P1 ARDR epic: StarLang-defined LLM agents compiled to Prolog-RLM #197

New issue

Open

opened 2026-09-02 17:26:15 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-02 17:26:15 +00:00

Owner

Mission

Research and design first-class explicit LLM agents in StarLang, with V1 compiling the StarLang agent control-plane into lost-rob0t/prolog-rlm specifically.

ARDR result — 2026-08-29

Canonical Org research/design have now been materialized:

roam/research/star-lang/STAR-LANG-RESEARCH-029-explicit-agents-prolog-rlm.org
roam/design/star-lang/STAR-LANG-013-explicit-agents-prolog-rlm.org

Research state: READY_FOR_DESIGN.
Design state: DESIGN_READY_FOR_OPERATOR_REVIEW.
Implementation authorization: PENDING / AWAITING_OPERATOR_IMPLEMENTATION_APPROVAL.

Current-source finding

Current Prolog-RLM documents a supported prolog/rlm entrypoint and public module graph covering provider/model abstraction, opaque context operations, capability-gated tools, prompt compilation, completion/planning, recursion, supervised agents, graph/durable state, MCP, and structured traces. Its runtime boundary is already the desired ownership model: the model proposes semantic strategy while Prolog validates capabilities/budgets and owns execution semantics.

Current StarLang search found no first-class agent declaration. Existing StarLang actor/domain-server and expert-system work therefore does not already own this explicit-agent scope.

Design decision

ordinary StarLang actors/domain servers remain deterministic and cannot acquire an LLM/provider effect implicitly;
deterministic expert systems remain available without a model;
only an explicit StarLang agent declaration may enter Prolog-RLM agent/provider semantics;
StarLang owns parser/static validation/closed AgentIR/typed schemas/effect boundary/version compatibility;
Prolog-RLM owns V1 planning, prompt/context compilation, provider calls, recursion/subagents, memory, budgets, graph state and traces;
Common Lisp owns only a typed policy-free bridge to normal local/remote StarLang capability dispatch;
capability visibility never grants execution authority;
generated artifacts contain no arbitrary Prolog/host code or credentials;
V1 targets Prolog-RLM specifically, not generic Prolog or a multi-backend agent framework.

Dependency-ordered proposed implementation slices

All remain awaiting operator implementation approval:

StarLang agent parser + closed AgentIR.
Explicit LLM/effect + authority validation.
Current Prolog-RLM public ABI inventory + deterministic artifact emitter/loader.
Typed Common Lisp bridge using fake action source first.
Local/remote equivalence + cancellation/deadline/correlation.
Prolog-RLM supervised agent integration with provider-free fixtures.
Memory/delegation/verifier mapping.
Explicit OSINT researcher proof using deterministic Source Acquisition/Identity/Search capabilities.

RED-first targets

valid minimal agent declaration fails on untouched StarLang;
ordinary actor attempting LLM effect is rejected;
generated capability schema matches StarLang domain message types;
incompatible Prolog-RLM ABI fails explicitly;
visible but unauthorized capability produces denial and zero handler invocation;
typed action -> StarLang request -> typed result round trip;
local/remote equivalence;
cancellation/deadline fencing;
supervised child result/resume;
deterministic domain server operates with zero provider/model dependency.

Adversarial review

Rejected generic compile-to-Prolog abstraction, raw Prolog escapes, Common Lisp RLM rewrite, model calls inside deterministic domains, visibility-as-authority, bridge-owned planning, duplicate subagent/memory systems, arbitrary host/predicate calls, credential leakage, silent model fallback, local/remote divergence, source-tree coupling, and treating every actor as an agent.

Remaining design refinement

Before implementation Slice 3, inspect current Prolog-RLM source/tests and pin the exact stable programmatic predicates/modules used for agent materialization. The design intentionally does not invent predicate signatures from README-level evidence.

Implementation approval

PENDING / AWAITING_OPERATOR_IMPLEMENTATION_APPROVAL

Issue state, research readiness, completed design, adversarial review, or this handoff do not authorize RAGE implementation.

## Mission Research and design first-class explicit LLM agents in StarLang, with V1 compiling the StarLang agent control-plane into `lost-rob0t/prolog-rlm` specifically. ## ARDR result — 2026-08-29 Canonical Org research/design have now been materialized: - `roam/research/star-lang/STAR-LANG-RESEARCH-029-explicit-agents-prolog-rlm.org` - `roam/design/star-lang/STAR-LANG-013-explicit-agents-prolog-rlm.org` Research state: `READY_FOR_DESIGN`. Design state: `DESIGN_READY_FOR_OPERATOR_REVIEW`. Implementation authorization: **PENDING / AWAITING_OPERATOR_IMPLEMENTATION_APPROVAL**. ## Current-source finding Current Prolog-RLM documents a supported `prolog/rlm` entrypoint and public module graph covering provider/model abstraction, opaque context operations, capability-gated tools, prompt compilation, completion/planning, recursion, supervised agents, graph/durable state, MCP, and structured traces. Its runtime boundary is already the desired ownership model: the model proposes semantic strategy while Prolog validates capabilities/budgets and owns execution semantics. Current StarLang search found no first-class `agent` declaration. Existing StarLang actor/domain-server and expert-system work therefore does not already own this explicit-agent scope. ## Design decision - ordinary StarLang actors/domain servers remain deterministic and cannot acquire an LLM/provider effect implicitly; - deterministic expert systems remain available without a model; - only an explicit StarLang agent declaration may enter Prolog-RLM agent/provider semantics; - StarLang owns parser/static validation/closed AgentIR/typed schemas/effect boundary/version compatibility; - Prolog-RLM owns V1 planning, prompt/context compilation, provider calls, recursion/subagents, memory, budgets, graph state and traces; - Common Lisp owns only a typed policy-free bridge to normal local/remote StarLang capability dispatch; - capability visibility never grants execution authority; - generated artifacts contain no arbitrary Prolog/host code or credentials; - V1 targets Prolog-RLM specifically, not generic Prolog or a multi-backend agent framework. ## Dependency-ordered proposed implementation slices All remain awaiting operator implementation approval: 1. StarLang agent parser + closed AgentIR. 2. Explicit LLM/effect + authority validation. 3. Current Prolog-RLM public ABI inventory + deterministic artifact emitter/loader. 4. Typed Common Lisp bridge using fake action source first. 5. Local/remote equivalence + cancellation/deadline/correlation. 6. Prolog-RLM supervised agent integration with provider-free fixtures. 7. Memory/delegation/verifier mapping. 8. Explicit OSINT researcher proof using deterministic Source Acquisition/Identity/Search capabilities. ## RED-first targets - valid minimal `agent` declaration fails on untouched StarLang; - ordinary actor attempting LLM effect is rejected; - generated capability schema matches StarLang domain message types; - incompatible Prolog-RLM ABI fails explicitly; - visible but unauthorized capability produces denial and zero handler invocation; - typed action -> StarLang request -> typed result round trip; - local/remote equivalence; - cancellation/deadline fencing; - supervised child result/resume; - deterministic domain server operates with zero provider/model dependency. ## Adversarial review Rejected generic compile-to-Prolog abstraction, raw Prolog escapes, Common Lisp RLM rewrite, model calls inside deterministic domains, visibility-as-authority, bridge-owned planning, duplicate subagent/memory systems, arbitrary host/predicate calls, credential leakage, silent model fallback, local/remote divergence, source-tree coupling, and treating every actor as an agent. ## Remaining design refinement Before implementation Slice 3, inspect current Prolog-RLM source/tests and pin the exact stable programmatic predicates/modules used for agent materialization. The design intentionally does not invent predicate signatures from README-level evidence. ## Implementation approval `PENDING / AWAITING_OPERATOR_IMPLEMENTATION_APPROVAL` Issue state, research readiness, completed design, adversarial review, or this handoff do not authorize RAGE implementation.