Agent infrastructure
Four skills built on Liminate.
Liminate now does more than run readable programs. It is also the small, inspectable language underneath four Agent Skills.
Liminate now acts as a bounded substrate for four portable Agent Skills: session contracts, prompt-to-intent compilation, deterministic context paging, and continuity handoff packets.
liminate-session-contracts
Persistent truth layer
Tracks what an agent has actually verified, what it is inferring, which decisions are locked, which questions remain open, and which user corrections should persist.
Its benchmark headline is simple: 252 cross-session continuity probes, zero fabrications.
It adds a session pack with claim, source, decision, cite, and verify; cite is checked by the Liminate interpreter as a runtime substring constraint.
prosecode-prompt-compiler
Request meaning layer
Compiles a user request into a compact Intent IR before the response is written, so the agent knows the verb, slots, missing information, contradictions, and response posture.
The user does not see the scaffold. They get an answer aimed at the real request.
The IR has seven verbs: explain, create, transform, analyze, decide, plan, and fix.
prosecode-context-pager
Attention management layer
Profiles long conversation history, scores each block against the active intent, then records whether to retain, page, or evict it.
Nothing is dropped silently. The choice is written down.
Its score combines keyword overlap with recency; the default retain and evict thresholds are deterministic and tunable without embeddings or network calls.
prosecode-handoff-packet
Continuity transfer layer
Packages the working state so another agent, model, or tool can continue without flattening progress into a vague summary.
It preserves what was verified, what was inferred, what remains open, what the user corrected, and what the next agent may honestly claim.
It emits a structured Markdown packet with fourteen required sections, including verified sources, inferred claims, locked decisions, open questions, active corrections, and epistemic boundaries.
The chain
The prompt compiler decides what the agent is trying to do. The context pager keeps the useful context in view. Session contracts record what was verified, kept, paged, or corrected. The handoff packet carries that state to the next agent without pretending uncertainty disappeared.
prosecode-prompt-compiler produces intent slots; prosecode-context-pager scores history against those slots; liminate-session-contracts persists the resulting claims, citations, paging directives, and corrections as Liminate-readable state; prosecode-handoff-packet exports the continuity boundary as a Markdown transfer object.
This is the same thesis in a different layer: bounded readable language as a trust surface.