Agent Onboarding Path v0.1
Last updated: 2026-02-22
Purpose: provide a minimal, stable retrieval path for AI coding agents. This file is the single agent-first onboarding spine.
Canonical Read Order (Required)
docs/INDEX.mddocs/spec/GC_AGENT_CORE_CARD_v0.3.mddocs/spec/GC_AGENT_PROFILE_v0.3.jsondocs/spec/GC_AGENT_TASK_CARDS_v0.3.mdselected through declared task intentdocs/spec/GC_AGENT_SYMBOL_INDEX_v0.3.jsonthrough exactagent-index --symbollookupdocs/spec/GC_DIAGNOSTIC_CATALOG_v0.1.jsonthrough exactagent-index --diagnosticlookupexamples/canonical_language/v0.1/suite.jsonfor signed valid/invalid executable pairsdocs/spec/GC_AGENT_BENCHMARK_SCORING_v0.1.jsonfor model-agnostic evaluationdocs/spec/GC_AGENT_BENCHMARK_RUN_v0.1.schema.jsonanddocs/spec/GC_AGENT_MODEL_RUNNER_EFFECT_v0.1.jsonfor reproducible model/run provenancedocs/spec/GENESISBENCH_PROTOCOL_v0.1.jsonandguides/genesisbench.qmdfor benchmark context, tools, contamination, and eligibilitydocs/spec/CLI_TOOLING_BUNDLE_v0.1.mddocs/spec/GCPM_BUNDLE_v0.1.mddocs/spec/HOST_RUNTIME_BUNDLE_v0.1.mddocs/spec/TESTING_BUNDLE_v0.1.mddocs/spec/AGENT_AUTHORING_BUNDLE_v0.1.mddocs/spec/WRITE_GENESISCODE_SKILL_PACK_v0.1.mddocs/spec/WRITE_GENESISCODE_SKILL_DISTRIBUTION_v1.mdROADMAP.mdfor strategic work orupgrade_plan.mdfor active P0/P1 defects
Agent-First Onboarding Spine (Required)
- Language semantics:
docs/spec/GC_AGENT_CORE_CARD_v0.3.md,docs/spec/GC_AGENT_PROFILE_v0.3.json, intent-selecteddocs/spec/GC_AGENT_TASK_CARDS_v0.3.md, exact symbol anddocs/spec/GC_DIAGNOSTIC_CATALOG_v0.1.jsonrecords,docs/spec/CLI_TOOLING_BUNDLE_v0.1.md,docs/spec/HOST_RUNTIME_BUNDLE_v0.1.md - Runtime profiles:
docs/spec/RUNTIME_BACKEND_PROFILES_v0.1.md,docs/spec/TESTING_BUNDLE_v0.1.md - Packaging and deployment:
docs/spec/GCPM_BUNDLE_v0.1.md,docs/spec/GCPM_WORKFLOW_REPORTS_v0.1.md - Assurance and traceability:
docs/spec/ASSURANCE_ARTIFACTS_v0.1.md,docs/spec/ASSURANCE_PROFILE_PACKS_v0.1.md,docs/spec/ASSURANCE_STANDARDS_CROSSWALK_v0.1.md - Active execution risk:
ROADMAP.md,upgrade_plan.md,docs/status/REDTEAM_REPORT.md,.genesis/perf/selfhost_readiness_report.json
Canonical Domain Sources
- CLI/tooling contracts:
docs/spec/CLI_TOOLING_BUNDLE_v0.1.md - Project/package manager contracts:
docs/spec/GCPM_BUNDLE_v0.1.md - Runtime/capability/host contracts:
docs/spec/HOST_RUNTIME_BUNDLE_v0.1.md - GPU/gfx contracts:
docs/spec/GPU_GFX_BUNDLE_v0.1.md - GPU compute contracts:
docs/spec/GPU_COMPUTE_BUNDLE_v0.1.md - GPU/XR productization templates + lanes:
docs/spec/GPU_COMPUTE_BUNDLE_v0.1.md(Productization Kits (Non-Gfx + XR)section) - GFX runtime contracts:
docs/spec/GFX_RUNTIME_BUNDLE_v0.1.md - Test/perf/release lanes:
docs/spec/TESTING_BUNDLE_v0.1.md - Agent authoring entrypoint:
docs/spec/AGENT_AUTHORING_BUNDLE_v0.1.md - Public 27-case authoring benchmark:
benchmarks/agent_tasks/v0.1/suite.json - Model-agnostic benchmark scoring:
docs/spec/GC_AGENT_BENCHMARK_SCORING_v0.1.json - Reproducible benchmark run contract:
docs/spec/GC_AGENT_BENCHMARK_RUN_v0.1.schema.json - GenesisBench profile and eligibility rules:
docs/spec/GENESISBENCH_PROTOCOL_v0.1.jsonandguides/genesisbench.qmd - Local benchmark model effect:
docs/spec/GC_AGENT_MODEL_RUNNER_EFFECT_v0.1.json - Versioned skill pack:
docs/spec/WRITE_GENESISCODE_SKILL_PACK_v0.1.md - Executable skill distribution kit:
docs/spec/WRITE_GENESISCODE_SKILL_DISTRIBUTION_v1.md - Active risk backlog:
upgrade_plan.md - Cross-language feature delta:
feature_matrix.md
Retrieval Rules (Agent-Facing)
- Prefer bundle docs before split docs.
- Prefer bundle docs before any non-bundle spec page, even if a direct symbol/path match exists.
- Treat split docs with a legacy banner as detail references, not first retrieval targets.
- Do not use superseded top-level docs as normative sources; consult
docs/DEPRECATION_MAP_v0.1.md. - Treat
experimental-syntax,host-only-operation,unavailable-target,nondeterministic-facility, andout-of-profile-capabilityas mandatory fail-closed classes. Use each record’ssafeAlternative; profile/capability negotiation is explicit and cannot be inferred from availability. - Route failures on exact
genesis/diagnostic/v1/...IDs or cataloged codes, never prose. PincatalogIdentitySha256for training/evaluation and useagent-index --diagnostic <exact-code>rather than loading the full catalog. - For language generation and repair, retrieve the nearest
GC-CANONICAL-EXAMPLES-v0.1pair, execute both recorded scenarios, and apply only its declared one-site repair. Invalid fixtures are counterexamples, never standalone templates. - For public development evaluation, select the exact task and context tier from
GC-AGENT-TASK-BENCHMARK-v0.1; its references are public oracles and cannot support a held-out claim. - Score candidate quality only with
GC-AGENT-BENCHMARK-SCORING-v0.1: semantics, obligations, effects, patch minimality, deterministic resource units, and policy scope share a closed 10,000-basis-point result. Keep latency, API cost, energy, and provider queue time in the separate model/run record. - Validate the separate
genesis/agent-benchmark-run-v0.1record read-only. It must bind immutable model/runtime artifacts, exact prompt/card/context assembly, integer decoding and retry policy, every attempt and candidate artifact, the canonical score, normalized host facts, and a complete content-addressed inventory. - Validate
GenesisBench-v0.1before comparing runs. Its frozen Git tree is additionally bound by a complete SHA-256 content manifest; context modes, MCP tools, capabilities, attempts, model disclosure, visibility, scoring, contamination, and eligibility are closed authorities. Public references aredeclared-contaminatedand unranked. Newness never proves cleanliness, and judge-model preference never changes quality. - Fully local benchmark models use the pinned
genesis.agent-model-runner.v0.1/infereffect only. Preserve request, response, tool transcript, and.gclog; replay must succeed without the model or weights and must never reinvoke either. - For held-out evaluation, retrieve only
GC-AGENT-HELD-OUT-v0.1public commitments. Never load.genesis/private/agent-evaluationinto training, retrieval, prompts, logs, or distributed artifacts; an evaluator with custody verifies the private pack separately and labels missing training provenance asunknowncontamination.