establish the requested route. Unknown effort is recorded as unknown, not guessed.

A target coordinator receiving cross-project context selects against its own project policy, budget, model access and worker capabilities. Source preferences are inputs, not remote authority. Record distinct source/target repository, canonical identity and actual route on acceptance. An issue/card or UAR direct coordinator schema neither binds a model nor starts remote execution.

Reviewer independence​

Use separate reviewer context and record actual producer/critic identities. This project's KBD policy requires a different model family: same-family native fallback, a different alias or a different effort setting does not satisfy it. A gateway name comparison is insufficient if aliases share the same underlying provider model. Preserve actual route evidence and report unverified identity. If no eligible reviewer is available, leave review pending until a valid route or explicitly authorized exception exists. Final integration and review still wait for all production work in the phase.

Verification limits​

Historical release notes reported exercises covering mixed-complexity planning, native Codex tool reads, an OpenCode worker through liter-llm, unresolved routes, legacy plans, changed tasks and repeated task numbers. Synthetic model evidence tested selection behavior, not provider rankings. The gateway's underlying provider identity and cross-model reviewer independence were not independently verified. Claude Code, DeepSeek Harness and Kimi Code execution were documentation/capability-checked only; Windows runtime was not exercised.

Those historical reports are not acceptance evidence for the current candidate; generated payloads, installed routes and final local gates require their own receipts.

See OpenSpec lifecycle updates for project refresh and recovery.