Dogfood Workflow Matrix
The dogfood matrix tests whether AgentDocs improves agent-mediated developer experience, not only whether builds and readiness audits succeed.
For the public interpretation of these runs, see Real-World Results and Findings by Target.
Every prepared target must run the standard regression capture:
pnpm regression:dogfood -- <target-directory>Workflow-specific searches are added with repeatable --query <label=query> options. The runner saves the top five results for every query, verifies a second build has the same generated-artifact hash, and keeps agent_task_passed as an explicit human judgment. Task-pack routing goals are added with --routing-goal <label=goal> and can be made strict with --expect-route <label=task-pack-id>.
Committed offline fixtures now gate the context boundaries exposed by this matrix. Run pnpm regression:fixtures to verify version, framework, router, locale, and content-type filters; mixed-context warnings; tolerant MDX ingestion; and quickstart generation. The June 16, 2026 workflow-layer rerun rebuilt all documented prepared targets and then checked status, handoff, and verify-context. The June 20, 2026 Phase 5 rerun rebuilt all documented prepared targets again and populated strict routing metrics. Live website recrawls remain opt-in.
Current Status
| Workflow | Prepared locally | Regression | Agent task | Current finding |
|---|---|---|---|---|
| AgentDocs self-dogfood | Yes, docs-only source | Passed: 13 pages, 4 packs, readiness 79 | Passed | Phase 5 rerun stable; setup goal routes to installation. |
| Hono dependency-user flow | Yes, local repo and prepared website crawl | Local: 85 pages, 7 packs, 93; prepared crawl: 100 pages, 4 packs, 79 | Pending | Phase 5 rerun stable; Cloudflare Workers routes to deployment, but quickstart routing still misses on local and prepared-crawl targets. |
| Fastify versioning flow | Yes, local repo and prepared website crawl | Local: 43 pages, 5 packs, 91; prepared crawl: 100 pages, 6 packs, 83 | Pending | Phase 5 rerun stable; local schema validation and migration route exactly; prepared crawl migration routes exactly. |
| Prisma local-docs monorepo | Blocked | Not run | Pending | Upstream repository contains Windows-invalid filenames; sparse checkout did not materialize the intended docs subtree |
| Supabase large-MDX stress test | Yes | Passed: 737 pages, 11 packs, readiness 79 | Pending | Phase 5 rerun stable; auth/RLS routes to authentication, with source coverage gaps explicit. |
| TanStack Query multi-framework test | Yes | Passed: 411 pages, 9 packs, readiness 79 | Pending | Phase 5 rerun stable; React mutation invalidation routes to query-invalidation. |
| Next.js large-site crawl | Yes, prepared crawl | Passed: 100 pages, 8 packs, readiness 88 | Pending | Phase 5 rerun stable; App Router POST route routes to route-handlers. |
Octokit REST is an additional prepared local-docs target. Its Phase 5 regression passed with 14 pages, 4 packs, readiness 93, and a stable repeated build. The auth request routing goal is captured report-only and selects authentication.
All completed post-hardening and workflow-layer regressions produced stable repeated-build hashes. Regression output is stored under each prepared target's ignored results/ directory and summarized in .dogfood/regression-summary.csv.
Completed Parser-Format Candidates
These targets test the expanded format parsers (Sphinx/reST and AsciiDoc/Antora) and transclusion resolution plumbing on large-scale real-world documentation estates.
| Repository | Ecosystem | Docs Shape | Pages | Task Packs | Readiness | Status |
|---|---|---|---|---|---|---|
django/django | Python | Sphinx/reST (.txt files) | 671 | 8 | 92 | Passed |
python/cpython | Python / systems | Sphinx/reST (.rst Doc tree) | 556 | 8 | 79 | Passed |
spring-projects/spring-framework | Java | AsciiDoc/Antora (.adoc) | 469 | 6 | 79 | Passed |
apache/airflow | Python / data | Mixed reST (.rst, .txt docs) | 1,617 | 10 | 79 | Passed |
Expansion Candidates
The current test bed covers useful cases, but broad confidence needs multiple documentation systems, multiple implementation languages, multiple scale profiles, and multiple kinds of ambiguity. These targets remain candidate regression additions. They should be added as bounded prepared targets or opt-in live runs before the product is called polished across large real-world documentation estates.
| Repository | Ecosystem | Docs shape | Why it matters | What should pass |
|---|---|---|---|---|
kubernetes/website | Go / cloud-native | Dedicated docs repo; large Hugo/Docsy site | Massive, versioned, highly linked docs with many examples and multiple contributor patterns | Correct version scoping, sane crawl bounds, strong task packs for deployment/networking/auth, low internal-link breakage |
dotnet/docs | C# / .NET | Giant docs-only repo | Multi-language platform docs with broad conceptual and reference content | Strong retrieval precision, bounded task packs, robust chunking at scale |
rust-lang/rust | Rust | Huge mixed code and docs repo | Very large codebase with compiler, standard library, and docs in one place | Good repo-source inclusion boundaries, no over-indexing irrelevant internals, useful task packs from docs |
apache/airflow-site | Python / data | Docs site | Split-docs-and-code pattern website | Successful per-site compile, correct cross-link handling, task packs for DAG authoring and deployment |
hashicorp/terraform | Go / IaC | Large product plus extensive documentation | Strong versioning and operational workflows, excellent fit for migration/provider tasks | Good tutorial/reference distinction, provider/version disambiguation, high-value task packs |
microsoft/TypeScript-Website | TypeScript | Docs-focused monorepo | Markdown/MDX-heavy website with handbook-style docs and generated schema content | Correct monorepo scoping, strong task packs for config/compiler API concepts, bounded chunking |
fastapi/fastapi | Python | Code plus docs; modern framework docs | Good control sample against the current Fastify/Hono/TanStack set | Clear docs/task retrieval for auth, dependencies, background tasks, and OpenAPI usage |
The first pass for each target should record preparation commands, source scope, page/chunk/entity/task-pack counts, readiness, repeated-build hash status, query captures, and at least one human-judged agent task. Large repositories should prefer scoped local preparation first, then opt-in website crawls once local ingestion behavior is understood.
The June 23, 2026 candidate expansion rerun verified the Sphinx and AsciiDoc format parsers on django, cpython, spring-framework, and airflow. All targets compiled with stable repeated builds. See Candidate Expansion Metrics for detailed counts.
Workflow-Layer Checks
The June 16, 2026 rerun also exercised the new utility layer:
agentdocs status
agentdocs handoff "<goal>"
agentdocs verify-context --task "<goal>"All rerun targets reported fresh. Verification passed where generated task families matched the goal, such as Hono deployment and Supabase auth/RLS. Other verification failures were mainly missing_task_pack for exact implementation goals such as Fastify schema routes, React mutation invalidation, and Next.js App Router POST routes. Those are now tracked as task-routing/product coverage gaps rather than hidden inside a successful build.
Phase 3 adds a deterministic routing benchmark to the same dogfood runner. See Routing Benchmarks Phase 3 and the Evaluation Metrics Reference for the reporting contract.
Workflow Commands
Hono
pnpm regression:dogfood -- .dogfood/hono-website \
--name hono-local-docs \
--query middleware=middleware \
--query cloudflare-workers="Cloudflare Workers" \
--query migration=migrationPass criteria:
- quickstart evidence identifies
npm create hono@latest; - routing, middleware, validation, and deployment/runtime topics are captured;
- Cloudflare Workers search returns runtime/deployment docs;
- migration context links to actual migration material;
- the agent task builds GET and POST routes, middleware, supported typed validation, and a Cloudflare Workers deployment using only generated context.
Fastify
pnpm regression:dogfood -- .dogfood/fastify \
--name fastify-local-docs \
--query schema-validation="schema validation" \
--query plugin=plugin \
--query migration=migrationPass criteria:
- readiness and generated context state the indexed Fastify version;
- v4 and v5 guidance are not mixed silently;
- plugin, schema, quickstart, and error-handling context is useful;
- the agent task builds a Fastify v5 server with a route, JSON schema validation, a plugin, and structured error handling without v4-only APIs.
Prisma
Prepare https://github.com/prisma/web with a local Markdown source at ./apps/docs/content/docs. The regression must confirm that unrelated blog and design-system content is excluded. The agent task defines a User model, creates a migration, instantiates Prisma Client, and queries users.
Supabase
Prepare https://github.com/supabase/supabase with a local Markdown source at ./apps/docs/content. The regression must inspect MDX component stripping, Row Level Security, auth middleware, and type generation. The agent task must not expose secret keys to the browser.
TanStack Query
Prepare https://github.com/TanStack/query with the intended documentation path only. The regression must verify React, Svelte, Vue, and Solid examples are not mixed. The agent task implements a React mutation with invalidation using React-specific evidence only.
Next.js
Collect https://nextjs.org/docs into a dedicated target. Search route handlers, server actions, and MDX. The agent task builds a current App Router POST route handler using only generated context.
Evaluation Rule
Readiness and search metrics are supporting evidence. A workflow passes only after its agent task is completed successfully and the result is recorded with:
--agent-task-passed true