Wire warm context, cold-history boundary, and lazy retrieval #106

Merged

lost-rob0t merged 32 commits from agent/conversation-cold-context into main

2026-08-19 04:13:57 +00:00

lost-rob0t commented

2026-08-18 20:13:00 +00:00

(Migrated from github.com)

What changed

Implements the managed hot/warm/cold conversation path for unbounded sessions.

durable transcript remains authoritative and unmodified;
existing warm artifacts are automatically loaded/ranked when warm_store/1 is configured;
adds a synthetic token-accounted managed_cold_history_boundary context unit when history lies outside the guaranteed hot tail;
the boundary tells the planner that older turns may be absent and shows the exact bounded search, slice, and peek forms for recovering them;
boundary instructions are projection state only: old user/assistant/tool messages are never rewritten;
adds a generic trusted external-context adapter and conversation-backed opaque cold handle;
managed turns no longer materialize terms(FullTranscript);
public rlm conversation pack/ledger/turn APIs route through rlm_conversation_runtime;
automatic warm production/compaction remains explicit and is not triggered under token pressure.

Canonical context contract

complete durable transcript
        |
        +--> bounded rolling hot context
        +--> existing warm artifacts
        +--> synthetic cold-history boundary
        `--> opaque lazy cold-history context
             peek / slice / search
        |
        v
shared hard token budget

Cold-history boundary

When the projected conversation extends beyond min_recent_turns, the managed runtime adds a mandatory, provider-visible boundary that reports the older prefix as outside the guaranteed hot tail and instructs the planner to recover original history before guessing.

Representative operations:

context(input(context), search("query"), Result).
context(input(context), slice(Start, Length), Result).
context(input(context), peek(item(Index)), Result).

The boundary is counted in the normal token ledger and exposed in result metadata as cold_history_boundary. Specialized callers can disable it with cold_history_boundary(false).

Tests

The managed-runtime tests prove:

warm context loads without caller-provided context_units/1;
managed turns reuse warm context without publishing new warm artifacts;
the cold-history boundary is mandatory and token-accounted;
the boundary text reaches the planner;
original old message refs/content remain unchanged after the turn;
the boundary can be explicitly disabled.

The existing cold-context end-to-end test also proves an old turn absent from active attention can be recovered through RLM search.

GitHub Actions on the exact head is authoritative.

## What changed Implements the managed hot/warm/cold conversation path for unbounded sessions. - durable transcript remains authoritative and unmodified; - existing warm artifacts are automatically loaded/ranked when `warm_store/1` is configured; - adds a synthetic token-accounted `managed_cold_history_boundary` context unit when history lies outside the guaranteed hot tail; - the boundary tells the planner that older turns may be absent and shows the exact bounded `search`, `slice`, and `peek` forms for recovering them; - boundary instructions are projection state only: old user/assistant/tool messages are never rewritten; - adds a generic trusted external-context adapter and conversation-backed opaque cold handle; - managed turns no longer materialize `terms(FullTranscript)`; - public `rlm` conversation pack/ledger/turn APIs route through `rlm_conversation_runtime`; - automatic warm production/compaction remains explicit and is not triggered under token pressure. ## Canonical context contract ```text complete durable transcript | +--> bounded rolling hot context +--> existing warm artifacts +--> synthetic cold-history boundary `--> opaque lazy cold-history context peek / slice / search | v shared hard token budget ``` ## Cold-history boundary When the projected conversation extends beyond `min_recent_turns`, the managed runtime adds a mandatory, provider-visible boundary that reports the older prefix as outside the guaranteed hot tail and instructs the planner to recover original history before guessing. Representative operations: ```prolog context(input(context), search("query"), Result). context(input(context), slice(Start, Length), Result). context(input(context), peek(item(Index)), Result). ``` The boundary is counted in the normal token ledger and exposed in result metadata as `cold_history_boundary`. Specialized callers can disable it with `cold_history_boundary(false)`. ## Tests The managed-runtime tests prove: - warm context loads without caller-provided `context_units/1`; - managed turns reuse warm context without publishing new warm artifacts; - the cold-history boundary is mandatory and token-accounted; - the boundary text reaches the planner; - original old message refs/content remain unchanged after the turn; - the boundary can be explicitly disabled. The existing cold-context end-to-end test also proves an old turn absent from active attention can be recovered through RLM search. GitHub Actions on the exact head is authoritative.