feat(migrate): recommend Claude Sonnet 5 / Opus 4.8 as Anthropic defaults

[herosjourney]

Summary

• Bumps new-migration recommend defaults from Claude Sonnet 4.6 / Opus 4.7 → Claude Sonnet 5 / Claude Opus 4.8, aligning clarify/design/scoring with models already marked Active in the pricing cache.
• Keeps Haiku 4.5 for cost/speed; Fable 5 stays opt-in only (frontier / Mythos pricing).
• Sonnet 4.6 remains Active in the pricing cache and static fallback table for existing workloads.
• Comparison tables keep steady-state $3/$15 for Sonnet 5 (post-intro), with callouts that intro pricing is $2/$10 through Aug 31, 2026.

Why

The catalog already knew about Sonnet 5 / Opus 4.8 / Fable 5, but Q16/Q17, mapping guides, agent-advisor scoring, and report fixtures still called Sonnet 4.6 “latest” and recommended it by default. That is inconsistent for customers starting migrations now.

Test plan

• Confirm Bedrock inference-profile IDs (anthropic.claude-sonnet-5, anthropic.claude-opus-4-8, prefer us. prefix) and regional availability
• Confirm agentic/tool-use claims still hold for Sonnet 5 (guides previously asserted Sonnet 4.6 GDPval leadership)
• python3 -m pytest skills/agent-advisor/scripts/test_scoring.py (57 passed locally)
• Spot-check a gcp-to-aws AI-only clarify → design path defaults to Sonnet 5
• After Aug 31, 2026: refresh Sonnet 5 static fallback from intro $2/$10 → $3/$15

Out of scope

• Classic Bedrock Agents EOL (fix(gcp-to-aws): stop recommending classic Bedrock Agents (closed to new customers 2026-07-30) #161)
• Stale Sonnet 4 example IDs in llm2bedrock agents (fix(llm-to-bedrock): sweep stale Sonnet 4 example model IDs to Sonnet 4.6 #164)
• Making Fable 5 a default

[herosjourney]

Addressed both merge-blocking review items:

1. Dropped anthropic.claude-opus-4-8-20250610-v1:0 / us.… from STATIC_FALLBACK — kept suffix-less Opus 4.8 IDs only.
2. Fixed gemini mapping guide Sonnet 5 context 200K → 1M.

Ready for another look.

[herosjourney]

Resolved the merge conflict with main (f064392). One conflicted hunk, in clarify-ai.md's Q17 feature table, and it was a semantic cross rather than a textual one: this PR bumps the agentic-workflows model (Sonnet 4.6 → Sonnet 5) while main independently moved the agentic runtime (Bedrock Agents → AgentCore Harness, from the Bedrock Agents EOL work this PR explicitly scoped out to #161). Resolution combines both intents: Claude Sonnet 5 with AgentCore (Harness). Post-merge verification on the combined tree: lint:md 0 errors (855 files), shared:check OK, agent-advisor scoring 230 passed, llm-to-bedrock scripts 160 passed, plugin tests 81 passed. Also confirmed the merge kept main's clarify changes intact around this PR's edits (Q15 provider-neutral wording, Q14/Q19 detect-confirm cards) with zero stray Sonnet 4.6 references left in clarify-ai.md.

[leon1418]

[🤖 AI review 🤖]

Summary

This PR does a mechanical text-substitution of Sonnet 4.6 → Sonnet 5 and Opus 4.7 → Opus 4.8 across 32 files (documentation, fixtures, scoring code, and tests). The change is well-scoped and internally consistent in the places it touches. However:

1. Partial supersedence by PR #195: PR #195 (Model Recommend phase) removes the model_recommendation field and _MODEL_PRIORITY / _FEATURE_OVERRIDE / _MIGRATE_FAMILY dictionaries from scoring.py entirely (−93 lines), and removes the corresponding tests. 7 files overlap. If #195 merges first, all scoring.py and related test changes in this PR become a merge conflict on dead code. Conversely if #167 merges first, #195 gets a trivial conflict (just different string values in the block it deletes).

2. Independent value: 25 of 32 files are NOT touched by #195 — the pricing guides, mapping tables, reference docs, lifecycle table, and llm-to-bedrock scripts are independent contributions that have value regardless of #195's fate.

3. PR #169 (discover-preview fix + breadcrumbs) is OPEN but touches discover-preview.md — this PR also touches that file. Textual conflict is likely but trivial (different hunks).

Findings

• 2 Nits (stale context-window claims after bumping to Sonnet 5 1M)
• 1 FYI (Opus 4.6 vs 4.8 inconsistency in clarify tables)
• 0 blocking issues

Merge Recommendation

MERGEABLE with the nits addressed, but coordinate with #195 author — whoever merges second gets conflicts in scoring.py. The cleanest path: merge #167 first (since it's simpler), then #195 rebases and just deletes the updated strings.

[leon1418]

[🤖 AI review 🤖]

Nit: This sentence now says "GPT-5.4 → Claude Sonnet 5" but keeps "Both have ~200K+ context" — the Gemini guide table (this PR) updates Sonnet 5's context column to 1M. Should be "Both have 200K+ context" → "Sonnet 5 has 1M context" or "~1M+ context" to stay consistent with the Bedrock capabilities table change at ai-gemini-to-bedrock.md line 44.

[leon1418]

[🤖 AI review 🤖]

Nit: Similarly, "longer context (200K vs 128K)" under the GPT-4/4 Turbo → Claude Sonnet 5 header should now reflect the 1M context window stated elsewhere in this PR for Sonnet 5. The 200K figure is Sonnet 4.6's context — Sonnet 5 is 1M per the capabilities table this PR introduces.

[leon1418]

[🤖 AI review 🤖]

FYI: Several rows in Q17/Q22 tables (clarify-ai.md, clarify.md) still say "Opus 4.6 for hardest" or "Opus 4.6 for most demanding" in lines this PR updates. The PR description explicitly bumps Opus 4.7 → 4.8 and the scoring code uses claude_opus_4_8, but the extended-thinking and complex-reasoning rows in the clarify tables reference Opus 4.6. This appears intentional per the anthropic-to-bedrock note ("Opus 4.6 remains a valid alternative where batch or regional availability matters"), but if the intent is to recommend 4.8 as the new default for hardest-reasoning, these rows should also say 4.8. No action required if the mixed references are deliberate.

[herosjourney]

Rebased onto current main (post-#195) and addressed the review nits.

Merge conflicts with #195 (Model Recommend):

• Dropped this PR's scoring.py / test_scoring.py model-selection block — feat(agent-advisor): Model Recommend phase (Anthropic + OpenAI→Bedrock) + seeded runs #195 already removed that path and ships Sonnet 5 / Opus 4.8 via anthropic_model_recommendation.py
• Kept feat(agent-advisor): Model Recommend phase (Anthropic + OpenAI→Bedrock) + seeded runs #195's model-selection.md and migration-plan Step 3.5 (advisor-wins) prose
• Combined clarify Q17 agentic row as Claude Sonnet 5 with AgentCore (Harness)
• Kept pricing-table Sonnet 5 / suffix-less Opus 4.8 entries; retained main's verified-2026-08-04 note; dated Opus IDs remain dropped
• POC example model id bumped to anthropic.claude-sonnet-5 inside feat(agent-advisor): Model Recommend phase (Anthropic + OpenAI→Bedrock) + seeded runs #195's CRIS/probe prose

Review nits: ai-openai-to-bedrock.md GPT-5.4 and GPT-4/4 Turbo sections now say Sonnet 5 has 1M context (was stale 200K).

FYI on Opus 4.6 in clarify tables: left as-is — intentional alternate where batch/regional availability matters; hardest-reasoning rows already say Opus 4.8.