Conditional AI and project toolchain

This page embeds the current generated execution matrix. The public MCP action inventory is separate from the tool-requirement inventory.

Current counts are derived release facts, not permanent ceilings.

Hardware acceleration providers

The default GPU memory ceiling is 80% with explicit headroom. It limits admitted VRAM; it never forces a utilization percentage. A missing/incompatible runtime, exhausted budget, thermal/throttle signal, or ineligible action yields a receipted CPU fallback.

Accelerators are not included in the tool count or MCP action count.

Derived tool count: 119. Public MCP action count is separate.

Host plane: Codex Desktop, Codex CLI, and Codex VM only. ChatGPT is a separate future plane.

Execution law: tools run only when the active lane/action/source type selects them. Presence never means run everything. SQLite remains durable authority; analytical, vector, graph, and web tools produce bounded evidence for SQLite persistence.

Eighteen project-sector lanes

Named/root authority surfaces

Git, search, JSON, CI, and public adapter

Hidden runtime, MCP, API, and tunnel

Authority, indexing, and retrieval

Graphs, AST, and reconciliation

Internal governance, build, and quality

Documents, OCR, and media

Data, Excel, and databases

Web, research, and source intake

RAG and workflow composition

Evaluation and testing

Observability and telemetry

Derived control facts

• Codex host profiles: CODEX_DESKTOP, CODEX_CLI, CODEX_VM.
• ChatGPT plane mixed: false.
• Public action count fixed by this matrix: false.
• Skill count fixed by this matrix: false.
• Hook class count fixed by this matrix: false.
• Every retained tool requires a package/runtime identity and an execution test before local install.
• License-classified tool requirements: 119 of 119.
• Exact copied runtime distribution licenses are materialized before tunnel startup; MCP actions remain a separate inventory.

Source-bound workflow map

This page is projected from the same current executable snapshot as the rest of the documentation set. The map is deliberately two-directional: each horizontal district shows peer stages while vertical edges show ownership and state progression.

flowchart TB
    subgraph InputDistrict["Input and classification"]
      direction LR
      A["Condition-true capability need"] --> B["Tool requirement classification"] --> C["One of 119 tool requirements"]
    end
    subgraph ExecutionDistrict["Selection and execution"]
      direction TB
      D["Primary and eligible fallback order"] --> E["Local, SDK, MCP, tunnel, or external execution"] --> F["Availability, license, and result validation"]
    end
    subgraph EvidenceDistrict["Evidence and outcome"]
      direction LR
      G["Tool execution receipt"] --> H["Owning action phase"]
      G -. mismatch .-> I["Never confuse tools with 91 MCP actions"]
    end
    C --> D
    F --> G

Contract and readback

Canonical source owners

• toolchains/tool-requirement-matrix.v1.json
• toolchains/tool-execution-routing.v1.json
• toolchains/tool-license-inventory.v1.json

Exact backend readback

Cross-surface invariants

• The current snapshot contains 91 public actions, 26 skills, 11 hook events / 44 handlers, 119 tool requirements, 18 sector lanes, and 11 named authorities. These are derived counts, not fixed ceilings.
• Executable ownership stays one-way: skills select, MCP exposes, the outer SDK routes, the internal SDK executes, ENV selects, UOP governs, tools perform bounded work, hooks emit receipts, and the owning authority validates effects.
• Any missing identity, schema, grant, capability, dependency, receipt, or authority proof must fail closed at its owning phase; a later green check cannot retroactively authorize the skipped boundary.
• A changed route refreshes every dependent schema, manifest, generator, test, diagram, and documentation reference; the superseded executable route is directly purged in the same Delta.
• Tests, Git, CI, installation, restart, deployment, discussion, or a rendered page never imply Project HIL, Learning HIL, Goal completion, or pointer movement.

This page is a Git-tracked documentation projection. Executable source, SQLite authorities, installed-runtime receipts, and explicit human gates remain the governing evidence.