Agently 4.1.4.4 Release Notes

Agently 4.1.4.4 strengthens two existing owner boundaries: ModelRequest remains responsible for structured-output prompting and validation, while TriggerFlow and its configured persistence provider remain responsible for recoverable execution snapshots. No new facade or parallel runtime was added.

Structured output constraints and correction

Pydantic v2 field constraints are now projected into supported structured-output prompt formats, including requiredness, nullability, aliases, enum/literal values, string and collection lengths, numeric ranges, patterns, and formats. The original model remains the final acceptance authority.

from pydantic import BaseModel, Field


class Ticket(BaseModel):
    title: str = Field(min_length=3, max_length=80)
    priority: int = Field(ge=1, le=3)
    labels: list[str] = Field(min_length=1, max_length=5)


ticket = (
    agent
    .input("Turn this incident report into a triage ticket.")
    .output(Ticket, format="json")
    .get_result()
    .get_data_object()
)

If parsed data fails Pydantic validation, Agently sends bounded field-level correction feedback into the existing retry path. An invalid dictionary is not returned as successful business data, and an accepted retry remains reusable through object, data, and text result readers.

Recovery-aware TriggerFlow snapshots

Issue #331 showed that repeated save/load cycles could retain large completed signal values in every recovery snapshot. 4.1.4.4 adds an opt-in projection policy for eligible terminal values:

execution.set_snapshot_projection_policy(
    terminal_value_mode="digest",
    min_value_bytes=4096,
)

Pending interrupts and state required for recovery remain complete. Projected terminal values retain a canonical SHA-256 digest and encoded size so duplicate or conflicting resume requests can still be checked. Schema-v2 snapshots can load prior schema-v1 full-value snapshots.

The built-in local RecordStore now keeps the latest three execution snapshot versions per run_id by default:

record_store = RecordStore(
    "./recovery",
    snapshot_retention={"keep_last": 5},
)

execution.set_snapshot_retention_policy(keep_last=2)
report = await execution.async_prune_recovery_snapshots(keep_last=1)

Use {"keep_last": None} to disable automatic pruning. Projection is owned by TriggerFlow because it understands recovery semantics; physical retention is owned by the persistence provider. Generic put_checkpoint(...) writes are not automatically pruned.

Release validation policy

Default pytest no longer depends on a locally running Ollama service or a pre-pulled Ollama model. Deterministic OpenAI-compatible protocol tests remain in the normal suite. Real-model release evidence is run separately against an explicitly configured online model so the provider, model, request count, and observed result can be recorded honestly.

Core changes and upgrade impact

Area	What changed	Recommended usage	Compatibility and risk	Evidence
Structured output	Supported Pydantic field constraints enter prompts and validation failures enter bounded retries.	Keep the BaseModel class as the .output(...) contract.	Additive correction of previously under-specified prompts; custom validators remain host-side only.	Prompt-generator, validation, result-reuse, and typing tests.
Snapshot projection	TriggerFlow can digest eligible terminal interrupt values and completed resume metadata.	Opt in for executions whose completed values dominate snapshot size.	Full values remain the default; pending recovery data is never projected.	Issue #331 A/B run: 1,307,086 B default versus 106,782 B digest projection, a 91.83% reduction.
Snapshot retention	The built-in local provider keeps the latest three versions by default.	Configure provider defaults or an execution-level override; use explicit prune for maintenance.	Intentional local-provider default change. keep_last=None preserves all versions.	Retention, override, save/load, registry, and provider-port tests.
Model validation	Local Ollama calls are removed from default pytest.	Use explicit bounded online-model experiments for release evidence.	Test-policy change only; Ollama remains usable as a configured OpenAI-compatible endpoint.	Deterministic mock coverage plus the release evidence record.
Deferred	Whole-snapshot byte ceilings, artifact-reference offloading, and distributed-provider retention implementations are not included.	Keep large business artifacts outside recovery snapshots and implement retention at each persistence provider boundary.	No claim of a universal snapshot size limit.	Explicit limitation from the #331 experiment and owner-boundary review.
Deferred	Concrete gVisor and Seatbelt providers tracked by #324 remain contributor-owned.	Use the released provider-neutral ExecutionResource seam or an explicitly authorized provider.	Not a 4.1.4.4 core-runtime blocker; no unreviewed sandbox implementation is bundled.	Issue #324 and PR status review.

Validation

Observed release-candidate validation:

• source Pyright over agently/, tests/, and examples/: 0 errors;
• clean-worktree default suite: 2,438 passed and 27 skipped; 25 skips are maintainer-local spec-runner checks and all 25 passed separately with the nested spec repository mounted, while the remaining two require an optional Anthropic Skills checkout;
• all three release-pinned deterministic usage scripts passed;
• the TriggerFlow durable-recovery example preserved load, latest-N retention, explicit prune, idempotent resume, and durable-event effects;
• wheel and source distribution built successfully; a fresh Python 3.10 environment installed the wheel, found py.typed, exercised the structured missing-dependency error, and passed an installed-package Pyright smoke;
• one bounded DeepSeek deepseek-v4-flash request returned the declared Pydantic model in one request, preserved priority and labels, and satisfied all length/count/range constraints. Deterministic tests remain the evidence for invalid-first-attempt correction and retry reuse.

Compatibility

• Package version: 4.1.4.4.
• Release manifest: compatibility/releases/4.1.4.4.json.
• Python: >=3.10.
• Recommended DevTools version remains agently-devtools >=0.1.10,<0.2.0.
• Skills authoring protocol remains agently-skills.authoring.v2.