This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Topics
+13 more
Summary
AI summaryAdds declared multi-agent topology with reusable role bundles and per-stage agent references.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Feature | Medium |
Adds top-level `agents:` map for reusable role bundles with persona and model settings. Adds top-level `agents:` map for reusable role bundles with persona and model settings. Source: llm_adapter@2026-06-15 Confidence: high |
— |
| Feature | Medium |
Adds per-stage `agent:` reference on `infer` and `repair` stages, resolved into concrete inference fields. Adds per-stage `agent:` reference on `infer` and `repair` stages, resolved into concrete inference fields. Source: llm_adapter@2026-06-15 Confidence: high |
— |
| Feature | Medium |
Adds inline `system:` persona field on `infer`/`repair` stages, prepended as a system message when needed. Adds inline `system:` persona field on `infer`/`repair` stages, prepended as a system message when needed. Source: llm_adapter@2026-06-15 Confidence: high |
— |
| Feature | Medium |
Adds per-role `agent` field on stage trace and event records for attribution. Adds per-role `agent` field on stage trace and event records for attribution. Source: llm_adapter@2026-06-15 Confidence: high |
— |
| Feature | Medium |
Updates `inspect --format json` to report the resolved `agent` for each stage. Updates `inspect --format json` to report the resolved `agent` for each stage. Source: llm_adapter@2026-06-15 Confidence: high |
— |
| Feature | Low |
Provides offline multi‑agent examples: generator/critic/reviser, planner/executor, debate/judge, and triage/specialist handoff. Provides offline multi‑agent examples: generator/critic/reviser, planner/executor, debate/judge, and triage/specialist handoff. Source: llm_adapter@2026-06-15 Confidence: high |
— |
Full changelog
llmff v1.2.0
llmff v1.2.0 adds declared multi-agent topology to manifests: reusable agent
role bundles and a per-stage agent: reference that expands at inspect time into
a bounded, inspectable DAG. It names the roles in a pipeline without becoming an
agent orchestrator.
Added
- Top-level
agents:map of reusable role bundles, each a persona (system)
plus model and sampling settings (model,temperature,top_p,
max_tokens,seed,stop,sampler,response_format). - Per-stage
agent: <name>reference oninferandrepairstages, resolved
byManifest::resolve_agentsinto concrete inference fields. Resolution
recurses into loop and map bodies, and stage-level fields always override the
referenced agent's defaults. - Inline
system:persona field oninfer/repairstages, prepended as a
system message when the assembled messages do not already lead with one. - Per-role
agentfield on stage trace and event records so a supervisor can
attribute work to a named role. inspect --format jsonstage views now report the resolvedagentfor each
stage.- Offline multi-agent examples: generator/critic/reviser, planner/executor,
debate/judge, and a triage/specialist bounded handoff overroute. - Compatibility with the existing Python subprocess supervisor, batch supervisor,
Node.js streaming supervisor, agent runner adoption guide, OpenTelemetry bridge,
and ecosystem readiness gates.
Boundary
Declared multi-agent topology is pure expansion sugar over the existing graph.
agents: names the roles; it never decides which role runs next at runtime.
Dynamic dispatch — a role choosing an undeclared or unbounded successor, spawning
sub-agents, or driving open-ended agent-to-agent handoff — stays above the
boundary with the host. llmff remains a bounded execution runner, not a
multi-agent orchestrator, scheduler, memory system, or autonomous planner.
Install
cargo install --git https://github.com/syndicalt/llmff --tag v1.2.0 llmff
scripts/smoke-install.sh --git https://github.com/syndicalt/llmff --tag v1.2.0
Release preflight
Run the release preflight gate before creating or pushing the tag:
scripts/release-preflight.sh v1.2.0
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About Llmff
All releases →Related context
Related tools
Beta — feedback welcome: [email protected]