docs: validate adapter agents with mock upstream fixtures by bokelley · Pull Request #3826 · adcontextprotocol/adcp

bokelley · 2026-05-02T14:55:16Z

Summary

Adds docs/building/validate-with-mock-fixtures.mdx — a four-step recipe adopters can follow to validate AdCP agents that wrap an upstream platform (DSP, SSP, retail data warehouse, creative server, signal marketplace), in any language:

Boot a published mock upstream for the specialism
Point your agent at the mock as its upstream
Run the storyboard runner for AdCP wire-contract conformance
Assert traffic counters at `/_debug/traffic` for façade resistance

Why this guide

Storyboards alone validate wire shape but cannot detect façade adapters — agents whose handlers return shape-valid AdCP responses without integrating with the upstream. Empirically observed (during SDK dogfood): an LLM-built adapter wrote OAuth client code, declared the import, and never called it from any handler. The handlers returned hardcoded shape-valid responses for every call. Storyboard graded `passing`. Real upstream traffic: zero requests.

Traffic counters close the hole. Both signals matter; one without the other is incomplete.

Why this lives in the spec repo

The pattern is language-agnostic:

The mock upstream contracts (auth shape, lookup endpoint, tenant scoping, traffic counter discipline) are spec-level. Any AdCP SDK in any language can ship equivalent mocks.
The four-step recipe is a CI workflow shape, not an SDK feature.
The known limitations (storyboard under-coverage, cascade fragility) are spec-level concerns.

The reference TS SDK (`@adcp/client`) ships four mock specialisms today; the recipe doesn't depend on the TS SDK specifically. Python, Go, Rust adopters can either point at the @adcp/client CLI for the mock-server step or implement equivalents in their language.

Honesty about limitations

Called out prominently rather than buried:

Storyboards under-cover payload variety (Storyboards are façade-permissive: payload variety doesn't force adapters to confront upstream contracts #3785)
Storyboard cascade silently skips downstream steps on early-step shape gaps (Storyboard runner: stateful step cascade silently skips downstream steps when prior response is shape-valid but missing extractor-required fields #3796)
Mock seeds may not match storyboard fixture inputs (Skill ↔ storyboard contradiction: sales-guaranteed create_media_buy contract #3822)
Traffic gate is necessary, not sufficient — sophisticated façades with synthetic placeholder data still pass

Discoverability

Linked from `validate-your-agent.mdx` via a single `` callout in the preamble. Adapter authors find it without needing to know the new file exists; the existing storyboard-only path remains intact for non-adapter agents.

Test plan

MDX renders (matches existing file's frontmatter + Mintlify component style)
All cross-references resolve (validate-your-agent, comply-test-controller, build-an-agent, compliance-catalog, schemas-and-sdks)
No duplication with existing storyboard-runner content; this is the complement for adapter agents

Refs

fix(matrix): blind-fixture testing + façade detection (caught real façade in validation) adcp-client#1239 (origin of the traffic-counter pattern in the reference SDK)
Storyboards are façade-permissive: payload variety doesn't force adapters to confront upstream contracts #3785, Storyboard runner: stateful step cascade silently skips downstream steps when prior response is shape-valid but missing extractor-required fields #3796, Skill ↔ storyboard contradiction: sales-guaranteed create_media_buy contract #3822 (limitations called out in guide)

🤖 Generated with Claude Code

…oryboard + traffic Makes the three validation gates we just used by hand into a CI test that runs on every commit: 1. `npx tsc` with --strict + noUncheckedIndexedAccess + exactOptionalPropertyTypes + noPropertyAccessFromIndexSignature + 2 other hardening flags. Catches latent type errors that default `--strict` misses. 2. Boot the published mock-server signal-marketplace as upstream, boot the example agent, run `adcp storyboard run signal_marketplace`, assert `summary.steps_failed === 0`. Catches AdCP wire-shape regressions. 3. After the storyboard run, GET /_debug/traffic and assert each of `_lookup/operator`, `/v2/cohorts`, `/v2/activations` was hit ≥1 time. Catches façade regressions where the adapter returns shape-valid responses without exercising upstream. Why these three together: tsc catches type errors at build time; storyboard catches wire-shape errors when exercised; traffic catches shape-valid-but-fake responses. Each is a distinct failure mode the others miss. Adversarial validation: sabotaged the example (replaced the upstream listCohorts call with `[]`) — gates 2 and 3 fail with distinct, actionable messages ("storyboard reported N failed steps" + "These upstream routes had zero hits — the adapter is a façade for them"). Gate 1 still passes because the regression is runtime-only, which is correct: the three gates cover orthogonal failure modes. CI wall: 18s (tsc 6.3s + storyboard 3.7s + traffic 0.04s + boot/teardown). Acceptable for a per-commit gate. Wired into ci.yml after `npm test`. Separate `test:examples` script keeps it isolatable for local iteration. Pretest:examples runs `npm run build` so the dist/ artifacts are present. This is the pattern the validate-with-mock-fixtures spec guide (adcontextprotocol/adcp#3826) describes — applied here to our own reference adapter so what we tell adopters to do is what we ourselves do. Future hello_seller_adapter_<specialism> additions each get their own test/examples/*.test.js with the same shape.

…ference (#1274) * docs(examples): hello_seller_adapter_signal_marketplace — worked passing reference A complete v6 typed-platform AdCP signals adapter that wraps an upstream signal-marketplace platform via HTTP. Demonstrates the full adapter pattern: account.operator → upstream X-Operator-Id resolution, get_signals → /v2/cohorts, activate_signal → /v2/cohorts/{id} + /v2/destinations + POST /v2/activations, with both `platform` and `agent` destination branches and proper schema-conforming response shapes. Designed as the starting point adopters fork. SWAP markers throughout mark every place the example assumes the published mock as upstream; adopters replace those with calls to their real backend's HTTP/SDK client. The AdCP-facing platform methods stay identical. ## Why this exists Every seller agent is an adapter — even ones whose "upstream" is internal infrastructure are translating AdCP wire calls into backend calls. Reading the spec + skill and synthesizing a correct agent from scratch is hard; empirically (from blind LLM-built adapters in our SDK dogfood) most first-pass attempts had at least one missing field, missing error code, or misshapen response that cascaded through the storyboard. Giving adopters a working starting point removes most of that friction. ## Validated end-to-end Against the published `adcp mock-server signal-marketplace`: steps_passed: 7 steps_failed: 0 steps_skipped: 4 (all missing_tool — tools this specialism doesn't claim) Traffic counters confirm real upstream integration: GET /_lookup/operator: 6 GET /v2/cohorts: 4 GET /v2/cohorts/{id}: 2 GET /v2/destinations: 2 POST /v2/activations: 2 This is the first clean traffic-verified pass on any specialism in the matrix-blind-fixtures lineage. Documents what "this is what a clean E2E pass looks like" actually means. ## Spec/schema gotchas surfaced while building Each documented inline so the next reader doesn't re-discover them: - `signal_ids` is `signal_id[]` (array of provenance objects), NOT array of strings — schema says so explicitly. Comparing against bare strings silently returns no matches. - `ActivationKey` is a oneOf: `type: 'key_value'` puts `key` and `value` at the TOP level of the activation_key object, NOT nested under a `key_value` sub-field. `value` MUST be string. - The mock's `agent_activation_key` is `{ agent_segment: <string> }` (an object), not a bare string — extract `.agent_segment` for the ActivationKey.value field. - `signal_spec` is a free-text semantic brief; passing it to the upstream's substring-match `q` filter loses real-platform-style semantic matching and breaks against literal-text mock filtering. Pass it to a real semantic index in production; for the mock, fetch the full catalog and filter client-side via signal_ids. ## Run # Demo: boot mock + agent npx @adcp/sdk@latest mock-server signal-marketplace --port 4150 UPSTREAM_URL=http://127.0.0.1:4150 \ npx tsx examples/hello_seller_adapter_signal_marketplace.ts # Validate adcp storyboard run http://127.0.0.1:3001/mcp signal_marketplace \ --auth sk_harness_do_not_use_in_prod curl http://127.0.0.1:4150/_debug/traffic ## Refs - Linked from the validate-with-mock-fixtures spec guide (adcontextprotocol/adcp#3826) - Companion to skills/build-signals-agent/ (skill should reference this example as the starting point for adapter-style agents) * fix(example): tighten hello signals adapter to pass strictest typecheck After review on PR #1274 the example was 435 lines and failed under `--strict --noUncheckedIndexedAccess --exactOptionalPropertyTypes --noPropertyAccessFromIndexSignature`. Eight unique errors including two real bugs (Account.authInfo missing, capabilities missing creative_agents/channels/pricingModels). The example "passed" only because the storyboard never exercised the missing branches. Cleanup: * Add `authInfo: ctx?.authInfo ?? { principal: 'anonymous' }` to accounts.resolve return — the field is required on Account and the framework strips it before emitting on the wire (typed intentionally so adapters thread the principal through to resource handlers, but signals-only adapters that don't authorize against it still need to pass it). * Declare `creative_agents: [] as const, channels: [] as const, pricingModels: ['cpm'] as const` on capabilities. Empty arrays are correct for signals-only platforms (don't sell media, don't compose with creative agents). Filed adcp-client#1278 for the underlying DX issue: these required fields are surprising for signals-only platforms; should be optional or specialism- conditional. * Narrow signal_ids filter on the `source: 'catalog'` discriminator before reading `data_provider_domain`. SignalID is a oneOf — catalog vs agent — and only catalog has provenance. * `process.env['FOO']` instead of `.FOO` (TS4111). * Cast ctx.account in resolveSessionKey since framework types it as broadly typed there. * Build RequestInit with conditional body-set instead of explicit `body: undefined` (exactOptionalPropertyTypes). * Extract httpJson<T>(method, path, opts) helper — collapses 5 UpstreamClient methods into one generic + 5 typed entry points. Each entry point still annotated with `// SWAP:` so adopters see exactly what to replace when wiring to a real backend. * Tighten comments — keep spec-gotcha learnings inline (the four "things schema validators catch but type checkers don't"), drop paragraph-length prose where one line works. Result: 360 lines (was 435), 0 errors under the strictest realistic adopter config (--strict + 4 extra flags), runtime gates unchanged (7/11 passed, 0 failed, 4 skipped — all `missing_tool`; traffic counters verify all 5 upstream routes hit). The two real bugs (authInfo + capabilities required fields) are now documented inline so adopters who fork this skip both. The discriminated-union narrowing is also documented inline. Each is the kind of friction that would cascade-fail an adopter's first run; the example now sidesteps all of them. * test(ci): three-gate test for hello signals adapter — strict tsc + storyboard + traffic Makes the three validation gates we just used by hand into a CI test that runs on every commit: 1. `npx tsc` with --strict + noUncheckedIndexedAccess + exactOptionalPropertyTypes + noPropertyAccessFromIndexSignature + 2 other hardening flags. Catches latent type errors that default `--strict` misses. 2. Boot the published mock-server signal-marketplace as upstream, boot the example agent, run `adcp storyboard run signal_marketplace`, assert `summary.steps_failed === 0`. Catches AdCP wire-shape regressions. 3. After the storyboard run, GET /_debug/traffic and assert each of `_lookup/operator`, `/v2/cohorts`, `/v2/activations` was hit ≥1 time. Catches façade regressions where the adapter returns shape-valid responses without exercising upstream. Why these three together: tsc catches type errors at build time; storyboard catches wire-shape errors when exercised; traffic catches shape-valid-but-fake responses. Each is a distinct failure mode the others miss. Adversarial validation: sabotaged the example (replaced the upstream listCohorts call with `[]`) — gates 2 and 3 fail with distinct, actionable messages ("storyboard reported N failed steps" + "These upstream routes had zero hits — the adapter is a façade for them"). Gate 1 still passes because the regression is runtime-only, which is correct: the three gates cover orthogonal failure modes. CI wall: 18s (tsc 6.3s + storyboard 3.7s + traffic 0.04s + boot/teardown). Acceptable for a per-commit gate. Wired into ci.yml after `npm test`. Separate `test:examples` script keeps it isolatable for local iteration. Pretest:examples runs `npm run build` so the dist/ artifacts are present. This is the pattern the validate-with-mock-fixtures spec guide (adcontextprotocol/adcp#3826) describes — applied here to our own reference adapter so what we tell adopters to do is what we ourselves do. Future hello_seller_adapter_<specialism> additions each get their own test/examples/*.test.js with the same shape. * refactor(examples): use new SDK ergonomics — drop optional empty arrays + authInfo + use public mock-server export

bokelley · 2026-05-02T17:12:31Z

Pushed 47b548605 addressing three-reviewer feedback (docs-expert + ad-tech-protocol-expert + adtech-product-expert) before requesting spec-side review. Substantive changes, not nits — recap below.

What changed

Convergent block (3 reviewers): non-normative framing

Original implied /_lookup/<resource> + /_debug/traffic were spec-level conventions. Three reviewers independently flagged that this risked promoting a TS-SDK debug surface to spec normative language without spec-side agreement. Fixed:

Added top <Note> explicitly disclaiming compliance-tier status: "Non-normative. The mock-fixture conventions described here are reference implementations in @adcp/client; alternative SDK implementations may diverge. Treat as harness convention, not protocol contract."
Reframed as "pre-staging gate" rather than "validation" — complements staging integration tests rather than replacing them.
Stripped _lookup route convention claims to be portable across SDK implementations.

Convergent block: language-agnostic claim vs Python-only sketch

Original opened "language-agnostic" then proved it with ~40 lines of Python only. Two reviewers (docs + product) flagged the self-contradiction. Fixed:

Dropped the Python sketch entirely. Replaced with a <Info> callout noting the sidecar pattern: agent stays in its native language, the TS CLI runs as a service container in CI.
Added a concrete GitHub Actions YAML example showing exactly that pattern (mock + agent + storyboard + traffic curl + jq assertion).

Reviewer-specific issues addressed

Reframed "façade resistance" → "integration gaps" / "integration-honesty gate" (product expert: the original lands as paranoid to senior eng managers; "integration honesty" is CI-owner vocabulary).
Threshold guidance added: ≥ 1 per headline route is the right floor; richer assertions live in the storyboard's payload coverage.
Mock specialism table: added caveat that account.advertiser/account.operator/account.publisher are SDK conventions for binding AdCP requests to upstream tenants, not normative AdCP terms (protocol expert: these don't appear in cached schemas). Also added a sentence that the four are representative, not exhaustive.
Limitations section restructured into three sub-headings (storyboard / runner-tooling / traffic-gate). Cascade fragility (Storyboard runner: stateful step cascade silently skips downstream steps when prior response is shape-valid but missing extractor-required fields #3796) is now under "runner/tooling caveats" not spec issues (protocol expert: implementation issue, not spec).
Added two new traffic-gate limitations (protocol expert): idempotency replay isn't exercised (façade that doubles writes passes hit count) and outbound webhook delivery isn't on upstream traffic counters (lives on the storyboard runner's --webhook-receiver).

Discoverability (docs expert)

Two new inbound links added:

build-an-agent.mdx — "If your agent wraps an upstream platform, see…" appended to the existing storyboard validation paragraph. Adapter-shape decisions are made here, so adopters find the recipe at the right moment.
compliance-catalog.mdx — <Tip> after the source-of-truth section, framing the recipe as complementary to the catalog itself.

Tone

Removed "Empirically observed (from blind LLM-built adapters during SDK dogfood)" — Slack voice. Replaced with prose that names the failure modes directly.
Replaced bare adcp ... invocations with npx @adcp/client@latest ... — matches the convention used elsewhere in adcp docs.

Open questions reviewers raised that need spec-side input

These are reasonable to leave for spec-maintainer judgment rather than self-resolve:

Should /_lookup/<resource> and /_debug/traffic become normative spec (in /docs/protocol/), or stay @adcp/client convention? Doc currently treats them as the latter; a spec section would change that. Protocol expert flagged.
Are AdCP-side principal field names (account.operator, account.advertiser, account.publisher) in any cached schema, or SDK-only? If schema-defined, the doc should link to them; if SDK-only, the current "SDK conventions" caveat is correct.
AAO-Verified / Practitioner certification implications — if either consumes this recipe as a gate, that turns harness convention into cert requirement and means non-TS SDKs must reproduce these conventions. Doc currently disclaims via top <Note> but the spec maintainer should confirm.

Diff is ~100 lines smaller, ~50 lines added — the non-normative framing replaced repeated implications throughout the doc rather than adding decorations.

Ready for spec-side review.

Adds `docs/building/validate-with-mock-fixtures.mdx` — a four-step recipe adopters can follow to validate AdCP agents that wrap an upstream platform (DSP, SSP, retail data, creative server, signal marketplace), in any language: 1. Boot a published mock upstream for the specialism 2. Point your agent at the mock as its upstream 3. Run the storyboard runner for AdCP wire-contract conformance 4. Assert traffic counters at /_debug/traffic for façade resistance Storyboards alone validate wire shape but cannot detect façade adapters that return shape-valid AdCP responses without integrating with the upstream. The traffic counter gate closes that hole. Both signals matter; one without the other is incomplete. The recipe is language-agnostic. The reference TS SDK (@adcp/client) ships four mock specialisms that other-language SDKs can also implement; the mock upstreams expose plain HTTP, so the agent under test can be in any language. Honest about limitations: storyboards under-cover payload variety (#3785), cascade fragility silently skips downstream steps on early-step shape gaps (#3796), mock seeds may not match storyboard fixture inputs, and the traffic gate is necessary but not sufficient (sophisticated façades with synthetic placeholder data still pass). Linked from validate-your-agent.mdx via a single Info callout in the preamble so adapter authors find it without needing to know the new file exists.

…/webhook caveats Three-way reviewer feedback (docs-expert, ad-tech-protocol-expert, adtech-product-expert): 1. Make explicit non-normative + harness-convention framing (top <Note>) — the original implied that /_lookup and /_debug/traffic are spec-level. Now framed as @adcp/client reference implementation; alternatives may diverge. 2. Reframe as 'pre-staging gate' (not 'validation/replacement'). Recipe complements storyboards + staging integration, doesn't replace either. 3. Drop Python-only sketch. The original violated its own 'language-agnostic' claim. Replaced with a sidecar pattern note: agent stays in its language, TS CLI runs as service container. 4. Bare 'adcp' → 'npx @adcp/client@latest'. Matches existing adcp docs convention. 5. Tone: removed 'dogfood' + 'empirically observed' phrasings. Now docs voice. 6. Added concrete GitHub Actions YAML example with the traffic-counter assertion. 7. Added threshold guidance (= 1 per route is the right floor; richer assertions are encoded in storyboard payloads, not the gate). 8. Mock specialism table: added caveat that AdCP-side identifier names (account.advertiser/operator/publisher) are SDK conventions for tenant binding, not normative AdCP terms. Plus a note that the four specialisms are not exhaustive. 9. Limitations section restructured into three sub-headings: - Storyboard limitations (#3785 payload variety) - Runner/tooling caveats (#3796 cascade — moved here, was wrongly tagged as spec issue) - Traffic gate limitations — added two from protocol-expert review: * idempotency replay isn't exercised (façade that doubles writes passes hit count) * outbound webhook delivery isn't on upstream traffic counters 10. Added inbound links from build-an-agent.mdx (so adapter-shape decisions trigger the discovery) and compliance-catalog.mdx (so adopters reading the storyboard taxonomy see the complementary gate). Reviewers' concerns about certification implications — addressed by the top <Note> explicitly disclaiming compliance-tier status.

…at new paths PR #4022 (IA Phase 2) relocated the linked pages after this branch diverged. Rebase silently dropped the three discoverability callouts because their source paths no longer existed. Re-author at the new paths: - Move docs/building/validate-with-mock-fixtures.mdx into docs/building/verification/ alongside the other validation pages. - Add discovery sentence to docs/building/by-layer/L4/build-an-agent.mdx. - Add Tip block to docs/building/verification/compliance-catalog.mdx. - Add Info callout to docs/building/verification/validate-your-agent.mdx. - Update internal links inside the moved doc to canonical post-IA paths. - Register the page in docs.json nav (both top-level and L4 builders nav). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The Info callout added to dist/docs/3.0.7/building/validate-your-agent.mdx in #3826 linked to /docs/building/validate-with-mock-fixtures, but the page actually lives at /docs/building/verification/validate-with-mock-fixtures (the IA Phase 2 path that #3826 itself migrated to in the live docs/ copy). The link in dist/ was missed during that migration, fails CI's broken-links check on every PR that touches docs/ or addie source code, and is currently red on origin/main. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The previous fix(docs) commit tried to correct the link path. But the target page (validate-with-mock-fixtures) doesn't exist in the dist/docs/3.0.7/ snapshot at all — it was added to the live docs/ tree after 3.0.7 was cut. Either spelling of the link breaks one of two checks: broken-links (target missing) or check-dist-links (unversioned link in a versioned snapshot). The 3.0.7 snapshot should be immutable; PR #3826 violated that by adding the Info callout pointing at content that didn't exist when 3.0.7 shipped. Restore the snapshot to its pre-3826 state. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…truth (#4500) * feat(agents): rescue long advisors via -deep suffix, unify source of truth `.agents/roles/` is now the single source for all subagent prompts, with two forms per role: - `{name}.md` — short triage checker (terse, PR-bound, used by triage routine) - `{name}-deep.md` — long-form design advisor (full reasoning) for open-ended work like MCP tool surface review or threat modeling Nine `-deep` counterparts are now reachable that had been silently shadowed by the short `.claude/agents/` versions for ~7 months: MCP agentic-API design review, context-window efficiency lens, threat modeling depth, curriculum architecture, etc. `.claude/agents/` and `.codex/agents/` are fully generated by `scripts/import-claude-agents.mjs`. The script gained a filename-matches-frontmatter-name check and a duplicate-name guard. Addie's expert-panel system-prompt section filters out `-deep` variants so each persona appears once (server/src/addie/rules/index.ts), with a regression test. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(agents): address expert review on PR #4500 * fix(docs): correct broken-link in 3.0.7 snapshot validate-your-agent.mdx The Info callout added to dist/docs/3.0.7/building/validate-your-agent.mdx in #3826 linked to /docs/building/validate-with-mock-fixtures, but the page actually lives at /docs/building/verification/validate-with-mock-fixtures (the IA Phase 2 path that #3826 itself migrated to in the live docs/ copy). The link in dist/ was missed during that migration, fails CI's broken-links check on every PR that touches docs/ or addie source code, and is currently red on origin/main. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: align schema registry adcp_version with package.json The legacy adcp_version alias in static/schemas/source/index.json had drifted to 3.0.12 (a non-existent release tag) while published_version and package.json are at 3.0.3. The pre-push verify-version-sync hook caught this; origin/main is in the same state but CI doesn't run the check, so it went unnoticed. Regenerated via `node scripts/build-schemas.cjs`; the only change is the one-line legacy alias. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(docs): remove invalid Info block from 3.0.7 snapshot The previous fix(docs) commit tried to correct the link path. But the target page (validate-with-mock-fixtures) doesn't exist in the dist/docs/3.0.7/ snapshot at all — it was added to the live docs/ tree after 3.0.7 was cut. Either spelling of the link breaks one of two checks: broken-links (target missing) or check-dist-links (unversioned link in a versioned snapshot). The 3.0.7 snapshot should be immutable; PR #3826 violated that by adding the Info callout pointing at content that didn't exist when 3.0.7 shipped. Restore the snapshot to its pre-3826 state. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

bokelley mentioned this pull request May 2, 2026

examples: hello_seller_adapter_signal_marketplace — worked passing reference adcontextprotocol/adcp-client#1274

Merged

6 tasks

bokelley and others added 3 commits May 13, 2026 13:27

bokelley force-pushed the bokelley/validate-with-mock-fixtures branch from 47b5486 to 86f2601 Compare May 13, 2026 17:35

bokelley merged commit 540f5a9 into main May 13, 2026
11 of 13 checks passed

bokelley deleted the bokelley/validate-with-mock-fixtures branch May 13, 2026 17:35

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

docs: validate adapter agents with mock upstream fixtures#3826

docs: validate adapter agents with mock upstream fixtures#3826
bokelley merged 3 commits into
mainfrom
bokelley/validate-with-mock-fixtures

bokelley commented May 2, 2026

Uh oh!

bokelley commented May 2, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

bokelley commented May 2, 2026

Summary

Why this guide

Why this lives in the spec repo

Honesty about limitations

Discoverability

Test plan

Refs

Uh oh!

bokelley commented May 2, 2026

What changed

Convergent block (3 reviewers): non-normative framing

Convergent block: language-agnostic claim vs Python-only sketch

Reviewer-specific issues addressed

Discoverability (docs expert)

Tone

Open questions reviewers raised that need spec-side input

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant