Potpie Source Ingestion

Potpie Source Ingestion

Use this skill for explicit ingestion requests. The harness is the intelligence: it gathers source data, reads it, decides what is durable, resolves identity, and writes semantic graph mutations with evidence. Potpie validates and stores; it does not decide what source material means.

Non-Negotiables

• Use a todo/checklist for every repository or multi-source ingestion. Do not jump directly from a README to graph writes.
• Local inspection is required for repo understanding. Scanner-driven graph updates are forbidden. Inspect files with rg, rg --files, git, and structured tooling; do not run legacy or deterministic ingestion/scanner commands that walk the tree and write graph facts.
• Use subagents only for read-only discovery slices. The main agent owns source selection, identity resolution, mutation proposals, commits, and final synthesis.
• Do not write until each required discovery lane is complete, explicitly unavailable, or intentionally scoped out by the user.
• Every write needs source refs, source authority, truth class, confidence, compact summary, and retrieval-grade description.

Phase 0: Scope And Preflight

1. Define source kind, pot/project, repo/path/URL, time window, and target memory shape: baseline, history, docs, infra, debug memory, preferences, or all.
2. Verify Potpie scope and graph availability:

potpie --json pot info
potpie --json source list
potpie --json graph status
potpie --json graph catalog --task "harness-led source ingestion"

3. If the repo is not registered, register metadata only. Use explicit --pot when pot scope is ambiguous:

potpie source add repo . --pot <pot-id-or-name>

4. Describe the views you expect to write/read before authoring mutations:

potpie --json graph describe features --view feature_context --examples
potpie --json graph describe infra_topology --view service_neighborhood --examples
potpie --json graph describe recent_changes --view timeline --examples
potpie --json graph describe decisions --view preferences_for_scope --examples
potpie --json graph describe debugging --view prior_occurrences --examples

5. If the CLI is unavailable or broken, continue discovery, build the proposed evidence matrix, and stop before committing graph writes.

Phase 1: Todo Plan

Create and maintain todos with at least these lanes for repository ingestion:

• Scope, pot, source registration, and graph contract preflight.
• Product/docs: README, docs, ADRs, runbooks, public docs, linked websites.
• Local repo map: manifests, packages/apps, entrypoints, route/API surfaces, tests, framework config, major modules, generated/API specs.
• Runtime/deploy: Dockerfiles, compose, Kubernetes, Terraform, CI workflows, deploy scripts, environment templates, feature flags.
• API/data/integrations: service clients, adapters, datastores, models, queues, auth providers, external APIs.
• GitHub/history: repo metadata, topics, releases/tags, recent merged PRs, open issues, linked tickets/docs, CI/deploy records.
• Preferences/workflows: explicit coding style, test commands, local dev setup, release/deploy/runbook workflows.
• Synthesis: evidence matrix, candidate graph facts, identity resolution, proposal, verified commit, and gate-driven follow-up checks.

Update todos as lanes finish. Preserve uncertain findings for the inbox instead of forcing them into canonical graph claims.

Phase 2: Parallel Discovery

Parallelize independent read-only slices when tools allow it. Recommended subagent prompts:

• Docs/product: "Read README, docs, ADRs, runbooks, package metadata, and linked product pages. Return product purpose, features, explicit decisions, preferences, workflows, and source refs. Do not write mutations."
• Local architecture: "Inspect manifests, top-level apps/packages, entrypoints, route/API surfaces, tests, and framework config. Return service/module map, likely features, explicit source files, uncertainty, and no mutations."
• Runtime/deploy: "Inspect Docker/compose/Kubernetes/Terraform/CI/deploy/env templates. Return environments, deploy shape, config variables, workflows, datastores, and source refs. Do not write mutations."
• API/data/integrations: "Inspect route specs, client/adapters, models, datastores, queues, auth, and external integrations. Return candidate APIContract/DataStore/Adapter/Dependency facts with evidence. No mutations."
• GitHub history: "Use GitHub tools/CLI for repo metadata, releases, recent merged PRs, open issues, linked tickets/docs, and CI signal. Return timeline, fixes, decisions, bug patterns, and source refs. No mutations."
• Preferences: "Find explicit preferences in docs, config, tests, contribution guides, PR templates, and comments. Return only explicit reusable policies with evidence. No mutations."

The main agent should continue non-overlapping discovery while subagents run.

Phase 3: Local Repo Inspection Targets

Use structured, bounded local inspection. Examples:

rg --files -g 'README*' -g 'docs/**' -g '*ADR*' -g 'package.json' -g 'pyproject.toml' -g 'Cargo.toml' -g 'go.mod'
rg --files -g 'Dockerfile*' -g 'docker-compose*' -g '.github/workflows/**' -g '*.tf' -g 'k8s/**' -g '.env*'
rg -n "FastAPI|APIRouter|express\\(|router\\.|Django|Flask|NextResponse|route\\(" .
rg -n "postgres|mysql|redis|mongo|s3|kafka|rabbit|queue|oauth|stripe|slack|github|linear|jira" .
rg -n "pytest|vitest|jest|playwright|make test|npm test|uv run|cargo test" README* docs .github pyproject.toml package.json Makefile

Do not infer durable facts from filenames alone. Use files to locate sources of truth, then read the relevant snippets or authored docs.

Phase 4: Hosted/GitHub Hydration

Use the agent's integration tools/connectors, including GitHub app/MCP/CLI tools when available, not Potpie queue ingestion commands. For GitHub-backed repo ingestion, gather the items below. Do not use Potpie CLI queue ingestion.

• Repository metadata: full name, default branch, description, topics, visibility, homepage, license, archived/fork status.
• Documentation: README, contributing guide, CODEOWNERS, PR/issue templates, docs linked from the repo.
• Releases/tags when they describe shipped behavior or deploy cadence.
• Recent merged PRs: title, body, author, merged date, branch, labels, linked issues, changed filenames; fetch patches only when author text is insufficient.
• Open/high-signal issues: title, body, labels, state, author, comments when available; issues can record requests/bugs but do not prove fixes.
• CI/workflows: workflow files and failed/passing runs only when relevant to durable workflows, release process, or recurring failures.
• Linked systems: Linear/Jira/Confluence/Slack/docs mentioned in repo, PRs, or issues when corresponding tools are available.

Respect API limits and user scope. Prefer recent/high-signal items over exhaustive pagination unless the user explicitly asks for full history.

Phase 5: Evidence Matrix

Before writing, build a compact matrix:

Guidelines:

• Use authoritative_fact for explicit docs/config/code ownership facts.
• Use source_observation for observed source material that may not be policy.
• Use agent_claim for lower-authority synthesis from multiple weak signals.
• Use preference only for explicit reusable project preferences.
• Use timeline_event for source-time activity from PRs, tickets, releases, or deployments.
• Put uncertain but potentially useful findings into graph inbox add.

Phase 6: Identity Resolution

Resolve before linking. Use specific filters when known:

potpie graph search-entities "<repo service feature dependency>" --limit 10
potpie graph search-entities "<service>" --type Service --environment prod --limit 10
potpie graph search-entities "<github-or-ticket-id>" --source-ref <github-or-ticket-ref> --limit 10

Reuse canonical keys. If duplicate candidates appear, stop and use inbox or a review-required correction flow; do not create near-duplicate entities.

Phase 7: Write

Author semantic mutation JSON from the live graph catalog and graph describe examples. graph mutation-template is only a skeleton helper, not the source of truth.

potpie --json graph propose --file mutation.json

Inspect proposal status, diff, warnings, rejected operations, conflicts, and review flags:

• invalid or rejected operations: fix the mutation or skip the weak fact.
• conflict or duplicate risk: resolve identity or use inbox.
• review_required: ask for approval or commit only with the required --approved-by value when policy allows.
• validated / low-risk: commit with --verify.

potpie --json graph commit <plan_id> --verify
potpie --json graph history --plan <plan_id>

For large batches of agent-authored mutations, use graph bulk apply with dry-run, chunking, manifest, and verify. Bulk apply must only apply facts the harness already selected.

Phase 8: Verify And Quality Gate

graph commit --verify reads back committed claims and checks quality. When it warns or fails, drill down with affected reads and quality reports:

potpie graph read --subgraph features --view feature_context --scope anchor_entity_key:<repo-key> --limit 50
potpie graph read --subgraph infra_topology --view service_neighborhood --scope service:<service> --depth 2 --direction both --limit 50
potpie graph read --subgraph recent_changes --view timeline --scope repo:<repo> --limit 50 --format table
potpie --json graph quality duplicate-candidates --limit 20
potpie --json graph quality low-confidence --limit 20
potpie --json graph quality conflicting-claims --limit 20
potpie --json graph quality orphan-entities --limit 20

If the verified commit misses expected facts, fix the mutation or record an inbox item. Report what was ingested, what was skipped, and what remains uncertain.

Repository Baseline

For repository ingestion, run baseline before change history. Use potpie-repo-baseline in deep mode when the user asks to ingest or understand a repo deeply. Record source-backed purpose, application type, features, service/module map, API contracts, datastores, integrations, environments, deploy shape, ownership, and explicit project preferences. Then use potpie-change-timeline for recent or historical activity. Do not infer baseline architecture from PR titles or issue status.

Represent capabilities as Feature entities. Link repositories or services to features with PROVIDES, and use IMPLEMENTED_IN only when a source locates the implementation.

Source Rules

• Tickets and issues can record timeline events, bug patterns, decisions, and docs. They do not prove a fix unless tied to a merged PR, commit, deployment, or explicit shipped-resolution source.
• Documents can record preferences, decisions, runbook notes, service notes, and infra facts only when they explicitly say them.
• Logs and transcripts can record diagnostic signals, investigations, fixes, and verifications. Keep raw logs out of descriptions except for short distinctive error text.