graft

What this page covers

This page describes the cloud built-in graft skill — the combo agent platform Agent uses unified_search in the cloud process to graft across sessions.

For local IDE (Cursor / Claude Code / Codex) access to cloud sessions over HTTPS, see skills/coding-agents/graft-comboagent.
For mechanisms and underlying design (storage model, permission isolation, context injection strategy, etc.), see Core concepts · Graft cross-session grafting.

Why graft exists

When a complex project spans multiple independent analyses (e.g. ASPICE v1.14 baseline tracing, v1.17 baseline tracing, change analysis, coverage audit), cramming everything into one Session causes:

Context explosion and lost detail during compaction
Conflicting Agent configs / skill sets
One mistake poisoning the whole run
Hard to divide work and review

graft fixes this: each Session completes its own analysis; when collaboration is needed, structured outputs are grafted in on demand.

Graft = attach another session’s execution result to the current session’s context, use it, then leave — without polluting the host.

Core operations: three actions

The graft skill aligns with the existing kb_ids / doc_ids retrieval pattern. It is invoked through unified_search with three actions:

① `list_sessions` — discover

unified_search(action="list_sessions", query="cooling")

Returns sessions visible to the same tenant (yours + shared), with optional name keyword filter. Fields:

Field	Description
`session_id`	Unique session ID
`name`	Human-readable session name
`agent_name`	Agent hosting the session
`updated_at`	Last update time
`visibility`	`private` / `shared`

② `get_digest` — see the big picture

unified_search(action="get_digest", session_id="Cooling system analysis")

Returns the session’s full execution digest table: what happened each round, what was produced, status, artifacts. Think table of contents before entering someone else’s session.

session_id accepts both ID and name (same _resolve_kb_ids pattern); names are resolved automatically.

③ `get_round` — fetch data

unified_search(action="get_round", round_id=5, session_id="Cooling system analysis")

Grafts the exact inputs and outputs of a specific round from a specific session into the current context. The Agent can then:

Compare conclusions across sessions
Continue downstream work on intermediate data from another session
Cross-check multi-perspective results

Standard workflow

Rendering diagram…

If the user already gave the session name, you can skip Step 0 and start at Step 1.

Real-world examples

Example 1: ASPICE version change impact

Context: v1.14 baseline tracing and v1.17 baseline tracing finished in two separate sessions; you now need a diff analysis.

/graft
Change-analysis session references both baseline sessions:

  unified_search(action="get_digest", session_id="v1.14-baseline-trace")
  unified_search(action="get_digest", session_id="v1.17-baseline-trace")
  → Full picture of nodes / trace chains for both baselines

  unified_search(action="get_round", round_id=<trace-rebuild-round>, session_id="v1.14-baseline-trace")
  unified_search(action="get_round", round_id=<trace-rebuild-round>, session_id="v1.17-baseline-trace")
  → Exact 303 nodes + 396 trace links

  → 178 change items + 313 affected nodes

Savings: the two baseline documents need not be re-read — the Agent reuses structured JSON already produced. See ASPICE case scenario 2.

Example 2: Test coverage audit cross-check

Context: Scenario 1 completed requirement–test trace audit; scenario 4 re-audits from a “test” angle and compares with scenario 1.

/graft
Coverage-audit session references:

  Scenario 1 → 303-node trace relationships
  Scenario 3 → newly generated CAPL cases

  → “34 / 14 / 79” three-way consistency breakdown
  → Discrepancies flagged for manual recheck

Value: reduces single-view false positives; output is more trustworthy than single-session analysis. See ASPICE scenario 4.

Example 3: Patent drafting reusing prior novelty search

Context: A novelty search for a patent direction finished three days ago; you are now writing the specification.

/graft reference last week’s “motor control patent novelty search”

  → Agent pulls that session’s search results (prior art list from 13 sources)
  → Jump straight to claims drafting without redoing novelty search

Design principles

Principle	Description
Explicit user activation	Must be triggered via `/graft` or clear natural language — the Agent must not graft across sessions without permission
Isolation within tenant	Cross-tenant is strictly forbidden; zero data leakage across tenants
On-demand reference	Do not copy the whole session — only graft the rounds / data needed
Full audit ledger	Every graft is logged: who, in which session, referenced which round of which session
Permission inheritance	You can only reference sessions you already have access to (yours + shared)

Dimension	graft	Memory (LTM)
Granularity	Session-level (multi-round)	Knowledge-unit level (single item)
Trigger	Explicit user intent	Agent retrieval on demand
Scope	All sessions same tenant	Scope set at write time
Use case	Reuse a full analysis run	Long-term knowledge base

Core concepts · Graft — storage, isolation, injection
ASPICE engineering loop case — full graft walkthrough
memory-sdk — long-term memory (works with graft)
Session — graft operates on sessions