Release Agently 4.1.4.4

[Maplemx]

Release

• Version: 4.1.4.4
• Level: patch
• Dev head: 395ffa67
• Release-candidate commit: 704fcc11
• Current PyPI package before promotion: 4.1.4.3
• Status: draft pending human release-note review and final promotion confirmation

User-visible changes

Structured output

• Project Pydantic v2 requiredness, nullability, enum/literal, string/list length, numeric range, pattern, and format constraints into supported structured-output prompts.
• Feed bounded field-level Pydantic validation failures into the existing retry path.
• Preserve accepted typed results across object, data, and text readers.
• Correct public overload typing for optional ensure_keys.

TriggerFlow recovery

• Add opt-in terminal digest projection for eligible completed interrupt/resume values while keeping pending recovery complete.
• Keep projection separate from RuntimeEvent compaction.
• Make the built-in local RecordStore keep the latest three execution snapshot versions per run_id by default.
• Add provider configuration, execution-level override, and explicit execution-bound prune.
• Preserve schema-v1 load compatibility, restart recovery, and resume-request idempotency.

Release validation

• Remove live Ollama dependencies from default pytest.
• Keep deterministic OpenAI-compatible protocol tests.
• Use explicitly configured online models for final release evidence; local Ollama is supplementary only.

Coverage-first acceptance

Requirement	Evidence	Result
Pydantic constraints reach prompts and final validation	prompt-generator, validation, result-reuse, and typing tests	accepted
Invalid typed data enters bounded correction	deterministic invalid-first-attempt/retry tests	accepted
#331 snapshot growth is materially reduced without losing recovery	1,307,086 B default vs 106,782 B digest projection; load/resume/idempotency checks	accepted, opt-in
Physical snapshot versions are bounded/configurable	retention/override/provider tests and durable-recovery example	accepted
Default pytest is independent of Ollama	clean-worktree full suite	accepted
Real online structured output satisfies the declared contract	one DeepSeek deepseek-v4-flash request	accepted
Docs, package, compatibility, and companion guidance agree	bilingual notes/README/manifests plus AgentEra/Agently-Skills#8	accepted

Validation

• Python: 3.10.13
• Source Pyright over agently/, tests/, and examples/: 0 errors
• Clean-worktree default pytest: 2,438 passed, 27 skipped in 245.49 s
  • 25 skips are maintainer-local nested-spec runner checks; all 25 passed separately
  • 2 skips require an optional Anthropic Skills checkout
• Focused release, compatibility, public-typing, and workflow gates: 20 passed
• Release-pinned deterministic usage scripts: 3 passed
• TriggerFlow durable-recovery example: passed; load, latest-N retention, explicit prune, idempotent resume, and durable-event effects observed
• Source distribution and wheel: built successfully
• Wheel SHA-256: a8978a1f2088ef8bcbaaa2f3cd3f5fdb4299d271f1e181d3611b5c79bf8cda49
• sdist SHA-256: 08d81e8c17269b81fc2ddd7b96aa8df5abcfb9c1add1a2cd8b66f209907f1669
• Fresh Python 3.10 wheel install:
  • ordinary Agently, Agent, and TriggerFlow startup passed
  • py.typed present
  • intentional missing optional dependency returned lazy_import.dependency/v1
  • installed-package Pyright: 0 errors

Online-model evidence

Accepted bounded run:

• Provider/model: DeepSeek / deepseek-v4-flash
• Model requests: 1
• External requests: 1
• Tool requests: 0
• Elapsed: 2.806 s
• Usage: 255 prompt + 241 completion = 496 tokens
• Parsed result: original Pydantic ReleaseTicket
• Priority, labels, title length, summary length, collection count, and numeric range: all satisfied
• The output separated the prior snapshot-duplication problem from the implemented resolution and preserved requested priority/labels

Deterministic tests, not this single model run, remain the acceptance evidence for invalid-first-attempt correction and retry reuse.

Compatibility and companions

• Release manifest: compatibility/releases/4.1.4.4.json
• Python: >=3.10
• DevTools recommendation remains agently-devtools >=0.1.10,<0.2.0
• Runtime/observation protocol is unchanged; no Agently-Devtools code change is required
• Skills authoring protocol remains agently-skills.authoring.v2
• Companion guidance PR: Align guidance with Agently 4.1.4.4 Agently-Skills#8

Issue and PR disposition

• TriggerFlow 4.1.4.3 snapshots multiply completed interrupt/resume payloads and provide no bounded retention API #331: reproduced and fixed in this candidate; keep open without solved until PyPI/GitHub release verification
• Should local sandbox capabilities use system adaptation? 本地沙箱能力是否应该使用系统适配？ #324: remains open as enhancement + help wanted for contributor-owned gVisor/Seatbelt providers; not a core release blocker
• CCS 协议整合 — GenAI 应用开发的安全基础设施 #319: closed as not planned after owner-boundary/evidence review
• Your project is now listed on CodeGuilds #307: closed as a listing notification
• Correct MiniMax Anthropic-compatible base URLs #321: changes requested because the proposed MiniMax base URL would drop the required /v1; excluded
• Feature/add sandbox plugin #323, Feature/gvisor docker runtime #325, Feature/seatbelt provider #327: conflicting contributor work; excluded

Deferred scope and residual risk

• Snapshot projection is opt-in; full values remain the compatibility default.
• last_signal, pending payloads, and application state may still be large; there is no whole-snapshot byte ceiling.
• Distributed providers must implement and prove retention separately.
• Two optional Anthropic Skills checkout tests were not run.
• PyPI 4.1.4.3 exists without a matching GitHub tag/release; promotion bookkeeping should explicitly reconcile that historical gap.

Required human review before merge

• Confirm the English and Chinese 4.1.4.4 release notes describe the actual shipped scope and limitations.
• Confirm the intentional local RecordStore default retention change is acceptable.
• Confirm the historical 4.1.4.3 GitHub release/tag gap handling.
• Confirm companion PR Bugfix #8 promotion timing.
• Merge only when publishing 4.1.4.4 is intended: the main-branch version change triggers the PyPI workflow.

Post-merge

1. Verify the PyPI publish workflow and installed 4.1.4.4 package.
2. Create/verify the GitHub v4.1.4.4 tag and release.
3. Merge/promote companion PR Bugfix #8.
4. Comment on TriggerFlow 4.1.4.3 snapshots multiply completed interrupt/resume payloads and provide no bounded retention API #331 with the published version/commit/validation and add solved.