TLDRGraph builds a source-backed feature and workflow catalog for this repository. It does not build an architecture graph, infer workflows heuristically, or run AI itself.
Run tldrgraph init to create a catalog. Run tldrgraph refresh when an
existing catalog needs updating after source changes. If either command returns
needs_feature_workflows, identify feature outcomes and immediately write the .tldrgraph/features.yaml v4 catalog index
with the reported source_hash. Then spawn one fresh source-reading subagent
for each indexed feature. Each worker writes only its assigned complete
.tldrgraph/workflows/<feature_id>.yaml v4 file; it never edits features.yaml
or another feature's workflow. Run the same command again to validate the direct
artifacts and generate the explorer.
Every capability and workflow step must cite a verified repository-relative
file, symbol, and line range. Define capabilities as user, admin, developer, or
operator outcomes. Use generated only for a proven end-to-end journey,
partial with missing_coverage for a proven fragment, and pending with no
steps when no reliable sequence can be established.
Model each feature as a directly renderable flowchart: initiating action,
proven process steps, decision branches, joined continuation, and final result.
When a workflow step has mutually exclusive paths, modes, or choices, make the
step the decision node and represent each labeled, evidence-backed branch as a
step options entry that rejoins the following proven step.
The full schema is in .tldrgraph/AGENT_CONTRACT.md.