fix(llm): drop deepseek v4 pro from all harness fallback chains by shunkakinoki · Pull Request #2360 · shunkakinoki/dotfiles

shunkakinoki · 2026-08-12T15:29:10Z

What

Removes deepseek-v4-pro from every automatic model path and documents the resulting layout in a new MODELS.md.

Fallback chain is now uniform across all fallback-capable harnesses:

deepseek-v4-flash -> gemma-4-31b-it -> glm-4.7 -> minimax-m3 -> free

File	Change
`config/opencode/opencode-fallback.tpl.jsonc`	dropped `shunkakinoki/__DEEPSEEK_PRO__`
`config/openclaw/openclaw.tpl.json`	dropped `cliproxy/__DEEPSEEK_PRO__`; normalized hardcoded `gemma-4-31b-it`/`glm-4.7` to `__GEMMA__`/`__GLM__`
`config/hermes/config.tpl.yaml`	dropped the pro `fallback_providers` entry
`config/factory/settings.tpl.json`	primary `__DEEPSEEK_PRO__` -> `__DEEPSEEK_FLASH__`

Why

The flash migration (#2350, #2354, #2357) only moved primaries and the MoA reference model. Fallback chains were never touched, so every harness fell back to pro on the first hop, and Factory still had pro as its outright default.

Evidence from kyber and matic cliproxy logs:

Host	Finding
kyber	`/api/show {"name":"deepseek-v4-pro"}` -> 404. Pro is not resolvable via the model-metadata endpoint, so anything probing it before use fails.
kyber	`POST /v1/chat/completions` -> 400 `unknown provider for model cliproxy/deepseek-v4-flash` (x2, Aug 12).
matic	503 `auth_unavailable: no auth available (providers=openai-compatible-opencode,openai-compatible-openrouter, model=deepseek-v4-flash)` (x10, Aug 10). Aliyun not in the tried set despite being configured at priority 200.

Error-log model counts: kyber 26 flash / 14 pro, matic 40 flash / 0 pro. A live probe on both hosts confirmed pro and flash each route past model lookup to provider auth, so pro is not broken, just redundant.

Notes for reviewers

deepseek-v4-pro stays in the cliproxy catalog (config/cliproxyapi/config.tpl.yaml) and in the opencode/openclaw model listings. It remains addressable by explicit request. It is only off the automatic paths.
The openclaw placeholder normalization is cosmetic: generated openclaw.template.json output is byte-identical apart from the removed pro line.
MODELS.md also records two things that were previously implicit: the l-suffixed fish shortcuts (ocxel, coxel, pixel) run local Qwen via LM Studio rather than the cloud default, and Handy's glm-4.7 is transcript post-processing rather than transcription.
OMP and Pi are documented under "No fallback chain". The runtime-fallback work on claude/pi-fallback-support-335ac0 is not in this tree; MODELS.md will need an edit once that lands.

Not addressed here

Two issues surfaced by the logs that are separate from chain composition:

The unknown provider for model cliproxy/<model> 400 is a prefix bug. OpenCode absorbs it via retryable_error_patterns; openclaw and hermes have no equivalent matcher and hard-fail on it.
Matic's 503 shows aliyun missing from the provider set for flash despite priority: 200.

Testing

shellspec full suite: 2064 examples, 0 failures. Updated assertions in spec/llm_update_spec.sh, spec/openclaw_hydrate_spec.sh, spec/hermes_hydrate_spec.sh.

make check (nix flake check) was not run locally; CI covers it.

Summary by cubic

Drops deepseek-v4-pro and minimax-m3 from all automatic fallback chains and switches Factory’s session default to deepseek-v4-flash, eliminating first-hop failures and enforcing a flash‑first policy. Pro remains available by explicit request; minimax-m3 remains in Hermes’s MoA reference set.

Shared chain is now deepseek-v4-flash -> gemma-4-31b-it -> glm-4.7 -> free; no pro or minimax on any fallback path.
OpenCode, OpenClaw, and Hermes configs remove pro and minimax; OpenClaw normalizes Gemma/GLM to placeholders; Factory default moves to flash.
Adds MODELS.md (linked from README.md), completes the alias reference to match models.json, and documents derived forms and provider-specific overrides; updates hydrate and LLM update specs.
OMP: task role and code-explorer subagent now use gpt-5.6-luna; removes the gpt-5.3-codex-spark alias from models.json.

Rollout

No migration required. If any job relied on pro or minimax via defaults or fallback, specify deepseek-v4-pro or minimax-m3 explicitly.
If OMP workflows expected gpt-5.3-codex-spark, use gpt-5.6-luna instead.

^{Written for commit 96a21d7. Summary will update on new commits.}

Flash is the primary everywhere, but the fallback chains still hopped to pro first and Factory still defaulted to it outright. Kyber cliproxy returns 404 for deepseek-v4-pro on /api/show, so anything probing model metadata before use fails on that hop. Chain is now uniform: flash -> gemma-4-31b-it -> glm-4.7 -> minimax-m3 -> free. Pro stays in the cliproxy catalog and is addressable by explicit request. Also normalizes openclaw's hardcoded gemma/glm entries to placeholders (generated output unchanged). Entire-Checkpoint: cb2c92cc5d56

Documents the models.json -> placeholder -> llm-update.sh pipeline, the shared fallback chain and the rules governing it, per-harness model assignments (opencode, openclaw, hermes, omp, codex, pi, factory, ccs, aichat, llm, handy), fish shortcut models, and cliproxy provider priority. Entire-Checkpoint: ca7fa7ea303f

indent-zero · 2026-08-12T15:29:13Z

Warning

Indent Zero is shutting down on August 7th. Please migrate over to Indent 2.0 to continue getting PR reviews.

PR Summary

Trims every automatic fallback chain to only vendor-diverse hops that CLIProxy reliably serves. deepseek-v4-pro was removed because Kyber CLIProxy 404s on its /api/show probe (hard-failing OpenClaw and Hermes, which lack OpenCode's error-pattern matcher); minimax-m3 is now also dropped from the fallback path but kept in model catalogs and the Hermes MoA reference set. gpt-codex-spark is retired entirely from models.json since its only OMP consumers (task role, code-explorer override) are moved to gpt-5.6-luna.

Fallback chain across OpenCode / OpenClaw / Hermes is now deepseek-v4-flash → gemma-4-31b-it → glm-4.7 → free (3 hops), applied to opencode-fallback.tpl.jsonc, openclaw.tpl.json, and hermes/config.tpl.yaml plus their generated twins.
Factory sessionDefaultSettings.model flipped from deepseek-v4-pro to deepseek-v4-flash.
Normalizes two remaining hardcoded model strings in openclaw.tpl.json to __GEMMA__ / __GLM__ placeholders (generated output unchanged).
OMP retires gpt-codex-spark: modelRoles.task and agentModelOverrides.code-explorer in config/omp/config.tpl.yml now use __GPT_LUNA__, and the gpt-codex-spark alias is deleted from models.json. Only remaining gpt-5.3-codex-spark strings are in a commented-out example block in config/cliproxyapi/.
spec/llm_update_spec.sh, spec/hermes_hydrate_spec.sh, spec/openclaw_hydrate_spec.sh updated to match the new chain; the MoA-scoped minimax-m3 assertion in hermes_hydrate_spec.sh:198 is intentionally preserved.
Adds MODELS.md documenting the models.json → placeholder → llm-update.sh pipeline: complete 19-row alias table matching models.json key-for-key (__GPT_IMAGE__ included), the derived _PRETTY_/_NONDOT_ forms, the two provider-specific overrides declared in scripts/llm-update.sh (__GPT_IMAGE_OPENROUTER__, __DEEPSEEK_FLASH_0731__), the shared fallback chain, per-harness assignments, CCS profiles, fish shortcuts, and a provider-prefix table clarifying that shunkakinoki/, cliproxy/, and cliproxyapi/ all resolve through CLIProxy but map to different remote/local endpoints per harness. Linked from README.md.

Issues

All clear! No issues remaining. 🎉

1 issue already resolved

MODELS.md "No fallback chain" table drops the provider prefix for OpenCode's small_model (glm-4.7) and code-reviewer (deepseek-v4-flash), while every other row in the same table carries its prefix; the actual template values are shunkakinoki/glm-4.7 and shunkakinoki/deepseek-v4-flash (OpenRouter preset via shunkakinoki, not cliproxy), so a reader is misled about which upstream serves those roles. (fixed by commit b239a0a)

CI Checks

Waiting for CI checks...

coderabbitai · 2026-08-12T15:29:22Z

Warning

Review limit reached

@shunkakinoki, you've reached your PR review limit, so we couldn't start this review.

Next review available in: 15 minutes

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

How can I continue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews.

How do review limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 39ad69df-b00e-49de-8d51-2abfc4074346

📥 Commits

Reviewing files that changed from the base of the PR and between 5b82760 and 96a21d7.

📒 Files selected for processing (16)

MODELS.md
README.md
config/factory/settings.json
config/factory/settings.tpl.json
config/hermes/config.template.yaml
config/hermes/config.tpl.yaml
config/omp/config.tpl.yml
config/omp/config.yml
config/openclaw/openclaw.template.json
config/openclaw/openclaw.tpl.json
config/opencode/opencode-fallback.jsonc
config/opencode/opencode-fallback.tpl.jsonc
models.json
spec/hermes_hydrate_spec.sh
spec/llm_update_spec.sh
spec/openclaw_hydrate_spec.sh

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

gpt-5.3-codex-spark was only reachable through two OMP slots (the task role and the code-explorer subagent). Both now use gpt-5.6-luna, matching the other fast-path slots, and the alias is removed from models.json. The remaining gpt-5.3-codex-spark strings in the cliproxyapi configs are inside a commented-out upstream example block and are left as synced. Entire-Checkpoint: 2c17b23a9772

cubic-dev-ai

All reported issues were addressed

_{Reply with feedback, questions, or to request a fix.

Re-trigger cubic}

Chain is now flash -> gemma-4-31b-it -> glm-4.7 -> free across opencode, openclaw, and hermes. minimax-m3 stays in the model catalogs and in the Hermes MoA reference set, which is not a fallback path. Also documents the provider prefixes in MODELS.md and qualifies the two OpenCode rows that were listed without one. Entire-Checkpoint: 999cd0db5586

Adds the missing __GPT_IMAGE__ row, documents the _PRETTY_/_NONDOT_ derived forms, and documents the two provider-specific overrides declared in llm-update.sh rather than models.json. The alias table now matches models.json key-for-key. Entire-Checkpoint: 46bcbe487816

shunkakinoki added 2 commits August 13, 2026 00:28

indent-zero Bot reviewed Aug 12, 2026

View reviewed changes

Comment thread MODELS.md

cubic-dev-ai Bot reviewed Aug 12, 2026

View reviewed changes

Comment thread MODELS.md

shunkakinoki added 2 commits August 13, 2026 00:36

shunkakinoki merged commit a5a9251 into main Aug 12, 2026
29 of 30 checks passed

shunkakinoki deleted the claude/deepseek-v4-fallback-refactor-ee254e branch August 12, 2026 15:39

shunkakinoki mentioned this pull request Aug 12, 2026

fix(nix): unbreak t3code build and shell-test on main #2363

Merged

indent-zero Bot mentioned this pull request Aug 12, 2026

chore: upgrade #2365

Closed

Navigation Menu

fix(llm): drop deepseek v4 pro from all harness fallback chains - #2360

fix(llm): drop deepseek v4 pro from all harness fallback chains#2360
shunkakinoki merged 5 commits into
mainfrom
claude/deepseek-v4-fallback-refactor-ee254e

shunkakinoki commented Aug 12, 2026 •

edited by cubic-dev-ai Bot

Loading

Uh oh!

indent-zero Bot commented Aug 12, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Aug 12, 2026 •

edited

Loading

Review limit reached

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

shunkakinoki commented Aug 12, 2026 • edited by cubic-dev-ai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Why

Notes for reviewers

Not addressed here

Testing

Summary by cubic

Uh oh!

indent-zero Bot commented Aug 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

coderabbitai Bot commented Aug 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review limit reached

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

shunkakinoki commented Aug 12, 2026 •

edited by cubic-dev-ai Bot

Loading

indent-zero Bot commented Aug 12, 2026 •

edited

Loading

coderabbitai Bot commented Aug 12, 2026 •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading