Skip to content

Releases

haechi 0.3.0

Choose a tag to compare

@jisuzzzz jisuzzzz released this 19 Nov 04:32
adc807c

Summary

Haechi 0.3.0 adds a shared implementation completion check and repeatable verification after context compaction.

Changes

  • Use /haechi:complete to check for a passing review and executed verification tied to the current Git diff. The check holds completion when evidence is missing, stale, failed, or incomplete.
  • Run the Codex compaction replay with --complete to exercise the completion check in the recovered turn.
  • Expand Work Graph recovery coverage for interruptions, retries, and intervening cancellation.

Validation

The live Codex integration replay completed all four fixture tests, reviewed the final diff, and passed the completion check after automatic compaction in the same turn. The release commit passed GitHub CI.

haechi 0.2.0

Choose a tag to compare

@jisuzzzz jisuzzzz released this 11 Nov 00:51
3c252b1

Repository setup is optional

Start a task using your existing repository instructions and scripts. Run /haechi:setup when you want to check your installation or discover candidate commands.

Migrate existing repository settings

Support for .haechi.json ends in this release.

  • Put shared verification rules, forbidden commands and Git policy in AGENTS.md or your existing repository instructions. Keep executable commands in project scripts.
  • Pass Python source-to-test mappings with --python-test-map and a custom single-test command with --test-one when using the default verifier.
  • Pass additional protected branches with --protected and, if needed, a main-worktree Node.js cleanup script with --teardown. Pass generated-file and test exclusions to the logic-line counter as Git pathspecs.
  • Keep personal preferences in ~/.haechi/user.json.

The Claude Code and Codex packages both ship at version 0.2.0.

haechi 0.1.2

Choose a tag to compare

@jisuzzzz jisuzzzz released this 10 Nov 23:35
4e11e4f

Changes

  • Codex restores workflow rules, saved skill choices and shared work at session start and after compaction. Ordinary prompts repeat a component when its content or repository changes. Setup warnings, work graph warnings and subagent guidance stay visible. PR #31
  • Work graph create and update requests can be saved and recovered with a scoped request ID. Recovery preserves newer changes, and captured local evidence is checked by SHA-256. PR #30
  • The Codex and Claude packages ship as 0.1.2. Node.js 24.22.2 or newer is required.

Verification

GitHub CI passed for both pull requests and the release commit. The generated command tests cover startup order, compaction, resume, context changes and warnings. Work graph tests cover interrupted writes, duplicate requests, newer cancellations and damaged evidence.

Full changelog

haechi 0.1.1

Choose a tag to compare

@jisuzzzz jisuzzzz released this 10 Nov 05:06
d54f6d0

Changes

  • Clarify that next-action skill selection also applies while following a repository skill or task plan. Preserve existing selection conditions and reuse valid evidence and completed work.
  • Include the README discovery and workflow example updates merged since 0.1.0.
  • Publish both Claude Code and Codex distributions as version 0.1.1. Node.js 24 LTS (24.22.2 or newer) remains required.

Verification scope

Source type checking, generated-file consistency, and the build and rules regression tests pass. The pull request CI checks are verified before publication.

This release clarifies existing guidance. No model-behavior comparison was run, so reduced skill-selection omissions are not claimed.

Full changelog: v0.1.0...v0.1.1

haechi 0.1.0

Choose a tag to compare

@jisuzzzz jisuzzzz released this 07 Nov 22:49
8e8e82d

haechi 0.1.0

This minor pre-1.0 release combines common work obligations, preview updates, request reminders and clearer shared-work recording.

Upgrade requirement

  • Use Node.js 24 LTS, version 24.22.2 or newer, and keep it on current security patches. The minimum rises from Node 20; .nvmrc pins development to 24.22.2.
  • Existing work-graph JSON records and native session links keep their schema and location. No SQLite or data migration is introduced.

Changes

  • Common pre-action, implementation, verification/review and completion obligations have one canonical owner; task methods and explicit commands remain compatible.
  • Later requests receive an actionable reminder, and Codex full rules recover after compaction or invalid/missing state, with session/repository isolation.
  • Worker packets preserve required guidance and source hashes while grouping duplicate source headers and reusing applicable verification evidence.
  • Preview remains supporting guidance, adds shared UI-design principles, serves disposable prototypes on localhost and retains selected references for implementation.
  • Work-graph guidance clarifies registration/update timing, concise resumption context and preservation of valid user instructions and durable results.

Validation and scope

  • Node 24.22.2: type checking and 247 generated files checked against source; all three manifests identify 0.1.0.
  • All 30 test files executed separately on macOS: 367 passed, zero failures, two Windows-only skips. CI also checks the release candidate before publication.
  • Earlier controlled model and installed Codex observations are documented in docs/workflow-validation.md; they are not a new universal compliance or Claude-native guarantee. No runtime completion gate or automatic context-compaction policy is added.

PRs: #24 Node LTS; #25 common work, preview and reminders; #26 work-graph guidance; #27 version and release notes.

Co-authored-by: Codex codex@openai.com

haechi 0.0.23

Choose a tag to compare

@jisuzzzz jisuzzzz released this 06 Nov 14:18
d6b9180

Haechi now explicitly revisits playbook selection before implementing a supplied brief or repository procedure, and when a conversation becomes a behavior or contract change.

  • Read applicable guidance before the first source modification for the new work, including test-source changes.
  • Combine compatible repository procedures and playbooks while preserving ownership, scope and settled decisions.
  • Keep the existing local nonbehavior exception and performance-playbook restriction.

Both Claude Code and Codex distributions include the same two-clause clarification.

Validation: Node.js 24 build and generated-file check; 11 rules tests and 13 build tests passed. Limited Opus 5.5 evidence supports the approved omission-rate screen on partially reconstructed tasks; the stock evaluation report remains inconclusive. This release does not establish general selection reliability or full application/UI correctness.

PR: #23

haechi 0.0.22

Choose a tag to compare

@jisuzzzz jisuzzzz released this 05 Nov 13:58
027085d

Changes

  • Select playbooks and supporting skills proactively from their descriptions, without duplicating individual selection conditions in shared rules.
  • Strengthen the optional spec recommendation for interacting workflows and requirements synthesis. Feature recommends it more strongly; other playbooks can use it when helpful.
  • Keep settled requirements usable without a compulsory spec stage, saved document, repeated questions or new approval. Preserve performance authorization and local-copy exceptions.
  • Update both Claude Code and Codex distributions to 0.0.22.

Validation and limits

Repository checks and targeted tests passed. A 48-run initial-stage comparison retained observed primary playbook, preview and delegate reads; complex Feature spec reads increased from 0/2 to 2/2 while narrow settled Feature tasks proceeded directly.

Overall automatic-selection preservation and complete workflow compliance remain unverified. The whole-case comparison was rejected on optional spec-consideration and performance-plan confirmation criteria; failed results remain documented. This release does not claim general selection-rate improvement or verified behavior across all hosts.

See the evaluation record for the full results and limitations.

Merged in PR #22.

haechi 0.0.21

Choose a tag to compare

@jisuzzzz jisuzzzz released this 04 Nov 15:20
d3f89d5

Changes

  • Require preview before implementing visible UI or interaction changes, including requests that do not explicitly name preview. Show connected component behavior through its immediate result without expanding unrelated surrounding behavior.
  • Replace /haechi:plain with shared writing guidance and the optional clear-writing principle. Chat, progress updates and documents share these rules; unchanged guidance can be reused from context.
  • Select request for cross-team change requests, fact questions and bug reports; keep replies and unresolved questions in the same document.
  • Run mutation-check for explicit test hardening or a concrete coverage concern, without requiring it for every added test.
  • Include the previously merged strict TypeScript tool-source migration and retirement of the Python example (#20). Installed distributions still contain generated JavaScript.
  • Regenerate Claude and Codex distributions and update plugin metadata to 0.0.21.

Validation

  • Node.js 24.22.1: generated-file checks and 34 relevant tests passed locally, with no skipped tests.
  • PR #21 CI passed with 355 tests passing and two Windows-only cases skipped on Ubuntu; merged main CI also passed.
  • A frozen 32-execution comparison passed all 16 candidate trials; repeatable improvement was observed in request selection. Preview selection and writing already passed the baseline cases.
  • Four additional UI trials confirmed real browser interaction checks before the first product edit, with independent judgment and product-hash evidence.
  • Evidence scope: synthetic CLI tasks with available, permitted browser tools. Installed desktop plugin delivery, universal routing reliability and human readability have not been established. Concise reports and evidence-retention hashes are included in the repository; raw runs remain local.

haechi 0.0.20

Choose a tag to compare

@jisuzzzz jisuzzzz released this 04 Nov 02:05
f79c80b

Changes

  • Explain bug-fix alternatives by the code being changed: the reported action, the cause, a shared processing point, or data representation and ownership (#18).
  • Keep cause fixes as the default, compare meaningful alternatives, and state containment limits and authorization requirements for persisted-data changes.
  • Select screenshot and GIF tools by actual session capabilities and evidence needs; use connected browser exports when suitable and existing helpers or Playwright otherwise (#19).
  • Preserve controlled comparisons, original recordings, timing limitations, and saved-file verification.
  • Regenerate Claude and Codex distributions and bump plugin manifests to 0.0.20.

Validation

  • Generated distributions match source (npm run check).
  • Relevant build, rules, agent-procedures, and document-template tests passed locally.
  • Pull request CI passed before merge.
  • These changes update instructions. Actual host-agent adherence and live browser recording with the new guidance have not been evaluated.

haechi 0.0.19

Choose a tag to compare

@jisuzzzz jisuzzzz released this 03 Nov 07:35
723110c

Changes

  • Use a compact preview header with borderless monochrome pill controls, English labels, and embedded Pretendard for offline previews (#16).
  • Open multiple variants side by side by default and preserve intrinsic component dimensions.
  • Update preview and feature guidance to show interactive proposals for visible UI changes and keep one set of controls.
  • Include the Pretendard font license and regenerate both Claude and Codex distributions.
  • Bump both plugin manifests to 0.0.19 (#17).

Validation

  • npm run build regenerated 243 files; npm run check confirmed they match source.
  • 18 build and preview-scaffold tests passed; git diff --check passed.
  • Chromium checks covered badge states, variant/state switching, comparison, and mobile overflow.
  • Release PR CI passed.