Potpie Graph

Potpie Graph Workbench

The graph is project memory: preferences, prior bugs and their fixes, infra topology, decisions, and a timeline of changes. You are the intelligence that reads it before acting and writes durable learnings after. Potpie validates, lowers, commits, audits, and ranks. It does not scan a repository or infer rich facts from prose for you.

Use text output for routine context reads. Add --json when a workflow needs exact machine parsing, mutation plans, commits, history verification, or full evidence/debug payloads.

1. Check Status And Discover The Contract

potpie graph status
potpie graph catalog --task "<task>" --profile read

Returns contract + ontology versions, the readable views, and active match_mode (vector vs lexical). Start graph-aware work here instead of reading docs. Use full JSON catalog output when you need mutation operation partitions, entity types, predicates, or exact machine parsing. Trust the catalog's current operation partition over any example in a skill file.

Describe the subgraph/view before a non-trivial read or write:

potpie graph describe debugging --view prior_occurrences --examples

2. Read - graph read --subgraph --view

potpie graph read --subgraph debugging --view prior_occurrences --query "refund race timeout" --limit 8
potpie graph read --subgraph debugging --view prior_occurrences --query "refund race timeout" --query-threshold 0.55 --limit 8
potpie graph read --subgraph decisions --view preferences_for_scope --scope repo:acme/x,path:src/payments/client.py
potpie graph read --subgraph recent_changes --view timeline --time-window 7d --limit 20 --format table
potpie graph read --subgraph recent_changes --view timeline --source-ref <github-pr-or-issue-ref> --format table
potpie graph read --subgraph infra_topology --view service_neighborhood --scope service:payments-api --depth 2 --direction out --environment prod
potpie graph neighborhood --entity service:payments-api --predicate USES --detail summary --limit 20

Views and what they answer:

Text reads return compact summaries for fast orientation. Timeline reads should use --format table or --format jsonl for bounded event rows. Use --json --detail full --relations full --format raw only when you need the underlying relation payloads for debugging or exact machine processing. Inspect coverage, freshness, and quality before relying on results.

Query expansion is your job

The local embedder is small; recall depends on the query. Expand the user's words before reading: "add retry to the payments client" → also carry "timeout, flaky, tenacity, backoff, external call". That expansion is in-session reasoning, not something the daemon does.

3. Resolve identity — graph search-entities

Before linking or asserting against an existing entity, find its canonical key:

potpie graph search-entities "payments api" --type Service --limit 10
potpie graph search-entities "github issue 881" --source-ref <github-pr-or-issue-ref> --limit 10

Reuse the returned key. Inventing a near-duplicate key (service:payments vs service:local:payments-api) fragments the graph and breaks future reads.

4. Write - graph propose then verified graph commit

Writes are semantic operations (never raw graph CRUD). First create a server-held plan with propose; then commit exactly that plan_id.

potpie graph mutation-template --kind repo-baseline   # schema-only skeleton to fill
potpie --json graph propose --file mutation.json
potpie --json graph commit mutation-plan:01JY8T5C --verify
potpie --json graph history --plan mutation-plan:01JY8T5C

mutation-template kinds: repo-baseline, feature, preference, preference-policy, infra-snapshot, bug-fix, decision, timeline-event, timeline-change — placeholders only; you fill them from sources you actually read. Use the use-case templates for durable memory:

• preference-policy writes structured policy fields (policy_kind, prescription, strength, audience) and can target a CodeAsset.
• infra-snapshot writes environment-qualified service, adapter, config, and deployment-target facts (USES_ADAPTER, CONFIGURES, DEPLOYED_WITH).
• timeline-change writes source-time activity events; occurred_at is the PR/ticket/deploy time, not ingestion time.
• bug-fix writes the symptom, known fix, and optional verification edge so debugging.prior_occurrences can return the fix inline.

Repo/service functionality is first-class: assert PROVIDES (repo/service → Feature) and IMPLEMENTED_IN (feature → repo/ service/CodeAsset), each Feature carrying a compact summary and a retrieval-grade description.

Payload is always batch-shaped:

{
  "graph_contract_version": "v1.5",
  "pot_id": "local/default",
  "idempotency_key": "mutation:bug:settle-deadlock",
  "created_by": {"surface": "cli", "harness": "claude"},
  "operations": [
    {
      "op": "assert_claim",
      "subgraph": "debugging",
      "subject": {"key": "bug_pattern:settle-deadlock", "type": "BugPattern",
                  "properties": {"name": "settle deadlock under concurrent refund"}},
      "predicate": "REPRODUCES",
      "object": {"key": "service:local:payments-api", "type": "Service"},
      "truth": "agent_claim",
      "confidence": 0.9,
      "description": "Concurrent refund + settle on the same order deadlocks the payments DB; shows up as 'refund race timeout' / 'payment deadlock on concurrent settle' under load in prod.",
      "evidence": [{"source_ref": "github:pr:412", "authority": "external_system"}]
    }
  ]
}

Use only operations advertised by graph catalog. Common operations include upsert_entity, link_entities, assert_claim, append_event, end_relation_validity, retract_claim, and audited correction operations when the catalog marks them applicable. Never hard-delete a claim: use validity, retraction, supersession, or merge operations according to the catalog policy.

5. Capture uncertainty - graph inbox

Use the inbox when you have evidence that may matter, but you cannot safely pick the canonical graph update yet. Inbox items are pending work only; they do not appear in ordinary graph reads as facts.

potpie --json graph inbox add --summary "Refund retry PR may relate to the prior timeout bug" --evidence github:pr:acme/payments:955 --subgraph debugging
potpie --json graph inbox list --status pending --limit 20
potpie --json graph inbox claim graph-inbox:abc123 --by user:alice
potpie --json graph inbox mark-applied graph-inbox:abc123 --plan mutation-plan:01JY8T5C --mutation mutation-1 --by user:alice
potpie --json graph inbox mark-rejected graph-inbox:abc123 --reason "not enough evidence" --by user:alice

Processing an inbox item is normal graph work: inspect the catalog/contract, read the relevant views, resolve identity with search-entities, propose and commit a mutation if warranted, then mark the inbox item applied or rejected.

6. Inspect quality - graph quality

Quality reports are read-only. They surface graph maintenance work but never repair semantic facts directly.

potpie --json graph quality summary
potpie --json graph quality duplicate-candidates --limit 20
potpie --json graph quality stale-facts --subgraph infra_topology --limit 20
potpie --json graph quality conflicting-claims --limit 20
potpie --json graph quality orphan-entities --limit 20
potpie --json graph quality low-confidence --threshold 0.75 --limit 20
potpie --json graph quality projection-drift --limit 20

If a finding changes canonical meaning, repair it through graph propose and graph commit --verify. If the evidence is uncertain, create a graph inbox add item instead. Reserve graph repair for operator projection maintenance such as index or summary rebuilds.

Truth classes

Pick the truth class honestly — it feeds the ranker:

authoritative_fact (explicit source of truth) · source_observation (observed source data read by the harness) · user_decision (a person decided) · preference · agent_claim (you inferred it; default when unsure) · timeline_event (something happened) · quality_finding. Durable writes need evidence or an explicitly low-authority truth class.

Do not use the graph as a deterministic code scanner. If a repo, PR, ticket, log, or document should become memory, the harness reads that source, decides what is worth recording, resolves identity, and writes a semantic mutation.

For GitHub, Linear, Jira, and similar hosted integrations, use the agent's integration tools/connectors to pull and hydrate source records first. Do not use Potpie CLI queue ingestion as the graph update path; after reading the integration data, write durable facts with graph propose / graph commit --verify or capture uncertainty with graph inbox.

Retrieval-grade descriptions (the one rule that matters most)

Every entity and claim carries a description — a natural-language retrieval card the local embedder indexes. Write it for search, not display: include the symptoms, synonyms, and scope a future searcher would type. Validation only warns on a weak description, but a vague card means the fact never resurfaces. Compare:

• Weak: "deadlock fix"
• Strong: "Concurrent refund + settle deadlocks payments DB under load; seen as 'refund race timeout' and 'payment deadlock on concurrent settle' in prod; fixed by ordering lock acquisition in services/payments/settle.py"

Responding To Nudges

A Potpie hook may call graph nudge and inject its result into your session. The hook never reasons — you do.

• inject_context → treat the injected facts as graph truth for this task; they were ranked for your current scope, so use them rather than re-fetching.
• instruction (e.g. "you resolved <error> after editing <files> — record the bug+fix if non-obvious", or "capture durable learnings") → a prompt to decide, not an auto-write. Decide the truth class, resolve identity with graph search-entities, write a retrieval-grade description, then graph propose and graph commit --verify. If the learning is useful but uncertain, create a graph inbox add item instead. If nothing durable was learned, do nothing.

Writes are idempotent by idempotency_key, so a nudge-driven capture you've already made will not duplicate.