AICoder 1.4.0 released — Teamrun bugs fixed, provider-aware execution and persistent memory

AICoder 1.4.0 is released. This release is the result of a long end-to-end debugging session around Team Runtime, MCP authentication, provider limits and persistent project memory.

The goal was simple: a real multi-agent Teamrun must be able to plan, research, implement, test, merge and finally write a verified result back to the source workspace — without getting stuck in read loops, losing tools or hammering one provider until it rate-limits.

What was broken — and what is fixed

1. Coding candidates could get trapped in read-only loops

The Team Runtime could repeatedly call file_tree, file_read and git, hit the semantic-stall guard, resume, and then repeat the same inspection cycle before ever reaching file_edit.

Fix: the implementation nudge now fires before the semantic-stall pause, giving the coding candidate a real chance to mutate the workspace. The repair phase also keeps the required write/repair tools available.

2. Parallel team roles could overload the same provider

Research, brainstorm and coding candidates were started in parallel without provider-aware throttling. Two roles using the same backend could therefore hit Groq, Cloudflare or another provider at the same time and trigger avoidable 429/rate-limit failures.

Fix: Team Runtime now serializes workers per provider while still allowing different providers to execute in parallel. The limit is configurable and emits queue events so waiting roles are visible instead of looking frozen.

3. Large tool catalogues could starve coding tools

With the full TriForce catalogue loaded, the agent could spend too much context on tool discovery and inspection.

Fix: Team Runtime progressively discloses the tool catalogue while preserving the write/repair path needed by coding candidates.

4. Fresh repositories could inherit the wrong project assumptions

AGENTS.md in a greenfield test repository could be interpreted like existing project metadata and push the agent toward the wrong assumptions.

Fix: greenfield metadata handling has been hardened.

5. MCP tools could fall back to the recovery catalogue

A valid AICoder account JWT reached /v1/mcp, but the outer TriForce auth middleware could reinterpret it as a normal MCP bearer before the route-owned dual-auth dependency saw it. That produced HTTP 401 and left AICoder with only the small recovery tool set.

Fix: the MCP endpoint now correctly accepts the AICoder JWT through the route-owned authentication layer. The same route-owned JWT handling was fixed for the shared notification/presence endpoints.

6. NVIDIA NIM returned 404 for valid models

TriForce sent model IDs such as nvidia/openai/gpt-oss-20b to NVIDIA NIM without stripping the internal provider prefix. NVIDIA expects openai/gpt-oss-20b.

Fix: the central model helper now strips the nvidia/ prefix. NVIDIA GPT-OSS requests also default to low reasoning effort for agent workloads so reasoning does not consume the entire response budget before final content is produced.

7. Teamrun state disappeared after the run

Project Memory and Claude-Mem already existed in AICoder 1.3.0, but Team Runtime was not writing its lifecycle into them.

Fix: every Teamrun now owns one canonical Project Memory entity. Stage checkpoints update that entity and accepted revisions are synchronized through TriForce into Claude-Mem. Memory is fail-open: a memory outage never aborts the actual coding run.

Verified end-to-end

The final smoke test completed the full pipeline:

plan_research
research
brainstorm
plan_code
code
merge_plan
merge
plan_tests
tests_function_ok
atomic_disk_write
run_completed

The coding candidate created the requested implementation and regression test, the isolated candidate passed verification with score 100, the integration workspace accepted it, final tests passed and the result was atomically written back to the source repository.

The memory test produced 12 accepted Teamrun revisions with 0 conflicts. Project Memory ended at version 12 / server sequence 14 and Claude-Mem received the same 12 append-only historical revisions.

Proven provider layout

Planner      → ChatGPT GPT-6-Sol
Coder        → ChatGPT GPT-6-Sol
Coordinator  → NVIDIA GPT-OSS 20B
Research     → Ollama Gemma 4 Cloud
Merge        → Ollama Gemma 4 Cloud
Tests        → Ollama Gemma 4 Cloud

This is not a hard-coded requirement; it is simply the provider layout that was used for the successful release verification. Provider-aware concurrency remains active so the system can safely mix additional backends.

Downloads

AICoder 1.4.0 is available for Linux, Debian/Ubuntu, Windows, macOS and Android/Termux.

GitHub Release: AICoder v1.4.0

Linux binary SHA256: b1490824a0ae3f5c477a0c7982fa385300541534d465275aabcf485397d6a2b0
Debian package SHA256: d1e6934e384990b259c68dd296767ff25b56b9e391ea3b097cea296097663bb8

What comes next

The next major Team Runtime step is resume-from-checkpoint: detect an unfinished Teamrun in Project Memory and continue from the last safe stage instead of starting again at research/planning.

For us, 1.4.0 is the point where Team Runtime stopped being a collection of promising pieces and completed a real, persistent, multi-provider coding workflow from plan to verified disk write.