Runtime Behavior

This document explains how Liminal behaves during real runs.

World Context Grounding

Root sessions inject a world-context block once per fresh session. It includes:

• current local date/time + timezone
• OS/shell/path conventions
• workspace root and project metadata
• git state
• available tools/signals
• memory/vault summaries
• repo map hints

This anchors model behavior to the actual machine and date.

Execution State and Long-Horizon Planning

Liminal tracks an ExecutionState model containing:

• mission
• milestones
• contracts
• commitments
• driftScore
• recoveryLog

During runs, contract transitions and drift events are emitted for observability and eval assertions.

Drift and Recovery

Drift score increases via periodic cadence and failure signals. When thresholds are crossed, replanning hints or state transitions can trigger.

When all tools in a round fail:

• error summary is appended
• recovery hints are injected
• recovery action is logged/emitted
• structured reflexion is attempted: calls the configured model to extract {lesson, root_cause, fix_pattern} JSON and stores the result as a reflection: typed memory note for future sessions. Falls back to plain-text if the model call fails. Controlled by AGENT_REFLEXION_SEMANTIC (defaults on).
• bumpRuleHits() is called with the error context to increment errors-prevented counters for any R-* rule IDs referenced in the failure.

Finalization and Critic Pass

Before turn completion, runtime may perform:

• path citation nudges
• synthesis checklist nudges for research outputs
• verify_result critic checks for heavy/path-rich completions
• memory auto-extraction
• vault auto-write (policy dependent)

These steps improve output quality and reduce silent regressions.

Vault Policy Semantics

Current default behavior:

• retrieval order guidance is advisory (memory -> vault -> web)
• auto-write mode defaults to research-oriented persistence (unless disabled)

Temporal Anchoring

Time-sensitive web queries (latest/news/current/update) are normalized to the current year unless the user explicitly asks for historical coverage.

This prevents silent drift into stale years.

Runtime Heartbeats

Each round can emit heartbeat telemetry including:

• round index
• uptime
• drift score
• active contract id

Heartbeats are used by UI traces and eval assertions for long-horizon health checks.

Harness protocol rule recall

On ReAct round 2 of each send(), unless AGENT_RULE_RECALL=0, the harness appends a system message listing named protocol rules (IDs such as R-PLAN-3STEPS, R-DEDUP-TOOLS, R-CLOSED-ARTIFACT, …) covering planning, citations, deduplicated retrieval, closed HTML artifacts, tool-error hygiene, research budgets, typography, and more. The text is defined in packages/core/src/harness_rules.ts and mirrors the spirit of the always-on protocol in packages/tools/src/systemPrompt.ts.

Purpose:

• reinforce habits that prevent broken half-files, duplicate memory_query loops, and ignored tool error messages
• surface chronically violated rules when .agent_rule_stats.json is populated (adaptive builder)

Full rule table and web_fetch Readability notes: Harness protocol.

Conversational Self-Management

The harness can interpret plain-language runtime directives (for example: "from now on use model X") with an internal JSON extraction pass.

Behavior:

• low-risk changes can be applied immediately
• risky policy relaxations require explicit confirmation
• approved changes persist to .agent_runtime_prefs.json at workspace root
• persisted preferences are loaded at startup by TUI/web/eval entrypoints

Preference precedence:

1. in-turn explicit user directive
2. persisted runtime preferences
3. environment configuration
4. hardcoded defaults