Crate av_atif 

ATIF — Agent Trajectory Interchange Format (Harbor), version 1.7.

Verified against the published specification at harborframework.com/docs/agents/trajectory-format (2026-08-10):

• root Trajectory with schema_version, optional run-scoped session_id (relaxed in v1.7), trajectory_id + subagent_trajectories (v1.7), agent, ordered steps, final_metrics, extra (v1.1);
• Step.step_id sequential from 1; source ∈ {system, user, agent}; agent-only fields (reasoning_content, reasoning_effort, model_name, tool_calls, metrics) rejected on non-agent steps; observation allowed on agent steps and, since v1.2, system steps; llm_call_count (v1.7) allowed on any step;
• Metrics carries prompt_tokens / completion_tokens / cached_tokens / cost_usd (+ logprobs, completion_token_ids v1.3, prompt_token_ids v1.4) — the cached-token preservation that makes an exported trajectory a replayable KV-cache checkpoint (brief Module H);
• validator collects errors in one pass (Harbor philosophy), not just the first — capped at MAX_VALIDATION_ISSUES with a truncation notice.

Reader accepts every published version v1.0–v1.7 with per-version field gating; the writer always emits v1.7 (inbound tolerant, outbound strict).

Re-exports

pub use model::Agent;

pub use model::FinalMetrics;

pub use model::Metrics;

pub use model::Observation;

pub use model::ObservationResult;

pub use model::ReasoningEffort;

pub use model::Source;

pub use model::Step;

pub use model::SubagentTrajectoryRef;

pub use model::ToolCall;

pub use model::Trajectory;

pub use model::ATIF_VERSION;

pub use model::SUPPORTED_VERSIONS;

pub use validate::validate_bytes;

pub use validate::validate_trajectory;

pub use validate::validate_value;

pub use validate::Mode;

pub use validate::ValidationIssue;

pub use writer::write_atomic;

pub use writer::TrajectoryBuilder;

Modules

model

ATIF data model (serde), matching Harbor’s published Pydantic models.

validate

ATIF validator, mirroring Harbor’s trajectory-validator semantics: validates required fields, types and constraints, sequential step ids (starting from 1), tool-call references in observations, ISO-8601 timestamps, agent-only field placement — and collects errors as it goes (not just the first), up to MAX_VALIDATION_ISSUES, after which a truncation notice is appended.

writer

Trajectory construction and crash-safe persistence.