Hare/yoi
1
1
Fork 0
Code Actions Packages Releases Activity
4302fbbf69
yoi/work-items/closed/20260530-054927-refresh-builtin-model-catalog/thread.md

8.2 KiB
Raw Blame History

Created

Created by tickets.sh create.

Plan

Preflight

Classification: research-first / implementation-ready after sources are recorded.

The work is mostly data/catalog maintenance. It should begin with current provider documentation/model-list research and a short source note before editing the catalog. Implementation should be limited to resources/models/builtin.toml and directly related docs/tests unless research proves a provider definition is wrong.

Critical risks:

  • Do not guess model IDs or context windows from memory.
  • Do not add models that the current provider client cannot address.
  • Do not churn provider definitions unless needed.
  • If changing the default profile model, explain the product reason and verify compaction/effective window metadata.

Decision

Research note for builtin catalog refresh:

Sources checked:

  • Anthropic Models overview (https://docs.anthropic.com/en/docs/about-claude/models/overview, redirected to https://platform.claude.com/docs/en/about-claude/models/overview): current comparison lists Claude Opus 4.8, Claude Sonnet 4.6, and Claude Haiku 4.5. API IDs: claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5-20251001; aliases include claude-haiku-4-5. Context windows: Opus 4.8 1M, Sonnet 4.6 1M, Haiku 4.5 200k. Opus 4.8 is described as the starting point for most complex tasks, but the table says Extended thinking: No, so the catalog gives it an explicit capability without reasoning = "budget_tokens".
  • OpenAI Models overview (https://platform.openai.com/docs/models, redirected to https://developers.openai.com/api/docs/models): recommends gpt-5.5 for complex reasoning/coding, with gpt-5.4 and gpt-5.4-mini as lower latency/cost variants. gpt-5.5 and gpt-5.4 have 1.05M context windows and 128k max output.
  • OpenAI model detail pages:
    • https://developers.openai.com/api/docs/models/gpt-5.5: model ID gpt-5.5, 1,050,000 context window, xhigh reasoning support, notes prompts over 272K input tokens are charged differently; local catalog retains max_context_window = 272000 for the existing backend/effective-window clamp decision.
    • https://developers.openai.com/api/docs/models/gpt-5.4: model ID gpt-5.4, 1,050,000 context window.
    • https://developers.openai.com/api/docs/models/gpt-5-codex: model ID gpt-5-codex, 400,000 context window, Responses API only, optimized for agentic coding in Codex/similar environments.
  • OpenRouter model list endpoint (https://openrouter.ai/api/v1/models): confirmed anthropic/claude-opus-4.8 (1M), anthropic/claude-sonnet-4.6 (1M), and openai/gpt-5.5 (1.05M) with tools/structured output/reasoning parameters. Dynamic ~...latest router aliases exist, but the builtin catalog uses concrete IDs to avoid unstable default behavior.
  • Ollama Library:
    • https://ollama.com/library/llama3.3: llama3.3 latest/70b has 128K context.
    • https://ollama.com/library/qwen3-coder: qwen3-coder latest/30b has 256K context and is positioned for agentic/coding tasks.

Selected changes:

  • Anthropic direct: replace stale claude-sonnet-4-5 / claude-opus-4-1 with claude-opus-4-8, claude-sonnet-4-6, and claude-haiku-4-5; update Sonnet context to 1M.
  • Codex OAuth/OpenAI: keep default codex-oauth/gpt-5.5, update advertised context to 1.05M while retaining the existing 272K effective clamp; replace older plain gpt-5 entry with gpt-5.4; keep gpt-5-codex because OpenAI documents it as a Codex/similar-environment Responses model.
  • OpenRouter: replace stale anthropic/claude-sonnet-4 / openai/gpt-5 with concrete current IDs anthropic/claude-opus-4.8, anthropic/claude-sonnet-4.6, and openai/gpt-5.5.
  • Ollama: replace llama3.1 / qwen2.5-coder with current generic local placeholders llama3.3 and qwen3-coder.
  • Provider definitions unchanged; no provider-level source indicated that resources/providers/builtin.toml is stale.
  • Default profile remains codex-oauth/gpt-5.5; this remains aligned with OpenAI’s current model recommendation and the existing effective-context clamp used by compaction safety.

Review: approve

External review by reviewer Pod builtin-catalog-reviewer-20260530: approve.

Reviewer summary:

  • Sources were recorded in the ticket thread before/with the catalog changes.
  • Catalog changes align with the recorded sources:
    • Anthropic entries now use Opus 4.8 / Sonnet 4.6 / Haiku 4.5, with Sonnet and Opus 1M context and Haiku 200k.
    • Codex OAuth/OpenAI entries use gpt-5.5, gpt-5.4, and gpt-5-codex; gpt-5.5 retains the existing max_context_window = 272000 effective clamp while advertising 1.05M.
    • OpenRouter uses concrete current IDs, avoiding unstable ~...latest aliases.
    • Ollama entries use llama3.3 and qwen3-coder.
  • Provider definitions and default profile were reasonably left unchanged.
  • Reported validation was adequate.

Blockers: none.

Non-blocking note addressed after review:

  • The provider catalog test name/assertion still implied provider-default capability fallback for a model that now has explicit capability. It was renamed/updated to describe provider+model catalog merge semantics.

Implementation report

Implementation report:

  • Refreshed resources/models/builtin.toml from recorded official/semiofficial sources.
  • Anthropic direct entries now cover claude-opus-4-8, claude-sonnet-4-6, and claude-haiku-4-5; Sonnet/Opus context windows are 1M and Haiku is 200k. Opus has explicit capability without reasoning = "budget_tokens" because the source table says Extended thinking is not supported.
  • Ollama local placeholders are now llama3.3 (128K) and qwen3-coder (256K).
  • Codex OAuth/OpenAI entries now cover gpt-5.5, gpt-5.4, and gpt-5-codex; gpt-5.5 advertises 1.05M while retaining the existing max_context_window = 272000 effective clamp.
  • OpenRouter entries now use concrete current IDs anthropic/claude-opus-4.8, anthropic/claude-sonnet-4.6, and openai/gpt-5.5; dynamic latest aliases were intentionally not added.
  • resources/providers/builtin.toml and resources/profiles/default.lua were left unchanged.
  • Updated provider catalog test expectations and renamed the affected test to avoid implying provider-default capability fallback for an explicitly cataloged model.

External review:

  • Reviewer Pod builtin-catalog-reviewer-20260530 approved with no blockers.
  • Reviewer non-blocking note about the stale test name/assert message was addressed.

Validation:

  • cargo fmt --check passed
  • cargo test -p provider passed
  • cargo test -p manifest model passed
  • cargo test -p manifest profile -- --nocapture passed
  • cargo check -p provider -p manifest passed
  • ./tickets.sh doctor passed
  • git diff --check passed

Closed

Refreshed the builtin model catalog from recorded official/semiofficial sources. Anthropic, OpenAI/Codex OAuth, OpenRouter, and Ollama entries now point at current concrete model IDs; default profile remains ; provider definitions were unchanged. External review approved and validation passed: cargo fmt --check, cargo test -p provider, cargo test -p manifest model, cargo test -p manifest profile, cargo check -p provider -p manifest, ./tickets.sh doctor, git diff --check.

Closed

Refreshed the builtin model catalog from recorded official/semiofficial sources. Anthropic, OpenAI/Codex OAuth, OpenRouter, and Ollama entries now point at current concrete model IDs; default profile remains codex-oauth/gpt-5.5; provider definitions were unchanged.

External review approved and validation passed:

  • cargo fmt --check
  • cargo test -p provider
  • cargo test -p manifest model
  • cargo test -p manifest profile -- --nocapture
  • cargo check -p provider -p manifest
  • ./tickets.sh doctor
  • git diff --check