diff --git a/.github/workflows/cloud-runtime-artifact.yml b/.github/workflows/cloud-runtime-artifact.yml
index c81cf655c..eb213bb3d 100644
--- a/.github/workflows/cloud-runtime-artifact.yml
+++ b/.github/workflows/cloud-runtime-artifact.yml
@@ -42,9 +42,22 @@ jobs:
         working-directory: kernel
         run: cargo build --locked --release -p relayflowd
 
+      - name: Install SDK dependencies
+        run: npm ci --prefix sdk
+
+      - name: Test SDK and type-level authoring contracts
+        working-directory: sdk
+        run: |
+          npm run build
+          npm run typecheck:tests
+          ./node_modules/.bin/vitest run \
+            tests/typed-output.test.ts \
+            tests/validate.test.ts \
+            tests/spec-parity.test.ts \
+            tests/deterministic-llm.test.ts
+
       - name: Build standalone flows CLI
         run: |
-          npm ci --prefix sdk
           mkdir -p dist/cloud-artifact-input
           bun build sdk/src/cli-executable.ts \
             --compile \
diff --git a/docs/SURFACE.md b/docs/SURFACE.md
index 4ea7f3131..a14ca5b59 100644
--- a/docs/SURFACE.md
+++ b/docs/SURFACE.md
@@ -91,6 +91,31 @@ No process runs between events: the handler wakes, executes to its next await, p
    **Project-config discovery:** starting in the flow file's directory, `flows check` walks parent directories through the filesystem root and selects the first readable `flows.json`. That nearest file is the whole project config; it is not merged with outer files. Its schema is `{ "cli"?: <non-empty string>, "executors"?: <non-empty string>[] }`; unknown keys fail closed as `config_invalid`. A nearer config therefore defines a self-contained nested project boundary and prevents accidental inheritance of outer credentials or executors. The selected path is printed with project-level resolutions and named in an unresolved-CLI refusal; if it declares no `cli`, outer configs remain shadowed. At gate 1, a trigger executor is considered registered only when its name is present in this author-written `executors` array; `flows check` does not yet contact a registry, broker, or RelayCron, and absence is `no_executor`.
 7. **Two dialects, one journal.** Declarative YAML — data, fully preflightable, sage's compile target, gate 9's self-authoring output. Imperative TS — journal-memoized function, maximum ergonomics. YAML is canonical; TS is the power tool. TS preflights its declared surface (agents, helpers, tools, identity), not arbitrary control flow — declared honestly per covenant 2.
 
+### Structured output declarations
+
+Declarative `llm` and `agent` steps may declare an `output` JSON Schema. This
+is authoring sugar for the existing kernel `json_schema` verification gate; the
+compiler removes `output` before the journal boundary and emits the schema as
+`verification.json_schema`. Authors must choose either `output` or an explicit
+`verification` block. Declaring both is ambiguous and fails closed.
+
+```yaml
+- id: extract
+  type: llm
+  prompt: Return the actionable request as JSON.
+  output:
+    type: object
+    required: [actionable, request]
+    properties:
+      actionable: { type: boolean }
+      request: { type: string }
+```
+
+The declaration does not add a kernel primitive and does not yet infer a
+TypeScript result type from arbitrary JSON Schema. Typed parsed values belong
+to the imperative `f.llm` / `f.agent` surface once that surface has a real
+consumer; the spec SDK does not publish an unchecked phantom type in advance.
+
 The authoring surface deliberately narrows `steps: []`: `flows check` refuses
 it as `invalid_spec`, while the kernel accepts it. This is a chosen
 authoring-time narrowing, not a kernel guarantee.
diff --git a/ops/reviews/20260902-1620-pr133-history.md b/ops/reviews/20260902-1620-pr133-history.md
new file mode 100644
index 000000000..5a2be39d2
--- /dev/null
+++ b/ops/reviews/20260902-1620-pr133-history.md
@@ -0,0 +1,444 @@
+# PR #133 fresh exact-head history / RFC review
+
+- Review lens: RFC and history fit, backwards compatibility, issue #132 acceptance scope, regression risk, evidence reproducibility
+- Repository: `AgentWorkforce/flows`
+- Base: `a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2`
+- Reviewed head: `5ab0fee366362de33ea1bd7d88a96daef1cad18a`
+- Date: 2026-09-02
+- Verdict: PASS
+
+## Findings
+
+No blocking findings.
+
+The change fits RFC-0001 decision 13: `output` is an authoring-surface field which is erased before the journal boundary and lowered to the existing kernel `json_schema` verification primitive. No kernel files, step verbs, or journal vocabulary change. The implementation fails closed on malformed or ambiguous declarations, including direct callers of the public `toKernelSpec` boundary.
+
+The repair commit removes the speculative `OutputFromSchema` / generic result-type API introduced at the first commit, and the test-only TypeScript project makes the negative `@ts-expect-error` assertion part of ordinary `npm test` and exact-head artifact CI. This matches the repository rule against speculative abstraction and provides a real type gate.
+
+Backwards-compatibility scope is preserved:
+
+- Existing explicit `verification: { type: json_schema, ... }` authoring still compiles.
+- A direct comparison below shows the legacy declaration and new `output` sugar produce byte-identical kernel specs.
+- The previous-generation `workflows/` tree is unchanged at base versus head. This review therefore establishes non-interference and continued acceptance of the legacy declaration; it does not claim to have executed every previous-generation v1 workflow through an external v1 runner.
+
+PR and issue scope are honest. Issue #132 remains open. PR #133 says `Refs #132`, not `Closes #132`, and its body explicitly limits the change to slice 1. The issue's full done condition still requires imperative typed `f.llm` / `f.agent` parsed values, a parsed-value gate in the research flow, the published surface package, direct-run input, parallel dispatch, and other listed slices. This PR adds declarative YAML/object `StepSpec.output` schema sugar only and documents that limitation.
+
+## Constitution read
+
+I read both required files completely before assessing code:
+
+```text
+$ wc -l AGENTS.md docs/RFC-0001-everything-is-a-relayflow.md
+      75 AGENTS.md
+     244 docs/RFC-0001-everything-is-a-relayflow.md
+     319 total
+
+$ shasum -a 256 AGENTS.md docs/RFC-0001-everything-is-a-relayflow.md
+3e542e190bd4375a105612cf119b18e39b7732be181126c5bcf75c37010b65b8  AGENTS.md
+cf8c0f41b12dc37699aa348d10c1907dae8b7e8550a34ad976aac1349639c7b7  docs/RFC-0001-everything-is-a-relayflow.md
+```
+
+The complete files were read with `sed -n '1,160p' AGENTS.md` and the three
+non-overlapping RFC commands `sed -n '1,90p'`, `sed -n '91,180p'`, and
+`sed -n '181,280p'`. The hashes above pin the exact contents read without
+substituting a prose summary for file output.
+
+Relevant RFC fit observed from that literal read:
+
+- RFC §1: `llm` is distinct from `agent`, has value output, and verification is the rail.
+- Covenant 2: malformed/ambiguous declarations must fail closed before a run.
+- Decision 13: surface vocabulary may be open, but it must compile to the closed kernel vocabulary.
+- The open versioning question requires compatibility rather than implicit deprecation.
+
+## Exact checkout and change inventory
+
+```text
+$ pwd
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt
+$ git rev-parse HEAD
+5ab0fee366362de33ea1bd7d88a96daef1cad18a
+$ git rev-parse a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2
+a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2
+$ git branch --show-current
+feat/v2-typed-outputs
+$ git log --oneline --decorate --no-merges a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a
+5ab0fee (HEAD -> feat/v2-typed-outputs, origin/feat/v2-typed-outputs) fix(sdk): enforce structured output contracts
+81c49df feat(sdk): compile typed outputs to json_schema
+```
+
+```text
+$ git diff --stat a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a
+ .github/workflows/cloud-runtime-artifact.yml |  15 +-
+ docs/SURFACE.md                              |  25 ++++
+ sdk/package.json                             |   3 +-
+ sdk/src/compile.ts                           |  19 ++-
+ sdk/src/index.ts                             |   1 +
+ sdk/src/output-schema.ts                     |  22 +++
+ sdk/src/spec.ts                              |  13 ++
+ sdk/src/validate.ts                          |   7 +-
+ sdk/tests/live-kernel.test.ts                |  36 ++++-
+ sdk/tests/typed-output.test.ts               | 209 +++++++++++++++++++++++++++
+ sdk/tsconfig.tests.json                      |  10 ++
+ testdata/hn-monitor.flow.yaml                |  33 ++---
+ 12 files changed, 365 insertions(+), 28 deletions(-)
+
+$ git diff --name-status a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a
+M       .github/workflows/cloud-runtime-artifact.yml
+M       docs/SURFACE.md
+M       sdk/package.json
+M       sdk/src/compile.ts
+M       sdk/src/index.ts
+A       sdk/src/output-schema.ts
+M       sdk/src/spec.ts
+M       sdk/src/validate.ts
+M       sdk/tests/live-kernel.test.ts
+A       sdk/tests/typed-output.test.ts
+A       sdk/tsconfig.tests.json
+M       testdata/hn-monitor.flow.yaml
+```
+
+I inspected the complete patch, including both commits' net diff and then the
+repair-only diff. The exact patch is reproducibly pinned by its literal hash,
+line count, and per-file numstat:
+
+```text
+$ git diff --find-renames --find-copies --no-ext-diff a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a | shasum -a 256
+cb334c3a6d3ae3dc9ecc007d06a599920e8111b8eec6d66038ee5881ddb685f2  -
+
+$ git diff --find-renames --find-copies --no-ext-diff a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a | wc -l
+     588
+
+$ git diff --numstat a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a
+14      1       .github/workflows/cloud-runtime-artifact.yml
+25      0       docs/SURFACE.md
+2       1       sdk/package.json
+18      1       sdk/src/compile.ts
+1       0       sdk/src/index.ts
+22      0       sdk/src/output-schema.ts
+13      0       sdk/src/spec.ts
+5       2       sdk/src/validate.ts
+30      6       sdk/tests/live-kernel.test.ts
+209     0       sdk/tests/typed-output.test.ts
+10      0       sdk/tsconfig.tests.json
+16      17      testdata/hn-monitor.flow.yaml
+
+$ git show --stat --oneline 81c49df32aebcbb1b21fc3f87eeae55cb2ed6ff1
+81c49df feat(sdk): compile typed outputs to json_schema
+ sdk/src/compile.ts             |  22 +++++-
+ sdk/src/index.ts               |   2 +
+ sdk/src/output-schema.ts       |  35 +++++++++
+ sdk/src/spec.ts                |  17 ++++-
+ sdk/src/validate.ts            |   7 +-
+ sdk/tests/live-kernel.test.ts  |  36 +++++++--
+ sdk/tests/typed-output.test.ts | 170 +++++++++++++++++++++++++++++++++++++++++
+ testdata/hn-monitor.flow.yaml  |  33 ++++----
+ 8 files changed, 294 insertions(+), 28 deletions(-)
+
+$ git show --stat --oneline 5ab0fee366362de33ea1bd7d88a96daef1cad18a
+5ab0fee fix(sdk): enforce structured output contracts
+ .github/workflows/cloud-runtime-artifact.yml | 15 ++++-
+ docs/SURFACE.md                              | 25 ++++++++
+ sdk/package.json                             |  3 +-
+ sdk/src/compile.ts                           | 13 ++---
+ sdk/src/index.ts                             |  1 -
+ sdk/src/output-schema.ts                     | 17 +-----
+ sdk/src/spec.ts                              | 14 ++---
+ sdk/tests/typed-output.test.ts               | 85 ++++++++++++++++++++--------
+ sdk/tsconfig.tests.json                      | 10 ++++
+ 9 files changed, 127 insertions(+), 56 deletions(-)
+```
+
+The repair-only patch was also inspected in full. It removes `OutputFromSchema` and generic `TOutput`, adds the dedicated test tsconfig and CI invocation, centralizes validation in `validateOutputDeclaration`, adds internal-normalization validation, and pins raw `toKernelSpec` success/conflict/malformed arms.
+
+## Issue #132 and PR claim check
+
+```text
+$ gh issue view 132 --repo AgentWorkforce/flows --json number,title,state,url
+{"number":132,"state":"OPEN","title":"v2 authoring ergonomics: close the gaps found in the research-flow / v1 / Smithers comparison","url":"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/AgentWorkforce/flows/issues/132"}
+
+$ gh pr view 133 --repo AgentWorkforce/flows --json number,title,state,baseRefName,headRefName,headRefOid,url
+{"baseRefName":"main","headRefName":"feat/v2-typed-outputs","headRefOid":"5ab0fee366362de33ea1bd7d88a96daef1cad18a","number":133,"state":"OPEN","title":"feat(sdk): compile structured outputs to json_schema","url":"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/AgentWorkforce/flows/pull/133"}
+```
+
+The literal issue comments establish the migration and slice boundaries:
+
+```text
+$ gh issue view 132 --repo AgentWorkforce/flows --comments
+author: kjgbot
+association: member
+edited: false
+status: none
+--
+Implementation ownership started on 2026-09-02.
+
+- Active first slice: typed `llm` / `agent` outputs on branch `feat/v2-typed-outputs`, owned by Agent Relay worker `flows-132-typed-output`, based on merged main `a0d42ff`. This must compile to the existing `json_schema` primitive; no kernel vocabulary expansion.
+- Next slices, each as a separately reviewed PR: publish `@relayflows/surface`; direct-run input; journaled `agents[].model` surface plus strict typo/model linting; gate-3 parallel dispatch; settle replayable data gates versus code gates; verb-specific unknown-field refusal.
+- Migration policy: v1 remains the default and supported runtime while these v2 authoring gaps close. Stored version remains authoritative for resume/schedules. Deprecation is a later explicit decision, not implied by this issue.
+
+The Cloud runtime artifact foundation merged separately in #131. A real Cloud v2 proof can proceed on the deterministic surface while this issue gates broader author migration and the research-flow completion criteria.
+```
+
+The current PR body was read in full. Its relevant scope language is literal:
+
+```text
+## Scope boundary
+
+This is issue #132 slice 1 only. It does not publish
+`@relayflows/surface`, implement imperative typed return values, or close #132.
+v1 explicit `verification:` remains supported.
+```
+
+That description matches the patch: no `research/` implementation or local shim is changed, no imperative `f.llm` / `f.agent` surface is published here, and issue #132 remains open.
+
+## Backwards compatibility evidence
+
+The previous-generation v1 workflow tree is byte-identical between base and head:
+
+```text
+$ if git diff --quiet a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a -- workflows; then echo V1_WORKFLOW_TREE_UNCHANGED=true; else echo V1_WORKFLOW_TREE_UNCHANGED=false; fi
+V1_WORKFLOW_TREE_UNCHANGED=true
+
+$ for f in workflows/review-swarm.yaml workflows/watchdog.yaml workflows/bootstrap-gate1.yaml workflows/drive.yaml; do printf '%s ' "$f"; git show a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2:"$f" | shasum -a 256 | awk '{print $1}'; done
+workflows/review-swarm.yaml 6343d8c9f1505a1cbd3660c507730aa371a3d72b6a4103fff815741f43377ca6
+workflows/watchdog.yaml e04a14e20dd1c3562e66901994dbf4533224581bc45262114c0c9c7aa72aa589
+workflows/bootstrap-gate1.yaml f7f74dd595f1a68c3d819fe7e01f2a4a07bb3d0bcbd314777f30c9d07a02035e
+workflows/drive.yaml b691305d132c2b2bacb5b4e44a4e6efa7b7872e9f52d4201bda1f00e09ae58c4
+
+$ for f in workflows/review-swarm.yaml workflows/watchdog.yaml workflows/bootstrap-gate1.yaml workflows/drive.yaml; do printf '%s ' "$f"; git show 5ab0fee366362de33ea1bd7d88a96daef1cad18a:"$f" | shasum -a 256 | awk '{print $1}'; done
+workflows/review-swarm.yaml 6343d8c9f1505a1cbd3660c507730aa371a3d72b6a4103fff815741f43377ca6
+workflows/watchdog.yaml e04a14e20dd1c3562e66901994dbf4533224581bc45262114c0c9c7aa72aa589
+workflows/bootstrap-gate1.yaml f7f74dd595f1a68c3d819fe7e01f2a4a07bb3d0bcbd314777f30c9d07a02035e
+workflows/drive.yaml b691305d132c2b2bacb5b4e44a4e6efa7b7872e9f52d4201bda1f00e09ae58c4
+```
+
+Legacy explicit verification and the new sugar lower identically:
+
+```text
+$ PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin node --input-type=module -e '<script compiling otherwise-identical explicit verification and output declarations>'
+LEGACY_VERIFICATION_ACCEPTED=object
+LOWERING_IDENTICAL=true
+{"json_schema":{"type":"object","required":["answer"],"properties":{"answer":{"type":"integer"}}}}
+```
+
+The exact script used was:
+
+```js
+import { compileYaml, toKernelSpec } from "./dist/compile.js";
+const legacy = `version: "0.1.0"
+steps:
+  - id: extract
+    type: agent
+    instruction: Return JSON.
+    verification:
+      type: json_schema
+      schema:
+        type: object
+        required: [answer]
+        properties:
+          answer: { type: integer }
+`;
+const sugar = `version: "0.1.0"
+steps:
+  - id: extract
+    type: agent
+    instruction: Return JSON.
+    output:
+      type: object
+      required: [answer]
+      properties:
+        answer: { type: integer }
+`;
+const legacyKernel = toKernelSpec(compileYaml(legacy));
+const sugarKernel = toKernelSpec(compileYaml(sugar));
+console.log(`LEGACY_VERIFICATION_ACCEPTED=${legacyKernel.steps[0].verification.json_schema.type}`);
+console.log(`LOWERING_IDENTICAL=${JSON.stringify(legacyKernel) === JSON.stringify(sugarKernel)}`);
+console.log(JSON.stringify(legacyKernel.steps[0].verification));
+```
+
+## Build, typecheck, and focused regression evidence
+
+The direct build, real test typecheck, and exact CI-selected suites pass:
+
+```text
+$ PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin ./node_modules/.bin/tsc && PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin node scripts/make-cli-executable.mjs && PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin ./node_modules/.bin/tsc -p tsconfig.tests.json && PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin ./node_modules/.bin/vitest run tests/typed-output.test.ts tests/validate.test.ts tests/spec-parity.test.ts tests/deterministic-llm.test.ts --reporter=verbose
+
+ RUN  v2.1.9 /Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk
+
+ ✓ tests/deterministic-llm.test.ts > compile: deterministic + llm flow (ladder rung b) > compiles a deterministic step and an llm step to spec JSON
+ ✓ tests/deterministic-llm.test.ts > compile: deterministic + llm flow (ladder rung b) > keeps the llm step workspace-free and value-output (distinct from agent)
+ ✓ tests/deterministic-llm.test.ts > compile: deterministic + llm flow (ladder rung b) > carries the json_schema verification gate for structured llm output
+ ✓ tests/deterministic-llm.test.ts > compile: agent step (ladder rung c, Appendix A surface) > compiles surfaces, recovery mode, and permissions
+ ✓ tests/deterministic-llm.test.ts > compile: agent step (ladder rung c, Appendix A surface) > defaults agent recoveryMode to reset when omitted (Appendix A rule 4)
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > typechecks output declarations on llm and agent steps only
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > compiles llm output sugar to the existing json_schema primitive
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > compiles agent output sugar to the existing json_schema primitive
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > accepts the output declaration in YAML
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > keeps the canonical hn-monitor kernel step unchanged
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed when output conflicts with verification ('output_contains')
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed when output conflicts with verification ('json_schema')
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a non-object output schema (null)
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a non-object output schema (%j)
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a non-object output schema ("not-a-schema")
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > compiles a raw llm spec directly through the public kernel boundary
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > compiles a raw agent spec directly through the public kernel boundary
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a conflicting raw spec at the public kernel boundary
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a malformed raw spec at the public kernel boundary
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > compiles hello-deterministic to the pinned canonical JSON
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > hashes hello-deterministic to the pinned spec_hash
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > round-trips hello-deterministic across the kernel dialect
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > compiles hello-ladder to the pinned canonical JSON
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > hashes hello-ladder to the pinned spec_hash
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > round-trips hello-ladder across the kernel dialect
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > compiles hello-llm to the pinned canonical JSON
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > hashes hello-llm to the pinned spec_hash
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > round-trips hello-llm across the kernel dialect
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > compiles hello-agent to the pinned canonical JSON
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > hashes hello-agent to the pinned spec_hash
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > round-trips hello-agent across the kernel dialect
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > normalizes empty triggers exactly as kernel serialization does
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > refuses a kernel retry policy the authoring dialect cannot represent
+ ✓ tests/spec-parity.test.ts > spec parity: one dialect at the SDK<->kernel boundary > round-trips flow, trigger, and step CLI declarations
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a non-object spec
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a missing version
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a version the kernel does not support
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a malformed optional name
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects an empty steps array
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects an unknown step type
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects duplicate step ids
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a deterministic step missing its command
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects an llm step missing its prompt
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects an agent step missing its instruction
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects an invalid recovery mode
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a custom exit_code expectation (v0 judges exit_code == 0)
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a timeout on a non-deterministic step (no dialect surface in v0.1.0)
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects an output_contains gate with no value
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects an unknown verification gate type
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a dependsOn referencing an unknown step
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a dependency cycle
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a float money budget (must be a decimal string)
+ ✓ tests/validate.test.ts > validate: rejects malformed specs > rejects a non-integer maxIterations
+ ✓ tests/validate.test.ts > validate: fail-closed on unknown keys (RFC covenant 2) > rejects the depends_on typo with the authoring-vocabulary suggestion
+ ✓ tests/validate.test.ts > validate: fail-closed on unknown keys (RFC covenant 2) > rejects unknown keys at the root level
+ ✓ tests/validate.test.ts > validate: fail-closed on unknown keys (RFC covenant 2) > rejects unknown keys at the budget level
+ ✓ tests/validate.test.ts > validate: fail-closed on unknown keys (RFC covenant 2) > rejects unknown keys at the step level, including per-type fields
+ ✓ tests/validate.test.ts > validate: fail-closed on unknown keys (RFC covenant 2) > rejects unknown keys at the verification level
+ ✓ tests/validate.test.ts > validate: fail-closed on unknown keys (RFC covenant 2) > rejects unknown keys at the surfaces level (object and items)
+ ✓ tests/validate.test.ts > validate: fail-closed on unknown keys (RFC covenant 2) > rejects unknown keys at the permissions level
+ ✓ tests/validate.test.ts > validate: fail-closed on unknown keys (RFC covenant 2) > names allowed keys when no valid key is close
+ ✓ tests/validate.test.ts > compile: surfaces validation failures as CompileError > throws CompileError with the offending messages for a malformed YAML spec
+ ✓ tests/validate.test.ts > compile: surfaces validation failures as CompileError > throws CompileError for YAML that is not a mapping
+ ✓ tests/validate.test.ts > validate: accepts the legal zero-agent flow > accepts the kernel-supported schema version
+ ✓ tests/validate.test.ts > validate: accepts the legal zero-agent flow > accepts a spec without a name because the kernel treats it as optional
+ ✓ tests/validate.test.ts > validate: accepts the legal zero-agent flow > accepts a pure-deterministic spec — zero agents/llm is legal (RFC §1)
+ ✓ tests/validate.test.ts > validate: accepts the legal zero-agent flow > accepts a deterministic + llm spec
+ ✓ tests/validate.test.ts > validate: preflight declarations > accepts CLI defaults and inert trigger data
+ ✓ tests/validate.test.ts > validate: preflight declarations > rejects duplicate trigger ids
+ ✓ tests/validate.test.ts > validate: preflight declarations > rejects malformed and unknown trigger/CLI fields fail-closed
+
+ Test Files  4 passed (4)
+      Tests  70 passed (70)
+   Start at  17:23:51
+   Duration  2.05s (transform 563ms, setup 0ms, collect 1.62s, tests 528ms, environment 4ms, prepare 1.67s)
+```
+
+The absence of output before Vitest is the literal successful output of the two `tsc` invocations and CLI build helper.
+
+## Full SDK / live-kernel evidence
+
+The repository's default test preparation could not build a fresh local daemon because this host has no rustup default. This is preserved as a failed command, not represented as a passing build:
+
+```text
+$ PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin sh ../ops/cargo.sh build && find ../testdata/preflight -name '*-cli' -type f -exec chmod +x {} + && PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin ./node_modules/.bin/tsc && PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin node scripts/make-cli-executable.mjs && PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin ./node_modules/.bin/tsc -p tsconfig.tests.json && PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/usr/bin:/bin ./node_modules/.bin/vitest run
+error: rustup could not choose a version of cargo to run, because one wasn't specified explicitly, and no default is configured.
+help: run 'rustup default stable' to download the latest stable release of Rust and set it as your default toolchain.
+```
+
+The documented prebuilt daemon exists:
+
+```text
+$ test -x /Users/khaliqgant/.relayflows-toolchain/target/1914866954/debug/relayflowd && echo PREBUILT_RELAYFLOWD_PRESENT || echo PREBUILT_RELAYFLOWD_MISSING
+PREBUILT_RELAYFLOWD_PRESENT
+$ ls -l /Users/khaliqgant/.relayflows-toolchain/target/1914866954/debug/relayflowd
+-rwxr-xr-x@ 1 khaliqgant  staff  22987640 Sep  2 11:27 /Users/khaliqgant/.relayflows-toolchain/target/1914866954/debug/relayflowd
+```
+
+Using that prebuilt daemon, the corrected full-suite command retained Node 22 and the installed analyzer CLI. It passed every suite, including the real analyzer and live journal assertions:
+
+```text
+$ PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/opt/homebrew/bin:/usr/bin:/bin RELAYFLOWD_BIN=/Users/khaliqgant/.relayflows-toolchain/target/1914866954/debug/relayflowd ./node_modules/.bin/vitest run
+
+ RUN  v2.1.9 /Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk
+
+ ✓ tests/preflight.test.ts (14 tests) 33ms
+ ✓ tests/backlog-picker.test.ts (14 tests) 346ms
+ ✓ tests/validate.test.ts (36 tests) 121ms
+ ✓ tests/journal-client.test.ts (13 tests) 232ms
+ ✓ tests/cli-hn-monitor.test.ts (16 tests) 239ms
+ ✓ tests/work-package-consumer.test.ts (13 tests) 968ms
+ ✓ tests/typed-output.test.ts (14 tests) 199ms
+ ✓ tests/hn-poller.test.ts (6 tests) 80ms
+ ✓ tests/dir-watcher-poller.test.ts (6 tests) 65ms
+ ✓ tests/backlog-picker-flow.test.ts (6 tests) 2338ms
+ ✓ tests/deterministic-llm.test.ts (5 tests) 288ms
+ ✓ tests/hello-deterministic.test.ts (5 tests) 199ms
+ ✓ tests/work-package-validator.test.ts (7 tests) 58ms
+ ✓ tests/cli.test.ts (50 tests) 2898ms
+ ✓ tests/parse-json-output.test.ts (7 tests) 23ms
+ ✓ tests/spec-parity.test.ts (15 tests) 578ms
+ ✓ tests/bin.test.ts (7 tests) 3329ms
+
+stdout | tests/live-kernel.test.ts > built flows CLI against live relayflowd > hn-monitor analyze-story reaches done through the real Claude analyzer CLI
+LIVE_ANALYZER ready: claude -p --model claude-haiku-4-5-20251001 round-trip OK
+
+stdout | tests/live-kernel.test.ts > built flows CLI against live relayflowd > hn-monitor analyze-story reaches done through the real Claude analyzer CLI
+LIVE_ANALYZER analysis: {"reasoning":"This story is directly about an autonomous agent performing software engineering tasks—specifically opening and reviewing pull requests without human intervention. This is highly relevant to AI agents and automation, representing a practical implementation of agent self-management and code review automation.","relevance_score":9,"story_title":"Show HN: an agent that opens and reviews its own pull requests [wake-nonce-7f3a91c4]"}
+
+stdout | tests/live-kernel.test.ts > surface resume after a real daemon kill > resumes a three-step run with each successful completion exactly once
+LIVE_KERNEL kill -9 pid=59567 run=01M1HBMX1RN2DHRKRAA3E33HEY while step=two state=Running
+
+ ✓ tests/live-kernel.test.ts (17 tests) 57410ms
+   ✓ built flows CLI against live relayflowd > runs rung (a), parks rung (b), and keeps JSON report-shaped 4594ms
+   ✓ built flows CLI against live relayflowd > allows a deterministic run to exceed the bounded request timeout 32816ms
+   ✓ built flows CLI against live relayflowd > follows a live worker dispatch through flows run 582ms
+   ✓ built flows CLI against live relayflowd > can always get a parked run to a late-attaching worker 5570ms
+   ✓ built flows CLI against live relayflowd > reports a real manual-recovery NeedsHuman state as parked 452ms
+   ✓ built flows CLI against live relayflowd > hn-monitor analyze-story reaches done through the real Claude analyzer CLI 10176ms
+   ✓ built flows CLI against live relayflowd > preflights before journaling and names an unreachable socket 523ms
+   ✓ surface resume after a real daemon kill > resumes a three-step run with each successful completion exactly once 834ms
+
+ Test Files  18 passed (18)
+      Tests  251 passed (251)
+   Start at  17:25:22
+   Duration  59.47s (transform 2.03s, setup 0ms, collect 7.29s, tests 69.40s, environment 16ms, prepare 8.16s)
+```
+
+An earlier full-suite attempt used a deliberately minimal PATH that omitted `/opt/homebrew/bin`, so the one real-analyzer test failed with `spawnSync claude ENOENT` while the other 250 tests passed. The corrected command above restores the analyzer path and passes 251/251. That first result was an invocation-environment failure, not counted as passing evidence.
+
+## Repository and hosted checks
+
+```text
+$ gh pr checks 133 --repo AgentWorkforce/flows
+CodeRabbit        pass    0               Review completed
+linux-x64-artifact pass   3m6s            https://github.com/AgentWorkforce/flows/actions/runs/33646896877/job/100303802460
+
+$ git diff --check a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a
+[no output]
+
+$ git rev-parse HEAD
+5ab0fee366362de33ea1bd7d88a96daef1cad18a
+```
+
+Hosted checks are supplementary only; the local commands above independently exercise the changed compiler, type gate, validator, parity boundary, full SDK suite, live kernel, and real analyzer path.
+
+## Residual risk and acceptance boundary
+
+- The surface accepts object-form JSON Schema, matching the existing `json_schema` verification contract. Deep JSON Schema validity remains the existing kernel/evidence-layer responsibility; this PR does not claim to add a new schema validator.
+- `toKernelSpec` is typed for `FlowSpec`. The new output-specific direct-call arms are runtime-pinned, but this review does not claim the public boundary validates every other possible `unknown as FlowSpec` corruption; that is pre-existing and outside this slice.
+- A fresh kernel build was not produced locally due the captured rustup host failure. No kernel source changed, the exact-head Linux artifact job built the kernel successfully, and the full local live suite passed against the named prebuilt daemon. This is sufficient for the SDK-only history review but must not be restated as a fresh local kernel build.
+- Issue #132 is not complete and must remain open. In particular, this PR does not provide imperative typed return values or rewrite the research lane gate to consume a parsed typed value.
+
+## Final verdict
+
+PR #133 at exact head `5ab0fee366362de33ea1bd7d88a96daef1cad18a` fits RFC-0001 and the repository history, preserves legacy/v1 surfaces by non-interference, keeps explicit verification supported, fails closed on the new declaration, avoids speculative public types, and scopes issue #132 honestly. No blocking finding remains.
+
+REVIEW_PASSED
diff --git a/ops/reviews/20260902-1620-pr133-structure.md b/ops/reviews/20260902-1620-pr133-structure.md
new file mode 100644
index 000000000..5f28e4460
--- /dev/null
+++ b/ops/reviews/20260902-1620-pr133-structure.md
@@ -0,0 +1,439 @@
+# Structure Review — PR #133 at `5ab0fee`
+
+- **Lens:** module boundaries, closed step unions, compiler/kernel boundary, fail-closed validation, package exports, journal behavior, and deterministic coverage
+- **Exact head:** `5ab0fee366362de33ea1bd7d88a96daef1cad18a`
+- **Base:** `a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2`
+- **Mode:** assessment only; no product code edited, committed, pushed, or merged
+- **Constitution:** `AGENTS.md` and `docs/RFC-0001-everything-is-a-relayflow.md` read completely before assessment
+
+## Verdict
+
+No blocking findings.
+
+The fixes at this exact head resolve the prior structural concerns. The public API now exposes only a plain `JsonOutputSchema = Record<string, unknown>` declaration type; the unused phantom parsed-result generic and extractor type are gone. The new test tsconfig includes the type assertions and is wired into both `npm test` and the exact-head Linux artifact workflow. Output validation is centralized in a 22-line single-purpose module, used by both `compileSpec` validation and direct `toKernelSpec` lowering. Direct raw `llm` and `agent` paths, direct conflict/malformed paths, unknown-key handling, and legacy explicit `verification:` are all pinned or independently exercised.
+
+The change stays on the authoring side of the journal boundary: no kernel source, kernel step verb, or kernel field changed. `output:` is accepted only on the existing `llm` and `agent` arms, normalized to the existing `json_schema` verification form, and removed before the kernel boundary. Live tests show the existing journal behavior remains authoritative: accepted parsed JSON is persisted with a passing schema verdict; rejected output is nulled with a failed schema verdict and declared completion reasons.
+
+Issue #132 remains open. The PR body accurately calls this “slice 1 only,” explicitly excludes imperative typed return values and `@relayflows/surface`, and does not claim to close the issue. The issue's current progress comment also says v1 remains the supported default and v2 remains explicit/version-pinned. The legacy authoring form continues to compile to the same kernel gate.
+
+## Exact review boundary
+
+Literal command:
+
+```sh
+pwd
+git rev-parse HEAD
+git rev-parse a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2
+git status --short
+git diff --stat a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a
+git diff --name-status a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a
+```
+
+Captured output:
+
+```text
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt
+5ab0fee366362de33ea1bd7d88a96daef1cad18a
+a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2
+?? ops/reviews/20260902-pr133-history.md
+?? ops/reviews/20260902-pr133-maintainability.md
+?? ops/reviews/20260902-pr133-structure.md
+?? sdk/node_modules
+ .github/workflows/cloud-runtime-artifact.yml |  15 +-
+ docs/SURFACE.md                              |  25 ++++
+ sdk/package.json                             |   3 +-
+ sdk/src/compile.ts                           |  19 ++-
+ sdk/src/index.ts                             |   1 +
+ sdk/src/output-schema.ts                     |  22 +++
+ sdk/src/spec.ts                              |  13 ++
+ sdk/src/validate.ts                          |   7 +-
+ sdk/tests/live-kernel.test.ts                |  36 ++++-
+ sdk/tests/typed-output.test.ts               | 209 +++++++++++++++++++++++++++
+ sdk/tsconfig.tests.json                      |  10 ++
+ testdata/hn-monitor.flow.yaml                |  33 ++---
+ 12 files changed, 365 insertions(+), 28 deletions(-)
+M	.github/workflows/cloud-runtime-artifact.yml
+M	docs/SURFACE.md
+M	sdk/package.json
+M	sdk/src/compile.ts
+M	sdk/src/index.ts
+A	sdk/src/output-schema.ts
+M	sdk/src/spec.ts
+M	sdk/src/validate.ts
+M	sdk/tests/live-kernel.test.ts
+A	sdk/tests/typed-output.test.ts
+A	sdk/tsconfig.tests.json
+M	testdata/hn-monitor.flow.yaml
+```
+
+The pre-existing untracked reports and dependency directory were not reviewed as PR content and were not modified by this assessment.
+
+Literal remote boundary command:
+
+```sh
+gh pr view 133 --repo AgentWorkforce/flows --json number,state,title,baseRefOid,headRefOid --jq .
+```
+
+Captured output:
+
+```json
+{"baseRefOid":"a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2","headRefOid":"5ab0fee366362de33ea1bd7d88a96daef1cad18a","number":133,"state":"OPEN","title":"feat(sdk): compile structured outputs to json_schema"}
+```
+
+## Full-diff structural assessment
+
+I inspected the complete 12-file diff from the exact base to the exact head.
+
+### Module boundaries and closed unions
+
+- `sdk/src/spec.ts` keeps `StepSpec` as the closed three-arm union `DeterministicStepSpec | LlmStepSpec | AgentStepSpec`, matching the RFC vocabulary.
+- `output?: JsonOutputSchema` exists only on the `llm` and `agent` arms. It is absent from `BaseStepSpec` and `DeterministicStepSpec`.
+- The public schema type is deliberately non-phantom: `Record<string, unknown>`. There is no speculative `TOutput` or `OutputFromSchema` API.
+- Runtime declaration validation lives in `sdk/src/output-schema.ts`, which has no dependency on compiler, validator, kernel, provider, tenant, or I/O modules.
+- The two large pre-existing files remain below the repository's 500-line design-smell threshold; this PR adds only focused delegations to them.
+
+Literal command:
+
+```sh
+wc -l sdk/src/compile.ts sdk/src/output-schema.ts sdk/src/spec.ts sdk/src/validate.ts sdk/tests/typed-output.test.ts
+```
+
+Captured output:
+
+```text
+     427 sdk/src/compile.ts
+      22 sdk/src/output-schema.ts
+     290 sdk/src/spec.ts
+     478 sdk/src/validate.ts
+     209 sdk/tests/typed-output.test.ts
+    1426 total
+```
+
+### `compileSpec` and `toKernelSpec`
+
+There are two intentionally supported lowering paths:
+
+1. `compileSpec` runs `validateSpec`, then `compileStep` replaces valid `output` sugar with authoring-form `{type: 'json_schema', schema}`. The normalized step contains no `output` property.
+2. A typed caller may pass a raw `FlowSpec` directly to exported `toKernelSpec`. `toKernelVerification` handles raw `output` for both `llm` and `agent`, delegates malformed/conflict checks to the same validator, and returns the existing `{json_schema: ...}` kernel form.
+
+No new field crosses the journal protocol boundary. The canonical hn-monitor kernel fixture is unchanged even though the authoring YAML now uses `output:`.
+
+Literal command:
+
+```sh
+git diff --check a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a
+printf 'diff_check_exit=%s\n' "$?"
+git diff --quiet a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a -- kernel
+printf 'kernel_diff_exit=%s\n' "$?"
+git diff --quiet a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a -- testdata/hn-monitor.spec.canonical.json
+printf 'canonical_fixture_diff_exit=%s\n' "$?"
+```
+
+Captured output:
+
+```text
+diff_check_exit=0
+kernel_diff_exit=0
+canonical_fixture_diff_exit=0
+```
+
+### Fail-closed behavior and v1 compatibility
+
+The step-key allowlists admit `output` only for `llm` and `agent`. Deterministic use and a misspelled `outputs` key are refused. Any simultaneous `output + verification` declaration is refused, including legacy `exit_code`, `output_contains`, and `json_schema` forms. Non-object declarations are refused. JSON Schema objects themselves remain open to JSON Schema vocabulary rather than being incorrectly checked against Relayflow keys.
+
+The legacy explicit `verification: {type: 'json_schema', schema}` path remains valid and lowers identically. This preserves the current v1 authoring form; the new syntax is additive sugar.
+
+Literal command:
+
+```sh
+/opt/homebrew/Cellar/node/26.7.0/bin/node --input-type=module -e "import {compileSpec,toKernelSpec,validateSpec} from './dist/index.js'; const schema={type:'object',required:['answer'],properties:{answer:{type:'number'}}}; const compiled=compileSpec({version:'0.1.0',steps:[{id:'new',type:'llm',prompt:'p',output:schema}]}); console.log('COMPILED='+JSON.stringify(compiled.steps[0])); console.log('KERNEL='+JSON.stringify(toKernelSpec(compiled).steps[0].verification)); console.log('LEGACY='+JSON.stringify(toKernelSpec(compileSpec({version:'0.1.0',steps:[{id:'old',type:'agent',instruction:'i',verification:{type:'json_schema',schema}}]})).steps[0].verification)); console.log('DETERMINISTIC_OUTPUT='+JSON.stringify(validateSpec({version:'0.1.0',steps:[{id:'d',type:'deterministic',command:'true',output:schema}]}))); console.log('MISSPELLED_OUTPUT='+JSON.stringify(validateSpec({version:'0.1.0',steps:[{id:'x',type:'llm',prompt:'p',outputs:schema}]}))); for (const [label,flow] of [['COMPILE_CONFLICT',{version:'0.1.0',steps:[{id:'c',type:'llm',prompt:'p',output:schema,verification:{type:'exit_code'}}]}],['DIRECT_CONFLICT',{version:'0.1.0',steps:[{id:'c',type:'agent',instruction:'i',output:schema,verification:{type:'output_contains',value:'x'}}]}]]) { try { label==='COMPILE_CONFLICT' ? compileSpec(flow) : toKernelSpec(flow); console.log(label+'=NO_ERROR'); } catch (error) { console.log(label+'='+error.name+': '+error.errors.join(' | ')); } }"
+```
+
+Captured output:
+
+```text
+COMPILED={"id":"new","type":"llm","verification":{"type":"json_schema","schema":{"type":"object","required":["answer"],"properties":{"answer":{"type":"number"}}}},"maxIterations":1,"prompt":"p"}
+KERNEL={"json_schema":{"type":"object","required":["answer"],"properties":{"answer":{"type":"number"}}}}
+LEGACY={"json_schema":{"type":"object","required":["answer"],"properties":{"answer":{"type":"number"}}}}
+DETERMINISTIC_OUTPUT={"ok":false,"errors":["spec.steps[0]: unknown key \"output\" (expected one of id | type | dependsOn | verification | maxIterations | timeoutMs | command)"]}
+MISSPELLED_OUTPUT={"ok":false,"errors":["spec.steps[0]: unknown key \"outputs\" — did you mean \"output\"?"]}
+COMPILE_CONFLICT=CompileError: spec.steps[0]: output already declares json_schema verification; remove verification
+DIRECT_CONFLICT=CompileError: step "c": output already declares json_schema verification; remove verification
+```
+
+### Package export surface
+
+`JsonOutputSchema` is available from the package root declaration. The source-only validation helper is not added as a second package subpath, and no unnecessary runtime public export was introduced.
+
+Literal commands:
+
+```sh
+rg -n 'JsonOutputSchema|output\?' dist/index.d.ts dist/spec.d.ts dist/output-schema.d.ts
+/opt/homebrew/Cellar/node/26.7.0/bin/node --input-type=module -e "import('./dist/index.js').then((m)=>console.log('runtime_exports='+['compileSpec','toKernelSpec','validateSpec'].filter((key)=>typeof m[key]==='function').join(',')))"
+```
+
+Captured output:
+
+```text
+dist/index.d.ts:1:export type { AgentStepSpec, AgentSurfaces, BaseStepSpec, BudgetSpec, DeterministicStepSpec, ExitCodeGate, FlowSpec, JsonSchemaGate, JsonOutputSchema, KernelAgentStep, KernelAgentSurfaces, KernelBudgetSpec, KernelDeterministicStep, KernelLlmStep, KernelPermissionsSpec, KernelRetryPolicy, KernelRunSpec, KernelStepCommon, KernelStepSpec, KernelTriggerSpec, KernelVerificationSpec, LlmStepSpec, OutputContainsGate, PermissionsSpec, RecoveryMode, StreamSurface, StepSpec, StepType, TriggerSpec, VerificationGateType, VerificationSpec, WorkspaceSurface, } from './spec.js';
+dist/output-schema.d.ts:2:export type JsonOutputSchema = Record<string, unknown>;
+dist/output-schema.d.ts:5:    output?: unknown;
+dist/spec.d.ts:1:import type { JsonOutputSchema } from './output-schema.js';
+dist/spec.d.ts:2:export type { JsonOutputSchema } from './output-schema.js';
+dist/spec.d.ts:109:    output?: JsonOutputSchema;
+dist/spec.d.ts:137:    output?: JsonOutputSchema;
+runtime_exports=compileSpec,toKernelSpec,validateSpec
+```
+
+## Deterministic/type coverage
+
+The added `sdk/tsconfig.tests.json` genuinely includes `tests/typed-output.test.ts`, so `expectTypeOf` and `@ts-expect-error` are checked by TypeScript instead of being runtime-only decoration.
+
+Literal command:
+
+```sh
+PATH=/opt/homebrew/Cellar/node/26.7.0/bin:/usr/bin:/bin /opt/homebrew/Cellar/node/26.7.0/bin/node node_modules/typescript/bin/tsc -p tsconfig.tests.json --listFilesOnly | rg 'typed-output|src/output-schema'
+printf 'list_files_probe_exit=%s\n' "$?"
+```
+
+Captured output:
+
+```text
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/backlog-picker.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/canonical.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/output-schema.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/spec.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/validate.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/compile.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/protocol.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/journal-client.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/worker.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/failure-kinds.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/preflight.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/cli/check.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/cli/run.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/hn-poller.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/cli/hn-monitor.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/cli.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/cli-executable.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/demo-hn-monitor.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/dir-watcher-poller.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/work-package-consumer.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/work-package-validator.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/src/index.ts
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/tests/typed-output.test.ts
+list_files_probe_exit=0
+```
+
+Literal build, typecheck, and focused test command:
+
+```sh
+PATH=/opt/homebrew/Cellar/node/26.7.0/bin:/usr/bin:/bin /opt/homebrew/Cellar/node/26.7.0/bin/node node_modules/typescript/bin/tsc && PATH=/opt/homebrew/Cellar/node/26.7.0/bin:/usr/bin:/bin /opt/homebrew/Cellar/node/26.7.0/bin/node scripts/make-cli-executable.mjs && PATH=/opt/homebrew/Cellar/node/26.7.0/bin:/usr/bin:/bin /opt/homebrew/Cellar/node/26.7.0/bin/node node_modules/typescript/bin/tsc -p tsconfig.tests.json
+printf 'build_and_typecheck_exit=%s\n' "$?"
+PATH=/opt/homebrew/Cellar/node/26.7.0/bin:/usr/bin:/bin /opt/homebrew/Cellar/node/26.7.0/bin/node node_modules/vitest/vitest.mjs run tests/typed-output.test.ts tests/validate.test.ts tests/spec-parity.test.ts tests/deterministic-llm.test.ts
+```
+
+Captured output:
+
+```text
+build_and_typecheck_exit=0
+
+ RUN  v2.1.9 /Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk
+
+ ✓ tests/deterministic-llm.test.ts (5 tests) 37ms
+ ✓ tests/validate.test.ts (36 tests) 37ms
+ ✓ tests/typed-output.test.ts (14 tests) 36ms
+ ✓ tests/spec-parity.test.ts (15 tests) 102ms
+
+ Test Files  4 passed (4)
+      Tests  70 passed (70)
+   Start at  17:26:56
+   Duration  1.18s (transform 373ms, setup 0ms, collect 982ms, tests 212ms, environment 2ms, prepare 915ms)
+```
+
+The host's `npm` launcher remained stuck without output when I first invoked the equivalent `npm run build && npm run typecheck:tests` chain, so I interrupted it after more than 60 seconds and do not represent that invocation as a pass. The evidence above invokes the exact underlying repository scripts through the installed Node binary and completed successfully. The repository's Linux artifact check also ran the npm script path successfully at this exact head:
+
+Literal command:
+
+```sh
+gh pr checks 133 --repo AgentWorkforce/flows
+```
+
+Captured output:
+
+```text
+CodeRabbit	pass	0		Review completed
+linux-x64-artifact	pass	3m6s	https://github.com/AgentWorkforce/flows/actions/runs/33646896877/job/100303802460
+```
+
+## Live journal behavior
+
+The focused live run used the explicitly named prebuilt daemon because the host cargo shim is known broken; no claim is made that this review rebuilt the kernel. The positive case asserts parsed JSON plus `{gate: 'json_schema', verdict: 'pass'}` in `step.completed`. The negative case asserts `output: null`, `{gate: 'json_schema', verdict: 'fail'}`, step `retries_exhausted`, and run `step_failed`.
+
+Literal command:
+
+```sh
+PATH=/opt/homebrew/Cellar/node/26.7.0/bin:/usr/bin:/bin RELAYFLOWD_BIN=/Users/khaliqgant/.relayflows-toolchain/target/1914866954/debug/relayflowd /opt/homebrew/Cellar/node/26.7.0/bin/node node_modules/vitest/vitest.mjs run tests/live-kernel.test.ts -t 'stub agent CLI|FAILS verification' --reporter=verbose
+```
+
+Captured output:
+
+```text
+ RUN  v2.1.9 /Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk
+
+stdout | tests/live-kernel.test.ts
+LIVE_KERNEL relayflowd=/Users/khaliqgant/.relayflows-toolchain/target/1914866954/debug/relayflowd
+LIVE_KERNEL flows=/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/dist/cli.js
+
+ ✓ tests/live-kernel.test.ts > built flows CLI against live relayflowd > runs hn-monitor analyze-story end-to-end via a stub agent CLI (gate 2 clause 2 demo)
+ ✓ tests/live-kernel.test.ts > built flows CLI against live relayflowd > hn-monitor analyze-story FAILS verification when the CLI omits required schema fields
+
+ Test Files  1 passed (1)
+      Tests  2 passed | 15 skipped (17)
+   Start at  17:24:06
+   Duration  1.88s (transform 262ms, setup 0ms, collect 394ms, tests 502ms, environment 0ms, prepare 230ms)
+```
+
+## Complete SDK compatibility run
+
+Literal command:
+
+```sh
+PATH=/opt/homebrew/Cellar/node/26.7.0/bin:$PATH RELAYFLOWD_BIN=/Users/khaliqgant/.relayflows-toolchain/target/1914866954/debug/relayflowd /opt/homebrew/Cellar/node/26.7.0/bin/node node_modules/vitest/vitest.mjs run
+```
+
+Captured output:
+
+```text
+ RUN  v2.1.9 /Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk
+
+ ✓ tests/preflight.test.ts (14 tests) 42ms
+SKIPPED_UNACTIONABLE=0
+ ✓ tests/backlog-picker.test.ts (14 tests) 255ms
+ ✓ tests/validate.test.ts (36 tests) 138ms
+SKIPPED_UNACTIONABLE=0
+ ✓ tests/journal-client.test.ts (13 tests) 270ms
+ ✓ tests/cli-hn-monitor.test.ts (16 tests) 191ms
+SKIPPED_UNACTIONABLE=0
+NO_ACTIONABLE_BACKLOG_ENTRY scanned=1 Scoped but unverifiable[missing_definition_of_done]
+ ✓ tests/work-package-consumer.test.ts (13 tests) 883ms
+   ✓ the Garden join: picker output feeds the consumer > accepts a package the picker actually emits, and refuses one lacking a definition of done 864ms
+NO_ACTIONABLE_BACKLOG_ENTRY scanned=0 
+ ✓ tests/typed-output.test.ts (14 tests) 145ms
+ ✓ tests/hn-poller.test.ts (6 tests) 65ms
+node:fs:539
+    return binding.readFileUtf8(path, stringToFlags(options.flag));
+                   ^
+
+Error: ENOENT: no such file or directory, open '.relayflow/backlog-picker-entry.json'
+    at Object.readFileSync (node:fs:539:20)
+    at [eval]:1:478
+    at runScriptInThisContext (node:internal/vm:219:10)
+    at node:internal/process/execution:483:12
+    at [eval]-wrapper:6:24
+    at runScriptInContext (node:internal/process/execution:481:60)
+    at evalFunction (node:internal/process/execution:315:30)
+    at evalTypeScript (node:internal/process/execution:327:3)
+    at node:internal/main/eval_string:71:3 {
+  errno: -2,
+  code: 'ENOENT',
+  syscall: 'open',
+  path: '.relayflow/backlog-picker-entry.json'
+}
+
+Node.js v26.7.0
+ ✓ tests/deterministic-llm.test.ts (5 tests) 167ms
+ ✓ tests/dir-watcher-poller.test.ts (6 tests) 41ms
+SKIPPED_UNACTIONABLE=0
+ ✓ tests/backlog-picker-flow.test.ts (6 tests) 2390ms
+   ✓ backlog-picker flow > emit-package describes the entry select-entry chose, even if the backlog changes between them 525ms
+   ✓ backlog-picker flow > does not emit a stale entry left by a previous run 1212ms
+   ✓ backlog-picker canonical spec > keeps directories and extensionless paths, and rejects prose 644ms
+ ✓ tests/hello-deterministic.test.ts (5 tests) 227ms
+ ✓ tests/work-package-validator.test.ts (7 tests) 65ms
+ ✓ tests/cli.test.ts (50 tests) 2891ms
+   ✓ flows check CLI > resolves a project CLI path relative to the flows.json that declares it 342ms
+ ✓ tests/parse-json-output.test.ts (7 tests) 39ms
+stdout | tests/live-kernel.test.ts
+LIVE_KERNEL relayflowd=/Users/khaliqgant/.relayflows-toolchain/target/1914866954/debug/relayflowd
+LIVE_KERNEL flows=/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/dist/cli.js
+
+ ✓ tests/spec-parity.test.ts (15 tests) 342ms
+ ✓ tests/bin.test.ts (7 tests) 3941ms
+   ✓ built flows binary > refuses through a symlink to the built artifact 697ms
+   ✓ built flows binary > refuses through a symlinked directory component 691ms
+   ✓ built flows binary > classifies a signal-terminated auth probe as probe_failed 874ms
+   ✓ built flows binary > classifies an unavailable PATH resolver as probe_failed 500ms
+   ✓ built flows binary > does not describe a present non-executable CLI as missing 423ms
+   ✓ built flows binary > runs one auth probe for three steps sharing a flow CLI 737ms
+stdout | tests/live-kernel.test.ts > built flows CLI against live relayflowd > hn-monitor analyze-story reaches done through the real Claude analyzer CLI
+LIVE_ANALYZER ready: claude -p --model claude-haiku-4-5-20251001 round-trip OK
+
+stdout | tests/live-kernel.test.ts > built flows CLI against live relayflowd > hn-monitor analyze-story reaches done through the real Claude analyzer CLI
+LIVE_ANALYZER analysis: {"reasoning":"This story directly describes an autonomous AI agent performing software development automation by independently opening and reviewing pull requests, which exemplifies practical AI agent capabilities in code review and CI/CD workflows.","relevance_score":9,"story_title":"Show HN: an agent that opens and reviews its own pull requests [wake-nonce-7f3a91c4]"}
+
+stdout | tests/live-kernel.test.ts > surface resume after a real daemon kill > resumes a three-step run with each successful completion exactly once
+LIVE_KERNEL kill -9 pid=55619 run=01M1HBJX8CJMNJRA2KGRMEF6MT while step=two state=Running
+
+ ✓ tests/live-kernel.test.ts (17 tests) 58934ms
+   ✓ built flows CLI against live relayflowd > runs rung (a), parks rung (b), and keeps JSON report-shaped 4505ms
+   ✓ built flows CLI against live relayflowd > allows a deterministic run to exceed the bounded request timeout 33265ms
+   ✓ built flows CLI against live relayflowd > follows a live worker dispatch through flows run 471ms
+   ✓ built flows CLI against live relayflowd > runs an agent CLI end to end through the SDK worker 374ms
+   ✓ built flows CLI against live relayflowd > can always get a parked run to a late-attaching worker 5564ms
+   ✓ built flows CLI against live relayflowd > reports a real manual-recovery NeedsHuman state as parked 426ms
+   ✓ built flows CLI against live relayflowd > hn-monitor analyze-story FAILS verification when the CLI omits required schema fields 537ms
+   ✓ built flows CLI against live relayflowd > AgentWorker exposes wake_context to the CLI via RELAYFLOW_WAKE_CONTEXT env var (real analyzer prerequisite) 314ms
+   ✓ built flows CLI against live relayflowd > hn-monitor analyze-story reaches done through the real Claude analyzer CLI 10834ms
+   ✓ built flows CLI against live relayflowd > preflights before journaling and names an unreachable socket 534ms
+   ✓ surface resume after a real daemon kill > resumes a three-step run with each successful completion exactly once 945ms
+
+ Test Files  18 passed (18)
+      Tests  251 passed (251)
+   Start at  17:24:16
+   Duration  60.58s (transform 1.58s, setup 0ms, collect 5.57s, tests 71.02s, environment 20ms, prepare 6.87s)
+```
+
+The ENOENT text above is output from an expected negative subprocess case; its owning tests and the overall suite exited successfully.
+
+## Scope accuracy for issue #132
+
+Literal command:
+
+```sh
+gh issue view 132 --repo AgentWorkforce/flows --json number,state,title,comments --jq '"number=\(.number)\nstate=\(.state)\ntitle=\(.title)\nlatest_comment=" + (.comments[-1].body // "<none>")'
+```
+
+Captured output:
+
+````text
+number=132
+state=OPEN
+title=v2 authoring ergonomics: close the gaps found in the research-flow / v1 / Smithers comparison
+latest_comment=Ownership/progress update (2026-09-02):
+
+- Typed-output slice: PR #133 is at exact head `5ab0fee366362de33ea1bd7d88a96daef1cad18a`. The repository artifact check is green; a fresh exact-head three-lens Agent Relay review is active. It remains unmerged until those final reports pass and the canonical review transcript gate is persisted.
+- Surface slice: PR #134 remains at `08d8133a61285428e1d04fd52ca1f57e2f771309` and is not mergeable. All three fresh reviews found the missing repository-owned package gate; history review also found that `flow()` discards the authored header/body and therefore cannot honestly be called executable. Agent Relay owner `flows-pr134-repair-r2` is repairing the CI/tarball boundary and either adding the smallest real journal bridge or narrowing the PR claim so this issue stays open.
+- Declared CLI/model slice: implementation ownership is now active under Agent Relay owner `flows-132-model-lint-r2`, isolated at `/Users/khaliqgant/AgentWorkforce/flows-132-model-wt` on `feat/v2-declared-model` from merged main. Acceptance is fail-closed model resolution: typos/unknown models are typed preflight refusals, not warnings or regex-only guesses; the chosen model lowers into the existing agent-step model field and is journaled.
+- Remaining issue slices still open: direct-run input, true research/sales surface migration, kernel parallel dispatch, data/code gate decision, and broad verb-specific optional-field lint.
+- Migration rail is unchanged: v1 remains supported and default; v2 is explicit and version-pinned. Deprecation is a later explicit gate after Cloud proof and workload-by-workload migration.
+
+Captured state:
+
+```text
+$ gh pr view 133 --repo AgentWorkforce/flows --json headRefOid,state,mergeStateStatus
+{"headRefOid":"5ab0fee366362de33ea1bd7d88a96daef1cad18a","mergeStateStatus":"CLEAN","state":"OPEN"}
+
+$ gh pr view 134 --repo AgentWorkforce/flows --json headRefOid,state,mergeStateStatus
+{"headRefOid":"08d8133a61285428e1d04fd52ca1f57e2f771309","mergeStateStatus":"CLEAN","state":"OPEN"}
+
+$ git -C /Users/khaliqgant/AgentWorkforce/flows-132-model-wt branch --show-current
+feat/v2-declared-model
+$ git -C /Users/khaliqgant/AgentWorkforce/flows-132-model-wt rev-parse HEAD
+a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2
+```
+````
+
+This review does not claim that PR #133 satisfies issue #132's full “done when.” It validates only the declared structured-output lowering slice at the requested head.
+
+REVIEW_PASSED
diff --git a/ops/reviews/20260902-1625-pr133-maintainability.md b/ops/reviews/20260902-1625-pr133-maintainability.md
new file mode 100644
index 000000000..10ce005e8
--- /dev/null
+++ b/ops/reviews/20260902-1625-pr133-maintainability.md
@@ -0,0 +1,412 @@
+# PR #133 — maintainability re-review (exact-head, post-fix)
+
+- **Verdict:** REVIEW_PASSED
+- **Reviewer lens:** maintainability, API/type honesty, fail-closed behaviour, test quality, documentation, prior-blocker repair
+- **Date:** 2026-09-02
+- **Head reviewed:** `5ab0fee366362de33ea1bd7d88a96daef1cad18a`
+- **Base:** `a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2`
+- **Worktree:** `/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt`
+- **Prior reviews at `81c49df`:** FAILED (maintainability), PASSED (history, structure)
+
+---
+
+## Environment disclosure
+
+Four `npm`-invoked subprocesses hung on this host and were terminated externally (exit 144):
+
+```
+npm run build          (background bulb52l28)
+npm run typecheck:tests (background b65vncakw)
+npm run typecheck:tests (background bchj1nyj0)
+npx tsc -p tsconfig.tests.json (background bdskhnlg2)
+```
+
+All evidence below was gathered via direct binaries with:
+`PATH=/Users/khaliqgant/.local/share/mise/installs/node/22.23.2/bin:/opt/homebrew/bin:/usr/bin:/bin`
+
+---
+
+## Exact-head verification
+
+Command:
+
+```sh
+pwd && git rev-parse HEAD && git log --oneline -5
+```
+
+Captured output:
+
+```text
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt
+5ab0fee366362de33ea1bd7d88a96daef1cad18a
+5ab0fee fix(sdk): enforce structured output contracts
+81c49df feat(sdk): compile typed outputs to json_schema
+a0d42ff feat(release): build verified Cloud v2 runtime artifact (#131)
+51415d9 feat(gate2): real Claude analyzer for hn-monitor, with a declared model (#130)
+7728565 feat(factory): concurrent Claude Code authoring driver over ops/factory/queue.md (#126)
+```
+
+Command:
+
+```sh
+git diff a0d42ffbdc7fb60b42c0b5bea4f58408249b08a2..5ab0fee366362de33ea1bd7d88a96daef1cad18a --stat
+```
+
+Captured output:
+
+```text
+ .github/workflows/cloud-runtime-artifact.yml |  15 +-
+ docs/SURFACE.md                              |  25 ++++
+ sdk/package.json                             |   3 +-
+ sdk/src/compile.ts                           |  19 ++-
+ sdk/src/index.ts                             |   1 +
+ sdk/src/output-schema.ts                     |  22 +++
+ sdk/src/spec.ts                              |  13 ++
+ sdk/src/validate.ts                          |   7 +-
+ sdk/tests/live-kernel.test.ts                |  36 ++++-
+ sdk/tests/typed-output.test.ts               | 209 +++++++++++++++++++++++++++
+ sdk/tsconfig.tests.json                      |  10 ++
+ testdata/hn-monitor.flow.yaml                |  33 ++---
+ 12 files changed, 365 insertions(+), 28 deletions(-))
+```
+
+Command:
+
+```sh
+git diff --quiet a0d42ff..5ab0fee -- kernel; printf 'kernel_diff_exit=%s\n' "$?"
+git diff --quiet a0d42ff..5ab0fee -- testdata/hn-monitor.spec.canonical.json; printf 'canonical_diff_exit=%s\n' "$?"
+```
+
+Captured output:
+
+```text
+kernel_diff_exit=0
+canonical_diff_exit=0
+```
+
+Both zero: no kernel file changed, canonical JSON fixture is unchanged.
+
+---
+
+## Prior blocking findings — repair assessment
+
+### F1 (BLOCKING at `81c49df`) — Type-level test cannot fail
+
+The prior review proved that `expectTypeOf` in `tests/typed-output.test.ts` was decoration
+because `tsc` never included test files (`tsconfig.json` explicitly excludes `tests/`).
+The fix commit adds `sdk/tsconfig.tests.json` and a `typecheck:tests` script, and wires
+both into `npm test` and the CI pipeline.
+
+**Repair verified.**
+
+Command:
+
+```sh
+./node_modules/.bin/tsc -p tsconfig.tests.json; echo "TYPECHECK_EXIT: $?"
+```
+
+Captured output:
+
+```text
+TYPECHECK_EXIT: 0
+```
+
+Command:
+
+```sh
+./node_modules/.bin/tsc -p tsconfig.tests.json --listFiles | grep tests
+```
+
+Captured output:
+
+```text
+/Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk/tests/typed-output.test.ts
+```
+
+The test file is now in the tsc program. TypeScript 5.x reports `error TS2578: Unused
+'@ts-expect-error' directive` when a directive suppresses nothing — so the directive at
+`typed-output.test.ts:41` (asserting no `output` field on `DeterministicStepSpec`) is
+a real enforcement gate, not a comment. Exit 0 means `DeterministicStepSpec` genuinely
+lacks `output`, and every `expectTypeOf` assertion is type-correct.
+
+### F2 (BLOCKING at `81c49df`) — Phantom types with no consumer (AGENTS.md rule 6)
+
+The prior review found `JsonOutputSchema<TOutput>` (a generic carrying a phantom TS type)
+and `OutputFromSchema` were exported with no consumer in `src/` and no enforced test —
+building for the next rung before this one was standing.
+
+**Repair verified.**
+
+Command:
+
+```sh
+grep -rn "OutputFromSchema\|TOutput" sdk/src/output-schema.ts sdk/src/spec.ts sdk/src/index.ts
+```
+
+Captured output:
+
+```text
+(no output)
+```
+
+`JsonOutputSchema` is now a plain alias `Record<string, unknown>`. `OutputFromSchema` and
+`TOutput` generics are absent from every source and export file. The package surface carries
+exactly what this slice needs.
+
+---
+
+## Non-blocking findings from prior review — repair assessment
+
+### F3 — One rule, two implementations, second path untested
+
+Prior finding: `toKernelVerification` duplicated `validateOutputDeclaration` with different
+message phrasing, and the direct `toKernelSpec` path was never tested.
+
+**Repair verified.**
+
+`sdk/src/compile.ts:402–410` now delegates to `validateOutputDeclaration`:
+
+```typescript
+function toKernelVerification(step: StepSpec): KernelVerificationSpec {
+  const output = step.type === 'deterministic' ? undefined : step.output;
+  if (output !== undefined) {
+    const errors = validateOutputDeclaration(step, `step "${step.id}"`);
+    if (errors.length > 0) throw new CompileError(errors);
+    return { json_schema: output };
+  }
+  ...
+}
+```
+
+The fix commit adds three tests that call `toKernelSpec` directly on a raw FlowSpec
+carrying `output` (two type arms), on a conflicting `output + verification`, and on a
+malformed `output`. All three use regex matchers that accommodate the `step "id"` prefix
+format the direct path produces.
+
+### F4 — `typedOutputVerification` silently dropped `verification`
+
+Prior finding: `typedOutputVerification` would silently discard `verification` if `output`
+was present, while `toKernelVerification` would throw — two opposing behaviours for the
+same input.
+
+**Repair verified.** `sdk/src/compile.ts:135–142` now calls `validateOutputDeclaration`
+before returning the verification gate and throws `CompileError` on any error. The
+fail-open path is gone.
+
+### F5 — `output:` sugar undocumented
+
+Prior finding: `docs/SURFACE.md` had no mention of the `output:` sugar.
+
+**Repair verified.** `docs/SURFACE.md` now contains:
+
+```text
+### Structured output declarations
+
+Declarative `llm` and `agent` steps may declare an `output` JSON Schema. This
+is authoring sugar for the existing kernel `json_schema` verification gate; the
+compiler removes `output` before the journal boundary and emits the schema as
+`verification.json_schema`. Authors must choose either `output` or an explicit
+`verification` block. Declaring both is ambiguous and fails closed.
+
+...
+
+The declaration does not add a kernel primitive and does not yet infer a
+TypeScript result type from arbitrary JSON Schema. Typed parsed values belong
+to the imperative `f.llm` / `f.agent` surface once that surface has a real
+consumer; the spec SDK does not publish an unchecked phantom type in advance.
+```
+
+The documentation is accurate and honest about what is not yet built.
+
+---
+
+## Independent test suite execution
+
+CI-specified deterministic test files (four files, direct binary):
+
+Command:
+
+```sh
+./node_modules/.bin/vitest run \
+  tests/typed-output.test.ts \
+  tests/validate.test.ts \
+  tests/spec-parity.test.ts \
+  tests/deterministic-llm.test.ts
+```
+
+Captured output:
+
+```text
+ RUN  v2.1.9 /Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk
+
+ ✓ tests/typed-output.test.ts (14 tests) 78ms
+ ✓ tests/deterministic-llm.test.ts (5 tests) 84ms
+ ✓ tests/validate.test.ts (36 tests) 86ms
+ ✓ tests/spec-parity.test.ts (15 tests) 173ms
+
+ Test Files  4 passed (4)
+      Tests  70 passed (70)
+   Start at  17:33:53
+   Duration  1.66s (transform 485ms, setup 0ms, collect 1.30s, tests 421ms, environment 2ms, prepare 1.15s)
+
+EXIT: 0
+```
+
+Verbose run of `typed-output.test.ts` naming all 14 tests:
+
+Command:
+
+```sh
+./node_modules/.bin/vitest run tests/typed-output.test.ts --reporter=verbose
+```
+
+Captured output:
+
+```text
+ RUN  v2.1.9 /Users/khaliqgant/AgentWorkforce/flows-132-typed-output-wt/sdk
+
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > typechecks output declarations on llm and agent steps only
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > compiles llm output sugar to the existing json_schema primitive
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > compiles agent output sugar to the existing json_schema primitive
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > accepts the output declaration in YAML
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > keeps the canonical hn-monitor kernel step unchanged
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed when output conflicts with verification (output_contains)
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed when output conflicts with verification (json_schema)
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a non-object output schema (null)
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a non-object output schema (%j)
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a non-object output schema ("not-a-schema")
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > compiles a raw llm spec directly through the public kernel boundary
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > compiles a raw agent spec directly through the public kernel boundary
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a conflicting raw spec at the public kernel boundary
+ ✓ tests/typed-output.test.ts > typed llm and agent outputs > fails closed on a malformed raw spec at the public kernel boundary
+
+ Test Files  1 passed (1)
+      Tests  14 passed (14)
+   Start at  17:35:31
+   Duration  1.30s (transform 179ms, setup 0ms, collect 261ms, tests 42ms, environment 0ms, prepare 232ms)
+
+EXIT: 0
+```
+
+Four tests added relative to the prior round (10 → 14): the type-declaration check
+(previously a vitest-runtime no-op, now type-enforced by tsc), and three direct
+`toKernelSpec` boundary tests.
+
+---
+
+## hn-monitor canonical roundtrip
+
+Command (direct Node, `dist/` rebuilt from prior session):
+
+```sh
+node -e "
+const { compileYamlToCanonicalJson } = require('./dist/compile.js');
+const { readFileSync } = require('fs');
+const yaml = readFileSync('../testdata/hn-monitor.flow.yaml', 'utf8');
+const compiled = JSON.parse(compileYamlToCanonicalJson(yaml));
+console.log(JSON.stringify(compiled.steps, null, 2));
+"
+```
+
+Captured output (abridged to verification field):
+
+```json
+[
+  {
+    "cli": "preflight/analyze-story-claude-cli",
+    "depends_on": [],
+    "id": "analyze-story",
+    "model": "claude-haiku-4-5-20251001",
+    "recovery_mode": "reset",
+    "type": "agent",
+    "verification": {
+      "json_schema": {
+        "properties": {
+          "reasoning": { "type": "string" },
+          "relevance_score": { "maximum": 10, "minimum": 1, "type": "integer" },
+          "story_title": { "type": "string" }
+        },
+        "required": ["story_title", "relevance_score", "reasoning"],
+        "type": "object"
+      }
+    }
+  }
+]
+```
+
+The canonical fixture (git diff exit 0) and the compiled YAML produce identical steps.
+`output:` in the authoring YAML compiles to `verification.json_schema` in the kernel
+dialect with no `output` key on the kernel step. ✓
+
+---
+
+## v1 compatibility and issue #132 scope
+
+The `toKernelVerification` fallthrough path (`gate = step.verification; ... return
+{ json_schema: gate.schema }`) is unchanged and handles all pre-existing explicit
+`verification:` declarations. The 15 spec-parity tests and 5 deterministic-llm tests
+cover v1 paths and all pass.
+
+Issue #132 remains OPEN. PR #133 uses `Refs #132`, the PR body says "issue #132 slice 1
+only" and explicitly disclaims `f.llm` / `f.agent` runtime API, `@relayflows/surface`,
+direct-run input, and parallel execution. `docs/SURFACE.md` echoes this scope boundary:
+"Typed parsed values belong to the imperative `f.llm` / `f.agent` surface once that
+surface has a real consumer; the spec SDK does not publish an unchecked phantom type in
+advance." Claims are scoped honestly.
+
+---
+
+## CI pipeline change
+
+`npm ci --prefix sdk` is moved before the kernel build and is run once. The new step:
+
+```yaml
+- name: Test SDK and type-level authoring contracts
+  working-directory: sdk
+  run: |
+    npm run build
+    npm run typecheck:tests
+    ./node_modules/.bin/vitest run \
+      tests/typed-output.test.ts \
+      tests/validate.test.ts \
+      tests/spec-parity.test.ts \
+      tests/deterministic-llm.test.ts
+```
+
+runs before the `bun build` artifact step. Tests now gate the release artifact. The
+ordering is correct; no duplication.
+
+---
+
+## Remaining pre-existing observation (not introduced by this PR)
+
+Error message prefix inconsistency: `validateOutputDeclaration` called from `validateSpec`
+uses `spec.steps[N]` as the location prefix; called from `toKernelVerification` it uses
+`step "id"`. The three new direct-boundary tests use regex patterns that match both forms.
+This is pre-existing vocabulary and not a blocker, but worth a follow-up issue.
+
+---
+
+## What I verified as correct (summary)
+
+| Claim | Result |
+|---|---|
+| Exact HEAD `5ab0fee366...` in correct worktree | ✓ |
+| Kernel files unchanged | ✓ (git diff exit 0) |
+| Canonical hn-monitor fixture unchanged | ✓ (git diff exit 0) |
+| `output:` YAML compiles to identical canonical JSON | ✓ |
+| Phantom types (`OutputFromSchema`, `TOutput`) removed | ✓ (grep: no output) |
+| `tsc -p tsconfig.tests.json` exits 0 | ✓ (TYPECHECK_EXIT: 0) |
+| Test files included in tsc program | ✓ (listFiles confirms) |
+| 4 CI test files pass (70/70) | ✓ |
+| 14 named typed-output tests pass | ✓ |
+| Direct `toKernelSpec` boundary pinned by tests | ✓ (3 new tests) |
+| `typedOutputVerification` now throws on conflict | ✓ (F4 fixed) |
+| `docs/SURFACE.md` documents `output:` sugar | ✓ (F5 fixed) |
+| v1 `verification:` unchanged | ✓ |
+| Issue #132 claims scoped honestly | ✓ |
+| No product code edited, committed, pushed, or merged | ✓ |
+
+---
+
+REVIEW_PASSED
diff --git a/sdk/package.json b/sdk/package.json
index 4532bf635..05763cb7a 100644
--- a/sdk/package.json
+++ b/sdk/package.json
@@ -23,8 +23,9 @@
     "demo:hn": "npm run build && node dist/demo-hn-monitor.js",
     "prepare": "npm run build",
     "typecheck": "tsc --noEmit",
+    "typecheck:tests": "tsc -p tsconfig.tests.json",
     "test:prep": "( cd ../kernel && sh ../ops/cargo.sh build ) && ( [ ! -d ../testdata/preflight ] || find ../testdata/preflight -name '*-cli' -type f -exec chmod +x {} + )",
-    "test": "npm run test:prep && npm run build && vitest run",
+    "test": "npm run test:prep && npm run build && npm run typecheck:tests && vitest run",
     "test:watch": "vitest"
   },
   "license": "UNLICENSED",
diff --git a/sdk/src/compile.ts b/sdk/src/compile.ts
index aeecde366..281a2b041 100644
--- a/sdk/src/compile.ts
+++ b/sdk/src/compile.ts
@@ -30,6 +30,7 @@ import type {
 } from './spec.js';
 import { SPEC_SCHEMA_VERSION } from './spec.js';
 import { canonicalize, specHash } from './canonical.js';
+import { validateOutputDeclaration } from './output-schema.js';
 import { validateSpec, type ValidationResult } from './validate.js';
 
 export class CompileError extends Error {
@@ -84,11 +85,12 @@ export function compileSpec(spec: unknown): FlowSpec {
 
 function compileStep(step: StepSpec): StepSpec {
   const maxIterations = step.maxIterations ?? 1;
+  const verification = typedOutputVerification(step);
   const base = {
     id: step.id,
     type: step.type,
     ...(step.dependsOn !== undefined ? { dependsOn: step.dependsOn } : {}),
-    ...(step.verification !== undefined ? { verification: step.verification } : {}),
+    ...(verification !== undefined ? { verification } : {}),
     maxIterations,
     ...(step.timeoutMs !== undefined ? { timeoutMs: step.timeoutMs } : {}),
   };
@@ -130,6 +132,15 @@ function compileStep(step: StepSpec): StepSpec {
   }
 }
 
+function typedOutputVerification(step: StepSpec): StepSpec['verification'] {
+  if (step.type !== 'deterministic' && step.output !== undefined) {
+    const errors = validateOutputDeclaration(step, `step "${step.id}"`);
+    if (errors.length > 0) throw new CompileError(errors);
+    return { type: 'json_schema', schema: step.output };
+  }
+  return step.verification;
+}
+
 // Kernel defaults, materialized at compile time so the emitted spec is
 // byte-identical to the kernel's own serialization of it (spec.rs defaults).
 const KERNEL_RETRY_DEFAULTS = {
@@ -389,6 +400,12 @@ function requireNoTimeout(step: StepSpec): void {
 }
 
 function toKernelVerification(step: StepSpec): KernelVerificationSpec {
+  const output = step.type === 'deterministic' ? undefined : step.output;
+  if (output !== undefined) {
+    const errors = validateOutputDeclaration(step, `step "${step.id}"`);
+    if (errors.length > 0) throw new CompileError(errors);
+    return { json_schema: output };
+  }
   const gate = step.verification;
   // No gate / explicit exit_code both compile to {}: exit_code == 0 is the
   // kernel's implicit gate for deterministic steps (kernel DESIGN.md §4).
diff --git a/sdk/src/index.ts b/sdk/src/index.ts
index f6ba94455..4c4a42fd9 100644
--- a/sdk/src/index.ts
+++ b/sdk/src/index.ts
@@ -11,6 +11,7 @@ export type {
   ExitCodeGate,
   FlowSpec,
   JsonSchemaGate,
+  JsonOutputSchema,
   KernelAgentStep,
   KernelAgentSurfaces,
   KernelBudgetSpec,
diff --git a/sdk/src/output-schema.ts b/sdk/src/output-schema.ts
new file mode 100644
index 000000000..bf448b13a
--- /dev/null
+++ b/sdk/src/output-schema.ts
@@ -0,0 +1,22 @@
+/** JSON Schema accepted by the `output` authoring declaration. */
+export type JsonOutputSchema = Record<string, unknown>;
+
+/** Validate authoring sugar before it can be compiled or submitted. */
+export function validateOutputDeclaration(
+  step: { output?: unknown; verification?: unknown },
+  at: string,
+): string[] {
+  if (step.output === undefined) return [];
+  const errors: string[] = [];
+  if (!isObject(step.output)) {
+    errors.push(`${at}.output: expected a JSON Schema object`);
+  }
+  if (step.verification !== undefined) {
+    errors.push(`${at}: output already declares json_schema verification; remove verification`);
+  }
+  return errors;
+}
+
+function isObject(value: unknown): value is Record<string, unknown> {
+  return typeof value === 'object' && value !== null && !Array.isArray(value);
+}
diff --git a/sdk/src/spec.ts b/sdk/src/spec.ts
index 613293efe..ce0fd9942 100644
--- a/sdk/src/spec.ts
+++ b/sdk/src/spec.ts
@@ -9,6 +9,9 @@
 // Zero-agent flows are legal: a spec with only `deterministic` (and/or `llm`)
 // steps is valid. Nothing here requires an `agent` step.
 
+import type { JsonOutputSchema } from './output-schema.js';
+export type { JsonOutputSchema } from './output-schema.js';
+
 /** The three rungs of the ladder (RFC §1; AGENTS.md rule 7). */
 export type StepType = 'deterministic' | 'llm' | 'agent';
 
@@ -125,6 +128,11 @@ export interface LlmStepSpec extends BaseStepSpec {
   model?: string;
   /** Inert preflight declaration; overrides the flow/project CLI default. */
   cli?: string;
+  /**
+   * Structured-output authoring sugar. Compiles to the existing `json_schema`
+   * verification primitive and is removed before the kernel boundary.
+   */
+  output?: JsonOutputSchema;
 }
 
 /**
@@ -149,6 +157,11 @@ export interface AgentStepSpec extends BaseStepSpec {
   surfaces?: AgentSurfaces;
   recoveryMode?: RecoveryMode;
   permissions?: PermissionsSpec;
+  /**
+   * Structured-output authoring sugar. A successful CLI JSON object is the parsed
+   * value; the kernel persists it only after `json_schema` verification.
+   */
+  output?: JsonOutputSchema;
 }
 
 export type StepSpec = DeterministicStepSpec | LlmStepSpec | AgentStepSpec;
diff --git a/sdk/src/validate.ts b/sdk/src/validate.ts
index dbe0ff9fb..a0bd73405 100644
--- a/sdk/src/validate.ts
+++ b/sdk/src/validate.ts
@@ -17,6 +17,7 @@ import type {
   VerificationSpec,
 } from './spec.js';
 import { SPEC_SCHEMA_VERSION } from './spec.js';
+import { validateOutputDeclaration } from './output-schema.js';
 
 export interface ValidationResult {
   ok: boolean;
@@ -46,8 +47,8 @@ const BUDGET_KEYS = ['maxTokensIn', 'maxTokensOut', 'maxDollars'] as const;
 const STEP_COMMON_KEYS = ['id', 'type', 'dependsOn', 'verification', 'maxIterations', 'timeoutMs'] as const;
 const STEP_TYPE_KEYS: Record<StepType, readonly string[]> = {
   deterministic: ['command'],
-  llm: ['prompt', 'model', 'cli'],
-  agent: ['instruction', 'cli', 'model', 'surfaces', 'recoveryMode', 'permissions'],
+  llm: ['prompt', 'model', 'cli', 'output'],
+  agent: ['instruction', 'cli', 'model', 'surfaces', 'recoveryMode', 'permissions', 'output'],
 };
 const VERIFICATION_KEYS: Record<string, readonly string[]> = {
   exit_code: ['type', 'expect'],
@@ -248,8 +249,10 @@ class Validator {
       this.validateDeterministic(st as unknown as DeterministicStepSpec, at);
     } else if (type === 'llm') {
       this.validateLlm(st as unknown as LlmStepSpec, at);
+      for (const error of validateOutputDeclaration(st, at)) this.fail(error);
     } else {
       this.validateAgent(st as unknown as AgentStepSpec, at);
+      for (const error of validateOutputDeclaration(st, at)) this.fail(error);
     }
   }
 
diff --git a/sdk/tests/live-kernel.test.ts b/sdk/tests/live-kernel.test.ts
index 5ae7a9b01..7f8cf1845 100644
--- a/sdk/tests/live-kernel.test.ts
+++ b/sdk/tests/live-kernel.test.ts
@@ -370,8 +370,8 @@ steps:
     await worker.attach();
 
     // Load hn-monitor's canonical spec, then patch the analyze-story
-    // step to declare a real CLI. Everything else — triggers, dedupe,
-    // wake context — comes straight from the existing gate-2 spec.
+    // step to declare a real CLI. The canonical fixture is separately pinned
+    // to the authoring YAML, whose `output:` compiles to this json_schema.
     const spec = JSON.parse(
       readFileSync(join(TESTDATA, 'hn-monitor.spec.canonical.json'), 'utf8'),
     ) as { steps: { id: string; cli?: string }[] };
@@ -400,6 +400,18 @@ steps:
       state: 'done',
     });
     const finalEntries = (await client.journalRead(runId)).entries;
+    const stepCompleted = finalEntries.find(
+      (entry) => (entry as { entry_type: string; step_id?: string }).entry_type === 'step.completed'
+        && (entry as { step_id?: string }).step_id === 'analyze-story',
+    ) as { payload: { output: unknown; verification: unknown } } | undefined;
+    expect(stepCompleted?.payload).toMatchObject({
+      output: {
+        story_title: 'stub',
+        relevance_score: 5,
+        reasoning: 'stub agent runtime — deterministic output for gate-2 clause-2 demo',
+      },
+      verification: { gate: 'json_schema', verdict: 'pass' },
+    });
     const runCompleted = finalEntries.find(
       (entry) => (entry as { entry_type: string }).entry_type === 'run.completed',
     ) as { payload: { completionReason: string } } | undefined;
@@ -464,8 +476,15 @@ steps:
     // a "hasn't completed yet" absence.
     const deadline = Date.now() + 10_000;
     let runCompleted: { payload: { completionReason: string } } | undefined;
+    let stepCompleted: {
+      payload: { completionReason: string; output: unknown; verification: unknown };
+    } | undefined;
     while (Date.now() < deadline) {
       const entries = (await client.journalRead(runId)).entries;
+      stepCompleted = entries.find(
+        (entry) => (entry as { entry_type: string; step_id?: string }).entry_type === 'step.completed'
+          && (entry as { step_id?: string }).step_id === 'analyze-story',
+      ) as typeof stepCompleted;
       runCompleted = entries.find(
         (entry) => (entry as { entry_type: string }).entry_type === 'run.completed',
       ) as { payload: { completionReason: string } } | undefined;
@@ -476,10 +495,15 @@ steps:
       runCompleted,
       'run.completed entry never arrived within 10s — test cannot assert schema-live under a hung run',
     ).toBeDefined();
-    // step_failed is the outer reason (a step failed → run failed);
-    // the inner step.completed record carries verification_failed
-    // for schema-rejected outputs. Pinning the outer reason avoids
-    // depending on retry/backoff behavior for this test.
+    // step_failed is the outer reason (a step failed → run failed). With one
+    // allowed semantic execution, the step records retries_exhausted while
+    // its verification record names the json_schema rejection. The rejected
+    // parsed value is nulled before the completion is persisted.
+    expect(stepCompleted?.payload).toMatchObject({
+      completionReason: 'retries_exhausted',
+      output: null,
+      verification: { gate: 'json_schema', verdict: 'fail' },
+    });
     expect(runCompleted!.payload.completionReason).toBe('step_failed');
 
     await worker.close();
diff --git a/sdk/tests/typed-output.test.ts b/sdk/tests/typed-output.test.ts
new file mode 100644
index 000000000..91c73fd0b
--- /dev/null
+++ b/sdk/tests/typed-output.test.ts
@@ -0,0 +1,209 @@
+import { readFileSync } from 'node:fs';
+import { dirname, join } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { describe, expect, expectTypeOf, it } from 'vitest';
+import {
+  compileSpec,
+  compileYaml,
+  compileYamlToCanonicalJson,
+  CompileError,
+  toKernelSpec,
+} from '../src/compile.js';
+import type {
+  AgentStepSpec,
+  DeterministicStepSpec,
+  JsonOutputSchema,
+  LlmStepSpec,
+} from '../src/spec.js';
+
+const TESTDATA = join(dirname(fileURLToPath(import.meta.url)), '..', '..', 'testdata');
+
+const extractionSchema: JsonOutputSchema = {
+  type: 'object',
+  additionalProperties: false,
+  required: ['actionable', 'request'],
+  properties: {
+    actionable: { type: 'boolean' },
+    request: { type: 'string' },
+  },
+};
+
+describe('typed llm and agent outputs', () => {
+  it('typechecks output declarations on llm and agent steps only', () => {
+    expectTypeOf<LlmStepSpec['output']>().toEqualTypeOf<JsonOutputSchema | undefined>();
+    expectTypeOf<AgentStepSpec['output']>().toEqualTypeOf<JsonOutputSchema | undefined>();
+
+    const deterministic: DeterministicStepSpec = {
+      id: 'build',
+      type: 'deterministic',
+      command: 'true',
+      // @ts-expect-error structured output sugar is llm/agent-only.
+      output: extractionSchema,
+    };
+    void deterministic;
+  });
+
+  it.each(['llm', 'agent'] as const)(
+    'compiles %s output sugar to the existing json_schema primitive',
+    (type) => {
+      const source = {
+        version: '0.1.0',
+        steps: [{
+          id: 'typed',
+          type,
+          ...(type === 'llm' ? { prompt: 'Return JSON.' } : { instruction: 'Return JSON.' }),
+          output: extractionSchema,
+        }],
+      };
+
+      const compiled = compileSpec(source);
+      expect(compiled.steps[0]).not.toHaveProperty('output');
+      expect(compiled.steps[0]?.verification).toEqual({
+        type: 'json_schema',
+        schema: extractionSchema,
+      });
+      expect(toKernelSpec(compiled).steps[0]?.verification).toEqual({
+        json_schema: extractionSchema,
+      });
+    },
+  );
+
+  it('accepts the output declaration in YAML', () => {
+    const compiled = compileYaml(`
+version: '0.1.0'
+steps:
+  - id: extract
+    type: llm
+    prompt: Return JSON.
+    output:
+      type: object
+      required: [answer]
+      properties:
+        answer: { type: number }
+`);
+
+    expect(compiled.steps[0]?.verification).toEqual({
+      type: 'json_schema',
+      schema: {
+        type: 'object',
+        required: ['answer'],
+        properties: { answer: { type: 'number' } },
+      },
+    });
+  });
+
+  it('keeps the canonical hn-monitor kernel step unchanged', () => {
+    const yaml = readFileSync(join(TESTDATA, 'hn-monitor.flow.yaml'), 'utf8');
+    const compiled = JSON.parse(compileYamlToCanonicalJson(yaml)) as { steps: unknown[] };
+    const canonical = JSON.parse(
+      readFileSync(join(TESTDATA, 'hn-monitor.spec.canonical.json'), 'utf8'),
+    ) as { steps: unknown[] };
+
+    // Trigger key normalization is a separate pre-existing surface gap; this
+    // assertion is deliberately about the step boundary changed in this PR.
+    expect(compiled.steps).toEqual(canonical.steps);
+  });
+
+  it.each([
+    { type: 'output_contains', value: 'done' },
+    { type: 'json_schema', schema: extractionSchema },
+  ])('fails closed when output conflicts with verification ($type)', (verification) => {
+    expect(() => compileSpec({
+      version: '0.1.0',
+      steps: [{
+        id: 'ambiguous',
+        type: 'agent',
+        instruction: 'Return JSON.',
+        output: extractionSchema,
+        verification,
+      }],
+    })).toThrowError(CompileError);
+
+    try {
+      compileSpec({
+        version: '0.1.0',
+        steps: [{
+          id: 'ambiguous',
+          type: 'agent',
+          instruction: 'Return JSON.',
+          output: extractionSchema,
+          verification,
+        }],
+      });
+    } catch (error) {
+      expect((error as CompileError).errors).toContain(
+        'spec.steps[0]: output already declares json_schema verification; remove verification',
+      );
+    }
+  });
+
+  it.each([null, [], 'not-a-schema'])(
+    'fails closed on a non-object output schema (%j)',
+    (output) => {
+      const result = (() => {
+        try {
+          compileSpec({
+            version: '0.1.0',
+            steps: [{ id: 'bad', type: 'llm', prompt: 'Return JSON.', output }],
+          });
+          return null;
+        } catch (error) {
+          return error as CompileError;
+        }
+      })();
+
+      expect(result).toBeInstanceOf(CompileError);
+      expect(result?.errors).toContain('spec.steps[0].output: expected a JSON Schema object');
+    },
+  );
+
+  it.each(['llm', 'agent'] as const)(
+    'compiles a raw %s spec directly through the public kernel boundary',
+    (type) => {
+      const step: LlmStepSpec | AgentStepSpec = type === 'llm'
+        ? {
+            id: 'direct',
+            type: 'llm',
+            prompt: 'Return JSON.',
+            output: extractionSchema,
+          }
+        : {
+            id: 'direct',
+            type: 'agent',
+            instruction: 'Return JSON.',
+            output: extractionSchema,
+          };
+      const kernel = toKernelSpec({
+        version: '0.1.0',
+        steps: [step],
+      });
+
+      expect(kernel.steps[0]?.verification).toEqual({ json_schema: extractionSchema });
+    },
+  );
+
+  it('fails closed on a conflicting raw spec at the public kernel boundary', () => {
+    expect(() => toKernelSpec({
+      version: '0.1.0',
+      steps: [{
+        id: 'direct-conflict',
+        type: 'agent',
+        instruction: 'Return JSON.',
+        output: extractionSchema,
+        verification: { type: 'output_contains', value: 'done' },
+      }],
+    })).toThrowError(/output already declares json_schema verification/);
+  });
+
+  it('fails closed on a malformed raw spec at the public kernel boundary', () => {
+    expect(() => toKernelSpec({
+      version: '0.1.0',
+      steps: [{
+        id: 'direct-malformed',
+        type: 'llm',
+        prompt: 'Return JSON.',
+        output: null,
+      } as unknown as Parameters<typeof toKernelSpec>[0]['steps'][number]],
+    })).toThrowError(/expected a JSON Schema object/);
+  });
+});
diff --git a/sdk/tsconfig.tests.json b/sdk/tsconfig.tests.json
new file mode 100644
index 000000000..2af505b54
--- /dev/null
+++ b/sdk/tsconfig.tests.json
@@ -0,0 +1,10 @@
+{
+  "extends": "./tsconfig.json",
+  "compilerOptions": {
+    "noEmit": true,
+    "rootDir": ".",
+    "types": ["node", "vitest"]
+  },
+  "include": ["src/**/*.ts", "tests/typed-output.test.ts"],
+  "exclude": ["node_modules", "dist"]
+}
diff --git a/testdata/hn-monitor.flow.yaml b/testdata/hn-monitor.flow.yaml
index ec7d9003d..37e73cf21 100644
--- a/testdata/hn-monitor.flow.yaml
+++ b/testdata/hn-monitor.flow.yaml
@@ -24,20 +24,19 @@ steps:
       is relevant to AI agents/automation. Output a JSON summary with: story
       title, relevance score (1-10), and reasoning.
     recoveryMode: reset
-    verification:
-      type: json_schema
-      schema:
-        type: object
-        required:
-          - story_title
-          - relevance_score
-          - reasoning
-        properties:
-          story_title:
-            type: string
-          relevance_score:
-            type: integer
-            minimum: 1
-            maximum: 10
-          reasoning:
-            type: string
+    # Surface sugar: compiles to the kernel's existing json_schema gate.
+    output:
+      type: object
+      required:
+        - story_title
+        - relevance_score
+        - reasoning
+      properties:
+        story_title:
+          type: string
+        relevance_score:
+          type: integer
+          minimum: 1
+          maximum: 10
+        reasoning:
+          type: string