Graph-backed work objects: repo, issue, release, docs, and trace graphs

[Hmbown]

Problem

Agents repeatedly rebuild local understanding from raw text: files, search results, GitHub issues, release notes, docs, tests, and logs. This wastes cache, bloats transcripts, and hides the relationships that matter for real work.

Refactoring is the obvious example, but the same shape exists in every workflow:

• GitHub triage needs issue/PR/milestone/contributor graphs.
• Release prep needs version/changelog/check/artifact graphs.
• Docs renovation needs claim/source/support-surface graphs.
• Debugging needs symptom/log/code/test graphs.
• Model Lab needs trace/task/tool/evidence graphs.

Desired behavior

Add durable graph-backed work objects that tools can build, update, query, and hand to RLM or sub-agents by handle.

Initial object types:

RepoGraph
IssueGraph
ReleaseGraph
DocsGraph
TraceGraph
RefactorPlan
SlopLedger
EvidenceLedger

Each object should support:

• Stable IDs and handles.
• Bounded handle_read projections.
• Source-linked nodes and edges.
• Freshness / invalidation rules.
• Redaction/export policy.
• Optional RLM semantic analysis over slices.
• Task/handoff persistence.

Acceptance criteria

• Design doc defines shared graph primitives: node, edge, source span, confidence, freshness, owner, evidence.
• Prototype RepoGraph and one non-code graph, preferably IssueGraph or ReleaseGraph.
• Graph objects are queryable without pasting full state into the parent transcript.
• RLM can load a graph slice and return structured classifications by handle.
• Existing task/evidence/handoff surfaces can link to graph object IDs.
• Tests cover graph creation, projection, invalidation, redaction, and handle reads.

Related

• RLM/log workbench: route large local logs and structured payloads out of the parent transcript #2023 RLM/log workbench.
• GitHub triage scout: delegate issue, PR, and milestone maintenance passes #2025 GitHub triage scout.
• Model Lab: connect CodeWhale traces to open-weight fine-tuning/eval services #1977 Model Lab.
• EPIC: v0.9.0 Phase 3 — graduate memory from flat markdown to a typed, graph-structured store #534 typed graph-structured memory.

[Hmbown]

Addendum: graphs are shared working memory for distributed agents

The graph-backed work-object layer should be treated as the shared memory bus for distributed intelligence.

A parent model, sub-agent, RLM session, micro-op, and verifier should all be able to point at the same object handle and ask bounded questions without replaying the full transcript. That changes coordination from "everyone rereads files and writes summaries" to "everyone operates over the same typed graph."

Additional design points:

• graph objects need ownership and freshness metadata so agents know when a slice is stale;
• each node/edge should support source spans and evidence IDs, not just labels;
• graph projections should be small enough for cheap model calls and RLM sub-queries;
• graph deltas should be append-only where possible, preserving cache affinity;
• graph objects should be portable into handoffs and Model Lab traces;
• verifiers should inspect graph deltas, not only final diffs.

This is where slop becomes queryable: old API paths, compatibility shims, stale docs, and unverified assumptions become edges in a graph rather than vague memory.