Portfolio→Flagship project
Trajectory Fidelity Bridge
A provider-neutral reference implementation for moving structured coding-agent trajectories without collapsing tool calls, observations, timestamps, and transformation loss into a prose summary.
Renamed September 28, 2026: formerly Agent Session Bridge. The project was renamed after an older, unrelated public project using the former name was identified. The projects are unaffiliated. New releases use atomicdjt-trajectory-fidelity-bridge; the previous distribution remains available at 0.4.0 for existing users. See the migration guide.
What it solves
Coding agents accumulate structured state: system and user messages, agent turns, tool calls, correlated tool results, timestamps, metadata, and provider-specific records. A conventional handoff often reduces that history to a summary. Trajectory Fidelity Bridge instead uses the public Agent Trajectory Interchange Format (ATIF) v1.7 as the portable trajectory layer and records unsupported or degraded source information explicitly.
The project is deliberately not a competing interchange standard. Its role is provider parsing, normalization, redaction, fidelity accounting, and target-specific mapping.
Current implementation
- Imports supported Claude Code JSONL into validated ATIF v1.7 trajectories.
- Preserves supported messages, timestamps, tool calls, arguments, and call-correlated observations.
- Records unsupported/degraded content in a namespaced Trajectory Fidelity Bridge fidelity extension.
- Applies best-effort secret redaction before export; human review remains required.
- Provides an Antigravity derived-log reference mapping while explicitly refusing to claim unsupported native historical-session resumption.
- Optionally projects historical ATIF trajectories into OpenTelemetry/OpenInference-compatible observability without presenting reconstructed spans as original runtime instrumentation.
Why the boundary matters
A portable transcript is not automatically a resumable session. Target runtimes own their internal persistence, validation, and security boundaries. Trajectory Fidelity Bridge therefore stops when a documented target import boundary does not exist instead of writing into opaque internal state.
The useful technical question is not whether every field can be copied. It is which state is genuinely portable, which state is transformed, and which provider-specific semantics must be declared lost.
Technical writing
From Claude Code JSONL to ATIF v1.7: What Actually Survives an Agent Handoff? walks through the preservation, normalization, loss-accounting, redaction, and target-ingestion boundaries behind the implementation.
Discovery path
This portfolio page describes the project and its technical significance. GitHub remains the authoritative source, documentation, issue, and verification surface.