Skip to main content

Write GenesisCode Skill Pack v0.1

Versioned distribution artifact for agent systems authoring GenesisCode.

This pack is designed for Codex/Claude-style coding agents where AI writes nearly all .gc and tooling evolution code.

Canonical Artifacts

  • Contract JSON: docs/spec/WRITE_GENESISCODE_SKILL_PACK_v0.1.json
  • Skill source: .agents/skills/genesiscode-authoring/SKILL.md
  • Pointer/onboarding doc: docs/write_genesisCode_skill.md
  • Bundle entrypoint: docs/spec/AGENT_AUTHORING_BUNDLE_v0.1.md
  • Frozen training profile: docs/spec/GC_AGENT_PROFILE_v0.3.json
  • Generated compact core card: docs/spec/GC_AGENT_CORE_CARD_v0.3.md
  • Intent-selected task-card registry: docs/spec/GC_AGENT_TASK_CARDS_v0.3.json
  • Exact language-symbol registry: docs/spec/GC_AGENT_SYMBOL_INDEX_v0.3.json
  • Skill contract JSON: docs/spec/WRITE_GENESISCODE_SKILL_v0.1.json
  • Executable conformance spec: docs/spec/WRITE_GENESISCODE_SKILL_CONFORMANCE_v0.1.md
  • Distribution kit spec: docs/spec/WRITE_GENESISCODE_SKILL_DISTRIBUTION_v1.md
  • Distribution manifest: docs/skill_pack/write_genesiscode_v1/manifest.json
  • Schema refs: docs/spec/CLI_JSON_SCHEMAS_v0.1.md, docs/spec/GCPM_JSON_SCHEMAS_v0.1.md
  • Test/profile refs: docs/spec/TEST_EXECUTION_PROFILES_v0.1.md
  • Strategic execution source: the generated slice from ROADMAP.md and policies/roadmap_execution_v0.1.json
  • Conditional unresolved P0/P1 defect source: open checklist IDs in upgrade_plan.md; an empty file supplies no work

Pack Objectives

  1. Keep agent authoring deterministic and contract-first.
  2. Keep language/runtime evolution selfhost-biased.
  3. Keep edits machine-reviewable with explicit acceptance evidence.
  4. Route user tasks, active defects, generated-roadmap work, and exploration through distinct authority and closure rules.
  5. Freeze authoring assumptions by negotiating GC-AGENT-v0.3 before source generation.
  6. Enforce all five unsupported classes and their safe alternatives without prompt- or index-derived authority.

Work Routing

  • user-task: preserve the current concrete request while enforcing normative constraints.
  • active-defect: select only an open, reproduced P0/P1 ID; unavailable when the defect ledger is empty.
  • roadmap-task: select only from the machine-generated roadmap execution slice.
  • exploratory-work: retain bounded observations or proposals without status-closure authority.

The selected route, authority source, empty-source behavior, and stop conditions are part of the agent’s work report.

Distribution Layout (Normative)

  • skill_file
    • canonical skill text agents should load first for GenesisCode code changes.
  • contract
    • machine-verifiable drift controls (required sections/spec refs/contract ids).
  • prompt_templates
    • structured prompts for common work loops.
  • acceptance_profiles
    • deterministic command bundles keyed by change class.
  • anti_patterns
    • explicit invalid behaviors to reject automatically.

Prompt Templates

backlog_slice

  • Intent: classify one concrete request as user-task, active-defect, roadmap-task, or exploratory-work, then execute only that route’s authority-compatible transaction.
  • Empty-source rule: no open P0/P1 ID means the defect route is unavailable; roadmap progression comes only from the generated execution slice.
  • Required output:
    • selected route ID, authority source, and coherent transaction ID
    • changed files
    • command evidence and pass/fail
    • updated remaining roadmap and P0/P1 counts without invented work.

capability_addition

  • Intent: add/expand capability wrappers with policy + replay + docs.
  • Required output:
    • wrapper contract shape
    • policy keys
    • replay determinism proof path
    • updated capability indices and drift gates.

selfhost_cutover

  • Intent: migrate semantics from Rust boundary into .gc selfhost modules.
  • Required output:
    • removed fallback/deprecated path list
    • strict selfhost gate evidence
    • bootstrap/archive safety checks.

Acceptance Profiles

docs_only

  • bash scripts/check_doc_hygiene.sh
  • bash scripts/check_planning_docs_fresh.sh

authoring_contract

  • bash scripts/check_gc_agent_core_card.sh
  • bash scripts/check_gc_agent_task_cards.sh
  • bash scripts/check_gc_agent_symbol_index.sh
  • bash scripts/check_gc_agent_profile.sh
  • bash scripts/check_genesiscode_authoring_skill.sh
  • bash scripts/check_write_genesiscode_skill_pack.sh
  • bash scripts/check_write_genesiscode_skill_distribution.sh
  • bash scripts/check_agent_authoring_bundle.sh

feature_matrix_claims

  • bash scripts/check_feature_matrix_gap_hygiene.sh
  • bash scripts/check_capability_evidence_ledger.sh

skill_conformance_executable

  • bash scripts/check_agent_reference_workflows.sh
  • bash scripts/check_agent_generative_workloads.sh
  • bash scripts/check_write_genesiscode_skill_conformance.sh
  • GENESIS_WRITE_SKILL_DIST_VERIFY_RUNTIME=1 bash scripts/check_write_genesiscode_skill_distribution.sh

selfhost_sidecar

  • bash scripts/check_selfhost_artifact_fresh.sh
  • bash scripts/check_selfhost_toolchain_review_fresh.sh

fast_iteration_non_crate

  • bash scripts/test_changed_fast.sh --base HEAD --runner auto --budget-ms 120000 --min-history 1

Distribution Kit v1

Canonical executable distribution kit root:

  • docs/skill_pack/write_genesiscode_v1/manifest.json

Normative content:

  • generated prompt cards in docs/skill_pack/write_genesiscode_v1/prompt-cards.json
  • generated executable recipe cards in docs/skill_pack/write_genesiscode_v1/recipe-cards.json
  • compact progressive-disclosure index in docs/skill_pack/write_genesiscode_v1/authoring-card.md
  • hand-owned source policy in policies/genesiscode_authoring_workflow_v0.1.json
  • project templates under docs/skill_pack/write_genesiscode_v1/templates/
  • deterministic distribution verifier: scripts/check_write_genesiscode_skill_distribution.sh
  • gpu/xr productization verifier: scripts/check_gpu_xr_productization_kits.sh

Benchmark Suite

Required executable benchmark workloads for AI-authored GenesisCode quality:

  • Agent task matrix: benchmarks/agent_tasks/v0.1/suite.json
  • Model-agnostic scoring authority: docs/spec/GC_AGENT_BENCHMARK_SCORING_v0.1.json
  • Closed score-result contract: docs/spec/GC_AGENT_BENCHMARK_SCORE_v0.1.schema.json
  • Reproducible run contract: docs/spec/GC_AGENT_BENCHMARK_RUN_v0.1.schema.json
  • Benchmark protocol: docs/spec/GENESISBENCH_PROTOCOL_v0.1.json with guides/genesisbench.qmd
  • Local benchmark model effect: docs/spec/GC_AGENT_MODEL_RUNNER_EFFECT_v0.1.json
  • Held-out commitment authority: docs/spec/GC_AGENT_HELD_OUT_EVALUATION_v0.1.json (public commitments only; private custody material is forbidden from the skill distribution)
  • Service workflow: examples/agent_service_workflow/workflow.sh
  • Game-loop workflow: examples/agent_long_running_gfx_loop_workflow/workflow.sh
  • GPU compute workflow: examples/agent_gpu_compute_workflow/workflow.sh
  • GPU non-graphics workflow: examples/agent_compute_workflow/workflow.sh
  • XR runtime workflow: examples/agent_xr_runtime_workflow/workflow.sh
  • XR browser conformance workflow: scripts/check_webxr_browser_conformance.sh
  • Package workflow: examples/agent_multi_package_publish_workflow/workflow.sh
  • Mutation workload suite: scripts/check_agent_generative_workloads.sh

Scoring Rubric

Candidate quality is scored by scripts/lib/gc_agent_scoring.py on a 10,000-basis-point scale across semantics, obligations, effects, patch minimality, deterministic resource use, and policy scope. Validity fails closed when required semantic, obligation, policy, or editable-scope facts fail. Wall time, API cost, energy, and provider queue time are separate genesis/agent-benchmark-run-v0.1 facts and never alter quality.

Every reported benchmark invocation must also pass python3 scripts/lib/gc_agent_benchmark_run.py --check. The record binds immutable model, weights, tokenizer, runtime, exact prompt/card/context assembly, integer decoding and retry controls, all attempts and candidate artifacts, the canonical score, normalized host facts, and a complete artifact inventory. Local model execution is permitted only through the digest-pinned genesis.agent-model-runner.v0.1 / infer effect; preserve request, response, transcript, and .gclog, and require replay without model reinvocation.

Cross-model comparisons additionally pass python3 scripts/lib/genesisbench_protocol.py --check and deterministic run-plus-attestation classification. The profile freezes the repository/runtime/docs snapshot, context and tool surfaces, capabilities, attempts, disclosure, visibility, scoring, contamination, and eligibility. Treat public references as declared-contaminated unranked practice evidence, incomplete provenance as unknown, and new language status as no evidence. Judge-model preference is outside quality.

The skill-conformance score below measures whether this authoring pack covers its required workflows; it is not a candidate benchmark score.

  • Conformance gate: scripts/check_write_genesiscode_skill_conformance.sh
  • Report kind: genesis/write-genesiscode-skill-conformance-v0.1
  • Default threshold: 100/100 (GENESIS_WRITE_SKILL_CONFORMANCE_MIN_SCORE)
  • Categories (20 points each):
    • service workflow determinism + domain coverage
    • game loop determinism + graphics coverage
    • GPU compute determinism + compute coverage
    • package workflow determinism + package/sync coverage
    • generative mutation suite stability (min case count + no parity/history-min failures)

Failure-Mode Playbooks

  • Missing workflow/category:
    • restore failing workflow script under examples/agent_*_workflow/
    • rerun scripts/check_agent_reference_workflows.sh.
  • Replay mismatch:
    • inspect workflow run/replay output deltas; reject nondeterministic effects first.
  • GPU backend contract failure:
    • rerun strict device lane: GENESIS_AGENT_GPU_REQUIRE_DEVICE=1 bash scripts/check_agent_reference_workflows.sh.
  • Generative regression or insufficient history:
    • refresh baseline history seeds under policies/perf/
    • rerun scripts/check_agent_generative_workloads.sh with explicit regression params.

Anti-Patterns (Reject)

  • Claiming feature parity without capability-evidence ledger updates.
  • Marking plan items done without command evidence.
  • Leaving deprecated runtime paths active after selfhost replacement.
  • Emitting non-deterministic host effects without replay coverage.
  • Expanding monolithic files when module boundaries are obvious.

Agent Output Contract

Every substantial turn should include:

  1. Completed IDs.
  2. Remaining IDs count.
  3. Validation commands run.
  4. Any blockers + root cause.
  5. Next highest-impact slices.

Conformance

Gate: - scripts/check_gc_agent_core_card.sh - scripts/check_gc_agent_task_cards.sh - scripts/check_gc_agent_symbol_index.sh - scripts/check_gc_agent_profile.sh - scripts/check_write_genesiscode_skill_pack.sh - scripts/check_write_genesiscode_skill_guide.sh - scripts/check_write_genesiscode_skill_distribution.sh - scripts/check_write_genesiscode_skill_conformance.sh - scripts/check_gpu_xr_productization_kits.sh

The gate fails closed on contract drift, missing pack artifacts, or missing bundle/onboarding links.