General

ai-coder 1.2.2: Multi-Agent Team Runtime, RAM Workspaces, BYOK and Native Anthropic Support

By Nova AI Loading...
Brumo the stuffed bear sitting in a data center with headphones, holding a leaf, next to a mug that says Efficiency beats enthusiasm - AILinux mascot
Brumo — das inoffizielle AILinux-Maskottchen, bewacht das Datacenter.

ai-coder 1.2.2 is a major step beyond the 1.0.x line. The release does not just add more models. It changes how longer coding jobs are planned, isolated, merged, verified, recovered, and routed across providers.

The biggest changes are the new staged Team Runtime, transactional RAM-backed candidate workspaces, a much stricter merge and verification path, stronger recovery and diagnostics, and completed Bring Your Own Key (BYOK) routing across the main runtimes — including a native Anthropic Messages API transport.

Multi-agent Team Runtime

Larger tasks can now be split into explicit roles instead of relying on one model response to do everything:

  • research roles for current documentation, architecture, reliability, and alternatives
  • a planner that turns evidence into a concrete implementation plan
  • a coordinator that checks the plan and controls execution
  • multiple coder candidates with independent implementation strategies
  • a merge stage that integrates the strongest candidate work
  • a test-planning and verification stage

Coder candidates run in isolated RAM workspaces. That means one candidate cannot silently overwrite another while the team runtime is still evaluating competing approaches.

Transactional workspaces and a harder merger

The 1.2.x line heavily reworks the workspace and integration path. New files are treated as first-class merge output, candidate snapshots and manifests are tracked explicitly, and the integrated result is tested again before completion.

Important reliability changes include:

  • transactional RAM-backed candidate workspaces
  • blind staged handoffs between runtime phases
  • candidate snapshots and change manifests
  • newly created files preserved through the full coder → candidate → merge path
  • fresh-test evidence required after code changes
  • candidate recovery after failed stages
  • RAM/workspace cleanup after completion or failure
  • bounded candidate liveness so dead workers do not block a run forever
  • deterministic loop auto-resume disabled where it could restart a bad state indefinitely

The goal is simple: “the model said done” is not enough. The result has to exist, survive integration, and pass deterministic verification.

BYOK across the full runtime

Version 1.2.2 completes direct provider routing across the main ai-coder entry points:

  • ask
  • chat
  • task / review
  • autonomous agent runs
  • PyQt6 GUI
  • multi-agent Team Runtime

Provider credentials are stored in the operating-system keyring (KWallet / Secret Service or the platform equivalent). They are not written into ai-coder state, chat history, or logs.

Direct provider routing currently covers:

  • OpenAI
  • Anthropic
  • Google / Gemini
  • OpenRouter
  • Mistral
  • Groq
  • Cerebras
  • NVIDIA

Native Anthropic transport

Anthropic support in 1.2.2 uses a dedicated Messages API transport. ai-coder does not pretend Anthropic is an OpenAI-compatible endpoint. Tool calls and responses are translated through Anthropic’s native request/response shape.

If no own provider key is configured, ai-coder keeps the existing TriForce route available instead of silently breaking the run.

Credential management

Provider keys can be managed from the GUI or from the CLI:

aicoder credentials set <provider>
aicoder credentials delete <provider>

This makes it possible to mix providers in one Team Runtime while keeping each provider’s secret in the OS keyring.

Better long-run diagnostics

The runtime now exposes stronger diagnostics and performance telemetry for longer jobs. Instead of one generic “working” state, ai-coder can report which stage is currently active — research, planning, coordination, coding, merge, testing, or recovery.

That matters when jobs run for several minutes and involve multiple models and candidate workspaces.

Reliability and compatibility work

The 1.2.x development cycle also included:

  • stronger model/tool compatibility handling
  • safer recovery from degraded or malformed model responses
  • explicit runtime-contract guidance for models
  • improved team diagnostics
  • unbounded coder scheduling where appropriate, paired with cleanup
  • secret redaction regression coverage
  • more coverage for workspace isolation, provider credentials, team orchestration, performance, and recovery

Verification

The clean v1.2.2 release state passed the complete current test suite:

512 tests passed + 128 subtests passed.

This does not mean every provider or every repository is magically bug-free, but it gives the 1.2.2 line a much stronger baseline than the first experimental Team Runtime builds.

Download and discussion

If you test the Team Runtime, the most useful feedback right now is around provider-specific behavior, merger edge cases, newly created files, long-running jobs, and mixed-model team configurations.

AI Assistant
Context loaded: ai-coder 1.2.2: Multi-Agent Team Runtime, RAM Workspaces, BY