Runtime profile test cases

Use these cases when Agent Context is used by Agent Runtime, Lime AgentRuntime Profile, or an equivalent runtime spine. The goal is to prove that context facts can be referenced by runtime turns without turning Agent Context into a runtime state owner.

Canonical boundary

ContextSurface / ContextItem / SourceRef / Selection / Assembly / Compaction / MissingContext
  -> context refs on RuntimeEvent, ThreadReadModel, TaskSnapshot, or EvidencePack
  -> Agent UI and Agent Evidence consume refs and summaries

Agent Context owns what context was available, selected, assembled, injected, omitted, compacted, redacted, or missing. Agent Runtime owns turn status, task status, retries, tool execution, action waits, and recovery.

Runtime correlation fields

Compatible context events and envelopes SHOULD preserve these ids when available:

Field	Purpose
runtime_id	Runtime installation or runtime instance.
session_id	Durable work container that requested context.
thread_id	Ordered execution context.
turn_id	Turn whose context was selected or injected.
task_id / run_id / attempt_id	Managed task attempt using the context.
step_id / tool_call_id	Runtime step or tool call that consumed context.
action_id	Human/policy action that requested more context.
evidence_id	Evidence pack or source map that references the context.

Test cases

ID	Case	Input facts	Expected result
ACTX-AR-ID-001	Turn context snapshot keeps ids	Context envelope for a submitted turn	Envelope preserves session_id/thread_id/turn_id; runtime can reference context_id.
ACTX-AR-ID-002	Task context keeps attempt ids	Selection for task retry	Selection records preserve task_id/run_id/attempt_id; old attempt context is not overwritten.
ACTX-AR-SEL-001	Selected vs omitted are explicit	Candidate context set exceeds budget	Selection records include selected and omitted items with reasons.
ACTX-AR-BUDGET-001	Budget/window decision is traceable	Context window is constrained	Budget record explains truncation; runtime does not guess from prompt text.
ACTX-AR-ASM-001	Injection is auditable	Model request uses assembled context	Assembly/injection record points to source refs and context item ids.
ACTX-AR-MISSING-001	Missing context becomes requestable fact	Runtime cannot access needed file/source	Missing-context record includes owner, requested action, and safe user-facing summary.
ACTX-AR-COMPACT-001	Compaction preserves boundary	Old thread is compacted	Summary references source window and does not replay old reasoning as final answer.
ACTX-AR-POLICY-001	Redaction is a context fact with policy refs	Sensitive item is redacted	Context item preserves redaction reason and policy_refs; raw secret is not injected.
ACTX-AR-EVID-001	Evidence can cite context refs	Evidence pack exports source map	Evidence refs point to context/source refs instead of copying all context payloads.
ACTX-AR-GOV-001	Context does not own runtime status	Context event contains runtime ids	No context record declares turn completed, failed, approved, or tool succeeded.

Failure cases

These are incompatible with the boundary:

1. A context record declares task completion or tool success.
2. A missing-context record is converted into a permission approval without Agent Policy or runtime confirmation.
3. Compaction output is treated as final answer text.
4. Runtime ids are dropped, making evidence unable to join context to turn or task.
5. Redacted context is reintroduced through summaries, diagnostics, or UI projection.

Minimum validation set

For Agent Runtime-compatible integration, run at least:

1. ACTX-AR-ID-001 and ACTX-AR-ID-002.
2. ACTX-AR-SEL-001, ACTX-AR-BUDGET-001, and ACTX-AR-ASM-001.
3. ACTX-AR-MISSING-001 and ACTX-AR-COMPACT-001.
4. ACTX-AR-POLICY-001 and ACTX-AR-EVID-001.

Runtime proves the turn/task state. Agent Context proves which context facts were available and why.