BaseInfinity · BaseInfinity · May 4, 2026 · May 5, 2026 · May 4, 2026
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
@@ -13,7 +13,7 @@
       "name": "sdlc-wizard",
       "source": ".",
       "description": "SDLC enforcement for AI agents — TDD, planning, self-review, CI shepherd",
-      "version": "1.65.0",
+      "version": "1.66.0",
       "author": {
         "name": "Stefan Ayala"
       },

diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "sdlc-wizard",
-  "version": "1.65.0",
+  "version": "1.66.0",
   "description": "SDLC enforcement for AI agents — TDD, planning, self-review, CI shepherd",
   "author": {
     "name": "Stefan Ayala",

diff --git a/.reviews/research-95-nous.md b/.reviews/research-95-nous.md
@@ -0,0 +1,69 @@
+# ROADMAP #95 Research: Nous Research vs SDLC Wizard
+
+**Verdict: NO-GO. Different layer of the stack.**
+
+**Date:** 2026-05-04
+**Source:** [nousresearch.com](https://nousresearch.com), [github.com/NousResearch](https://github.com/NousResearch)
+
+## Original question
+
+> Evaluate `nousresearch.com` — what are they building, how does it compare to SDLC wizard's approach? Are they doing agent enforcement, testing methodology, or something orthogonal? Competitive analysis.
+
+## What Nous Research actually builds
+
+From their own copy: *"We train world-class open source language models and build infrastructure to coordinate distributed, unbiased training."*
+
+Product surface:
+
+| Product | What it is |
+|---------|-----------|
+| **Hermes** (Hermes 4 et al.) | Open-weights base / instruct LLMs |
+| **Hermes Agent** | Generic agent framework (autonomous task execution) |
+| **hermes-agent-self-evolution** | DSPy + GEPA pipeline that evolves the agent's own skills / prompts / code |
+| **atropos** | RL environments for collecting + evaluating LLM trajectories |
+| **Hermes-Function-Calling** | Function-calling tooling for the Hermes models |
+| **Psyche** | Distributed-training network infrastructure |
+| **Nous Chat / API Portal / Simulators** | End-user + developer access surfaces |
+
+## Layer-comparison
+
+| Layer | Nous Research owns | SDLC Wizard owns |
+|-------|--------------------|------------------|
+| Pre-training / fine-tuning | ✓ (Hermes models, Psyche) | — |
+| RL eval environments | ✓ (atropos) | — |
+| Agent framework (model executes tasks) | ✓ (Hermes Agent) | — |
+| Agent self-evolution (model rewrites its own prompts/skills) | ✓ (hermes-agent-self-evolution) | — |
+| **SDLC process enforcement on a human-driven coding agent** | — | ✓ |
+| TDD red-before-green hook | — | ✓ |
+| Planning gate / confidence stating | — | ✓ |
+| Cross-model adversarial review | — | ✓ |
+| CI shepherd loop | — | ✓ |
+
+**They build the engines. We enforce the build pipeline that uses an engine.** No surface overlap.
+
+## "But what about Hermes Agent — isn't that an agent like Claude Code?"
+
+Hermes Agent is an autonomous agent framework — the model loop itself, plus function-calling, plus a self-evolution layer. That's the same layer Claude Code, Codex CLI, and OpenCode occupy. The wizard doesn't compete with any of those — the wizard runs **inside** an agent host and enforces SDLC discipline on what the host does.
+
+If anything, Hermes Agent is a *target* for the wizard, not a competitor: the OpenCode port (ROADMAP #9) is exactly the play to make the wizard agent-host-agnostic, which would let it run on any OpenAI-compat backend including Hermes via Nous's API portal. That work is already tracked.
+
+`hermes-agent-self-evolution` is interesting (model-driven self-improvement of skills/prompts/code) but it operates at the agent's *own* layer — the agent rewrites itself. Our self-evolution loop operates at the *process* layer — the harness rewrites the wizard. Conceptually parallel, mechanically separate.
+
+## Pattern repeat — what this looks like next to prior NO-GOs
+
+This is the fourth external-methodology audit to land NO-GO:
+
+| # | Audit | Verdict | Reason |
+|---|-------|---------|--------|
+| #76 | Promptfoo | NO-GO | Already implements Promptfoo's best patterns; no statistical layer |
+| #77 | Constrain-to-Playbook | NO-GO | Already constrained on the prompts where it helps; PR review is open-ended on purpose |
+| #235 | Thoughtworks AI Evals | NO-GO | Methodology-only article; every layer already has an analog in our pipeline |
+| #95 | Nous Research | NO-GO (this) | Different layer of the stack — model R&D vs process enforcement |
+
+Pattern: external audits keep validating that the wizard is mature in its niche. Revisit only when an external source surfaces a *specific technique* we don't already have, not when it surfaces a different *product*.
+
+## Action
+
+- Mark ROADMAP #95 as DONE pointing at this research doc.
+- No code changes.
+- Keep an eye on the OpenCode port (#9) — that's the right vehicle if anyone ever asks "can the wizard run against Hermes / a self-hosted Nous endpoint?" The answer becomes "yes, via OpenCode."
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -4,6 +4,28 @@ All notable changes to the SDLC Wizard.
 
 > **Note:** This changelog is for humans to read. Don't manually apply these changes - just run the wizard ("Check for SDLC wizard updates") and it handles everything automatically.
 
+## [1.66.0] - 2026-05-04
+
+### Closed (paperwork-stale roadmap rows)
+
+- **ROADMAP #95 — Nous Research competitive audit** ✅ DONE 2026-05-04 with NO-GO verdict. Research write-up at `.reviews/research-95-nous.md`. Nous Research builds open-weights LLMs (Hermes), agent frameworks (Hermes Agent), RL environments (atropos), and distributed-training infra (Psyche) — different layer of the stack from SDLC enforcement. Hermes Agent is the same layer as Claude Code / Codex CLI / OpenCode (a *target* for the wizard, not a competitor); the OpenCode port (#9) is the right vehicle if anyone ever wants the wizard to run against a self-hosted Nous endpoint. Pattern continues with #76 (Promptfoo NO-GO), #77 (constrain-to-playbook NO-GO), #235 (Thoughtworks AI Evals NO-GO): external-product audits keep validating the wizard's niche.
+
+- **`docs/codex-near-top` cross-reference** ✅ DONE 2026-05-04 in PR #309. Surfaced `codex-sdlc-wizard` sibling at the top of README.md (after the tagline, before Install) and `CLAUDE_CODE_SDLC_WIZARD.md` (after the "What This Is" intro), so users on OpenAI's Codex CLI find the alternative without scrolling 250+/500+ lines to the Ecosystem section. Two new doc-consistency tests (head -30 / head -50 grep) keep the callout from drifting out of the top fold. OpenCode sibling intentionally not mentioned yet (per maintainer — bootstrap shipping in a different session).
+
+- **GitHub issue #308 (API features review)** ✅ CLOSED 2026-05-04. 4/4 entries audited (`gh issue view 308#issuecomment-4375055751`): zero wizard changes needed. Sonnet 1M-beta retirement (2026-04-30) only affects Sonnet 4.5 and Sonnet 4 — wizard's `sonnet[1m]` mixed-mode tier resolves to Sonnet 4.6 which keeps 1M GA per [API release notes](https://platform.claude.com/docs/en/release-notes/api.md). Rate Limits API + Memory for Managed Agents are different products. Haiku 3 retirement is irrelevant — `grep -ri "haiku.3\|claude-3-haiku"` returned zero hits across the wizard.
+
+### Files
+
+- `.reviews/research-95-nous.md` (new — research write-up, force-added past `.reviews/` gitignore matching #206 + #235 precedent)
+- `ROADMAP.md` (#95 marked DONE with verdict reference)
+- `README.md` + `CLAUDE_CODE_SDLC_WIZARD.md` (codex sibling callout near top, shipped in PR #309)
+- `tests/test-doc-consistency.sh` (2 new top-of-doc grep tests for codex callout, shipped in PR #309)
+- `CHANGELOG.md`, `SDLC.md`, `skills/update/SKILL.md`, `package.json`, `.claude-plugin/plugin.json` + `marketplace.json`, `CLAUDE_CODE_SDLC_WIZARD.md` (1.65.0 → 1.66.0)
+
+### Notes
+
+Zero code changes (research + cross-reference docs only). Backlog after this release: 1 open issue (#302 user-level setup skill — design-blocked) + ROADMAP top items #212 (multi-day), #9 OpenCode (separate session per maintainer).
+
 ## [1.65.0] - 2026-05-04
 
 ### Closed (paperwork-stale roadmap rows)

diff --git a/CLAUDE_CODE_SDLC_WIZARD.md b/CLAUDE_CODE_SDLC_WIZARD.md
@@ -2976,7 +2976,7 @@ If deployment fails or post-deploy verification catches issues:
 
 **SDLC.md:**
 ```markdown
-<!-- SDLC Wizard Version: 1.65.0 -->
+<!-- SDLC Wizard Version: 1.66.0 -->
 <!-- Setup Date: [DATE] -->
 <!-- Completed Steps: step-0.1, step-0.2, step-0.4, step-1, step-2, step-3, step-4, step-5, step-6, step-7, step-8, step-9 -->
 <!-- Git Workflow: [PRs or Solo] -->
@@ -4055,7 +4055,7 @@ Walk through updates? (y/n)
 Store wizard state in `SDLC.md` as metadata comments (invisible to readers, parseable by Claude):
 
 ```markdown
-<!-- SDLC Wizard Version: 1.65.0 -->
+<!-- SDLC Wizard Version: 1.66.0 -->
 <!-- Setup Date: 2026-01-24 -->
 <!-- Completed Steps: step-0.1, step-0.2, step-1, step-2, step-3, step-4, step-5, step-6, step-7, step-8, step-9 -->
 <!-- Git Workflow: PRs -->

diff --git a/ROADMAP.md b/ROADMAP.md
@@ -262,7 +262,7 @@ Living tracker of projects shipped using this wizard. **Rule:** only list projec
 | 92 | Research: Rigorous Autocompact Benchmarking | Nobody has published controlled benchmarks testing different autocompact thresholds. Opportunity to be THE authority. Test: quality metrics (task completion accuracy, code correctness) at 50/60/70/75/80/83% thresholds across 200K and 1M models. Measure: pre/post compaction context preservation, cost per session, degradation curves. Low priority but high differentiation potential |
 | 93 | ~~Node.js 20 EOL~~ DONE | Workflows already migrated to Node 24-native action versions: `actions/checkout@v5`, `actions/setup-node@v5`, `actions/upload-artifact@v6`. No `oven-sh/setup-bun` usage in repo. Verified by `tests/test-node24-compliance.sh` (13 tests, all green). Shipped pre-deadline (deadline was June 2, 2026) |
 | 183 | Advisor Tool A/B — Light Tier Validation (supersedes #94 research) | **RESEARCH COMPLETE.** aistupidlevel.info independent benchmarks (7-axis, CUSUM, 95% CI) show: Opus 4.6=47/VOLA, Sonnet 4.6=37/DEGR, Codex GPT-5.3=27/DEGR. Degradation confirmed by Anthropic staff (Boris Cherny, GH #42796, 274 comments). Root causes: adaptive thinking under-allocation, medium effort default, "simplest approach" system prompt (v2.1.64). Gap is multi-turn/planning specific — SWE-bench (single-shot) looks fine. **Tiered model strategy decided:** Critical (Opus executor + Codex reviewer), Standard (Opus + optional Codex), Light (Sonnet executor + Opus advisor — advisor compensates for weaker executor), Vibe (Sonnet only). Advisor tool (`advisor-tool-2026-03-01`) is API-only beta, not in CC CLI yet. Opus-advising-Opus is waste — advisor only valuable when executor is weaker. **Next:** build A/B workflow to validate Light tier (Sonnet+advisor vs Opus) on our E2E, integrate aistupidlevel monitoring. Blog: `claude.com/blog/the-advisor-strategy` |
-| 95 | Research: Nous Research Comparison | Evaluate `nousresearch.com` — what are they building, how does it compare to SDLC wizard's approach? Are they doing agent enforcement, testing methodology, or something orthogonal? Competitive analysis |
+| 95 | ~~Research: Nous Research Comparison~~ ✅ DONE 2026-05-04 — see [`.reviews/research-95-nous.md`](.reviews/research-95-nous.md) | **Verdict: NO-GO.** Different layer of the stack — Nous Research builds open-weights LLMs (Hermes), agent frameworks (Hermes Agent), RL environments (atropos), and distributed-training infra (Psyche). No surface overlap with SDLC enforcement: they build the engines, we enforce the pipeline that uses an engine. Hermes Agent is the same layer as Claude Code / Codex CLI / OpenCode — a *target* for the wizard, not a competitor. The OpenCode port (#9) is the right vehicle if anyone ever wants the wizard to run against a self-hosted Nous endpoint. Pattern continues with #76 (Promptfoo NO-GO), #77 (constrain-to-playbook NO-GO), #235 (Thoughtworks AI Evals NO-GO): external-product audits keep validating our niche; revisit only when an external source surfaces a *specific technique* we don't have, not when it surfaces a different *product*. **Original question:** evaluate `nousresearch.com` — what are they building, how does it compare to SDLC wizard's approach? |
 | 96 | ~~Audit: CC Degradation Detection~~ ✅ ALL PHASES DONE | Degradation infra shipped earlier. **Phase 1 DONE v1.57.0 (PR #290):** de-coached benchmark prompt. **Phase 2 DONE v1.58.0 (PR #293):** ground-truth gate. **Phase 3 PR 1 DONE v1.60.0:** wizard-installation lift-proof harness. **Phase 3 PR 2 DONE v1.61.0:** calibration scenario suite — `tests/e2e/scenarios/calibration-careful-read.md` is the first in a `calibration-*` family designed to reward self-review (parsePrice with 5 edge-case formats; rushed agent silently corrupts `'$1,000.00'` to `1`, a thousand-fold pricing bug). 6 new tests in `tests/test-calibration-scenarios.sh` (scenario format validator). End-to-end calibration verification (does low-effort agent actually score lower?) deferred to ROADMAP #212(i) Prove-It Gate paired runs. **The #96 series is now complete:** prompt de-coached → tests required → wizard contribution measurable → calibration signals embedded in scenario suite. |
 | 97 | Research: Anthropic Policy & Research Alignment | Audit Anthropic's policy/research pages for SDLC relevance: (a) Responsible Scaling Policy (`anthropic.com/responsible-scaling-policy`) — does our enforcement model align with their safety levels? (b) Transparency reports (`anthropic.com/transparency`) — can we reference/align? (c) Constitution (`anthropic.com/constitution`) — SDLC as constitutional AI for dev process? (d) Economic Futures (`anthropic.com/economic-futures`) — positioning for AI-assisted development economy (e) Research page (`anthropic.com/research`) — any papers relevant to agent enforcement, tool use safety, or evaluation methodology. Goal: align wizard's philosophy with Anthropic's public positions where genuine overlap exists |
 | 98 | ~~Audit: Community Feedback & Contribution Paths~~ DONE | PR #191. Added `.github/ISSUE_TEMPLATE/` (bug_report, feature_request, question) with GitHub-valid frontmatter + `config.yml` (blank issues disabled, contact links to Discussions + in-session `/feedback`). Added `.github/PULL_REQUEST_TEMPLATE.md` with Prove-It Gate checklist + test plan section. Enabled GitHub Discussions via `gh api ... -f has_discussions=true`. Added README "Feedback" section with live links to all three paths. `tests/test-community-paths.sh` — 14 tests, mutation-verified (deleting config.yml or stripping frontmatter `about:` trips CI loudly). Codex xhigh 2-round code review: 5/10 → 10/10 CERTIFIED |

diff --git a/SDLC.md b/SDLC.md
@@ -1,4 +1,4 @@
-<!-- SDLC Wizard Version: 1.65.0 -->
+<!-- SDLC Wizard Version: 1.66.0 -->
 <!-- Setup Date: 2026-01-24 -->
 <!-- Completed Steps: step-0.1, step-0.2, step-1, step-2, step-3, step-4, step-5, step-6, step-7, step-8, step-9 -->
 # SDLC Configuration
@@ -7,8 +7,8 @@
 
 | Property | Value |
 |----------|-------|
-| Wizard Version | 1.65.0 |
-| Last Updated | 2026-04-30 |
+| Wizard Version | 1.66.0 |
+| Last Updated | 2026-05-04 |
 | Claude Code Baseline | v2.1.111+ (required for Opus 4.7 / `opus[1m]`) |
 | Recommended Model | `opus[1m]` (Opus 4.7, 1M context) — run `/model opus[1m]` |
 | Recommended Effort | `max` (preferred) / `xhigh` (floor) — run `/effort max` |

diff --git a/package.json b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "agentic-sdlc-wizard",
-  "version": "1.65.0",
+  "version": "1.66.0",
   "description": "SDLC enforcement for Claude Code — hooks, skills, and wizard setup in one command",
   "bin": {
     "sdlc-wizard": "cli/bin/sdlc-wizard.js"

diff --git a/skills/update/SKILL.md b/skills/update/SKILL.md
@@ -93,9 +93,10 @@ Parse CHANGELOG entries between the user's installed version and latest. Present
 
 ```
 Installed: 1.42.0
-Latest:    1.65.0
+Latest:    1.66.0
 
 What changed:
+- [1.66.0] roadmap hygiene — closed #95 (Nous Research competitive audit — NO-GO; different layer of the stack, Nous builds open-weights LLMs + agent frameworks, we enforce SDLC process; pattern continues with #76 + #77 + #235 NO-GOs). Research write-up at `.reviews/research-95-nous.md`. Also rolls in PR #309 (codex sibling callout near top of README + wizard doc) and `gh issue close 308` (4/4 API entries audited, zero wizard changes — Sonnet 1M retirement only affects 4.5/4, our mixed-mode tier pins to 4.6).
 - [1.65.0] roadmap hygiene — closed paperwork-stale rows #210 (Node 24 false-green test, already shipped in PR #217) and #235 (Thoughtworks AI Evals research — NO-GO verdict, methodology already implemented under different naming; pattern continues with #76 + #77 NO-GOs). Research write-up at `.reviews/research-235-ai-evals.md`. No code changes.
 - [1.64.0] XDLC ecosystem cross-references — README, wizard doc, and ROADMAP now cross-reference all three sibling packages (`agentic-sdlc-wizard`, `codex-sdlc-wizard`, `claude-gdlc-wizard`). New "Ecosystem (Sibling Projects)" section in README. 3 new doc-consistency tests prevent drift.
 - [1.63.0] cache-cost observability closeout (#204 absorbed by #220) — `tests/test-token-spike.sh` gains explicit cache-miss regression test + negative-control test. SDLC skill + wizard doc gain "Cache-Cost Surprises" sections covering 10-20× silent cost blowups (mid-session CLAUDE.md edits, idle pruning, upstream cache bugs) and detection via `hooks/token-spike-check.sh`'s `costly_tokens` metric.
@@ -111,15 +112,7 @@ What changed:
 - [1.53.0] delete scan-community cron (#231 Phase 3c) — manual `claude --print` invocation of analyze-community.md
 - [1.52.0] delete community-e2e-test cron (#231 Phase 3b) — manual local-shepherd review of scan-community digest
 - [1.51.0] delete version-test cron (#231 Phase 3a) — manual local-Max replacement via npm i + local-shepherd
-- [1.50.0] local-shepherd.sh --strip-paths flag (#231 Phase 2 — replaces deleted prove-it-test cron)
-- [1.49.0] local-shepherd.sh --compare-baseline flag (#230)
-- [1.48.0] SKILL.md trim — token bloat audit phase 2 follow-up
-- [1.47.0] Codex review progress wrapper (#259)
-- [1.46.1] npx check surfaces dangling+enabled plugin state (#266)
-- [1.46.0] PreCompact dry-run env vars (#240)
-- [1.45.0] PreCompact path (c) — SHA-ancestry self-heal (#257)
-- [1.44.1] Autocompact compound-misconfig detection (#207)
-... (full entries from fetched CHANGELOG)
+... (older entries omitted — read the full CHANGELOG.md for anything pre-1.51.0)
 ```
 
 Read the actual entries from the fetched CHANGELOG; don't paraphrase. The user wants to see exactly what shipped.