build-agent: Model & Harness config not persisted for simple (no-tool) agents — tool-wired agents unaffected

[bekossy]

Summary

Every agent created by the build-agent skill has an empty harness and model config in the Agenta Playground, making it immediately broken. The skill's own success message is false.

Tested against: Agenta host https://bighetzner.agenta.dev, workspace mahmoud/Default.

---

Repro

1. Run the build-agent skill with any prompt (tested with multiple prompts across Claude Code and Codex).
2. Skill transcript claims agent was "created and registered" and may even claim it "updated the agent config to use the OpenAI-backed provider."
3. Open the Agenta Playground for the created agent.
4. Observe: Harness = unselected, Model = "Select a provider" — config is empty.
5. Try to invoke the agent — fails with pi_core: model authentication failed, which is misleading because the real cause is no model being configured at all.
6. Manually select Harness = Pi + any model in the UI → agent immediately works.

Confirmed agents:

• support-ticket-summarizer (app 019f33ef-0520-7a11-8011-c032968ff71b)
• complaint-reply-agent (app 019f3466-c979-7002-a34c-0cc00cdaaef2)
• A third agent (same pattern)

Reproduced: 3/3 times, across different prompts and different coding agents (Claude Code, Codex).

---

Required fix

The skill must explicitly set harness + model at creation time via the API, and must verify the resulting config is non-empty before declaring the agent "created successfully." As-is, the skill's own success message is false.

---

Environment

Agenta host	https://bighetzner.agenta.dev
Workspace	mahmoud/Default
Coding agents tested	Claude Code, Codex
Skill version	build-agent from Agenta-AI/agenta-skills (current main)

[bekossy]

Update — this does NOT reproduce universally.

Confirmed broken: simple, non-tool agents built on the 'Pi' harness (support-ticket-summarizer, complaint-reply-agent) — Model & Harness left completely unset.

Confirmed working: a tool-wired agent (support-ticket-digest, app 019f3498-c340-7242-90ec-fe11a426326f) built via the discover-tools.sh path, which correctly landed on the 'Claude Code' harness with Model=sonnet, self-managed/sandboxed.

So the bug appears specific to the simple-agent build path (build.sh with a plain provider/model JSON config), not tool-using agents. Retitling to reflect this narrower scope.

[bekossy]

4th independent reproduction, with a stronger detail: re-ran the exact original prompt ('build an Agenta agent that summarizes support tickets into 3 bullet points') in a fresh session. New agent ticket-bullet-summarizer (app 019f36aa-5f0d-7f13-95f0-3420913788a8, variant 019f36aa-5f1c-77c3-80d1-8347b4c9d60f) again came out with Harness unselected and Model = 'Select a provider' in the Agenta Playground.

Crucially, this run also proves the misleading-error problem goes further than we thought: the project vault already had both a valid OpenAI key (added earlier in this session) and a valid Anthropic key configured — confirmed directly in the Agenta LLMs settings page — yet the agent's test run still failed with model authentication failed, and the skill's own remediation message told the user to 'add your OpenAI key,' which was simply false; the key was already there. This confirms the root cause is exclusively the missing harness/model binding on the created agent — it has nothing to do with vault state at all, and the skill's error-remediation logic should stop asserting a specific missing-credential fix it hasn't actually verified. It should instead check (or at least caveat) whether the real issue might be an unconfigured harness/model before telling the user to touch the vault.

[bekossy]

Root cause confirmed by fetching the actual stored config for a broken agent directly via the API: the config paired harness.kind: "claude" (Claude Code harness, Anthropic-only) with llm.provider: "openai", llm.model: "gpt-4o-mini". That combination is accepted by the create/commit endpoints with no error, but never resolves to a working Model & Harness in the UI.

Opened #9 with a fix: create-agent.sh / update-agent.sh now validate this pairing (and, for creation, that the full boilerplate is present) before calling the API, and SKILL.md/config-schema.md now explain the coupling explicitly. Verified end to end against a live instance — see the PR for details.