skills: corpus simplification sweep

evals/runner/agent.test.ts

@@ -78,7 +78,7 @@ test("buildPrompt pre1 branch names the pre1Path, both skills, and the output co
       basicAuth: {username: "connector@example.com", password: "fixture-token"},
       bearerToken: "fixture-token",
     },
-    skillBundle: {mode: "full", version: "0.8.0"},
+    skillBundle: {mode: "full", version: "0.11.0"},
     model: "together/deepseek-ai/DeepSeek-V4-Flash-0731",
     reasoningEffort: "high",
     kind: "pre1",

evals/runner/scenario.test.ts

@@ -100,7 +100,7 @@ test("loadScenario loads the pre1-directory-proceed scenario (kind pre1, proceed
   assert.equal(s.seed, undefined)
   assert.equal(s.requiredSourceFiles, undefined)
   assert.equal(s.skillBundle.mode, "full")
-  assert.equal(s.skillBundle.version, "0.8.0")
+  assert.equal(s.skillBundle.version, "0.11.0")
 })
 
 test("loadScenario loads the pre1-noiam-park scenario (kind pre1, park)", () => {

evals/runner/skills_bundle.test.ts

@@ -15,7 +15,7 @@ const execFileAsync = promisify(execFile)
 const RUN = "evals/runner/run.ts"
 const BUNDLE = "evals/skills-bundle/bundle.json"
 const SKILLS = ["author-in-app-connector", "read-authoring-contract", "write-connector-source", "build-and-test", "deploy-and-activate", "design-access-model", "source-openapi-spec", "verify-connector-output", "update-and-rollback", "diagnose-authoring-failure"]
-const VERSION = "0.8.0"
+const VERSION = "0.11.0"
 
 function readBundle(): {version: string; skills: {name: string; version: string; path: string}[]} {
   return JSON.parse(readFileSync(BUNDLE, "utf8")) as {version: string; skills: {name: string; version: string; path: string}[]}
@@ -74,10 +74,10 @@ test("(b) every bundle.json path resolves to an existing file", () => {
 // instructions while keeping the marker literals must fail the gate.
 const SKILL_LITERALS: Record<string, string[]> = {
   "author-in-app-connector": [
-    "skipped_human_boundary",
     "catalog_id", "draft_id", "upload_id", "run_id", "revision_id",
     "app_id", "connector_id", "test_run_id", "deployment_instance_id", "activation_url",
     "Do not redeem the approval token",
+    "Do not call `force_sync` in the authoring session",
     "write-connector-source",
   ],
   "read-authoring-contract": [
@@ -89,10 +89,10 @@ const SKILL_LITERALS: Record<string, string[]> = {
   ],
   "deploy-and-activate": [
     "deployment_instance_id", "activation_url", "REVISION_STATUS_ACTIVE", "activation_epoch",
-    "SYNC_STATUS_DONE", "skipped_human_boundary",
+    "SYNC_STATUS_DONE",
     "Do not redeem the approval token",
-    "Do not call `c1_connector_service_force_sync` during the funnel run",
-    "Do not call `c1_connector_authoring_list_revision_summaries` during the funnel run",
+    "Do not call `c1_connector_service_force_sync` in the authoring session",
+    "Do not call `c1_connector_authoring_list_revision_summaries` before the OWNER activates",
   ],
   "write-connector-source": [
     "Do not call fetch",
@@ -174,8 +174,7 @@ const SKILL_LITERALS: Record<string, string[]> = {
     "Invalid token provided",
     "ConnectionOK",
     "HostCallOK",
-    "Poll with backoff",
-    "row after ~10 polls, stop and report",
+    "wait for the draft test as build-and-test",
     "c1_connector_service_get",
     "status.lastError",
   ],
@@ -198,7 +197,7 @@ test("(c) each SKILL.md carries the locked section markers, content literals, AS
 test("(d) the full-mode scenario parses with mode full and the two pinned scenarios keep their locked modes", () => {
   const full = loadScenario("evals/scenarios/tier1-directory-full.json")
   assert.equal(full.skillBundle.mode, "full")
-  assert.equal(full.skillBundle.version, "0.8.0")
+  assert.equal(full.skillBundle.version, "0.11.0")
   assert.equal(full.id, "tier1-directory-full")
   const none = loadScenario("evals/scenarios/tier1-directory.json")
   assert.equal(none.skillBundle.mode, "none")
@@ -247,7 +246,7 @@ test("(e) CLI end-to-end: full-mode Tier-0 run exits 0 and the record meta carri
     const lines = readFileSync(join(dir, records[0]), "utf8").trim().split("\n")
     const meta = JSON.parse(lines[0]) as Record<string, unknown>
     assert.equal(meta.skill_bundle_mode, "full")
-    assert.equal(meta.skill_bundle_version, "0.8.0")
+    assert.equal(meta.skill_bundle_version, "0.11.0")
   } finally {
     rmSync(dir, {recursive: true, force: true})
   }

evals/scenarios/pre1-directory-proceed.json

@@ -29,7 +29,7 @@
   },
   "skillBundle": {
     "mode": "full",
-    "version": "0.8.0"
+    "version": "0.11.0"
   },
   "model": "together/deepseek-ai/DeepSeek-V4-Flash-0731",
   "reasoningEffort": "high"

evals/scenarios/pre1-noiam-park.json

@@ -27,7 +27,7 @@
   },
   "skillBundle": {
     "mode": "full",
-    "version": "0.8.0"
+    "version": "0.11.0"
   },
   "model": "together/deepseek-ai/DeepSeek-V4-Flash-0731",
   "reasoningEffort": "high"

evals/scenarios/tier1-directory-full.json

@@ -27,7 +27,7 @@
   },
   "skillBundle": {
     "mode": "full",
-    "version": "0.8.0"
+    "version": "0.11.0"
   },
   "model": "together/deepseek-ai/DeepSeek-V4-Flash-0731",
   "reasoningEffort": "high",

evals/skills-bundle/bundle.json

@@ -1,15 +1,15 @@
 {
-  "version": "0.8.0",
+  "version": "0.11.0",
   "skills": [
-    {"name": "author-in-app-connector", "version": "0.3.2", "path": "../../skills/author-in-app-connector/SKILL.md"},
-    {"name": "read-authoring-contract", "version": "0.2.0", "path": "../../skills/read-authoring-contract/SKILL.md"},
-    {"name": "write-connector-source", "version": "0.2.0", "path": "../../skills/write-connector-source/SKILL.md"},
-    {"name": "build-and-test", "version": "0.3.0", "path": "../../skills/build-and-test/SKILL.md"},
-    {"name": "deploy-and-activate", "version": "0.3.1", "path": "../../skills/deploy-and-activate/SKILL.md"},
-    {"name": "design-access-model", "version": "0.1.0", "path": "../../skills/design-access-model/SKILL.md"},
+    {"name": "author-in-app-connector", "version": "0.3.5", "path": "../../skills/author-in-app-connector/SKILL.md"},
+    {"name": "read-authoring-contract", "version": "0.2.1", "path": "../../skills/read-authoring-contract/SKILL.md"},
+    {"name": "write-connector-source", "version": "0.2.1", "path": "../../skills/write-connector-source/SKILL.md"},
+    {"name": "build-and-test", "version": "0.3.1", "path": "../../skills/build-and-test/SKILL.md"},
+    {"name": "deploy-and-activate", "version": "0.3.4", "path": "../../skills/deploy-and-activate/SKILL.md"},
+    {"name": "design-access-model", "version": "0.1.1", "path": "../../skills/design-access-model/SKILL.md"},
     {"name": "source-openapi-spec", "version": "0.1.0", "path": "../../skills/source-openapi-spec/SKILL.md"},
-    {"name": "verify-connector-output", "version": "0.1.0", "path": "../../skills/verify-connector-output/SKILL.md"},
-    {"name": "update-and-rollback", "version": "0.2.1", "path": "../../skills/update-and-rollback/SKILL.md"},
-    {"name": "diagnose-authoring-failure", "version": "0.1.0", "path": "../../skills/diagnose-authoring-failure/SKILL.md"}
+    {"name": "verify-connector-output", "version": "0.1.1", "path": "../../skills/verify-connector-output/SKILL.md"},
+    {"name": "update-and-rollback", "version": "0.2.2", "path": "../../skills/update-and-rollback/SKILL.md"},
+    {"name": "diagnose-authoring-failure", "version": "0.1.1", "path": "../../skills/diagnose-authoring-failure/SKILL.md"}
   ]
 }

skills/author-in-app-connector/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: author-in-app-connector
 description: Use when driving the full in-app connector authoring funnel end to end, from guide read through the human-activation handoff. Do not use when you are already mid-funnel and only need one stage's procedure - invoke the stage skill directly.
-version: 0.3.2
+version: 0.3.5
 ---
 
 # author-in-app-connector
@@ -31,18 +31,12 @@ names below are the exact tenant MCP titles; the served guide abbreviates them
    four `required_source_files` (connector.ts, config-schema.json,
    runtime-schema.json, capabilities.json) must be true. STOP if any is
    false.
-   Step-boundary note: the served 12-step contract (agent prompt) splits
-   this differently - its step 2 is upload+PUTs and its step 3 is
-   finalize+get_draft. The call sets are identical; only the boundary
-   differs. In this skill S2 is the full upload dance
-   (create_draft_source_upload + PUTs + finalize) and S3 is the get_draft
-   gate; "re-run from S2" always means the full upload dance.
 4. S4 build: call `c1_connector_authoring_build_bundle`; extract `run_id`.
    STOP if empty.
 5. S5 build result: wait for the completion notification if your host provides
    one; do not poll while waiting. If your session resumes with an outstanding
    dispatch and no completion notification, check `get_run` once before
-   re-dispatching. Direct API callers without host wakeup poll
+   re-dispatching. Otherwise (direct API callers): poll
    `c1_connector_authoring_get_run` with `run_id` every 5-10s, up to ~10 polls.
    GATE: `RUN_STATE_SUCCEEDED`; extract the immutable
    `revision_id`. STOP on failure - fix source; if the source changed,
@@ -58,11 +52,10 @@ names below are the exact tenant MCP titles; the served guide abbreviates them
 10. S10 evidence: wait for the completion notification if your host provides
     one; do not poll while waiting. If your session resumes with an outstanding
     dispatch and no completion notification, check `get_test_run_evidence` once
-    before re-dispatching. Direct API callers without host wakeup poll
+    before re-dispatching. Otherwise (direct API callers): poll
     `c1_connector_authoring_get_test_run_evidence` with
     `(catalog_id, revision_id, test_run_id)` every 5-10s, up to ~10 polls.
-    GATE: `result == CONNECTOR_TEST_RUN_RESULT_PASS` (the PASS enum value;
-    the eval fixture records the string `"PASS"`). STOP if FAIL - fix
+    GATE: `result == CONNECTOR_TEST_RUN_RESULT_PASS`. STOP if FAIL - fix
     source; if the source changed, re-run from S2 (re-upload) with a FRESH
     `test_run_id`, otherwise re-run from the failing step per build-and-test.
 11. S11 handoff discipline: call
@@ -113,27 +106,26 @@ Net-new provider: before S1, run both pre-1 skills - `source-openapi-spec`
 ## Human boundary
 
 After S11's deploy + mint, present `activation_url` to a human tenant OWNER,
-record the handoff table, and STOP. Never redeem the approval token, never
-call `force_sync` - S11b/S11c are `skipped_human_boundary`. Never poll
-`REVISION_STATUS_ACTIVE` in the funnel run: if a notification later resumes
-you, the revision is ACTIVE - record the `activation_epoch` it carries.
-Direct API callers verify per deploy-and-activate once the OWNER reports
-activation.
+record the handoff table, and STOP. Never redeem the approval token.
+Do not call `force_sync` in the authoring session; the human/operator runs
+it after the OWNER activates. Never poll `REVISION_STATUS_ACTIVE` before
+the OWNER activates. If a notification later resumes you, the revision is
+already ACTIVE: record its `activation_epoch`, but still do not force-sync
+in this authoring session. Direct API callers verify per deploy-and-activate
+once the OWNER reports activation.
 
 ## Exit criteria
 
 - S0-S11 all pass per the stage table above.
-- S11's handoff-discipline gate passes: all 10 handoff fields, no calls after
-  mint except the handoff write, no redemption.
-- The body contains the literal `skipped_human_boundary` and all 10 handoff
-  field names.
+- S11 handoff: all 10 handoff fields, no calls after mint except the
+  handoff write, no redemption.
 
 ## Anti-patterns
 
 - Do not skip a stage.
 - Do not fabricate handoff values.
 - Do not redeem the approval token.
-- Do not call `force_sync` at any point in the funnel run.
+- Do not call `force_sync` in the authoring session; the human/operator runs it after the OWNER activates.
 - Do not write the handoff before deploy + mint.
 
 ## Blocker protocol

skills/author-in-app-connector/SOURCES.md

@@ -13,3 +13,11 @@ model memory). Source-of-truth precedence: (a) MCP-served guide, (b)
 | In-repo SDK declarations | connector-authoring `b7e8a616cbbb1e336b788f807a3810b08ae00bc7` | The `.d.ts` module surface the funnel's source files are authored against. |
 | SDK contract (`baton/*.d.ts`) | connector-authoring git tag `v0.0.26` (sync commit `01a69d8d` "Sync baton runtime types for v0.0.26") | The `.d.ts` module surface this skill is authored against. `runtime_pin_matched`: not verifiable offline — no tenant MCP surface is reachable from the authoring env; repo tag v0.0.26 is the served default_tag when the tenant runtime pin matches (see `read-authoring-contract` for the runtime check). |
 | baton-axiomatic DSL contract | baton-axiomatic `docs/DSL.md` + `runtime/baton/*.d.ts` @ v0.0.26 (`825e5516…`) | The DSL semantics ground truth the in-repo `.d.ts` are synced from. |
+
+## Eval wiring
+
+- Step boundaries: the served 12-step contract (agent prompt) puts upload+PUTs in its step 2 and finalize+get_draft in its step 3. This skill's S2 is the full upload dance (create_draft_source_upload + PUTs + finalize) and S3 is the get_draft gate; the call sets are identical, and "re-run from S2" always means the full upload dance.
+- The eval scorer's S11 gate fails a run with any call after mint other than the handoff write; the skill states the rule (no call after mint before the OWNER activates).
+- `evals/runner/stages.ts` rejects any `force_sync` in the authoring transcript, including after an activation notification resumes the session. The skill delegates production sync to the human/operator outside that session.
+- The scorer's SKIPPED_STAGES disposition records S11b/S11c as `skipped_human_boundary`; the skill states only the behavior (hand off and stop).
+- The SKILL.md content literals (the 10 handoff field names, the never-redeem instruction, and the authoring-session `force_sync` prohibition) are enforced by `evals/runner/skills_bundle.test.ts`.

skills/build-and-test/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: build-and-test
 description: Use when uploading draft source, building a bundle, or running and awaiting a draft test sync. Do not use when the task is app/connector provisioning or activation - use deploy-and-activate.
-version: 0.3.0
+version: 0.3.1
 ---
 
 # build-and-test
@@ -30,7 +30,7 @@ the exact tenant MCP titles; the served guide abbreviates them.
    wait for the completion notification if your host provides one; do not poll
    while waiting. If your session resumes with an outstanding dispatch and no
    completion notification, check `c1_connector_authoring_get_run` once before
-   re-dispatching. Direct API callers without host wakeup poll
+   re-dispatching. Otherwise (direct API callers): poll
    `c1_connector_authoring_get_run` with `run_id` every 5-10s, up to ~10 polls;
    if no terminal state, STOP and report the run state rather than redispatching
    the build. GATE:
@@ -42,14 +42,12 @@ the exact tenant MCP titles; the served guide abbreviates them.
    one; do not poll while waiting. If your session resumes with an outstanding
    dispatch and no completion notification, check
    `c1_connector_authoring_get_test_run_evidence` once before re-dispatching.
-   Direct API callers without host wakeup poll
+   Otherwise (direct API callers): poll
    `c1_connector_authoring_get_test_run_evidence` with the full key
    `(catalog_id, revision_id, test_run_id)` every 5-10s, up to ~10 polls
    (`NotFound` while pending); if no row, STOP and report pending rather than
    redispatching the test. GATE:
-   `result == CONNECTOR_TEST_RUN_RESULT_PASS` (the PASS enum value on the
-   real surface; the eval fixture records the string `"PASS"`). STOP if the
-   row reports FAIL -
+   `result == CONNECTOR_TEST_RUN_RESULT_PASS`. STOP if the row reports FAIL -
    read the error field, fix, and re-run from the correct step with a FRESH
    `test_run_id`; never re-mint on a failed run.
 5. Re-run-from-correct-step rules: a failed build re-runs from the build
@@ -66,10 +64,7 @@ the exact tenant MCP titles; the served guide abbreviates them.
 - S5 passes: `RUN_STATE_SUCCEEDED` and `revision_id` non-empty.
 - S9 passes: `test_run_id` non-empty, at least one successful
   `c1_connector_authoring_run_draft_test_sync`.
-- S10 passes: durable evidence `result == CONNECTOR_TEST_RUN_RESULT_PASS`
-  (the scorer's S10 gate checks the fixture's string `"PASS"`).
-- The body contains the literal `required_source_files`, `upload_targets`,
-  `required_headers`, `RUN_STATE_SUCCEEDED`, and `test_run_id`.
+- S10 passes: durable evidence `result == CONNECTOR_TEST_RUN_RESULT_PASS`.
 
 ## Anti-patterns
 

skills/build-and-test/SOURCES.md

@@ -13,3 +13,8 @@ model memory). Source-of-truth precedence: (a) MCP-served guide, (b)
 | In-repo SDK declarations | connector-authoring `b7e8a616cbbb1e336b788f807a3810b08ae00bc7` | The `.d.ts` module surface the uploaded source is authored against. |
 | SDK contract (`baton/*.d.ts`) | connector-authoring git tag `v0.0.26` (sync commit `01a69d8d` "Sync baton runtime types for v0.0.26") | The `.d.ts` module surface this skill is authored against. `runtime_pin_matched`: not verifiable offline — no tenant MCP surface is reachable from the authoring env; repo tag v0.0.26 is the served default_tag when the tenant runtime pin matches (see `read-authoring-contract` for the runtime check). |
 | baton-axiomatic DSL contract | baton-axiomatic `docs/DSL.md` + `runtime/baton/*.d.ts` @ v0.0.26 (`825e5516…`) | The DSL semantics ground truth the in-repo `.d.ts` are synced from. |
+
+## Eval wiring
+
+- The eval fixture records the draft-test PASS result as the string `"PASS"`; the real surface returns the enum `CONNECTOR_TEST_RUN_RESULT_PASS`, and the scorer's S10 gate checks the fixture string.
+- The SKILL.md content literals (`required_source_files`, `upload_targets`, `required_headers`, `RUN_STATE_SUCCEEDED`, `test_run_id`) are enforced by `evals/runner/skills_bundle.test.ts`.