[All Platforms][Inference] Deep Agents (dcode) onboard on Ollama truncates a colon model tag in config.toml — qwen2.5:7b becomes "7b" (name lost)

[hulynn]

Description

On nemoclaw v0.0.74, onboarding a Deep Agents (dcode / langchain-deepagents-code) sandbox on a local Ollama model whose tag contains a colon (e.g. qwen2.5:7b) writes a truncated model name into the agent config: dcode's /sandbox/.deepagents/config.toml ends up with default = "openai:7b" and models = ["7b"] — the part before the colon (qwen2.5) is dropped. The host-level model (nemoclaw list / status) correctly shows the full tag qwen2.5:7b; only the in-sandbox dcode config is truncated.

Impact is low on a single-model host (the short name still resolves), but on a host serving more than one Ollama model the truncated 7b is ambiguous and can resolve to the wrong model. Reported as low severity.

Platform scope: Config-generation defect, platform-independent — observed on Ubuntu 26.04 (qwen2.5:7b) and Ubuntu 24.04 (qwen3.5:9b) with v0.0.74.

Regression: Unknown — earlier versions not tested for this specific case.

Environment

Device:        Ubuntu 26.04 server (RTX PRO 6000 Blackwell)
OS:            Ubuntu 26.04 LTS
Architecture:  x86_64
Node.js:       v22.23.1
npm:           10.9.8
Docker:        29.5.3
OpenShell CLI: 0.0.72
NemoClaw:      v0.0.74
OpenClaw:      N/A (Deep Agents / langchain-deepagents-code sandbox)

Steps to Reproduce

1. Have a local Ollama model with a colon tag available on the host, e.g. qwen2.5:7b.
2. Onboard a Deep Agents sandbox on the local Ollama provider so it picks that model:

   NEMOCLAW_PROVIDER=ollama nemoclaw onboard --agent dcode --name {name} --non-interactive --fresh --yes

3. Compare the host-level model tag with what dcode wrote into its config:

   nemoclaw {name} status
   nemoclaw {name} exec -- grep -E "default|models" /sandbox/.deepagents/config.toml

Expected Result

dcode's config.toml keeps the full Ollama tag (qwen2.5:7b), matching the host-level model, so the model reference is unambiguous.

Actual Result

Host-level tag (nemoclaw list): qwen2.5:7b

In-sandbox /sandbox/.deepagents/config.toml:

default = "openai:7b"
models  = ["7b"]

The qwen2.5 before the colon is dropped; only 7b is kept.

Logs

host model tag: qwen2.5:7b
config.toml default = "openai:7b"
config.toml models  = ["7b"]

[yanyunl1991]

will take a look

[yanyunl1991]

Thanks for the detailed report — reproduced from source and fixed. Quick summary:

Root cause

agents/langchain-deepagents-code/generate-config.ts::modelNameForOpenAiProvider split the model id on the first : and returned the suffix, on the assumption that NEMOCLAW_MODEL might arrive in <provider>:<model> shape (e.g. openai:gpt-oss-120b) and the strip would avoid a double openai: prefix in the generated default. That assumption is fine for the openai:<model> shape, but for an Ollama tag like qwen2.5:7b the : is Ollama's name:tag separator — not a provider prefix — so the strip silently ate qwen2.5 and left 7b, producing default = "openai:7b" and models = ["7b"] on the reporter's box.

Fix

Narrow the strip to the exact openai: provider label (the sole provider emitted under [models.providers.openai]); any other colon is treated as part of the model id and passes through verbatim. This preserves both contracts:

NEMOCLAW_MODEL input	default line	models = [...]
qwen2.5:7b (Ollama tag; your case)	"openai:qwen2.5:7b"	["qwen2.5:7b"]
openai:gpt-oss-120b (provider-prefixed; pre-existing contract)	"openai:gpt-oss-120b"	["gpt-oss-120b"]
anthropic:claude-3-sonnet (other prefixes stay opaque)	"openai:anthropic:claude-3-sonnet"	["anthropic:claude-3-sonnet"]
nvidia/nemotron-3-super-120b-a12b (no colon; unchanged)	"openai:nvidia/nemotron-3-super-120b-a12b"	["nvidia/nemotron-3-super-120b-a12b"]

PR

#6328

Verification

Nine new regression tests in test/generate-dcode-config.test.ts cover the reporter's exact qwen2.5:7b case, an additional variant-qualified Ollama tag (llama3.1:8b-instruct-q4_0), non-colonized ids, the pre-existing openai:gpt-oss-120b strip contract, and a guard that openai: never appears more than once in the emitted default. The pre-existing test/langchain-deepagents-code-config.test.ts end-to-end spawn suite still passes (7/7). No live-GPU repro on our side — we do not own an x86_64 discrete-GPU Linux + Ollama test host, and the reporter's platform scope note ("config-generation defect, platform-independent") explicitly says the defect is not platform-conditional; the unit tests drive the exact buildConfig function the Dockerfile invokes at image-build time. CI on the branch is green.

To try it locally against your Ollama sandbox:

git fetch origin fix/dcode-model-colon-truncation-6325
git checkout fix/dcode-model-colon-truncation-6325
npm ci --ignore-scripts && npm run build:cli
NEMOCLAW_PROVIDER=ollama nemoclaw onboard --agent dcode --name <name> --non-interactive --fresh --yes
nemoclaw <name> exec -- grep -E "default|models" /sandbox/.deepagents/config.toml

Will land after CI is green and maintainer review. If the symptom persists on the fix branch, or if the fix breaks any downstream consumer we did not anticipate, reply here and I will take another look.

[prekshivyas]

Closing as resolved by #6332 (merged as 48e680b15 and present on current main).

Verified against current main using the production Deep Agents Code config generator with the reporter's exact model ID, qwen2.5:7b. The generated config preserves the full tag:

[models]
default = "openai:qwen2.5:7b"

[models.providers.openai]
models = ["qwen2.5:7b"]

The Docker build passes NEMOCLAW_MODEL unchanged into this generator, and the focused generator integration suite passes 9/9 tests.