megamen32/last-human-commit

Design a working solution when boundaries, data ownership or costly assumptions remain uncertain; test pivotal choices, independently challenge the design, then simplify into complete outcomes and dispatchable tasks.

Compatibility alias for older LHC assignments that request bugfix TDD. Prefer worker-bugfix for new Worker defect tasks; use this only when an existing task or caller explicitly names bugfix-tdd.

Lead-owned least-cost procedure for proving and handing off the accepted business result.

Independently challenge an important proposal or synthesis with counterexamples, alternative explanations and a minimal evidence-producing check.

Council (кворум, quorum, «созови кворум») - independent strong models propose and challenge solutions, a synthesis is rechecked and disputed assumptions are tested. Use when the owner asks for a quorum, when a decision is complex or architectural, or when a hard dead end persists.

Let the human-selected Lead coordinate a goal into coherent model-assigned tasks, dependency-aware parallel lanes and verified integration joins.

Compatibility alias for older LHC assignments that request feature implementation. Prefer worker-code for new Worker coding tasks; use this only when an existing task or caller explicitly names feature-implementation.

Run a focus group (фокус-группа) of Zero Knowledge testers on the real product - diverse personas on mobile and desktop pursue their own goals through a browser, report what they like and do not, are asked what each button does, and every finding is fixed, redeployed and retested in rounds without waiting.

Convert observed failures and successful methods into tested, retrievable skills, tools or routing changes, then verify reuse and retire ineffective guidance.

Automatically handle every authorized engineering request, including quickfixes, incident repairs, integrations, “try installing” requests and rough voice transcripts, from intent through a real result. Create or update the concise plan before implementation or delegation; the user need not ask for a plan. Questions, read-only research/status, testing without code, and an explicit owner no-LHC choice are outside plan creation. You are the Lead; Workers execute. Answer in the user's language.

Legacy recovery only. Use solely when the user explicitly selects old Fleet file rollout or rollback for Last Human Commit; never for ordinary plugin install, update, or delivery. Preserves deterministic preview, exact apply, verification and rollback receipts.

Update the shared LHC sources, generate and validate the native plugin, then publish and verify the selected installed loaders.

Before a PR into a repository you do not own, let that project's strict maintainer read the project and the diff and return a hard critique.

Fixed iteration loop for turning a reviewed defect list into a pushed, real-surface-proven result — parallel worker fixes, adversarial critique round, delta re-critique, converge at SHIP. Use when a review, audit, or incident produced concrete findings AND every finding is a bounded fix inside existing code. If the scope also creates a project, a service, a repository or an architecture decision, that is Full, not this loop.

Check existing model catalogs, read-only subscription quotas and assignment capabilities before council, delegation or focus acceptance; the allowed tier list governs those assignments while the route table is installation-specific evidence; never infer availability from a model name or telemetry status.

Choose and improve model assignments for Lead, executors and independent verification using task difficulty, actual capabilities, acceptance evidence and total cost.

The owner typed /plan or asked to update or check the plan. Relate the current state of the work to the plan and check that the plan is still true; if there is no plan, create it and check its decomposition.

Create or update LHC plan artifacts for every authorized engineering objective, including quickfixes, incident repairs, integrations and “try installing” requests; the user need not say “make a plan.” Before implementation or delegation, record the accepted goal and activate one executable leaf bound to the actual session. Read-only diagnosis may precede this. Status/read-only/research tasks and a direct owner no-LHC/STOP do not create a plan. Also use for explicit project-plan requests, task decomposition and plan refresh. The leaf is source of truth; graph, dashboard and Markdown are projections. Follow this procedure, not the tool source.

Lead-owned business-first procedure for choosing the least-cost route to the next real canary.

Mandatory final-gate procedure for proving a user-facing result on the real surface, with the interaction tool ladder (a11y tree -> browser MCPs -> CDP/Playwright -> XY clicks) and evidence rules.

Lead-owned claim-calibrated procedure for a controlled consequential release action.

Decompose large or stalled work into independently verifiable subtasks planned around 15 minutes where useful. Use when planning, assigning Workers, routes are entangled, or progress has produced little business delta.

Independently attempt a user goal through a real browser, device, application or CLI, returning observable usability and behavior evidence without reading implementation.

Worker-owned least-cost bug-fixing procedure for one bounded behavior failure. Use when a symptom, regression, broken business flow, failing test, or production-path defect must be reproduced, localized, and fixed. Reuse prior failure shields, prefer rg for fresh source tracing, choose the cheapest discriminating red proof, make the smallest coherent fix, and preserve a reusable root-cause map entry when recurrence is likely.

Worker-owned implementation procedure for writing the smallest coherent code change on the real business path. Use for a bounded feature, integration, endpoint, UI flow, adapter, or production-path implementation after the accepted outcome is known. Reuse verified project knowledge, confirm it with rg, ship the thinnest usable vertical, ask Lead non-blockingly at product decisions, and stop when the requested claim is proven.

Worker-owned least-cost codebase research for tracing real production paths, locating symbols and ownership, testing hypotheses, and handing Lead a decision-ready implementation route. Use for read-only repository investigation, architecture orientation, root-cause localization, the research stage of a bounded bugfix, or any task likely to repeat earlier code-location research. Search the reusable project code map first, prefer rg for fresh source truth, use Graphify only for genuinely multi-hop relationships, and use context-mode to process large outputs without flooding context.