Boot Sequence

If the OSTK_AGENTFILE environment variable is set, boot reads and parses the Agentfile at that path. Resolves:

• FROM <model> — which LLM the agent targets
• TOOL <name> — tool declarations (counted, not loaded)
• LIMIT, ISOLATION — resource and sandbox constraints

Missing Agentfile is non-fatal — boot continues with defaults. Present-but-unreadable is an error.

Hierarchical resolution loads operator identity and preferences. Resolution order:

1. ~/.HUMANFILE — global identity (MODEL, SIGN, TRUST directives are immutable from this layer)
2. .ostk/HUMANFILE — project-level config (may EXTENDS global)
3. Merge if both exist — project overlays global, but identity fields stay pinned to the global source

Directives resolved at this stage: MODEL, SIGN, PREPROCESS, TRUST, SECRET, EMBED, DRIVER, VERB, BOOT.

After loading, boot auto-populates the AVAILABLE model list by scanning for known API keys (ANTHROPIC_API_KEY, GEMINI_API_KEY, etc.) and initializes secret masking for any SECRET directives.

The identity and capability layer. Boot determines who you are and what you can do:

• Trust tier — GPG web-of-trust determines your tier (T0 through T3). Your GPG key is checked against the project's signing chain.
• ENTITYFILE — governance document loaded and GPG-verified. Invalid signatures are fatal (fail-closed). Missing ENTITYFILE is non-fatal (fresh-install path).
• Pin caps — capability pins loaded from .ostk/pin.caps. These constrain what tools agents can invoke.
• Drivers — declared drivers probed via Unix sockets. Boot never spawns drivers — it only checks what is already running. Inactive drivers are registered for demand-spawn on first use.
• .language — the verb registry is populated. Kernel services, MCP primitives, introspection tools, and HUMANFILE VERB directives are all registered into the .language file.

Before any of this, boot reaps dead agents from the process table and cleans up orphaned heartbeat files — housekeeping from the previous session.

Builds the context injection that every agent instance receives at session start. The preload is assembled from:

• Project structure scan (what files exist, language breakdown)
• .language register contents (all verbs, devices, services with their resolutions)
• boot.md swap — the compiled operator section replaces the template boot.md
• Identity blocks from HUMANFILE (communication style, scheduling preferences, corrections log)

The stable portion of this preload can be rendered once and uploaded as firmware — a deployment artifact that fleet members reference by ID rather than re-rendering independently.

Resolves the tack grammar and validates the dispatch loop:

• Tack grammar — HUMANFILE VERB definitions are validated against .language. Unresolvable verbs reduce boot confidence.
• INIT validation — the post-language INIT verb sequence is checked. Missing verbs in the INIT chain are flagged and penalize the boot confidence score.
• Verb momentum — recent tack resolution history from the audit log is loaded to seed the confidence gradient.

After this stage, the compressed register dump is printed — a one-line summary of .language state (verb count, fleet size, open needles).

The agent loop begins. Boot transitions from observation to action:

• Harness detection — environment is probed to determine tool routing (Claude Code, ostk-serve, CI, terminal). This controls whether the agent uses native tools or MCP primitives.
• Checkpoint verification — the most recent checkpoint's GPG signature is verified. Invalid signatures are fatal for governed agents, advisory for interactive sessions.
• Version pin — if .ostk/version exists, the running binary version is checked against the pin.
• Continuation prompt — if enabled, a dynamic prompt is generated from current workspace state and written to .ostk/prompts/continuation.md.

Tool calls from this point forward pass through the 7-step approval chain.