rickylabs · rickylabs · Aug 3, 2026 · Aug 3, 2026 · Aug 3, 2026 · Aug 3, 2026
@@ -0,0 +1,40 @@
+# 2026-08-04 — OpenAPI→MCP Wave-0 proofs
+
+Run dir: `.llm/runs/test-openapi-mcp-wave0-proofs--wave0/` · Branch `test/openapi-mcp-wave0-proofs`
+· draft PR [#1182](https://github.com/rickylabs/netscript/pull/1182)
+
+## What the proofs established
+
+P1 found a usable Aspire post-allocation callback seam: `onResourceEndpointsAllocated` plus
+`getValueAsync()` emitted a complete, identity-bound endpoint manifest with the allocated port. The
+full proof still correctly failed because the generated SQLite service exited without `--allow-ffi`,
+and a later HTTP 200 could not be attributed to that owned process. The qualified, revisitable
+arbitration is F1(b): productize a disposable startup-side writer after the generated runtime
+permission defect is fixed.
+
+P2 measured a real, attributed no-database scaffold: a 3657-byte OpenAPI 3.1.1 document with three
+dotted operation IDs, 73/89/88-byte discovery rows, no references, no non-2xx/common error envelope,
+and no observed truncation-limit exceedance. Its combined verdict is FAIL because the required
+DB-backed branch could not produce an attributable live spec; #1128 therefore remains open without a
+closing keyword.
+
+P3 independently confirmed the auth-guarded spec route's exact 401 and 403 envelopes plus an
+authorized 200, and ratified the locked `spec_unavailable` guidance without implying authenticated
+spec support exists in the future production feature.
+
+## Two things worth remembering
+
+1. A proof slice can finish successfully with an empirical FAIL. The formal Qwen IMPL-EVAL passed
+   this run because all required verdict artifacts were truthful and evidence-backed; it did not
+   reinterpret a failed runtime or missing branch as a product PASS.
+
+2. Ownership evidence is part of the result. An HTTP response without captured listener ownership
+   and ordering cannot rescue a failed owned process. Likewise, shared-host leak checks must name
+   specific owned roots and leave foreign or unproven AppHosts and containers untouched.
+
+## Deliberately not done
+
+- No package/plugin public-surface or generated-template productization.
+- No workaround for the generated SQLite `--allow-ffi` defect.
+- No closing keyword for #1128 while its DB-backed acceptance evidence is absent.
+- No `deno task e2e:cli`; merge-readiness E2E remains the orchestrator's decision.
@@ -374,6 +374,14 @@ legitimate result that activates option (b) (the `aspire-cli` source), not a blo
 verdict is the deciding artifact; #1131 and #1133 may not start before it exists. The original
 fork table is retained below as the decision record:
 
+**P1 verdict (2026-08-03): `FAIL`; F1(b) selected for the current decision record.** The
+post-allocation callback did emit a complete identity-bound manifest with the allocated
+`http://localhost:3001` endpoint, so the seam itself was not refuted. The locked coherent-owned-run
+bar failed because the generated SQLite service exited without `--allow-ffi`; a later HTTP 200 was
+unattributed and could not establish manifest/description/health agreement. F1(b) is therefore the
+first-class endpoint source now, while a future healthy owned-run proof may legitimately revisit F1.
+Both mechanisms remain additive implementations of the same endpoint-source port.
+
 | # | Fork | Options | Seed recommendation |
 | --- | --- | --- | --- |
 | F1 | Endpoint manifest mechanism | (a) generated run-mode **post-allocation** callback writes the run-state manifest (b) `aspire` CLI query adapter (c) MCP as Aspire-hosted HTTP resource | **(a), [P1]-arbitrated** — S-7 proved the naive helper-body write impossible, so (a) stands only if P1's artifact demonstrates the post-allocation seam; (b) is a first-class source in the same port contract otherwise; (c) rejected (no HTTP transport exists; port chicken-and-egg; `.mcp.json` churn) |
@@ -388,4 +396,5 @@ fork table is retained below as the decision record:
 document condenses the run's normative record (`design/canonical/00–06` rev 2, `plan.md` rev 2,
 `design/examples/`). Where this RFC and the run docs conflict, the run docs win until
 ratification, then GitHub wins. Refs #1117 #1102 #1072 #1071 #1093 — no closing keywords; the §7
-board is now live (epic #1126, children #1127–#1140); F1 remains open pending #1127's verdict.</sub>
+board is now live (epic #1126, children #1127–#1140); F1 was resolved to qualified F1(b) by
+#1127's committed proof verdict.</sub>
@@ -0,0 +1,27 @@
+use harness
+
+# S1 Fable re-review — amended P1 proof
+
+You are the separate opposite-family slice reviewer. Work only in
+`/home/codex/repos/ns005-proofs-impl`. Do not edit implementation or evidence files, do not run an
+AppHost or scaffold, do not delegate, and do not create a commit or contact GitHub.
+
+Read the locked plan and PLAN-EVAL, then the original review at `reviews/S1-fable.md`. Re-review
+the amended stable S1 artifacts, especially:
+
+- `proofs/P1-verdict.md`
+- `proofs/evidence/P1-runtime.json`
+- `proofs/evidence/P1-attempts.md`
+- `worklog.md`
+- `context-pack.md`
+- `drift.md`
+
+Verify each original finding M1, M2, M3, m1, m2, and m3 is resolved without weakening the locked
+D5/D6 gate or misrepresenting the unattributed HTTP 200. Confirm that P1 remains an explicit FAIL,
+that F1(b) is causally qualified and revisitable, that only DB-backed P2 is product-blocked, and
+that evaluator vocabulary and supervisor-owned RFC updates are correctly scoped.
+
+Write exactly one review artifact at `reviews/S1-fable-rereview.md`. Its first line must be exactly
+`APPROVED` or `CHANGES_REQUESTED`. Include a per-finding disposition table and list any new finding
+with severity and a concrete required action. If all original findings are resolved and no new
+blocking issue exists, approve. This is advisory slice review, not IMPL-EVAL.
@@ -0,0 +1,36 @@
+use harness
+
+# S1 opposite-family review — P1 lifecycle verdict
+
+## SKILL
+
+Read and follow `.agents/skills/netscript-harness`, `.agents/skills/netscript-doctrine`,
+`.agents/skills/netscript-tools`, `.agents/skills/aspire`, and `.agents/skills/netscript-pr`. This
+is an advisory `review_codex` lane, not formal IMPL-EVAL.
+
+You are the separate native Claude Fable 5/low reviewer for Codex-authored S1 in
+`test-openapi-mcp-wave0-proofs--wave0`. Work from `/home/codex/repos/ns005-proofs-impl` and review
+the stable uncommitted diff only.
+
+Read the run's `plan.md` D4–D6, `plan-eval.md`, `worklog.md`, `drift.md`, all `proofs/P1-*` and
+`proofs/experiments/p1-*` / `proofs/evidence/P1-*` files, RFC #1123 §4 and §9, and issue #1127's
+acceptance contract. Inspect the scratch manifest/AppHost only as needed; do not change it.
+
+Adversarially answer:
+
+1. Does the callback + `getValueAsync()` + atomic identity-bound manifest actually demonstrate the
+   RFC's post-allocation seam, or does D5 correctly require coherent Aspire description/health/live
+   request evidence and therefore make this an explicit FAIL selecting F1(b)?
+2. Is the later HTTP 200 after Aspire `Finished`/exit 1 adequately explained and safe to use, or is
+   the evidence contaminated/ambiguous?
+3. Are manifest identity, endpoint allocation, atomicity, multi-service completion, evidence
+   normalization, exact teardown, scope, and static gates truthful and sufficient?
+4. Does `FAIL_RESCOPE` improperly block independent P2/P3 proof work, given that a P1 FAIL is a
+   legitimate deliverable rather than a slice failure?
+5. Identify every correctness, evidence, path, wording, or acceptance-box defect that must be fixed
+   before supervisor sign-off.
+
+Write only `.llm/runs/test-openapi-mcp-wave0-proofs--wave0/reviews/S1-fable.md`. Begin with exactly
+`APPROVED` or `CHANGES_REQUESTED`, then list severity-ranked findings with file/line evidence and a
+specific required action. Do not edit any proof, plan, worklog, drift, product/template source, or
+scratch file; do not commit, push, post to GitHub, run a new AppHost, or spawn subagents.
@@ -0,0 +1,31 @@
+use harness
+
+# S1 Fable findings amendment
+
+## SKILL
+
+Continue following the already-loaded harness, tools, Aspire, CLI, doctrine, and PR skills. This is
+the same Codex implementation thread; do not self-sign off.
+
+Read `reviews/S1-fable.md` and resolve every M1, M2, M3, m1, m2, and m3 finding without rerunning an
+AppHost and without changing product/template source:
+
+- De-attribute the unexplained HTTP 200 in `P1-runtime.json`; it must not remain a satisfied
+  pass-bar condition. State that listener ownership/timing was not established and preserve the
+  observation only as ambiguous evidence.
+- Qualify the P1/F1(b) verdict and future RFC/epic wording: the post-allocation seam produced
+  correct allocated endpoint values, but the locked D5 coherent owned-run bar failed due service
+  health; a later healthy proof may legitimately revisit F1.
+- Distinguish the DB-backed P2 block from runnable no-DB P2 and P3. The product rescope
+  recommendation must not halt independent proof work.
+- Make `FAIL_RESCOPE` explicitly an implementation recommendation, not a self-issued evaluator
+  verdict.
+- Annotate RFC §9 as supervisor-applied after review in the S1 file list.
+- Rename/qualify `literalLoopbackUrl` because the raw host was the `localhost` name, not a literal
+  numeric loopback.
+- Update `worklog.md`, `context-pack.md`, and `drift.md` truthfully; preserve explicit P1
+  FAIL/F1(b), teardown, static-gate, and no-product-change evidence.
+
+Run scoped format/check/lint only for the touched run/proof artifacts. Do not commit, push, edit
+GitHub, rerun Aspire, or begin P2/P3. Stop with a concise handoff for the supervisor's Fable
+re-review.
@@ -0,0 +1,18 @@
+use harness
+
+# S2 Fable re-review — amended P2 evidence
+
+You are the same separate opposite-family review role in a fresh native session. Work read-only in
+`/home/codex/repos/ns005-proofs-impl`; do not edit, run runtime/scaffold work, delegate, commit,
+push, or contact GitHub.
+
+Read `reviews/S2-fable.md`, `briefs/S2-review-fixes.md`, and every amended artifact named there.
+Verify M1 and m1–m3 are resolved: `summary` and all actually-present standard keywords are captured;
+unknown/non-allowlisted keys remain auditable with paths; the context-blind limitation is honest;
+the retained raw spec exactly matches the recorded bytes/hash and regenerated evidence; lint-ignore,
+plan-command drift, and digest-algorithm records are corrected. Reconfirm the measurement script's
+claims, the explicit combined P2 FAIL, and that #1128 acceptance must remain unchecked.
+
+Write only `reviews/S2-fable-rereview.md`. First line exactly `APPROVED` or
+`CHANGES_REQUESTED`; include a per-finding disposition and any new blocking finding. This is
+advisory slice review, not IMPL-EVAL.
@@ -0,0 +1,35 @@
+use harness
+
+# S2 Fable review — P2 no-DB measurement and DB failure
+
+You are the separate opposite-family slice reviewer. Work only in
+`/home/codex/repos/ns005-proofs-impl`. Do not edit implementation/evidence files, run an AppHost or
+scaffold, delegate, commit, push, or contact GitHub.
+
+Read plan D3/D7/D8/D10/D12, `plan-eval.md`, issue #1128, RFC §4, the approved S1 re-review, and the
+stable S2 diff. Review at minimum:
+
+- `proofs/P2-verdict.md`
+- `proofs/experiments/p2-measure-live-spec.ts`
+- `proofs/evidence/P2-no-db.json`
+- `proofs/evidence/P2-db-failure.json`
+- `proofs/evidence/P2-runtime.json`
+- `proofs/evidence/P2-attempts.md`
+- the P1 evidence carried into the DB failure
+- `worklog.md`, `context-pack.md`, and `drift.md`
+
+Determine whether the explicit combined P2 `FAIL` is the only truthful D7/D12 verdict. Confirm the
+unattributed P1 HTTP 200 was not reused; the no-DB spec is attributed to one owned healthy real
+scaffold; all required operationId, compact-byte, discovery-row, request/response/error/all-schema,
+source/dereference, error-envelope, reference, keyword, and truncation-budget observations are
+measured rather than inferred; and the tool's algorithms support every claim.
+
+Adversarially check evidence consistency, normalization under D10 (including volatile PIDs/times),
+fixed-port ownership, teardown, lock/scope hygiene, static-gate claims, and whether any result could
+be misread as satisfying #1128 acceptance despite the missing DB measurement. Confirm no product
+or template workaround was made.
+
+Write exactly one artifact at `reviews/S2-fable.md`; first line exactly `APPROVED` or
+`CHANGES_REQUESTED`. Rank findings by severity with concrete required actions. State separately
+whether #1128's acceptance box may truthfully be checked. This is advisory slice review, not
+IMPL-EVAL.
@@ -0,0 +1,41 @@
+use harness
+
+# S2 implementation follow-up — P2 no-DB measurement and explicit DB failure
+
+Resume the existing tracked Codex implementation thread in
+`/home/codex/repos/ns005-proofs-impl`. Read the current run artifacts, the approved
+`reviews/S1-fable-rereview.md`, plan D3/D7/D8/D10/D12, issue #1128, and RFC §4 before acting.
+
+Execute S2 only, then stop for supervisor/Fable review. Do not start P3. Do not commit, push, edit
+GitHub, update the seed RFC, or change any product/template file.
+
+The supervisor accepts S1 and authorizes the independent no-database P2 half. The SQLite/DB P2
+half remains failed/rescoped because the generated service lacks `--allow-ffi`. Do not rerun,
+patch, wrap, or manually relaunch that DB scaffold, and do not reuse the unattributed P1 HTTP 200.
+Carry the normalized P1 runtime failure forward as the DB-side P2 evidence. Because D7 requires
+both scaffolds, the resulting P2 verdict must be explicit `FAIL`, never partial PASS or NOT_RUN.
+
+For the no-database half:
+
+1. Re-inventory shared AppHosts, containers, and relevant ports read-only; leave every foreign or
+   unproven resource untouched.
+2. Create a fresh local-source no-database scaffold under the approved `.llm/tmp/` root. Run at
+   most one owned AppHost and capture exact ownership/PID/port before fetching the live
+   `/api/openapi.json` document.
+3. Commit-ready experiment/evidence must record the D7 schema for the valid no-DB live spec:
+   compact UTF-8 spec bytes; every operationId and shape classification; discovery-row bytes;
+   request, response, error, and all-schema views in source and local-dereferenced form; every
+   non-2xx response/envelope observation; local/external/unresolved refs; and the recursively
+   observed OpenAPI/JSON-Schema keyword subset.
+4. Compare every measured array/string with `maxItems=50` and `maxStringLength=2000`, and state
+   explicitly that the current MCP path has no whole-result byte ceiling. Never infer a common
+   error envelope where the no-DB template lacks one.
+5. Normalize paths/process noise and record commands, versions, timestamps, teardown, root-lock
+   hash, and attribution evidence. Stop the exact owned tree and prove zero owned survivors before
+   writing the verdict.
+6. Run the scoped check/lint/fmt wrappers on owned TypeScript and scan for lint ignores. Update
+   worklog/context/drift. Leave a stable uncommitted S2 diff and stop.
+
+The verdict must distinguish measured no-DB results from the failed DB branch and explain that a
+skipped/blocked branch maps to FAIL under D7/D12. Do not claim #1128 acceptance or issue an
+IMPL-EVAL disposition.
@@ -0,0 +1,28 @@
+use harness
+
+# S2 review fixes — keyword fidelity and record hygiene
+
+Resume the same tracked Codex thread in `/home/codex/repos/ns005-proofs-impl`. Read
+`reviews/S2-fable.md` completely and resolve M1 plus m1–m3. Do not run a scaffold/AppHost, start
+P3, change product/template source, commit, push, or contact GitHub.
+
+Required amendments:
+
+1. Fix `p2-measure-live-spec.ts` so the observed OpenAPI 3.1 / JSON Schema 2020-12 keyword evidence
+   cannot silently omit present standard keywords such as `summary`. Use a defensible complete
+   standard-key set and also emit non-allowlisted object keys with paths (or an equally auditable
+   mechanism). Document the context-blind property-name limitation if it remains.
+2. Copy the retained, hash-matched, credential-free 3657-byte raw no-DB spec into normalized
+   committed evidence under `proofs/evidence/` using `apply_patch`, then rerun only the measurement
+   script against that retained spec to regenerate `P2-no-db.json`. Do not fetch the network.
+3. Update `P2-verdict.md` so its keyword claim exactly matches the amended evidence.
+4. Update the no-lint-ignore gate note to cover both P1 and P2 experiment files.
+5. Append drift for the split fetch/measurement command and its filename/permissions deviation.
+6. Name the lock-digest algorithm in P1/P2 evidence, or otherwise make the 40-vs-64-hex distinction
+   explicit without changing the historical values.
+7. Re-run the scoped wrappers, JSON assertions, raw-spec hash/byte checks, lint-ignore scan, and
+   diff/scope checks. Update worklog/context/drift, leave a stable uncommitted amendment, and stop
+   for Fable re-review.
+
+The combined P2 verdict remains `FAIL`; do not claim #1128 acceptance or issue an IMPL-EVAL
+disposition.
@@ -0,0 +1,32 @@
+use harness
+
+# S3 Fable review — P3 auth-guarded spec fixture and wording
+
+You are the separate opposite-family slice reviewer. Work only in
+`/home/codex/repos/ns005-proofs-p3b-impl`. Do not edit implementation/evidence files, delegate,
+commit, push, or contact GitHub. A read-only rerun of the single focused test is permitted; do not
+start Aspire, scaffold an app, or touch shared resources.
+
+Read plan D9/D10 and the skip-as-FAIL rule, `plan-eval.md`, issue #1129 and RFC #1123 §4 Wave 0, the
+existing fixture source, and the stable S3 diff. Review at minimum:
+
+- `packages/service/tests/auth/define-service-auth_test.ts`
+- `proofs/P3-verdict.md`
+- `proofs/evidence/P3-auth-fixture.json`
+- `worklog.md`, `context-pack.md`, and `drift.md`
+
+Adversarially verify that the selected fixture really performs all three requests against one live
+auth-guarded `/api/openapi.json` service; asserts the exact 401 and 403 JSON envelopes; asserts the
+authorized 200 without claiming an unasserted success-body shape; and was actually executed with the
+recorded focused command and exit-zero result. Check repository-head, fixture-blob, assertion line,
+runtime-version, and normalization claims where independently possible.
+
+Confirm the ratified `spec_unavailable` wording is byte-for-byte the D9 text, is justified by the
+measured 401/403/200 behavior, and does not weaken auth or imply authenticated-spec support exists.
+Check scope/lock hygiene: S3 may add only run artifacts, must not edit product/test source, must not
+add lint ignores, and must not represent a skipped branch as PASS.
+
+Write exactly one artifact at `reviews/S3-fable.md`; first line exactly `APPROVED` or
+`CHANGES_REQUESTED`. Rank findings by severity with concrete required actions. State separately
+whether #1129's acceptance box may truthfully be checked after the reviewed artifacts are committed.
+This is advisory slice review, not IMPL-EVAL.