Graft (cross-session)

What Graft is

Graft lets parallel Agent sessions reference one another’s results—like Git graft, stitching independent histories together.

Cross-session example: full V-Model chain

Cross-session auto-chaining for full V-Model management

Each role runs its own session; the platform chains a full V-Model path via Hooks + shared infrastructure:

Session A (requirements analyst): cross-vehicle CTS comparison → configuration delta analysis → requirement cross-references → option extraction
Hook handoff →
Session B (development engineer): consumes deltas → reviews code modules → review report → fix notes
Hook handoff →
Session C (test engineer): consumes deltas + review → auto-generates tests → coverage matrix → regression scope

Shared infrastructure:

Enterprise knowledge base (artifacts indexed automatically; downstream retrieval)
MemoryOS (cross-session memory: error_pattern / fact / preference)
Hook system (SessionStart injects upstream outputs + long-term memory)
Skill repository (one person captures; whole team reuses)

Standalone desktop AI cannot do this—every chat is an island. The platform lets each person focus on their step while upstream/downstream connects automatically.

Typical scenarios

Scenario	Description
Parallel agent merge	Session A finishes cooling-system analysis; Session B needs A’s conclusions for vehicle-level assessment
Incremental iteration	Last week’s session did baseline analysis; this week’s session deepens on top
Multi-user collaboration	Under the same tenant, user A’s session outputs are referenced by user B’s session (with authorization)

Security boundary

Users must explicitly enable via /graft; the Agent cannot decide cross-session access on its own.

This ensures:

By default the Agent does not know other sessions exist—no random browsing
Cross-session access reflects explicit user intent
Every cross-session action is auditable

Design notes

Graft ships as a builtin skill under skills/builtin/graft/:

Activation (Skill): controls whether the Agent knows cross-session capability exists
Capability (Tool): session_id parameter, permission checks, cross-session search

Without /graft → SKILL.md is not injected → the Agent does not know the session_id parameter exists → no misuse.

Real execution example

Below is a real session flow—customer demo executive summary DEMO-E used 4-worker parallel Graft to read outputs from prior sessions DEMO-A/B/C/D and produce a demo landing page.

⚡ Parallel execution · 4 workers: read findings from four upstream sessions at once

Graft with 4 workers in parallel reading DEMO-A/B/C/D

Each Graft worker (DEMO_C_Graft, DEMO_B_Graft, DEMO_A_Graft, DEMO_D_Graft) reads its upstream session independently, saves slim artifacts such as A-findings.json / B-findings.json / C-findings.json, then the main Agent in this session aggregates them.

Aggregated output: OUTPUT_JSON records graft sources and final artifacts

OUTPUT_JSON after graft: four upstream sessions summarized

The Agent’s OUTPUT_JSON states in natural language that “this task grafted across four prior sessions (DEMO-A/B/C/D) …”—traceable provenance is a core property of Graft.

Session
Memory
Skills
skills/agentic/graft — cloud builtin (platform Agent)
skills/coding-agents/graft-comboagent — local (Cursor / Claude Code / Codex)