Skip to content

Release v1.2.0 — Headroom, provider-aware defaults, workflow/observability fixes

Choose a tag to compare

View all tags
@Mehrad-shyftlabs Mehrad-shyftlabs released this 20 Jul 18:51
· 200 commits to main since this release
v1.2.0
5661abe
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

compression: Context & compression

  • Headroom Compression (v1.2.0, optional) — per-turn engine that shrinks bulky tool output (logs, tables, search/RAG dumps, prose) and collapses stale/redundant file reads before the model sees them. Two modes: local (in-process) or endpoint (sidecar). Includes CCR reversible retrieval (continuum_headroom_retrieve) to pull compressed content back on demand, forgery-guarded. Off by default; per-role token savings logged.

robot_face: Models & providers

  • Provider-aware default model (v1.2.0) — with DEFAULT_LLM_MODEL unset, the chat/meta-op default follows whichever key is set, so Anthropic-/Gemini-only deployments need no OpenAI key for chat. New ANTHROPIC_DEFAULT_MODEL / GEMINI_DEFAULT_MODEL.
  • Open LLM provider registry (v1.1.0) — register_provider(prefix, factory) extends model-name routing without editing core (longest-prefix match).
  • Configurable temperature across all workflows (v1.2.0) — temperature: float | None; None omits it (provider default).
  • Claude 4.6+ compatibility (v1.1.0) — auto-strip-and-retry when adaptive-thinking models reject temperature; correct 200k/1M context windows for new Claude/Gemini families (was a ~48× underestimate).

electric_plug: Connectors & infra

  • Connector module (continuum.connectors) (v1.1.0) — uniform pluggable layer for Redis / vector store / Temporal / Langfuse with one
    interface, mode inference (local_docker/cloud/custom)

floppy_disk: Sessions & resilience

  • In-memory session provider (provider="memory") (v1.-semantics mirror; never raises connection errors.
  • SESSION_FALLBACK_MODE (v1.1.0) — degrade (keep serving via in-memory when Redis is down) vs fail; plus lazy session init and
    persistence-degraded observability (flag, health chec
  • Loud session preflight (v1.2.0) — passing a session_id for an uncreated session now raises SessionNotCreatedError (or warns) instead of
    failing silently.

brain: Memory

  • MEMORY_MAX_QUERY_CHARS (v1.2.0) — bounds search queids hard-fail / silently-empty results at the tokencap).

🛠️ Reliability & correctness (highlights)

  • Workflow agents dispatched to execute() (v1.2.0) — nested workflows were being flattened into a single tool-less LLM call; run_stream()
    too.
  • Workflow ERROR no longer laundered into SUCCESS (v1.2.0) — circuit-breaker errors aren't chained forward or persisted.
  • Per-request trace grouping (v1.2.0) — one agent-runttach as events, not orphan error-UNKNOWN_ERROR traces.
  • Streaming parity (v1.1.0) — structured_output populutput PII scanners now run in streaming (closed aredaction bypass); handoff on_end hooks fire; return_to_parent loop guard.
Assets 2
Loading