Bump models (GLM-5.2 / MiniMax-M3 / DeepSeek-V4-Pro), add route preference, + GitHub Pages KB site

[jimstratus]

Summary

Updates reviewer models to the latest versions, adds an easy OpenRouter↔direct-API routing switch, and (incoming) a hosted GitHub Pages knowledge-base site.

Models

• glm-5.1 → glm-5.2 (z.ai direct + OpenRouter dual route)
• minimax-m2.7 → minimax-m3 (MiniMax direct + OpenRouter dual route)
• DeepSeek now defaults to deepseek-v4-pro (V3.2 demoted to custom_only, kept for benchmark-history continuity)
• New deepseek aichat client (api.deepseek.com) for the direct route
• Profiles repointed; new direct profile (direct-API subs only, no Gemini)

Routing — OpenRouter vs direct API

A single knob decides which provider a dual-route reviewer (glm-5.2, minimax-m3, deepseek-v4-pro) tries first; the other becomes the automatic fallback.

• defaults.route_preference: openrouter (public default — one OPENROUTER_API_KEY covers most reviewers) | direct
• Per-run override: --route-pref {openrouter,direct}, shorthands --prefer-direct / --prefer-openrouter, or ARGUS_ROUTE_PREF env. Precedence: CLI flag › env › config
• Wired through dispatch.py, verify.py, benchmark.py, estimate_cost.py via the single resolver _common.resolve_routes / resolve_route_preference
• CLI reviewers are never reordered — their paid CLI sub stays primary and OpenRouter remains a true fallback
• OR-balance pre-flight only fires when OpenRouter is the resolved primary, so a direct-preference run with a depleted OR balance isn't gated

This covers the "depleted OR balance → fall back to direct API for all but Gemini (skipped)" workflow: set route_preference: direct (or --prefer-direct) and use the direct profile.

Docs

• README / SKILL / CLAUDE / DEVELOPMENT / CONTRIBUTING updated (routing section, env vars incl. DEEPSEEK_API_KEY, flags, profiles, dual-route guidance)
• GitHub Pages KB site under docs/ — responsive, collapsible sidebar (mobile-collapsed / desktop-open), dark/light toggle (system default + persisted override), Mermaid diagrams, benchmark chart, onboarding (beginner + advanced), contributing/FAQ/glossary, per-page "View as Markdown" links, llms.txt, plus a Pages deploy workflow. (Landing in a follow-up commit on this branch.)

Tests

• New tests/test_routes.py (route classification, preference reordering, CLI-never-reordered invariant, precedence)
• Full suite green (37 passed), py_compile clean, config validates (15 reviewers, 8 profiles)

Note: live provider dispatch isn't exercised in CI (no network/keys); model slugs for the bumped versions follow the repo's existing forward-dated convention and may need a verify.py pass against live endpoints.

🤖 Generated with Claude Code

---

Generated by Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[jimstratus]

🔍 Self code-review (round 1)

Copilot review is unavailable (quota), so this is a manual self-review of the branch diff — covered correctness (line-by-line, removed-behavior, cross-file), reuse/simplification/efficiency, altitude, and CLAUDE.md conventions.

No correctness bugs. The route-resolution refactor is sound — the fallback flow is preserved in all four scripts (dispatch/verify/benchmark/estimate_cost), CLI reviewers correctly never reorder, and routing is centralized in one resolver (_common.resolve_routes), matching the repo's "single source of truth" convention.

Findings & fixes (both addressed in c978931)

#	Severity	Finding	Fix
1	low (accuracy)	hermes-4.3 is actually a 4th dual-route reviewer (Nous-direct ↔ OpenRouter), so it's reordered by route_preference too — but the docstring, README, and KB site claimed only glm-5.2/minimax-m3/deepseek-v4-pro. Behavior was already correct and benign (under the openrouter default it now skips a guaranteed-fail nous-direct attempt, since NOUSRESEARCH_API_KEY is unset).	Corrected the resolve_routes docstring, README routing section, and docs/ site content to list hermes-4.3 (custom-only) as dual-route.
2	low (duplication)	The 3-line "is the resolved primary OpenRouter?" check was copy-pasted into benchmark.py and estimate_cost.py.	Extracted to _common.primary_is_openrouter() and called from both.

No behavior change from the fixes. 37 unit tests pass, the site rebuilds cleanly, and there are no stale model names in docs/.

CI (lint-and-test 3.12/3.13 + fixture-integrity) is green on the prior commit and re-running on c978931.

(The copilot-pull-request-reviewer check showing "failure" is the unavailable Copilot bot, not this repo's CI.)

---

Generated by Claude Code

[jimstratus]

🔍 Self code-review — Pages-artifact tidy (4ea1d7a)

Focused review of the workflow change that slims the published tree (earlier rounds covered the scripts/config/site).

No correctness bugs. The staging step is sound and was verified locally with the same commands:

• cp -r docs/. _site/ copies contents including dotfiles, so .nojekyll is preserved.
• rm -f _site/*.py + rm -rf _site/__pycache__ drop the generator sources; rm -f is a no-op if the glob doesn't match.
• Resulting published tree = 11 HTML pages + assets/ + .nojekyll + llms.txt, and no build.py / build_content.py / __pycache__.
• Portable cp (not rsync, which isn't guaranteed on the runner); permissions / concurrency unchanged.

Finding & fix

#	Severity	Finding	Fix
1	low (tidiness)	The workflow creates _site/ at the repo root. It's never committed in CI, but a maintainer running the staging steps locally would get an untracked _site/ not covered by .gitignore — a footgun for an accidental commit of a generated tree.	Added _site/ to .gitignore (5868ff7).

Net: one low-severity fix; no behavior change. The repo CI (lint-and-test 3.12/3.13 + fixture-integrity) runs on the push; the Pages deploy runs after merge to master.

(Reminder: the copilot-pull-request-reviewer "failure" is the unavailable Copilot bot — quota — not this repo's CI.)

---

Generated by Claude Code