[leon1418]

Opened by the knowledge auto-update pipeline. Draft — a human decides.

What changed upstream

OpenAI GPT-6 Astra now supports UltraFast mode on Amazon Bedrock

Verdict: new_knowledge on bedrock.openai_models.gpt_6_astra_speed_tier

• was — No mention of GPT-6 Astra or an UltraFast speed tier in model-selection.md or the OpenAI model recommendation catalog (openai-bedrock-2026-08-21.json referenced).
• now — GPT-6 Astra now has a premium "UltraFast" speed tier on Bedrock: up to 6x faster inference, up to 300 tokens/sec, aimed at latency-sensitive workloads (real-time coding assistants, interactive agents, customer-facing experiences).

Still true: Existing OpenAI model routing (mantle_openai_responses / runtime_converse paths) and existing models remain valid; this adds a new speed-tier option rather than replacing any existing one.

This is why the old value is not simply overwritten.

Proposed edits

Each row states its justification. No CI check can catch a badly reworded judgment, so the
"why" column is the only protection a reviewer has — please read it rather than the diff alone.

agent-advisor/references/phases/model-recommend/model-recommend.md:95 — derived judgment

- OpenAI is handled by a dedicated provider module (`openai-bedrock-2026-08-21.json` catalog):
+ OpenAI is handled by a dedicated provider module (`openai-bedrock-2026-08-21.json` catalog, now including GPT-6 Astra's UltraFast speed tier for latency-sensitive workloads):

Why: The recommendation logic drawing from this catalog should account for the new premium speed tier option when matching latency-sensitive workload needs.

Evidence (verbatim from the announcement): "Use GPT-6 Astra Ultrafast for latency-sensitive applications such as real-time coding assistants, interactive agents, and customer-facing experiences that require fast, high-quality responses."

Filter false positives

The announcement scan flagged these files; the judge rejected them:

• gcp-to-aws/references/design-refs/ai-gemini-to-bedrock.md
• azure-to-aws/references/design-refs/ai.md

Known limits of this proposal

• Blast radius is not stable between runs. Two runs of the same input returned 9 and 13
  locations and neither was a superset of the other, so this list may be incomplete.
• Paths are relative to plugins/aws-startup-advisor/skills.
• The pipeline did not touch any file the judge did not name.

[herosjourney]

Reviewed: head 8e17915 (base ec0fa3d), 1 file (+1/−1) — agent-advisor only. Draft, knowledge auto-update pipeline.
Ran: diff against base; git grep of openai-bedrock-2026-08-21.json and agent-advisor for GPT-6 / UltraFast / speed tier (no hits on the base); git merge-tree --write-tree pr-369 pr-390 (content conflict on model-recommend.md). Primary sources read 2026-10-01: the what's new post and the GPT-6 Astra model card.
Did not run: mise run build, pytest (no test change), live Bedrock invocation.

Cross-plugin parity: n/a — one tree, one file, no twin.

Change narrative: The pipeline wants recommendation logic to know about GPT-6 Astra UltraFast. The diff adds a parenthetical claiming openai-bedrock-2026-08-21.json now includes that tier. Body vs diff: the evidence quote matches the what's new post; the claim about the catalog does not match the catalog.

The evidence sentence is verbatim from the what's new post. The post's "up to 6x faster" and "up to 300 tokens per second" are also in the body. The model card is more specific and is not reflected: UltraFast is service_tier: "ultrafast" on model id openai.gpt-6-astra, not a separate model; mantle UltraFast is us-east-1 only (us-west-2 is Standard only); runtime uses us.openai.gpt-6-astra or global.openai.gpt-6-astra; UltraFast prices are six times Standard. Priority and Flex are not supported.

Axis	Stance	One line
Parsimony	advisory	A parenthetical is smaller than a catalog edit and does not do what the Why asks.
Completeness	finding	The engine reads the JSON. The JSON is unchanged and has no Astra entry.
Security	advisory	The sentence can send a latency-sensitive workload to a tier priced at 6× Standard, with the region limit omitted.
Alternatives	advisory	Record the tier on the catalog entry #369 adds, after that PR.
Contradictions	finding	The prose says the 2026-08-21 catalog includes the tier. The catalog's verified_at is 2026-08-21 and it has no such entry.
Overlap	clear	No second copy of a price or model id.
Cross-plugin parity	clear	Single file.
Propagation & data	finding	Title names bedrock.openai_models.gpt_6_astra_speed_tier. The SSOT the engine loads (references/models/openai-bedrock-2026-08-21.json, model_recommendation.py) is not in the diff.

Rejected-file check: azure-to-aws design-refs/ai.md sends OpenAI sources to ai-openai-to-bedrock.md and is not the place for this fact. ai-gemini-to-bedrock.md has no GPT-6 / UltraFast hit on the base. Those two rejections are right.

Blocking

1. [Completeness] The sentence says the August catalog includes UltraFast. It does not, and nothing that recommends a model reads this sentence. — plugins/aws-startup-advisor/skills/agent-advisor/references/phases/model-recommend/model-recommend.md:95

   OpenAI is handled by a dedicated provider module (openai-bedrock-2026-08-21.json catalog, now including GPT-6 Astra's UltraFast speed tier for latency-sensitive workloads):

   references/models/openai-bedrock-2026-08-21.json on the base has eight entries (openai_gpt_5_6_sol, openai_gpt_5_5, openai_gpt_5_6_terra, openai_gpt_5_6_luna, openai_gpt_5_4, and three Anthropic ids). No GPT-6 Astra, no service_tier, no UltraFast. scripts/model_recommendation.py loads that file. model-selection.md says not to restate path rules beside the catalog: the dated catalog and the engine are the record.
   The next sentence still says "verified 2026-08-21". A reader will attach that date to the September 30 tier.
   Consequence: a latency-sensitive recommendation follows a parenthetical, misses service_tier: "ultrafast", misses "us-west-2 mantle is Standard only", and misses the 6× price on the model card.
   Fix: drop this parenthetical. Do not merge it onto the 2026-08-21 pointer.

2. [Propagation] This line conflicts with #369, which is the PR that actually adds GPT-6 Astra. — same file, line 95
   git merge-tree of pr-369 and pr-390 conflicts on model-recommend.md. #369 changes this line to openai-bedrock-2026-09-09.json and adds openai_gpt_6_astra (openai.gpt-6-astra) in that catalog. That catalog also has no UltraFast entry (grep on pr-369).
   Consequence: merging #390 first writes a false inclusion into the old filename. Merging it after #369 rewrites #369's catalog pointer.
   Fix: land #369 first. Then add the tier on openai_gpt_6_astra in openai-bedrock-2026-09-09.json: service_tier value ultrafast; mantle only in us-east-1; runtime via us.openai.gpt-6-astra or global.openai.gpt-6-astra; price six times the Standard row on the model card. Point the recommender at that field when the user asked for latency. Leave model-recommend.md as a pointer to the catalog, not a second copy of the fact.

Should-fix

None beyond the blocking pair.

Heads-up

The what's new post says "up to 300 tokens per second" and "up to 6x faster". The model card states the price multiplier and the request field, and does not repeat 300 tokens/sec. Keep the speed figure attributed to the what's new post / OpenAI, not as a Bedrock quota.

Alternatives

1. Catalog field plus engine, after #369 — wins: the recommender can actually select the tier, with region and price attached. Costs: waits on #369 and touches the JSON the engine reads. Recommendation: switch to that. This one-line splice does not.

Verdict

Request changes — the only edit attributes a September 30 service tier to a catalog snapshot that does not contain it, and it collides with #369 on that line. Plugin code owners decide. This PR is a draft; it should stay a draft until the catalog change exists.

[github-actions]

This PR has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.