Crate swink_agent_policies

Expand description

Policy implementations for swink_agent.

This crate provides all policy implementations built against swink-agent’s public policy trait API. Each policy is feature-gated independently:

budget: BudgetPolicy — stops the loop when cost or token limits are exceeded
max-turns: MaxTurnsPolicy — stops the loop after a configured number of turns
deny-list: ToolDenyListPolicy — rejects tool calls by name
sandbox: SandboxPolicy — restricts file paths to an allowed root directory
loop-detection: LoopDetectionPolicy — detects repeated tool call patterns
checkpoint: CheckpointPolicy — persists agent state after each turn

prompt-guard: PromptInjectionGuard — blocks prompt injection in user messages and tool results
pii: PiiRedactor — redacts personally identifiable information from assistant responses
content-filter: ContentFilter — keyword/regex blocklist for assistant output
audit: AuditLogger — records every turn to a pluggable sink
memory-nudge: MemoryNudgePolicy — heuristic detection of save-worthy content in agent turns

Structs§

AuditCost: Subset of cost relevant for audit records.
AuditLogger: PostTurnPolicy that builds an AuditRecord for every turn and writes it to the configured AuditSink.
AuditRecord: Summary of a single turn, suitable for serialization to an audit log.
AuditUsage: Subset of token usage relevant for audit records.
BudgetPolicy: Stops the agent loop when accumulated cost or tokens exceed configured limits.
CheckpointPolicy: Persists agent state after each turn via a CheckpointStore.
ContentFilter: Keyword/regex blocklist for assistant output.
FilterRule: A single filter rule consisting of a compiled regex and metadata.
JsonlAuditSink: Appends one JSON line per audit record to a file on disk.
LoopDetectionPolicy: Detects when the model is stuck in a cycle, calling the same tools with identical arguments repeatedly.
MaxTurnsPolicy: Stops the agent loop after a configured number of turns.
MemoryNudge: A structured payload emitted when save-worthy content is detected in a turn.
MemoryNudgePolicy: A PostTurnPolicy that detects save-worthy content via heuristic pattern matching.
PiiPattern: A named regex pattern used for PII detection.
PiiRedactor: Detects and optionally redacts PII in assistant output.
PromptInjectionGuard
SandboxPolicy: Rejects tool calls that reference file paths outside an allowed root directory.
ToolDenyListPolicy: Rejects tool calls whose names appear in the deny list.

ContentFilterError: Errors returned when constructing a ContentFilter.
LoopDetectionAction: What to do when a repeated tool call pattern is detected.
MemoryNudgeCategory: The category of save-worthy content detected by MemoryNudgePolicy.
NudgeSensitivity: Controls how aggressively MemoryNudgePolicy flags borderline content.
PiiMode: Behaviour when PII is detected.