CHANGELOG.md

Changelog

日本語

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog.

[0.45.0] - 2026-06-10

Added

• Phase-level usage events and a usage analysis script (#785). With observability.enabled: true and observability.usage_events_phase: true, each run writes per-phase token usage events to .takt/runs/<run>/logs/<session>-usage-events.phase.jsonl, as a separate stream from the existing logging.usage_events output. Events are grouped by workflow phase (phase1_execute, phase2_report, and the status-judgment variants phase3_structured / phase3_tag / phase3_fallback), and calls whose usage is unavailable are recorded with usage_missing: true instead of being counted as zero tokens. A new npm run analyze:usage script aggregates one or more event files or run directories into a Markdown or CSV table keyed by step × phase × provider × model, with token totals and per-call statistics. Documented in the new Observability guide.
• Clipboard image paste in interactive mode (#791). Pressing Ctrl+V (or running the new /paste-image command) during interactive input now attaches an image directly from the OS clipboard; previously paste only worked for terminals that emit inline-image (OSC 1337) escape sequences. Attached images also flow through the provider abstraction now: claude-sdk and codex receive them as native image input, while the other providers get the attachment file paths referenced in the prompt so the agent can open them with its own tools.

Changed

• BREAKING: Interrupted running tasks are no longer auto-requeued (#791). When takt run or takt watch was interrupted (process crash, kill), tasks left in running status used to be recovered to pending and re-executed on the next invocation. They are now marked failed with an explanatory error instead; requeue them explicitly to run them again.

Fixed

• Worktree-isolated clones no longer fail with missing-object errors when branching off a fetched base-branch commit (#791). The isolated clone now fetches the base branch's commits from the main repository before running git reset --hard, so the reset target is always present in the clone.
• OpenCode readonly permission mode now allows read tools (#797). The readonly mode denied every tool — including read, glob, and grep — so read-only steps such as reviews could not inspect the codebase at all. Those three read tools are now allowed while edit, bash, and network tools stay denied.

Internal

• Claude Agent SDK and Codex SDK dependency updates (#789, #795).
• Removed the takt-quality-check command gate from the implement-type steps in the repository's own .takt/config.yaml.

[0.44.0] - 2026-06-03

Added

• kiro provider added (#773). TAKT can now drive the Kiro CLI as an AI agent provider alongside Claude, Codex, OpenCode, Cursor, and Copilot. Select it with --provider kiro or in config. Authentication uses kiro_api_key in config (or the TAKT_KIRO_API_KEY env var), and the CLI binary can be overridden via kiro_cli_path / TAKT_KIRO_CLI_PATH.
• OpenTelemetry observability gained working exporters and richer spans (#753). Building on the span foundation shipped in 0.42.0, observability now emits a local monitor.json of per-run workflow metrics (enable with observability.monitor: true) and a shadow session log derived from OTel spans (observability.sessionLogExporter: true), and the span set was extended to cover phase execution and status-judgment (judge) phases. Exporters are routed per run id and the shadow session log keeps redaction parity with the canonical NDJSON session log, so sensitive agent output stays sanitized. Still off by default behind observability.enabled: true.

Changed

• Coding review extended to all builtin review and development workflows. The coding-review parallel sub-step (the coding-reviewer persona with the review-coding instruction and coding-review output contract) — previously only on default-peer-review — is now appended to every builtin review / review-fix workflow and to the parallel reviewer waves of the development workflows (backend, frontend, dual, terraform, and their variants). It is a general-purpose, near-instruction-less pass that flags implementation bugs, regressions, security risks, and missing tests using the model's own coding judgment. The intentionally-minimal *-mini and compound-eye variants are left unchanged.

Fixed

• Codex Reconnecting... events no longer abort the run (#775). A transient reconnect from the Codex SDK could surface as a fatal provider error and tear down the whole workflow; the Codex client now treats it as a recoverable reconnect and retries.
• Worktree-clone isolation hardened (#778). Fixes to the git clone --shared isolation path and clone execution, including normalized gitdir isolation handling, so worktree-isolated tasks stay properly isolated from the main repository.

Internal

• Repository quality gates wired into TAKT's own .takt/config.yaml so the dogfooded review/dev steps run the build, lint, unit, and mock-E2E checks via a command gate.

[0.43.0] - 2026-05-29

Added

• Image attachments (experimental, #751). Images can now flow through TAKT end to end. In interactive mode you can paste an image straight into the prompt — TAKT decodes terminal inline-image (OSC 1337) sequences into a pending attachment — and takt add plus retries carry image attachments with the task spec, so the agent receives the images alongside the text instruction. Works in assistant / passthrough / quiet / retry input modes, with a 10 MB per-image limit. Still under verification — behavior may change.
• takt resume command for direct runs (#759). A direct (one-shot, non-queued) run that fails or aborts can now be resumed with takt resume, which finds the latest failed/aborted direct run and continues it instead of starting over. Resume reuses the existing run directory, and a dedicated scoring prompt decides how to re-enter the workflow.
• Command quality gates (#761). Step quality_gates now accept machine-executed type: command entries in addition to AI directive strings. A command gate runs after an agent step completes and passes only when the command exits with code 0; on failure TAKT feeds the command metadata, cwd, exit code (or timeout / output-limit details), the output log path, and bounded sanitized stdout/stderr back into the same agent step for another attempt. Workflow-YAML command gates require workflow_command_gates.custom_scripts: true in config. system and workflow_call steps do not accept quality_gates.
• frontend-maintenance builtin workflow (experimental). A workflow for modifying existing frontend products rather than greenfield builds, shipping maintenance-scoped plan / implement / write-tests / fix / supervise instructions, an existing-system knowledge facet, an existing-system-respect policy that enforces respecting current conventions, and a maintenance-scope output contract. Experimental — it can currently be heavier-handed than intended, so treat it as a starting point for existing-product changes and tune to your codebase.
• Coding review added to the default peer review. The builtin default-peer-review workflow now includes a coding-review sub-step backed by a new coding-reviewer persona, a review-coding instruction, and a coding-review output contract, so general code quality is reviewed alongside the existing specialist reviewers.

Changed

• Review facets hardened against scope creep across the fix ↔ review loop (en + ja). The review baseline is now anchored to the task's base — every reviewer evaluates the entire cumulative diff from the merge-base, not just the increment from the most recent iteration. This stops unrequested changes that slipped in during earlier iterations (unrelated comment deletions, renames, reformatting, weakened tests) from being missed once the diff narrows to the latest fix. The ai-antipattern scope-creep check was made cumulative-diff based, and review and React facet guidance was refined alongside.

Fixed

• claude-terminal prompt detection now works with Claude Code v2.1 (#766, refs #765). The tail-line regex required the prompt row to be exactly ❯ / ❱ / >, but v2.1 renders it as ❯ Try "...", so waitForClaudeInputReady never matched and timed out after 60 s. The pattern was relaxed to accept any prompt character followed by whitespace or end-of-line, while the busy-state gate still prevents false positives.
• Codex Reconnecting... is no longer treated as a fatal error (#767). A transient Reconnecting... N/5 event from the Codex SDK was surfaced as a final provider_error and aborted the whole workflow; it is now handled as a recoverable reconnect and retried.
• team_leader part timeouts no longer abort the run (#764). When a worker part hit part_timeout or a feedback failure, TeamLeaderRunner aborted immediately; a timeout fallback now keeps the run going, and the decomposition instructions and facets were tuned so the leader is less likely to create one oversized part.
• Parallel review aggregation no longer fails the whole step when a single reviewer errors (#770). A provider_error in one reviewer's Phase 1 during a parallel reviewers step previously broke aggregation even when the other reviewers completed; ParallelRunner terminal-status handling was fixed so the step aggregates correctly.
• review-fix-takt-default now routes supervisor findings correctly. Findings raised by the review-fix supervisor were not routed back into the fix loop as intended; the workflow rules were corrected.

Internal

• Added a three-phase step model tutorial to the docs (#735).

[0.42.0] - 2026-05-20

Added

• claude-terminal provider added (#727). A new way to run Claude that drives an interactive Claude Code CLI session inside a tmux pane and reads results back from the session transcript, instead of calling the Anthropic SDK (claude-sdk) or the headless CLI (claude). Select it with --provider claude-terminal or in config. It supports structured output, MCP servers, and allowed-tools, and surfaces permission / ask-user-question prompts back through the terminal. Provider options live under provider_options.claude_terminal (backend: tmux, timeout_ms, keep_session, transcript_poll_interval_ms). Requires tmux to be installed; maxTurns is not supported, and API usage figures are unavailable because the terminal transcript does not expose them
• Opt-in OpenTelemetry observability added (#706, #745). Set observability.enabled: true in ~/.takt/config.yaml (global) or .takt/config.yaml (project) — also overridable via the TAKT_observability__enabled env var — to emit OTel spans for workflow execution. Each run produces a workflow.<name> span with child step.<name> spans carrying attributes such as workflow / step name, step type, iteration counts, the resolved provider / model (and their config source), and final status (including abort kind). Spans are emitted as a non-blocking "shadow" alongside normal execution and never alter run behavior. The foundation initializes the OTel Node SDK (service name takt) but ships no exporter, so you wire up your own collector via the standard OTEL_* environment variables. Off by default
• /accept interactive command added (#733). In interactive assistant mode, /accept takes the most recent assistant response verbatim and runs it as the task, without re-summarizing through /go. If there is no assistant response yet, it asks you to describe the task first
• Assistant init files added (#734). List project context files under assistant.init_files in .takt/config.yaml and they are loaded automatically into every interactive assistant conversation as an "Assistant Init Context" section, so project-specific context (architecture notes, conventions, custom instructions) is included without manual setup. Paths must be relative and inside the project; sensitive files (.env*, .pem, .key, .npmrc, .netrc, .git/, etc.) are rejected, with limits of 16 files, 256 KB per file, and 1 MB total
• GitHub PR review threads are now classified by resolution state (#746). When TAKT feeds PR review comments into a task, threads are split into Active, Outdated-but-unresolved, and Resolved/Outdated sections, each annotated with who resolved it and whether it is outdated. A review policy directs the agent to focus on active threads, re-check outdated-unresolved ones for current relevance, and skip resolved threads unless the same issue still persists in the code — so already-handled feedback is not re-litigated
• Enqueue effect base_branch can create the branch on demand (#725). The system-workflow enqueue effect's base_branch now accepts an object form { name, create_if_missing: { from, push } } in addition to a plain string. When the named base branch does not exist, TAKT creates it from from (and pushes it when push: true). The builtin auto-improvement-loop uses this to create its improve base branch from main automatically, so the loop runs without manual branch setup

Changed

• Builtin review facets refined (en + ja). CQRS-ES knowledge gained guidance on event evolution and abstraction boundaries; the AI-antipattern, coding, QA, review, and testing policies tightened their REJECT / APPROVE criteria; and frontend knowledge plus the frontend-review output contract gained canonical-state guidance. These sharpen what the builtin reviewers enforce without changing workflow structure

Fixed

• OpenCode responses no longer duplicate content when the SDK emits both incremental deltas and a full snapshot (#749). The two streams were previously concatenated, so the assistant text appeared twice in the response
• Provider rate-limit messages are now preserved instead of being flattened into a generic process error (#730). When a provider reports a rate limit, the original message survives through the response so the cause is visible

Internal

• Configuration docs and validation error messages aligned to snake_case (#747). The config reference and error text used camelCase names (workflowArpeggio, syncConflictResolver, taktProviders, …) that never matched the actual snake_case YAML keys (workflow_arpeggio, sync_conflict_resolver, takt_providers, …) the parser expects. Documentation and messages now show the keys TAKT actually reads. No behavior or schema change
• CodeRabbit integration added for repository reviews — .coderabbit.yaml configuration, TAKT facets referenced as code_guidelines, probe-based config tuning, and a sponsor mention (#737, #738, #742, #744)
• CI consolidated and retriggered on /review. The four issue_comment-driven workflows were merged into a single pr-comment-commands.yml, and the takt-review comment trigger moved from /takt-review to /review (#726, #728, #736)
• Documentation reorganized: added a Design Philosophy page and an External Integrations page, refreshed the workflows guide (including workflows.ja.md), and removed stale internal docs (data-flow, provider-sandbox, report-phase-permissions, agents) (#723, #729, #739)
• Removed brittle non-executable asset tests (README terminology / instruction-template checks) and added testing-policy guidance discouraging such tests (#730)

[0.41.0] - 2026-05-14

Added

• Step-level promotion field added (#349). Per-step execution-count or AI-judgment escalation for provider / model / provider_options. Each entry can specify at: <execution-count> (matches from that execution onward) and/or condition: ai("..."), plus the override target (provider, model, or provider_options.* leaf). Multiple entries are evaluated in declaration order with last-match wins. Useful for cases like "use the faster cheap model up to attempt 2, then escalate to Opus when the reviewer keeps rejecting". Promotion is the highest-priority source in model/provider resolution (see CLAUDE.md Model resolution priority order). Promotion is not supported on parallel sub-steps
• Rate-limit fallback chain added (#716). New rate_limit_fallback.switch_chain config (workflow workflow_config, project .takt/config.yaml, and global ~/.takt/config.yaml) lets a workflow continue across a Claude / Codex / OpenCode rate-limit hit by re-running the interrupted step on the next provider in the chain. The new session receives a fallback notice instruction (facets/instructions/_system/fallback-notice.md) describing why the previous session was interrupted, which step is being retried, and how to rebuild context from report_dir / commit diff. Attempts within a single fallback chain are tracked on workflow state and reset on successful step completion
• AI-generated GitHub Issue titles for auto-improvement-loop (#333). The follow-up-task / pr-followup-task structured output schema was extended with title, type, scope, summary, goals, acceptance_criteria, and labels. The planning instruction now requires the AI to emit a short, Issue-appropriate title (rejecting generic headings like # タスク指示書 / # Task Order) plus structured task metadata that TAKT renders into the Issue body using ## Summary / ## Goals / ## Acceptance Criteria. Title validation has fallback handling for missing / too-short / prohibited titles with a fallback_reason metric so degradations are observable
• OpenCode provider_options.opencode.variant added (#694). Pass-through string forwarded to the OpenCode prompt call as the model variant (e.g. high / low). Resolvable from step provider_options, workflow / persona / project / global config, and TAKT_PROVIDER_OPTIONS_OPENCODE_VARIANT env var
• PromptBasedStructuredCaller retry on malformed JSON output (#695). decomposeTask / requestMoreParts now wrap each call in withRetry (3 attempts, 1000 ms delay) so a transient \``json ... ```extraction failure, schema validation failure, orstatus: 'error'response from the provider no longer aborts the whole team-leader run. Retry attempts are logged vialog.infowithattempt/maxAttempts/errorso retry frequency is observable. The final attempt's error still propagates if all attempts fail, andphase:startis deduped across retries so thephase:start/phase:complete` event pair stays balanced

Changed

• Review-instruction observation lists removed; reviewers now read policy / knowledge facets directly (#718). The nine review instructions (review-arch / review-cqrs-es / review-frontend / review-qa / review-requirements / review-security / review-terraform / review-test / ai-antipattern-review) plus supervise and implement / implement-after-tests no longer carry per-instruction "review observation" enumerations. Instead, every review now runs a three-step procedure: Read the bound Knowledge and Policy Source Paths, enumerate every ## section in them, then check each section's criteria against the diff. Common review boilerplate (design-judgment lookup, prior-comment tracking, final-decision steps) is consolidated into policies/review.md as "Review basic procedure". This fixes the drift observed in #713 where ai-antipattern policy gained a dead-code detection chapter but the review instruction's observation list did not include it. INSTRUCTION_STYLE_GUIDE.md was updated to forbid observation enumeration in review-type instructions
• Phase 1 prompt template gains a "Judgment rules" section (en + ja). Two universal instructions are now injected into every step's main phase: do not infer or guess values that have not been confirmed, and do not trust "already-fixed" / "already-confirmed" memories from earlier iterations of the same session — re-verify against the current file/working tree immediately before judging. Targets context-rot on long-running sessions
• cqrs-es knowledge clarifies Aggregate decision boundaries. Adds a decision table that explicitly separates state recoverable by event replay (Aggregate responsibility) from format interpretation / ownership lookup of external identifiers (API / UseCase layer responsibility), with the principle that external-identifier interpretation does not belong inside the Aggregate
• Frontend / React knowledge reinforced (knowledge frontend.md / react.md, en + ja). Targeted additions to existing chapters to tighten review criteria

Fixed

• claude-sdk provider no longer flags every step as rate-limited when the organization does not provide an overage allowance. rate_limit_event is emitted as an informational stream event each call, and overageStatus = 'rejected' is the steady state for organizations without overage. The previous OR-judgment treated a standalone rejected overage status as an active rate limit, so each step aborted immediately. isRejectedRateLimitEvent now requires the base status === 'rejected' (and falsely-fixed tests were corrected) so only an actual rate-limit event triggers fallback

Internal

• delay() helper extracted to shared/utils/delay.ts and shared between ArpeggioRunner and PromptBasedStructuredCaller. Includes unit tests under src/__tests__/delay.test.ts

[0.40.0] - 2026-05-10

Added

• takt list failed-task action Requeue added (#435). Previously the only options were Retry (always conversation-driven) and Delete. Requeue directly returns the task to the pending queue without entering a conversation, useful for quickly resending a failed task while working on something else. The retry_note is auto-generated from the failure context (failed step name + error excerpt + a hint that the user has acknowledged the issue), so the next run still receives ## Requeue Notes context. Existing retry_note values are accumulated rather than overwritten
• AI antipattern review now runs on every reviewers cycle. Added ai-antipattern-review-2nd to the parallel reviewers step in default / default-mini / default-high / backend / backend-cqrs / dual / dual-cqrs / frontend / terraform / takt-default, so over-defensive code or ghost comments introduced by a fix pass are caught on every review pass instead of only the initial one. Split workflows (backend / dual / frontend series) place it on reviewers_1 only, since fix always returns to reviewers_1

Changed

• BREAKING: AI antipattern review/fix facet names unified under ai-antipattern-* and split into 1st/2nd. ai_review (standalone) -> ai-antipattern-review-1st, ai_review (parallel sub-step) -> ai-antipattern-review-2nd, ai_fix / ai_no_fix -> ai-antipattern-fix / ai-antipattern-no-fix, ai_fix_parallel -> ai-antipattern-fix-parallel. review-ai.md is consolidated into ai-antipattern-review.md and removed. loop-monitor-ai-fix.md is renamed to loop-monitor-ai-antipattern-fix.md. Custom workflows that referenced any of these step names, instruction files, or report formats need to be updated
• Review policy treats CHANGELOG / RELEASE_NOTES / MIGRATION as point-in-time history (#710). Reviewers no longer REJECT past entries solely because their config keys, API names, or behaviors no longer match current code. Reviewers can still REJECT factual errors in newly added entries (relative to the target release) and Markdown / formatting issues. Targets are identified by file name (CHANGELOG.md, etc.) or conventional headings (### Changed / ### Added / dated release headings)
• default / default-mini / default-high / takt-default workflows refactored to share two internal subworkflows: default-draft / default-peer-review (default series) and draft / peer-review (takt-default series). The four parent workflows are now workflow_call-based shells, and the subworkflows expose params (e.g. impl_knowledge, fix_knowledge, arch_knowledge) so each parent injects its own knowledge facets without duplicating step definitions. The subworkflows are visibility: internal and not selectable from the workflow UI. Observable behavior is unchanged
• Quiet / Passthrough mode show mode-specific intros (#593). Previously both modes printed the assistant-mode intro listing slash commands like /go / /cancel that those modes do not actually parse. Quiet mode now prints interactive.ui.introQuiet and Passthrough mode prints interactive.ui.introPassthrough, both English and Japanese, so the displayed instructions match what the mode actually does
• auto-improvement-loop structured-output schemas tightened for Codex compatibility. task_markdown and issue are now required on followup-task, and task_markdown is required on pr-followup-task. The planning instruction prompts now explicitly state how to populate these fields for each action (e.g. empty string for wait_before_next_scan / prepare_merge / reject_pr). Previously a Codex provider could produce a partial agent_message without these fields and the parent workflow had no schema enforcement

Fixed

• Codex provider structured output extraction uses the final agent_message text instead of the concatenated stream content (#707). When a Codex session emitted multiple agent_message text events (typical when intermediate JSON drafts precede the final answer), the previous implementation concatenated all of them and tried to parse the resulting JSONL-like blob as a single JSON object, causing Structured output response is missing aborts on workflows like auto-improvement-loop. The last agent_message.text is now parsed independently when an outputSchema is configured, so older intermediate drafts are not incorrectly adopted and the run no longer aborts

Internal

• CI: takt-review workflow now cancels stale runs when a new commit lands on the same PR (concurrency.cancel-in-progress: true), avoiding wasted reviewer runs against an outdated revision

[0.39.0] - 2026-05-02

Added

• Facet inheritance with {extends:<parent>} syntax added (#690). File-based facets can extend a parent facet of the same kind; the parent is resolved through the facet kind and layer order, with self-reference exclusion and circular detection both done by source path. Applies to instructions, policies, knowledge, output contracts, and loop monitor judge instructions. The parent name must be a bare facet name (no path or @scope references); persona facets do not support inheritance
• Step-level allow_git_commit field added (#587). Opt-in flag (default false) that removes the standard "do not run git add / commit / push" prohibition injected into Phase 1 / Phase 2 instructions. Useful for workflows that consume many issues per task and need to commit per work unit
• default-mini builtin workflow added: a lightweight variant of default with write_tests removed (plan -> implement -> AI antipattern review -> parallel review -> complete). Registered to both Quick Start and Mini categories and the builtin catalog
• Run checkpoint resume added (#568). .takt/runs/<slug>/meta.json now records currentStep / phase / iterations / resumePoint continuously during run execution, and tasks.yaml accepts start_step (with start_movement retained as a backward-compatible alias) so interrupted runs resume from the last known step on retry
• Agent failure categorization added (#678). Team leader part failures now distinguish external_abort / part_timeout / provider_error / stream_idle_timeout and surface the category in the trace report, session log, and aggregated error message instead of a generic "execution aborted". Codex client preserves abort cause (timeout / external) and propagates it through the failure detail

Changed

• System-mode workflow execution is no longer restricted to the project workflows root (#691). mode: system, workflow-level runtime.prepare, and step allow_git_commit: true were previously rejected unless the workflow lived under .takt/workflows/. Builtin workflows (e.g. auto-improvement-loop) and user workflows under ~/.takt/workflows/ are now accepted, so reusable orchestration workflows can be installed globally and invoked by name. The previous workflow trust boundary that re-classified workflows by file path is removed in favour of the loader-confirmed WorkflowTrustInfo
• auto-improvement-loop PR-branch action set simplified (#676). Removed comment_on_pr and noop; added reject_pr. The PR branch now chooses among enqueue_from_pr / prepare_merge / reject_pr. reject_pr closes the PR via close_pr effect without leaving a comment, deleting the branch, or re-enqueueing the task. The pr-followup-task schema enum was updated accordingly
• auto-improvement-loop issue and fresh-planning paths tightened (#685). Removed noop from plan_from_issue and plan_fresh_improvement; planning instructions now explicitly reject low-value, cosmetic-only, ambiguous, or duplicate tasks and require concrete deliverables and completion conditions. The followup-task schema action enum is reduced to enqueue_new_task / wait_before_next_scan
• Rate-limit cause is preserved through workflow abort (#569). When a Claude rate limit is observed in provider events, downstream phase / parallel / workflow errors now surface as Rate limit exceeded. Please try again later. instead of the generic Claude Code process exited with code 1. AgentResponse.errorKind is normalized at the provider boundary so session resume / report-phase retry paths no longer flatten the cause

Fixed

• OpenCode shared server processes no longer leak after takt exit (#550). The server kept running in the background after takt terminated, accumulating across runs. Cleanup now runs on takt exit so OpenCode provider sessions release their server processes deterministically
• Interactive /go works without prior conversation history (#680). /go could fail when there was no preceding dialog; the first input now flows directly into workflow execution while existing dialog-driven flows are preserved
• takt watch stdin handlers cleaned up after stop. The immediate-SIGINT-exit installer now returns a cleanup hook that runs after parseAsync and slash-fallback paths, so stdin handlers installed for watch are released and process termination is no longer blocked
• Interactive mode now shows progress feedback while the AI is responding. Assistant is thinking... is displayed after each user input is sent, and Creating instruction... is displayed while /go summarizes the conversation, so the user is no longer left wondering whether takt is hung. Stdin is paused for the duration of these calls. LF (\n) now submits the line in the multiline input editor in addition to CR (\r), so pasted text and \n-terminated inputs commit cleanly; Shift+Enter (CSI 13;2u) remains the way to insert a literal newline

[0.38.0] - 2026-04-25

Added

• persona_providers.<persona>.provider_options added (#623). Persona-level provider_options (e.g. claude.effort, codex.reasoning_effort) can now be configured alongside provider / model, removing the need to repeat the same options across every step. Resolution priority: step > workflow > persona > project > global > default
• Report handles {current_report} / {previous_report} / {report_history} / {peer_reports} added to instruction template variables (#627). Facets can now reference self/peer report files abstractly without hardcoding filenames; usable in reviewer / fix / supervise style steps
• Standardized verification evidence section added to review output contracts (#628). architecture-review, qa-review, testing-review, security-review, requirements-review now emit a common build / test / behavior-check section with target, content, and result (or explicit "not verified") so supervisors can evaluate evidence consistently
• default-high workflow added: full-spec general-purpose workflow combining team-leader implementation, 5-parallel review, AI antipattern review with arbitration, and supervision (plan -> write_tests -> team-leader implement -> AI review -> 5-parallel review -> fix -> supervise -> complete). Quick Start categories reorganized into default / default-high / frontend / backend / dual
• takt-default-refresh-all / takt-default-refresh-fast workflows added: session: refresh variants of the TAKT development workflow. refresh-all refreshes every step for full conversation-carry-over comparison; refresh-fast refreshes only context-heavy steps (write_tests, ai_review, reviewers, fix)
• takt watch --ignore-exceed option added (#651). Same semantics as takt run --ignore-exceed: ignores workflow max_steps and continues running tasks instead of marking them exceeded
• provider_options.*.effort value and source surfaced in console / NDJSON (#647). Active provider's effort (claude.effort, codex.reasoningEffort, copilot.effort) is shown alongside Provider: / Model: in console output, with (source: step|persona|global|...) suffix in debug / verbose mode. NDJSON step_start records always include providerOptions and providerOptionsSources for the full tracked path set
• Task-based PRs without an Issue now use order.md content for the ## Summary section of the PR body (#600). Previously the section was empty; now reviewers can read the full task instructions directly from the PR

Changed

• <!-- takt:managed --> hidden marker is now opt-in, not a default (#665). Normal --auto-pr / autoPr no longer attaches the marker; only orchestration-driven task creation (e.g. enqueue_task from auto-improvement-loop) adds it. PRs created by ordinary pipeline / task execution are now indistinguishable from human-authored PRs
• Workflow source / trust resolution unified at the loader entry (#660). Parser / normalizer / doctor / workflow_call child loads now consume the WorkflowTrustInfo produced by the loader instead of re-classifying workflows by file path. Builtin privileged workflows (e.g. auto-improvement-loop with kind: system steps) are now treated consistently across discovery / runtime / doctor entrypoints
• Interactive mode --pr source context separated from conversation history (#656). PR review comments fetched by --pr are no longer pushed into history as a hidden user message; they are kept as a separate "Source Context" block. /go summaries no longer treat the full PR comment thread as the user's request, preventing instruction bloat
• takt-default implement step (team leader / part workers) now receives the parent takt run PID as a protected PID along with a short process-safety policy (#603). Prevents AI cleanup logic from killing the parent run via pkill / killall / name-based kill when stray child processes are present
• Retry / re-execution now syncs the project-local .takt/ directory from the root repository before re-running on an existing worktree (#607). Root-side facet / workflow / output-contract updates now take effect on retry by default, matching first-run behavior
• PR sync system effects (sync_with_root, resolve_conflicts_with_ai) reorganized to operate on PR scope via a dedicated temporary worktree / checkout (#661). Effects no longer depend on the orchestration step's cwd, so prepare_merge from auto-improvement-loop succeeds deterministically regardless of where the orchestration runs

Fixed

• team_leader + output_contracts.report no longer aborts when the root step has no session (#655). runReportPhase() now accepts the team leader's aggregated lastResponse as a fallback and only fails when both root session and lastResponse are absent
• Builtin workflow source / trust is preserved during discovery loading inside the takt repo itself (#659). Stop-gap fix that prevented auto-improvement-loop and other builtin privileged workflows from being reclassified as project workflows when discovered from the takt repository root (root cause addressed in #660)
• provider_options.copilot.effort round-trip persistence fixed (#626). denormalizeProviderOptions() now writes back copilot.effort instead of silently dropping it on config save
• Removed the dependency on the takt-managed GitHub label for managed-PR identification. The hidden body marker <!-- takt:managed --> is now the single identifier

Internal

• SDK dependencies bumped: @anthropic-ai/claude-agent-sdk ^0.2.71 -> ^0.2.119, @openai/codex-sdk ^0.114.0 -> ^0.125.0 (bundled @openai/codex binary 0.114.0 -> 0.125.0), @opencode-ai/sdk >=1.2.10 <1.3.0 -> ^1.14.24 (v2 export retained)
• E2E test added for the claude provider verifying that provider_options.allowed_tools propagates to claude --allowed-tools so Bash(python3 -m pytest:*) runs without approval prompts in real claude provider runs
• Test policy updated: line count thresholds are no longer treated as test failures in the review workflow

Experimental

The following features are still being tuned. Behavior, schema, and naming may change in subsequent releases. Use only if you are willing to follow breaking changes.

• auto-improvement-loop builtin workflow (#653). An infinite-loop orchestration workflow that scans the repository on a schedule and routes work by priority (PR -> Issue -> fresh improvement -> wait -> route)
• max_steps: infinite (no step cap, no exceeded status). Currently used by auto-improvement-loop
• pr_list system input (open PRs filtered by author / base_branch / head_branch / draft)
• issue_list / issue_selection system inputs (#662). Repo-wide open Issue observation from orchestration workflows
• task_queue_context.items (queue contents observable from when:)
• when: array references with exists(list, item.field == "X") function. Initial scope: == and && only
• followup-task builtin structured-output schema with enqueue_new_task / comment_on_pr / enqueue_from_pr / prepare_merge / noop actions

[0.37.0] - 2026-04-20

Added

• callable subworkflow に params / returns / visibility: internal を追加 (#635)。親 workflow_call.args から子 workflow に引数を渡し、子は return で論理結果を親に返せるようになった。param の対応型は facet_ref / facet_ref[]、facet_kind は knowledge / policy / instruction / report_format。子 workflow 内では $param: で facet field を差し替え可能。visibility: internal で内部用 subworkflow をワークフロー選択 UI から隠せる
• provider / model の解決ソースを log に表示 (#370)。cli / persona_providers / step / project / global / default のどの層で確定したかを debug または verbose 時にコンソール表示し、NDJSON step_start イベントの providerSource / modelSource フィールドには常時記録。期待と異なる provider / model が選ばれたときの原因特定が容易になる
• provider_options.claude.effort に xhigh を追加。Opus 4.7 のみサポートされる reasoning level（high と max の中間）。Claude Agent SDK を 0.2.114 へバンプ。モデル能力テーブルで早期検証し、claude-opus-4-6 + xhigh のような非互換組み合わせは具体的なエラーメッセージで弾く。alias (opus / sonnet / haiku) と未知モデルは permissive に SDK 側へ委ねる
• takt run に --ignore-exceed オプションを追加 (#629)。指定時は workflow の max_steps 超過を無視して最後まで実行継続する。未指定時は従来通り exceeded 扱いで requeue される

Changed

• BREAKING: 旧用語 piece / movement のレガシー環境変数サポートを完全に削除 (#637)。TAKT_PIECE_* / TAKT_MOVEMENT_* 形式の環境変数はマッピングされなくなる。TAKT_WORKFLOW_* / TAKT_STEP_* など新名称の環境変数への移行が必須

Fixed

• Codex プロバイダーでタイムアウト（abortCause === 'timeout'）発生時に workflow が停止していた問題を修正 (#640)。タイムアウトをリトライ対象に追加し、最大 2 回までリトライする。外部からの明示的な中止（abortCause === 'external'）はリトライ不可のまま維持

[0.36.0] - 2026-04-15

Added

• サブワークフロー呼び出し（call: ステップ）を追加: ステップに call: <workflow-name> を指定すると、別のワークフローをサブルーチンとして実行可能。呼び出し先ワークフローは subworkflow: { callable: true } 宣言が必要。overrides: でプロバイダ/モデルの上書きも可能。再帰呼び出し検知・最大ネスト深度 5 (#153, #624)
• システムステップ（kind: system）を追加: AI エージェントを介さずに実行されるステップ。system_inputs: でランタイムコンテキスト（タスク、ブランチ、PR、Issue、タスクキュー）をワークフロー状態にバインドし、effects: で副作用（enqueue_task, comment_pr, sync_with_root, resolve_conflicts_with_ai, merge_pr）を実行 (#586, #622)
• 決定論的 when: ルール条件を追加: ルールに condition: の代わりに when: を指定すると、比較演算子（==, !=, >, <, >=, <=）やブール論理（&&, ||）、ワークフロー状態参照（context.*, structured.*, effect.*）で AI を介さずルーティング (#586, #622)
• ステップの構造化出力（structured_output:）を追加: structured_output: { schema_ref: "<name>" } でワークフロー定義の schemas: マップ内の JSON スキーマを参照し、エージェント出力をバリデーション・保存。他ステップから {structured:step.field} で参照可能 (#586, #622)
• インタラクティブモードにスラッシュコマンド補完メニューを追加: / 入力時にインライン補完ドロップダウンを表示。矢印キーで選択、Tab で適用、Enter で確定。コンテキストに応じて /retry /replay の表示を制御 (#580)
• Copilot プロバイダに effort（推論の深さ）設定を追加: provider_options.copilot.effort で low / medium / high / xhigh を指定可能 (#625)
• takt workflow init <name> コマンドを追加: ワークフローのスキャフォールドを生成。--template minimal|faceted、--steps <count>、--description <text>、--global オプション対応 (#597)
• takt workflow doctor [targets] コマンドを追加: ワークフロー YAML の定義を検証。ターゲット未指定時は .takt/workflows/ 内の全ワークフローを検証。ワークフロー名またはファイルパスを指定可能 (#597)
• 画面専用 API ポリシー（screen-api）を追加: 画面単位の専用エンドポイント、サーバー主導のページネーション、サーバーサイド集約、スコープ付きタブ通信を強制。全 dual 系ワークフローに適用
• AI アンチパターンポリシーに早期キャッシュ戦略の禁止ルールを追加: 明示的な要求や計測なしにキャッシュレイヤー・localStorage キャッシュ・過度な memoization を導入することを REJECT

Changed

• BREAKING: 旧用語 piece / movement を完全に廃止し、workflow / step に統一 (#602, #609)。ワークフロー YAML の piece_config: → workflow_config:、movements: → steps:、initial_movement: → initial_step:、max_movements: → max_steps: への移行が必須。ディレクトリも ~/.takt/pieces/ → ~/.takt/workflows/、.takt/pieces/ → .takt/workflows/ に変更。レガシー環境変数（TAKT_PIECE_*）は引き続きマッピングされる
• 非対応プロバイダ向けの provider-specific オプション（claude.allowed_tools、mcp_servers、team_leader.part_allowed_tools）をサイレントドロップするよう変更。ワークフローをプロバイダ非依存に保てるように改善
• 全レビュー系インストラクションのエビデンスガイダンスを統一: reopened ステータスの追加、検証ターゲット・確認内容・観測結果の記録を必須化、オープン指摘事項の脱落防止ルールを追加

Fixed

• 全 audit 系ワークフロー（7 種）で supervise ↔ review 間のデッドロックを修正。review ステップに output_contracts を追加し、ループモニターの閾値を 4→3 に調整、ジャッジの選択肢を 3 択（十分/進捗あり/停滞）に変更

Internal

• piece* → workflow* のファイル名一括リネーム（84 ファイル、テスト・E2E フィクスチャ・ドキュメント含む）
• WorkflowEngine をリファクタリング: WorkflowEngineSetup、WorkflowEngineStepCoordinator、WorkflowRunLoop に責務を分離
• セッションロガーを sessionLoggerPhaseTracker.ts、sessionLoggerRecordFactory.ts に分割
• CapabilityAwareStructuredCaller を追加し、プロバイダの構造化出力対応可否に応じたルーティングを実装
• ルール評価を 10 段階フォールバックに拡張（when: 条件の評価ステージを追加）
• workflow-state-access.ts でテンプレート参照（{context:*}、{structured:*}、{effect:*}）の統一解決を実装

[0.35.4] - 2026-04-11

Changed

• レビューポリシーにツール出力の信頼性検証ルールを追加: ツール出力が正常に読めることを確認してから指摘すること、検索失敗だけでコード不在と断定しないことをルール化

Fixed

• ターミナルの行数が少ない環境で選択メニューが正常に動作しない問題を修正。ビューポートベースのスクロールを追加 (#608)
• Windows で .cmd shim の spawn が失敗する問題を修正（Claude Headless、Cursor、Copilot プロバイダ）

Internal

• 選択メニューを select-menu.ts、select-viewport.ts に分離し、純粋関数化とビューポートテストを追加
• ワークフロードキュメントの例で古いフィールド名を使っていた問題を修正 (#619)

[0.35.3] - 2026-04-10

Added

• loop_monitors の judge に provider/model フィールドを追加。ジャッジムーブメントのプロバイダーとモデルを明示的に指定可能に (#599)
• takt list --action sync を非インタラクティブモードでサポート

Changed

• Codex プロバイダーのリトライ戦略を強化: 最大リトライ回数を 3→9 に増加、ベース遅延を 250ms→1000ms に変更、"at capacity" エラーを自動リトライ対象に追加 (#614)

Fixed

• ループモニタージャッジが常にデフォルトプロバイダーで実行される問題を修正。トリガー元ムーブメントのプロバイダー/モデル設定を継承するよう変更 (#599)
• 完了済みタスクのブランチ操作（merge/try-merge/diff）でルートブランチが欠落している場合にエラーとなる問題を修正。クローンから自動復元するよう改善 (#616)
• Phase 2 エラーイベント（phase:complete）が phase:start より先に発火されることがある問題を修正

Internal

• テストを多数追加（codex-client-retry, engine-loop-monitors, it-completed-task-root-branch, it-piece-loader, provider-resolution, taskBranchLifecycleActions 等）
• providerModelCompatibility をコアモジュール（src/core/piece/）に移動
• タスク実行後のプッシュ処理を簡素化（clone 内フォールバック push を廃止し、ルートリポジトリ経由に一本化）
• git コマンドのエラーメッセージに stderr 詳細を含めるよう改善

[0.35.2] - 2026-04-09

Added

• takt list でスタックした実行中タスクを強制失敗にできる「Mark as failed」アクションを追加 (#604)
• タスクレコードに run_slug を追加し、実行中タスクの現在のステップ情報を meta.json から取得可能に

Changed

• write_tests ムーブメントの出力契約を test-scope.md + test-decisions.md の2ファイルから test-report.md の1ファイルに簡素化

Internal

• CI auto-tag ワークフローから冗長な build・test ステップを削除
• RunMeta 型を src/core/piece/run/run-meta.ts に抽出し、currentStep/currentIteration トラッキングを追加

[0.35.1] - 2026-04-09

Added

• タスク実行中に常時スピナーを表示: TTY 環境でタスク実行中にアニメーションスピナーを表示し、処理中であることを可視化
• Claude Headless の thinking ストリーム表示: ヘッドレスモードで thinking トークンをリアルタイム表示。parallel モードでは色分け表示に対応

Changed

• SIGINT（Ctrl+C）でクローン作成中でも即座にプロセスを終了できるよう改善: run/watch コマンドで raw mode による即時検知を導入し、git サブプロセスを AbortSignal で中断可能に

Fixed

• Codex プロバイダが Git リポジトリ外での実行を拒否する問題を修正

Internal

• StatusLine の wrapWrite をアロー関数にリファクタリング
• StreamDisplay のハンドラーをムーブメントごとにキャッシュするよう最適化
• クローン作成関連のテスト追加・更新

[0.35.0] - 2026-04-07

Added

• Claude Headless プロバイダを追加: Claude CLI のヘッドレスモード（claude --print）をサブプロセスとして実行する新プロバイダ。claude プロバイダ名で利用可能 (#584)
• trace ログレベルを追加: poll_tick/no_new_tasks などの高頻度ログを抑制し、デバッグログの可読性を向上。logging.trace: true で有効化
• レガシー設定キーの非推奨警告: piece_*/movement* 系の旧キー使用時に deprecation warning を表示し、新しい workflow_*/step* キーへの移行を促進 (#581, #594)
• Mock プロバイダで delayMs と AbortSignal をサポート: SIGINT テスト等で利用可能に (#595)

Changed

• BREAKING: claude プロバイダがヘッドレス CLI モードをデフォルトに変更。従来の SDK ベースプロバイダは claude-sdk として引き続き利用可能 (#584)
• クローン環境からのプッシュに relay push パターンを導入: 一時 ref 経由でプッシュすることで、チェックアウト中のブランチへの直接プッシュによるデータ損失を防止 (#592)
• PR 由来タスクのメタデータ伝搬を改善: shouldPublishBranchToOrigin フラグにより、ローカルプッシュ失敗時にクローンから直接 origin へプッシュするフォールバックを追加 (#592)
• dual/backend ワークフローの max_steps を 60 に増加

Fixed

• pr_body_template が --task オプション実行時に無視される問題を修正 (#538)
• 既存 PR ブランチ更新時にリモートを正として worktree を作成するよう修正 (#557)
• SIGINT（Ctrl+C）が AI レスポンス待機中に正しくアボートされない問題を修正 (#595)
• mise 等のバージョン管理ツール環境下で Claude SDK サブプロセスの PATH が安定しない問題を修正 (#591)
• ジャッジムーブメントのプロバイダフォールバック解決が正しく機能しない問題を修正 (#577)
• exceeded 時のワークツリーパスが正しく解決されない問題を修正 (#575)
• PR 作成時のエラー詳細が失われる問題を修正
• non-fast-forward プッシュ拒否時にヒントメッセージを表示するよう改善

Internal

• .takt/config.yaml の非推奨キーをリネーム
• レガシーワークフローキー非推奨警告の重複排除 (#594)
• AgentSetup から claudeAgent/claudeSkill フィールドを削除（Headless プロバイダ移行に伴い不要に）
• 型定義ファイルから冗長な JSDoc コメントを整理
• review-fix-takt-default ワークフローの max_steps を増加

[0.34.0] - 2026-04-03

Added

• StructuredCaller インターフェースを導入: プロバイダーのネイティブ構造化出力（Structured Output）をサポートし、ステータス判定・条件評価・タスク分解でテキストパースに代わる JSON ベースの応答抽出が可能に (#570)
• Traced Config を導入: traced-config パッケージによる設定値の出所追跡（環境変数・設定ファイル・デフォルト値）をサポート (#558)
• 並列ステップに concurrency フィールドを追加: セマフォベースの同時実行数制御が可能に
• 無効なワークフロー YAML のロード時に警告を表示するようになった（スキーマバリデーションエラーの詳細を表示） (#540)
• 計画・レビュー用ファセットを強化: planner・requirements-reviewer・supervisor のペルソナ、plan・requirements-review・supervisor-validation の出力契約、coding ポリシーを新規追加
• takt add コマンドで --workflow オプションによるワークフロー指定に対応

Changed

• BREAKING: ワークフロー YAML のキーをリネーム: movements → steps、initial_movement → initial_step、max_movements → max_steps、piece_config → workflow_config。旧キーは互換エイリアスとして引き続き使用可能 (#576)
• BREAKING: ビルトインワークフローのディレクトリを builtins/{lang}/pieces/ から builtins/{lang}/workflows/ に移動。設定キーも piece_categories → workflow_categories、enable_builtin_pieces → enable_builtin_workflows にリネーム。旧キーは互換エイリアスとして引き続き使用可能 (#571, #561)
• BREAKING: CLI オプション -w, --piece を -w, --workflow にリネーム。--piece はレガシーエイリアスとして使用可能 (#576)
• BREAKING: ワークフロー YAML の instruction_template フィールドを削除。instruction フィールドを使用すること (#539)
• takt-default ワークフローの max_steps を 50 に増加（default は 30 のまま）
• 設定キーのエイリアス解決時に旧キーと新キーの両方が異なる値で存在する場合はエラーを発生させるよう変更

Fixed

• Claude SDK のエラーペイロードが正しく処理されない問題を修正
• ワークフロー用語の統一: CLI ヘルプ、エラーメッセージ、ドキュメントを workflow / step 用語に更新
• ワークツリーモードで PR の Issue 解決がプロジェクト cwd から正しく行われるよう修正
• Cursor Agent のヘッドレスワークツリー実行で --trust フラグが渡されるよう修正
• ワークツリー環境下で runtime.prepare が設定されている場合にセルフホスト GitLab の glab CLI 認証が失敗する問題を修正 (#563)
• ピースプロバイダー解決の統一化

Internal

• Zod スキーマを schemas.ts から schema-base.ts・workflow-schemas.ts・config-schemas.ts に分割
• 環境変数オーバーライドを spec ベースの宣言的定義にリファクタリング
• レガシーの step_provider・step_model フィールドを削除
• TeamLeader のパートタイムアウト処理を簡素化
• yaml パッケージを v2.8.3 に更新
• ドキュメント（README、CLI リファレンス、設定ガイド等）をワークフロー用語に全面更新

[0.33.2] - 2026-03-26

Added

• 読み取り専用の監査ピースを追加: audit-architecture, audit-architecture-frontend, audit-architecture-backend, audit-architecture-dual, audit-e2e, audit-unit。コードを変更せずにモジュール境界やカバレッジギャップを列挙し、Issue 作成可能なレポートを出力

Changed

• security-audit ピースを audit-security にリネーム（監査ピース群の命名規則を統一）
• ビルトインピースカテゴリを再構成: 🧪 Testing カテゴリを廃止し、監査ピースを 🔍 Review カテゴリに統合
• fill-unit, fill-e2e ピースを削除（audit-unit, audit-e2e に置き換え）

Fixed

• GitLab セルフホスト環境で worktree（共有クローン）実行時に MR 作成が失敗するバグを修正。Git プロバイダーの cwd がクローンパスに正しく伝播するよう変更 (#552)

Internal

• Git プロバイダーの cwd 伝播に関するテストカバレッジを追加
• 設定ドキュメントから verbose オプションの記載を削除し、logging.level による設定方法に統一 (#543)

[0.33.1] - 2026-03-24

Changed

• ファイナルレビューとセキュリティレビューのガードレールを強化: supervisor のファセット、セキュリティナレッジ、レビューポリシー・インストラクションを拡充

Fixed

• GitLab セルフホスト環境で gitlab.com の認証がない場合にタスク完了後の MR 作成が必ず失敗するバグを修正。glab auth status がリモートのホスト名を指定して認証チェックするよう変更 (#545)

Internal

• GitLab プロバイダーのテストカバレッジを拡充（セルフホスト環境の認証チェック、ホスト名ベースの CLI ステータス検証）

[0.33.0] - 2026-03-22

Added

• GitLab VCS プロバイダーを追加: glab CLI を使った Issue 取得・マージリクエスト作成・レビューコメント取得に対応。git リモート URL からの自動検出をサポートし、vcs_provider: gitlab による明示的な設定も可能 (#512)
• インタラクティブモード用プロバイダー設定 (takt_providers.assistant) を追加: ピース実行とは独立したプロバイダー/モデルをインタラクティブモードに指定可能 (#483)

Changed

• BREAKING: ピース YAML の MCP サーバー設定をデフォルト拒否に変更。使用するには piece_mcp_servers でトランスポート別に明示的に許可が必要 (#524)
• BREAKING: ピース YAML の Arpeggio カスタムコード（カスタムデータソース、インライン JS、外部マージファイル）をデフォルト拒否に変更。使用するには piece_arpeggio で明示的に許可が必要 (#521)
• BREAKING: ピース YAML の runtime prepare カスタムスクリプトをデフォルト拒否に変更（ビルトインプリセットは常に許可）。使用するには piece_runtime_prepare.custom_scripts: true が必要 (#520)
• BREAKING: sync conflict resolver の自動ツール承認をデフォルト拒否に変更。使用するには sync_conflict_resolver.auto_approve_tools: true が必要 (#522)
• team leader のタスク分解における最大ターン数を 4 → 5 に引き上げ (#511)
• supervisor ファセットを強化: 要件カバレッジのエビデンスベース検証を追加
• ペルソナファセットからクロスエージェント参照を除去し、ピース横断での再利用性を向上

Fixed

• パイプラインモードで auto-commit の push 失敗時に PR 作成が無診断で失敗する問題を修正 (#532)
• .takt/.gitignore のファセットパスが実際のディレクトリ構造と不一致だった問題を修正 (#535)
• レビューピースの gather モードでブランチ検出が不正確だった問題を修正（完全一致を要求するよう変更） (#523)
• レビューピースで reject findings のフォーマットが正しく処理されない問題を修正 (#528)
• パイプラインモードでタスクブランチが PR 作成前に push されず、PR 作成が失敗する問題を修正

Internal

• GitLab プロバイダーのテストカバレッジを追加（issue, pr, provider, utils）
• VCS プロバイダーの自動検出・ファクトリ・フォーマットのテストカバレッジを追加
• MCP サーバー・Arpeggio・runtime prepare・conflict resolver のデフォルト拒否に関するテストカバレッジを追加
• ピースローダーのテストカバレッジを大幅に拡充
• プロジェクト設定・グローバル設定のテストカバレッジを追加
• MCP サーバーヘルパー、ポリシー正規化、conflict resolver ヘルパーのリファクタリング
• ドキュメント更新（レビューピース名の修正、ビルトインカタログ更新）
• ビルド/lint/テスト品質ゲートの追加と E2E テスト環境の CLAUDECODE 環境変数分離
• テスト契約チェックのビルトインファセット強化（review-test, write-tests-first, testing-review）
• タスク auto-PR の E2E テストを追加

[0.32.2] - 2026-03-17

Added

• 到達経路（Reachability）ファセットを追加: 新しい画面・機能を追加する際に、ルーティング・メニュー・ボタン等のエントリーポイントを同時に整備することを計画・実装・レビューの各段階で検証
• 再取得ループ防止ファセットを追加: useEffect の依存配列に不安定な Context/Provider 関数参照を含めることで起きる無限ループを検出・防止するナレッジとポリシー
• UIライブラリ統合ファセットを追加: サードパーティ UI コンポーネント（データグリッド、日付ピッカー等）導入時のバージョン互換性・実マウント検証のナレッジとポリシー
• React ナレッジファセット (react.md) を新規追加: Effects の再実行制御、Context/Provider の値安定性に関する判断基準テーブル付き
• デザイン計画ポリシー (design-planning.md) を新規追加: デザインリファレンスが存在する場合の要素インベントリ・スコープ決定の基準
• フロントエンド専用プランフォーマット (plan-frontend.md) を新規追加: デザイン要素の Keep/Change 判定テーブルを含むプラン出力契約

Changed

• フロントエンド系ピース（frontend, frontend-mini, dual, dual-mini, dual-cqrs, dual-cqrs-mini）の plan ムーブメントに design-planning ポリシーと react ナレッジを統合
• フロントエンド系ピースのプランフォーマットを plan から plan-frontend に変更
• frontend ピースの全ムーブメント（テスト、レビュー、修正等）に react ナレッジを追加

Internal

• @openai/codex-sdk を 0.112.0 → 0.114.0 に更新
• team leader worker pool の E2E テストを安定化

[0.32.1] - 2026-03-14

Fixed

• --pr 経由のタスクで autoPr が無効になっていたため origin push がスキップされる問題を修正
• PR レビューコメントの取得が gh pr view の reviews.comments に依存していたため、インラインコメントを取りこぼす問題を修正。GitHub REST API によるページネーション取得に変更 (#489)
• config のパス指定で ~ チルダ展開が効かない問題を修正（worktree_dir、*_cli_path、analytics.events_path 等） (#496)
• auto-commit 時に git hooks/filter がそのまま実行され、TAKT 管理下のコミットが意図しない hooks の影響を受ける問題を修正。デフォルトで無効化し、allow_git_hooks / allow_git_filters で opt-in に変更 (#503)
• インタラクティブモードで初回入力時に不要な AI 呼び出しが発生していた問題を修正 (#504)
• Cursor provider でプロンプト文字列が CLI オプションとして解釈される可能性がある問題を修正（-- セパレータを追加） (#500)
• snapshot ファイル名にムーブメント名がそのまま使われ、パストラバーサルが可能だった問題を修正 (#498)
• provider_options の優先順位で、環境変数・プロジェクト設定がムーブメント定義より低くなっていた問題を修正（セキュリティ設定がムーブメントで上書きされないよう変更） (#497)
• worktree パスの再利用時に、クローンベースディレクトリ外のパスが受け入れられる問題を修正 (#502)
• terraform ピースから不要な強制 full パーミッションを削除 (#507)

Internal

• テスト系ピース・ファセットの全面整備（e2e-test → audit-e2e、unit-test → audit-unit にリネーム、ナレッジ・ポリシー追加）
• デザイン忠実度ポリシーの追加とフロントエンド系ピースへの統合
• audit-security ピースの追加
• ファセットデプロイメントのリファクタリング（templates ディレクトリの廃止、facets ディレクトリへの統合） (#505)
• isPathInside ユーティリティを追加し、クローン削除・worktree 再利用のパス検証を強化
• ループモニターの閾値調整とレビューポリシーの改善

[0.32.0] - 2026-03-09

Added

• takt export-codex コマンド: ピース/ファセットを Codex スキルとしてエクスポート (~/.agents/skills/takt/) (#475)
• frontend / backend / backend-cqrs ピースにテストファースト（write_tests）ムーブメントを追加し、レビューを2段階化（Stage 1: 構造・実装品質 → Stage 2: 安全性・品質保証）
• セキュリティナレッジにログ・マスキングセクションを追加（パスワード露出、toString() によるフィールド漏洩の検出基準）
• CQRS+ES ナレッジにマスタデータと CRUD の使い分けセクションを追加（6つの判断基準テーブル付き）
• /ci コメントで PR の CI を手動トリガーするワークフローを追加
• devcontainer で worktree クローン先の親ディレクトリが書き込み不可の場合に .takt/worktrees/ へフォールバック
• インタラクティブモードのアシスタントが設計判断を勝手にしないようポリシーを追加

Changed

• BREAKING: ピース YAML の instruction_template フィールドを非推奨化。instruction に統一（後方互換あり、deprecated 警告を表示） (#476)
• レビュー系ピースの命名規則を review-{variant} / review-fix-{variant} に統一
• タスク分解の REJECT 基準をナレッジからポリシーに分離
• faceted-prompting を npm パッケージ (@anthropic-ai/faceted-prompting) に移行し、内蔵コードを削除

Fixed

• takt run の Slack 通知が当該 run で実行したタスクのみを送信するよう修正（従来は全タスクを通知していた）
• ProviderPermissionProfilesSchema に copilot が欠落していた問題を修正 (#487)
• PR fix フローで既存ブランチ存在時に baseBranch 検証をスキップするよう修正
• review-fix-takt-default の fix 後フローを takt-default と統一
• write-tests-first インストラクションからビルド検証手順を削除
• cc-resolve ワークフローに actions: write パーミッションを追加

Internal

• SDK 依存パッケージを最新化
• deploySkill のコア処理を deploySkillInternal に抽出し、deploySkillCodex と共有
• clone ブランチ解決をリモートブランチ対応に拡張（localBranchExists / remoteBranchExists に分離）
• README の起動フローを整理し「タスクにつむ」を通常フローとして記載

[0.31.0] - 2026-03-06

Changed

• dual ピースを大幅強化: テストファースト（write_tests）ムーブメント追加、implement を team_leader 化（FE/BE 分割）、レビューを2段階化（reviewers_1: arch/frontend/testing → reviewers_2: security/qa/requirements）
• takt-default-team-leader ピースを takt-default に統合し削除。takt-default の implement を team_leader 化
• quality_gates のペルソナ単位オーバーライドをサポート: piece_overrides.personas.<name>.qualityGates で特定ペルソナのムーブメントに品質ゲートを追加可能に (#472)
• Status 型を done / blocked / error の3値に整理し、ステータスハンドリングを厳格化。blocked / error 時は即座に ABORT するよう変更 (#477)

Fixed

• git check-ref-format コマンドから不要な -- を削除し、ブランチ名の検証が正しく動作するよう修正 (#481)
• log_level → logging.level の設定キー不整合を修正（E2E テスト全滅の原因）
• Phase 3 ステータス判定が失敗した際に Phase 1 のルール評価にフォールバックするよう修正（従来はエラーで中断していた） (#474)
• Parallel ムーブメントの Phase 3 判定失敗時も同様にフォールバック対応 (#474)
• タスクリトライ・追加指示時のピース名取得元を runInfo?.piece から task.data?.piece に変更（worktree 内で runInfo が常に null になる問題を修正）

Internal

• config 3層モデルの整理: PersistedGlobalConfig → GlobalConfig にリネーム、マイグレーション用フォールバック処理を削除、persisted-global-config.ts → config-types.ts にリネーム
• supervisor ペルソナからインラインの知識・ポリシーをファセットファイルに分離
• team leader の分解品質を改善するナレッジ（task-decomposition.md）とインストラクション（dual-team-leader-implement.md）を追加
• ~/.takt/config.yaml テンプレートに不足していた設定項目を追加
• Provider Sandbox & Permission ガイドのドキュメントを拡充

[0.30.0] - 2026-03-05

Added

• トレースレポートの自動生成: piece 実行完了時に movement の遷移・フェーズ・ルール評価結果を Markdown レポートとして .takt/runs/ に自動出力。logging.trace: true で全文モード、デフォルトは redacted モード (#467)
• 使用量イベントログ: プロバイダー呼び出しごとのトークン使用量を NDJSON 形式で記録。logging.usage_events: true で有効化 (#470)
• タスクリトライ時のピース再利用確認: takt list からリトライ・追加指示する際に、前回と同じピースを使うか選び直すかを選択可能に (#468)

Changed

• BREAKING: takt switch コマンドを削除。ピース選択はインタラクティブモード起動時（takt）に毎回行う方式に変更 (#465)
• Claude プロバイダーの allowed_tools をビルトインピースの YAML 定義からエグゼキューター側に移動し、ピース YAML の簡素化と保守性を向上 (#469)
• 設定構造をリファクタリング: globalConfig.ts を globalConfigCore.ts・globalConfigAccessors.ts・globalConfigResolvers.ts・globalConfigSerializer.ts に分割。プロジェクトローカル設定（.takt/config.yaml）のフォールバック優先度を明確化 (#460)
• observability モジュールを core/logging/ に再編成: providerEventLogger と usageEventLogger を統一的なログ基盤として整理 (#466)
• レビュアー全体に coder-decisions.md の参照を追加し、コーダーの設計判断を考慮したレビューで誤検知を抑制
• レビュー↔修正ループの収束を支援: レポート履歴の参照、ループモニター、修正方針のガイドラインを整備

Fixed

• runtime 環境の XDG_CONFIG_HOME 上書きで gh CLI の認証が失敗する問題を修正。GH_CONFIG_DIR を元の設定から保持するよう変更
• .takt/config.yaml に runtime.prepare を記述するとエラーになる問題を修正（プロジェクトレベルでの runtime 設定を許可） (#464)
• インタラクティブモードで iteration limit 到達時にプロンプトが表示されず、exceeded 状態が保持されない問題を修正
• PR 作成失敗時のタスクステータスを failed から pr_failed に分離し、実行成功だが PR 作成のみ失敗したケースを区別可能に
• リトライ時にタスクにピース情報が引き継がれるよう修正
• .gitignore の .takt/ ディレクトリ ignore を削除し .takt/.gitignore に委譲（プロジェクト設定ファイルの追跡を可能に）
• CI: push トリガーから takt/** を削除し二重実行を防止
• cc-resolve ワークフローで push 後に CI を自動トリガーするよう修正

Internal

• deprecated config マイグレーション処理を削除
• プロジェクトローカル設定の優先度に関する統合テストを追加
• テストヘルパーとテストセットアップの改善

[0.29.0] - 2026-03-04

Added

• レビュー＋修正ループピース群を追加: review-fix（多角レビュー）、frontend-review-fix、backend-review-fix、dual-review-fix、dual-cqrs-review-fix、backend-cqrs-review-fix および対応するレビュー専用ピース群を追加。コードレビューと自動修正を反復するワークフロー
• takt-default-review-fix ピースを追加: TAKT 自己開発向けのレビュー＋修正ループワークフロー
• quality_gates のグローバル/プロジェクトレベルオーバーライドをサポート: ~/.takt/config.yaml および .takt/config.yaml の piece_overrides.quality_gates でビルトインピースの品質ゲートを上書き可能に (#384)
• タスクの base_branch 設定: takt add 時に現在のブランチを base_branch として記録し、タスク実行時にそのブランチから分岐するよう設定可能に (#455)
• プロバイダー設定の統一: .takt/config.yaml で provider ブロックに type/model/プロバイダー固有オプション（network_access 等）をまとめて記述可能に (#457)
• ワーカープール超過時のリキュー: タスク実行がワーカー上限を超えた場合、タスクを自動的に再キューイングするよう対応 (#366)
• --pr インタラクティブモードで create_issue アクションを除外し、save_task 時に PR のブランチ名を base_branch として自動設定
• team_leader の decomposeTask/requestMoreParts/Phase 3 ステータス判定のプロバイダーイベントをロギング: provider-events.jsonl に記録されるようになり、デバッグ・分析が可能に

Fixed

• export-cc で facets/ のサブディレクトリ構造（personas/、policies/ 等）が出力先に再現されなかった問題を修正 (#8dcb23b)
• cc-resolve コマンドがコンフリクト解決後にマージコミットを生成するよう修正 (#1b1f758)
• グローバル設定 (~/.takt/config.yaml) の piece フィールドがピース解決チェーンで無視されるバグを修正 (#458)
• Codex プロバイダーでプロバイダー優先のパーミッションモード解決が機能しない問題と EPERM エラーの E2E テストを追加 (#d2b48fd)
• レビューコメントがない PR で --pr を使用した際にエラーになる問題を修正
• --auto-pr/--draft オプションをパイプラインモード専用に制限（インタラクティブモードでの誤用を防止）
• team_leader のストリーミングでバウンダリの先行フラッシュによる断片化を修正 (#769bd87, #bddb66f)
• team_leader のエラーメッセージが空文字列になるバグを修正 (#52968ac)
• decomposeTask/requestMoreParts の maxTurns を 2 から 4 に増加（複雑なタスク分解でタイムアウトしていた問題を緩和）
• Copilot プロバイダーのクライアント実装のバグを修正 (#434)

Internal

• E2E プロバイダー別テストをコンフィグレベル（vitest.config.e2e.provider.ts）で振り分けるよう変更。テストファイル内の skip ロジックを廃止し、JSON レポート出力を追加
• 共有ノーマライザを configNormalizers.ts に抽出してプロバイダー設定解析を整理
• agent-usecases/schema-loader を移動し pieceExecution の責務を分割
• check:release で全プロバイダー（claude/codex/opencode）の E2E を実行するよう変更
• CI: PR と push の重複実行を concurrency グループで抑制
• CI: feature ブランチへの push と手動実行に対応

[0.28.1] - 2026-03-02

Changed

• BREAKING: expert / expert-mini / expert-cqrs / expert-cqrs-mini ピースを dual / dual-mini / dual-cqrs / dual-cqrs-mini にリネーム。カスタマイズしている場合はピース名の更新が必要
• default-mini / default-test-first-mini ピースを default に統合。default ピースが「テスト優先モード」を内包するよう拡張
• coding-pitfalls ナレッジの主要項目を coding ポリシーに移動し、ポリシーとして実際に適用されるよう強化
• implement / plan インストラクションにセルフチェック・コーダー指針を追加

Removed

• passthrough ピースを削除
• structural-reform ピースを削除

Internal

• expert-supervisor ペルソナを dual-supervisor にリネーム
• ビルトインカタログに不足していた terraform、takt-default 系、deep-research を追加
• カテゴリ設定に deep-research を追加
• 全ドキュメントに copilot プロバイダーの説明を追加し、Claude Code 寄りの記述をプロバイダー中立に修正

[0.28.0] - 2026-03-02

Added

• GitHub Copilot CLI プロバイダーを追加: copilot プロバイダーとして GitHub Copilot CLI を利用可能に。セッション継続、パーミッション制御（readonly/edit/full）に対応。copilotCliPath / TAKT_COPILOT_CLI_PATH で CLI パスを指定、copilotGithubToken / TAKT_COPILOT_GITHUB_TOKEN で認証トークンを設定 (#425)
• --pr オプションを追加: PR のレビューコメントを取得してタスクとして実行。パイプラインモードとインタラクティブモードの両方で利用可能 (#421)
• takt add --pr N で PR のレビューコメントをタスクとして追加可能に。PR のブランチ名で worktree を自動作成し、レビュー指摘の修正タスクとしてキューイング (#426)
• takt list に「Pull from remote」アクションを追加: リモートの変更を worktree に取り込み、再プッシュ可能に (#395)
• プロジェクト単位の CLI パス設定: .takt/config.yaml で claudeCliPath / cursorCliPath / codexCliPath / copilotCliPath をプロジェクトごとに設定可能に (#413)
• インタラクティブモードのスラッシュコマンドを行末でも認識可能に（例: タスクの内容 /go）(#406)
• takt-default / takt-default-team-leader ビルトインピースを追加（TAKT 自己開発用のワークフロー定義）
• TAKT ナレッジファセット（takt.md）を追加: TAKT のアーキテクチャとコード規約を体系化
• ai-antipattern ポリシーに冗長な条件分岐パターン検出を追加: 同一関数を if/else で呼び分けるコードを検出し、三項演算子やスプレッド構文での統一を促す

Fixed

• 不正な tasks.yaml を検出した場合、ファイルを削除せず保持してエラーメッセージで停止するよう修正 (#418)
• shallow clone リポジトリで worktree 作成が失敗する問題を修正: --reference 付きクローンが失敗した場合に通常クローンへフォールバック (#376, #409)
• グローバル/プロジェクト設定の model がモデルログに反映されない不具合を修正 (#417)
• fork PR レビュー時に GH_REPO を設定して正しいリポジトリの issue を参照するよう修正
• takt-review ワークフローの PR コメント投稿ステップにも GH_REPO を設定

Internal

• resolveConfigValue の不要な defaultValue 引数を削除し、設定解決ロジックを簡素化 (#391)
• PRコメント /resolve でコンフリクト解決・レビュー指摘修正を行う GitHub Actions ワークフロー（cc-resolve）を追加
• takt-review ワークフローを pull_request_target に変更し、fork PR でもシークレットを利用可能に
• CI に ready_for_review / reopened トリガーを追加
• CONTRIBUTING にレビューモードの例を追加、日本語版（CONTRIBUTING.ja.md）を追加

[0.28.0-alpha.1] - 2026-02-28

Added

• GitHub Copilot CLI プロバイダーを追加: copilot プロバイダーとして GitHub Copilot CLI を利用可能に。セッション継続、パーミッション制御（readonly/edit/full）に対応。copilotCliPath / TAKT_COPILOT_CLI_PATH で CLI パスを指定、copilotGithubToken / TAKT_COPILOT_GITHUB_TOKEN で認証トークンを設定 (#425)
• --pr オプションを追加: PR のレビューコメントを取得してタスクとして実行。パイプラインモードとインタラクティブモードの両方で利用可能 (#421)
• takt add --pr N で PR のレビューコメントをタスクとして追加可能に。PR のブランチ名で worktree を自動作成し、レビュー指摘の修正タスクとしてキューイング (#426)
• takt list に「Pull from remote」アクションを追加: リモートの変更を worktree に取り込み、再プッシュ可能に (#395)
• プロジェクト単位の CLI パス設定: .takt/config.yaml で claudeCliPath / cursorCliPath / codexCliPath / copilotCliPath をプロジェクトごとに設定可能に (#413)
• インタラクティブモードのスラッシュコマンドを行末でも認識可能に（例: タスクの内容 /go）(#406)
• takt-default / takt-default-team-leader ビルトインピースを追加（TAKT 自己開発用のワークフロー定義）
• TAKT ナレッジファセット（takt.md）を追加: TAKT のアーキテクチャとコード規約を体系化
• ai-antipattern ポリシーに冗長な条件分岐パターン検出を追加: 同一関数を if/else で呼び分けるコードを検出し、三項演算子やスプレッド構文での統一を促す

Fixed

• 不正な tasks.yaml を検出した場合、ファイルを削除せず保持してエラーメッセージで停止するよう修正 (#418)
• shallow clone リポジトリで worktree 作成が失敗する問題を修正: --reference 付きクローンが失敗した場合に通常クローンへフォールバック (#376, #409)
• グローバル/プロジェクト設定の model がモデルログに反映されない不具合を修正 (#417)
• fork PR レビュー時に GH_REPO を設定して正しいリポジトリの issue を参照するよう修正
• takt-review ワークフローの PR コメント投稿ステップにも GH_REPO を設定

Internal

• resolveConfigValue の不要な defaultValue 引数を削除し、設定解決ロジックを簡素化 (#391)
• PRコメント /resolve でコンフリクト解決・レビュー指摘修正を行う GitHub Actions ワークフロー（cc-resolve）を追加
• takt-review ワークフローを pull_request_target に変更し、fork PR でもシークレットを利用可能に
• CI に ready_for_review / reopened トリガーを追加
• CONTRIBUTING にレビューモードの例を追加、日本語版（CONTRIBUTING.ja.md）を追加

[0.27.0] - 2026-02-28

Added

• Cursor Agent CLI プロバイダーを追加: cursor-agent CLI を介して Cursor を AI プロバイダーとして利用可能に。API キー（TAKT_CURSOR_API_KEY / cursor_api_key）または cursor-agent login セッションで認証、JSON 出力解析、セッション継続（--resume）、モデル指定（--model）、パーミッション制御（full → --force）に対応 (#403)
• Cursor プロバイダーの E2E テスト設定を追加（vitest.config.e2e.cursor.ts、npm run test:e2e:cursor）

Fixed

• Phase 1 が error または blocked を返した場合に Phase 2（レポート出力）をスキップするよう修正。Phase 1 失敗時に不要なレポート生成が実行される問題を解消
• Codex 互換性のため、runtime prepare で Gradle デーモンを無効化するよう修正

Internal

• エージェント/カスタムペルソナのドキュメントを整合

[0.26.0] - 2026-02-27

Added

• TeamLeader に refill threshold と動的パート追加を導入: 実行中のパートが refill_threshold 以下になると、リーダーが完了済みパートの結果を評価して追加パートを動的に生成。max_parts は同時並行数、refill_threshold で追加計画のタイミングを制御（最大合計 20 パートまで）
• deep-research ピースの dig ムーブメントに team_leader 設定を追加し、リサーチの並列実行が可能に
• TeamLeader が Phase 2（レポート出力）/ Phase 3（ステータス判定）を通常ムーブメントと同様にサポート（applyPostExecutionPhases の共通化）
• ParallelLogger が動的なサブムーブメント追加に対応（addSubMovement）し、TeamLeader の動的パート追加時にもストリーミング出力を表示
• LineTimeSliceBuffer を導入し、並列ストリーミング出力のバッファリングを時間スライスベースで最適化
• プロジェクト設定（.takt/config.yaml）で model 指定をサポート

Changed

• BREAKING: カスタムエージェント定義（~/.takt/personas/*.md）の provider / model を解釈しない方針とし、エージェントのプロバイダー・モデルはピース側の解決ロジック（CLI → persona_providers → ステップ → ローカル → グローバル）に統一 (#390)
• エージェントの provider/model 解決ロジックを resolveAgentProviderModel に一元化し、ムーブメント解決と同じ優先順位チェーンを使用するよう変更 (#386)
• movement:start イベントが providerInfo を含むよう変更し、表示側でのプロバイダー再解決を不要に (#390)
• takt list の「Sync with root」を「Merge from root」にリネーム (#394)
• インタラクティブモードの要約 AI がセッション非継承で実行されるよう修正し、会話コンテキストの汚染を防止 (#368)
• interactive policy のガイドラインを改善: ユーザーが「自分で調べて」と指示した場合と、ピースへの指示作成を区別するルールを明確化

Fixed

• default / default-test-first-mini ピースの write_tests ムーブメントで、テスト対象が未実装の場合にスキップして implement へ進むルールを追加（従来は ABORT になっていた）(#396)
• takt add の GitHub Issue タイトル抽出を改善: Markdown 見出し（h1-h3）を優先的にタイトルとして使用するよう変更（従来は先頭行がそのまま使われていた）(#368)
• quiet モードの要約 AI がセッションを引き継がない問題を修正 (#368)
• repertoire add の gh api 呼び出しにバッファサイズ上限（100MB）を設定し、大きなリポジトリでのバッファオーバーフローを防止
• E2E テストで gh ユーザー検索が無効な場合にローカルリポジトリへフォールバックするよう修正

Internal

• TeamLeaderRunner をリファクタリング: 実行ロジック（team-leader-execution.ts）、集約（team-leader-aggregation.ts）、共通ユーティリティ（team-leader-common.ts）、ストリーミング（team-leader-streaming.ts）に分離
• more-parts.json スキーマと loadMorePartsSchema ローダーを追加
• AGENTS.md を更新（プロジェクト構成とガイドラインの改訂）
• テスト拡充: provider/model 解決マトリクス、TeamLeader refill threshold / worker pool / aggregation / execution、OptionsBuilder、stream-buffer、conversationLoop resume、quietMode session、createIssueFromTask、schema-loader

[0.25.0] - 2026-02-26

Added

• Terraform/AWS ピース: IaC 開発用の完全なピースとファセット一式を追加。plan → implement → 並列3レビュー（architect/QA/security）→ supervise → complete の15ムーブメント構成（EN/JA）
• GitProvider 抽象化: Git/GitHub 操作を GitProvider インターフェースに統一し、将来の複数 Git プロバイダー対応の基盤を構築 (#375)
• プロジェクト設定で submodule の自動取得をサポート: submodules: all または submodules: [path1, path2] で指定可能に (#387)
• takt add で GitHub Issue 作成時にラベルをインタラクティブに選択可能に (#377, #111)
• deep-research ピースにデータ保存・レポート出力機能を追加（dig/analyze ムーブメントに Write・Bash ツール許可、supervise に research-report 出力契約）
• GitHub Discussions・Discord・X への一斉アナウンス GitHub Actions ワークフローを追加

Changed

• default ピースをテスト先行開発構成に変更: plan の後に write_tests ムーブメントを追加し、テストを先に書いてから実装する流れに。並列レビューに testing-review を追加（3→4 レビュアー）。レポートファイル名をセマンティック命名に統一（00-plan.md → plan.md 等）
• sync with root をピースエンジン経由からプロバイダー抽象化を利用した単発エージェント呼び出しに簡素化。コンフリクト解決プロンプトをテンプレートファイル化（EN/JA 分離）

Fixed

• lineEditor でサロゲートペア（絵文字等）のカーソル位置がずれる問題を修正。Ctrl+J による改行挿入を追加
• --task オプションでの直接実行時に tasks.yaml へ不要な記録がされる問題を修正
• --task でワークツリー作成時は tasks.yaml に記録するよう修正（takt list でのブランチ管理に必要）
• Provider resolution: removed implicit fallback to claude and switched to fail-fast when provider cannot be resolved (#386)
• Provider resolution: unified display and execution provider/model resolution via movement:start event providerInfo, ensuring displayed provider always matches execution provider (#390)
• E2E テスト config-priority の不安定性を修正 (#388)

Internal

• GitProvider 抽象化に伴うテスト追加（github-provider, taskGit）と既存テストのインポート更新
• CLAUDE.md 更新

[0.24.0] - 2026-02-24

Added

• AskUserQuestion support: AI agents can now ask interactive questions during execution with single-select, multi-select, and free-text input via TTY UI; automatically denied during piece execution to maintain agent autonomy (#161, #369)
• review builtin piece with 3-mode auto-detection: automatically selects PR mode (by PR number), branch mode (by branch name), or working diff mode (by free text) for multi-perspective parallel review
• testing-reviewer and requirements-reviewer builtin personas for specialized review perspectives
• testing policy: integration test requirement criteria (3+ module data flow, state merging into workflows, option propagation through call chains)
• gather-review instruction and review-gather output contract for the new review piece gather movement
• requirements-review instruction and output contract for requirements-focused review
• testing-review output contract for testing-focused review
• settingSources: ['project'] in SDK options: delegates CLAUDE.md loading to the Claude SDK for proper project-level settings resolution

Changed

• BREAKING: review-only piece renamed to review; review-fix-minimal piece removed — users referencing these piece names must update to review
• write-tests-first instruction now includes integration test decision criteria instead of a generic "Write E2E tests if appropriate"

Fixed

• Planner persona: added bug fix propagation check rule (grep for same pattern in related files) and prohibited deferring decidable questions to Open Questions

Internal

• Docs: fixed music metaphor origin description, catalog gaps, broken links, orphaned documents, event names, API Key references, eject descriptions, removed stale personas section map from YAML example, aligned legacy terminology with current codebase
• New test suites: StreamDisplay, ask-user-question-handler, pieceExecution-ask-user-question, review-piece, opencode-client-cleanup
• Removed legacy review-only-piece test and loadProjectContext from session module (CLAUDE.md loading now delegated to SDK)

[0.23.0] - 2026-02-23

Added

• default-test-first-mini builtin piece for test-first development workflow
• auto_fetch global config: opt-in remote fetch before cloning to keep clones up-to-date (default: false)
• base_branch config (global/project): specify the base branch for clone creation (defaults to remote default branch)
• model project config: override model at the project level (.takt/config.yaml)
• concurrency project config: set parallel task count per project for takt run
• --create-worktree support in pipeline mode for worktree-based execution
• skipTaskList option: interactive "Execute" action skips adding to tasks.yaml
• takt list now displays GitHub Issue numbers alongside task names
• Retry failed tasks now offers to reuse the previous piece before prompting piece selection
• Pipeline mode Slack notifications: sends run summary with task details, duration, branch, and PR URL
• CI workflow: lint, test, and e2e:mock checks run automatically on PRs (#364)

Changed

• Provider/model resolution unified via resolveProviderModelCandidates() — single resolution function used in both AgentRunner and resolveMovementProviderModel
• Pipeline execution refactored into thin orchestrator (execute.ts) + step implementations (steps.ts)
• Clone directory default changed from takt-worktree (singular) to takt-worktrees (plural) with auto-migration of legacy directory
• PR titles now include issue number prefix (e.g., [#6] Fix the bug)
• Task status now reflects PR creation failure — previously only piece execution success was tracked
• auto-tag.yml tags PR head SHA instead of merge commit for correct hotfix code publishing
• Session reader falls back to JSONL file scanning when sessions-index.json is missing or invalid
• ProjectLocalConfig type normalized to camelCase (auto_pr→autoPr, draft_pr→draftPr) — YAML snake_case preserved
• getLocalLayerValue simplified from switch-case to dynamic property lookup

Fixed

• repertoire add pipe stdin: multiple confirm() calls failed when reading from piped stdin due to readline destroying buffered lines (#334)
• Movement provider override precedence in AgentRunner: step provider was incorrectly overridden by global config
• Project-level model config was silently ignored — getLocalLayerValue was missing the model case
• PR creation failure now properly propagated as task failure with error message (#345)
• Claude session resume candidates now fall back to JSONL file scanning when sessions-index.json is unavailable

Internal

• CI: PR checks for lint, test, e2e:mock (ci.yml)
• Expanded e2e test coverage for repertoire (#364)
• New test suites: clone, config, postExecution, session-reader, selectAndExecute-skipTaskList, taskStatusLabel, pipelineExecution
• Refactored: project config case normalization (#358), clone manager (#359), pipeline steps extraction, confirm pipe reader singleton, provider resolution (#362)

[0.22.0] - 2026-02-22

Added

• Repertoire package system (takt repertoire add/remove/list): Import and manage external TAKT packages from GitHub — takt repertoire add github:{owner}/{repo}@{ref} downloads packages to ~/.takt/repertoire/ with atomic installation, version compatibility checks, lock files, and package content summary before confirmation
• @scope references in piece YAML: Facet references now support @{owner}/{repo}/{facet-name} syntax to reference facets from installed repertoire packages (e.g., persona: @nrslib/takt-fullstack/expert-coder)
• 4-layer facet resolution: Upgraded from 3-layer (project → user → builtin) to 4-layer (package-local → project → user → builtin) — repertoire package pieces automatically resolve their own facets first
• Repertoire category in piece selection: Installed repertoire packages automatically appear as subcategories under a "repertoire" category in the piece selection UI
• Build gate in implement/fix instructions: implement and fix builtin instructions now require build (type check) verification before test execution
• Repertoire package documentation: Added comprehensive docs for the repertoire package system (en, ja)

Changed

• BREAKING: Facets directory restructured: Facet directories moved under a facets/ subdirectory at all levels — builtins/{lang}/{facetType}/ → builtins/{lang}/facets/{facetType}/, ~/.takt/{facetType}/ → ~/.takt/facets/{facetType}/, .takt/{facetType}/ → .takt/facets/{facetType}/. Migration: move your custom facet files into the new facets/ subdirectory
• Contract string hardcoding prevention rule added to coding policy and architecture review instruction

Fixed

• Override piece validation now includes repertoire scope via the resolver
• takt export-cc now reads facets from the new builtins/{lang}/facets/ directory structure
• confirm() prompt now supports piped stdin (e.g., echo "y" | takt repertoire add ...)
• Suppressed poll_tick debug log flooding during iteration input wait
• Piece resolver stat() calls now catch errors gracefully instead of crashing on inaccessible entries

Internal

• Comprehensive repertoire test suite: atomic-update, repertoire-paths, file-filter, github-ref-resolver, github-spec, list, lock-file, pack-summary, package-facet-resolution, remove-reference-check, remove, takt-repertoire-config, tar-parser, takt-repertoire-schema
• Added src/faceted-prompting/scope.ts for @scope reference parsing, validation, and resolution
• Added scope-ref tests for the faceted-prompting module
• Added inputWait.ts for shared input-wait state to suppress worker pool log noise
• Added piece-selection-branches and repertoire e2e tests

[0.21.0] - 2026-02-20

Added

• Slack task notification enhancements: Extended Slack webhook notifications with richer task context and formatting (#316)
• takt list --delete-all option: Delete all tasks at once from the task list (#322)
• --draft-pr option: Create pull requests as drafts via --draft-pr flag (#323)
• --sync-with-root option: Sync worktree branch with root repository changes (#325)
• Model per persona-provider: Allow specifying model overrides at the persona-provider level (#324)
• Analytics project config and env override: Analytics settings can now be configured per-project and overridden via environment variables
• CI dependency health check: Periodic CI check to detect broken dependency packages

Changed

• Config system overhaul: Replaced loadConfig() bulk merge with per-key resolveConfigValue() resolution — global < piece < project < env priority with source tracking and OptionsBuilder merge direction control (#324)

Fixed

• Retry command scope and messaging: Fixed retry command to show correct available range and guidance text
• Retry task completed_at leak: Clear completed_at when moving a failed task back to running via startReExecution, preventing Zod validation errors
• OpenCode multi-turn hang: Removed streamAbortController.signal from OpenCode server startup so subsequent turns no longer hang; restored sessionId carry-over for multi-turn conversations
• Romaji conversion stack overflow: Prevented stack overflow on long task names during romaji conversion

[0.20.1] - 2026-02-20

Fixed

• Pin @opencode-ai/sdk to <1.2.7 to fix broken v2 exports that caused Cannot find module errors on npm install -g takt (#329)

[0.20.0] - 2026-02-19

Added

• Faceted Prompting module (src/faceted-prompting/): Standalone library for facet composition, resolution, template rendering, and truncation — zero dependencies on TAKT internals. Includes DataEngine interface with FileDataEngine and CompositeDataEngine implementations for pluggable facet storage
• Analytics module (src/features/analytics/): Local-only review quality metrics collection — event types (review findings, fix actions, movement results), JSONL writer with date-based rotation, report parser, and metrics computation
• takt metrics review command: Display review quality metrics (re-report counts, round-trip ratio, resolution iterations, REJECT counts by rule, rebuttal resolution ratio) with configurable time window (--since)
• takt purge command: Purge old analytics event files with configurable retention period (--retention-days)
• takt reset config command: Reset global config to builtin template with automatic backup of the existing config
• PR duplicate prevention: When a PR already exists for the current branch, push and comment on the existing PR instead of creating a duplicate (#304)
• Retry mode now positions the cursor on the failed movement when selecting which movement to retry
• E2E tests for run-recovery and config-priority scenarios

Changed

• README overhaul: Compressed from ~950 lines to ~270 lines — details split into dedicated docs (docs/configuration.md, docs/cli-reference.md, docs/task-management.md, docs/ci-cd.md, docs/builtin-catalog.md) with Japanese equivalents. Redefined product concept around 4 value axes: batteries included, practical, reproducible, multi-agent
• Config system refactored: Unified configuration resolution to resolveConfigValue() and loadConfig(), eliminating scattered config access patterns across the codebase
• takt config command removed: Replaced by takt reset config for resetting to defaults
• Builtin config templates refreshed with updated comments and structure
• @anthropic-ai/claude-agent-sdk updated to v0.2.47
• Instruct mode prompt improvements for task re-instruction

Fixed

• Fixed issue where builtin piece file references used absolute path instead of relative (#304)
• Removed unused imports and variables across multiple files

Internal

• Unified loadConfig, resolveConfigValue, piece config resolution, and config priority paths
• Added E2E tests for config priority and run recovery scenarios
• Added postExecution.test.ts for PR creation flow testing
• Cleaned up unused imports and variables

[0.19.0] - 2026-02-18

Added

• Dedicated retry mode for failed tasks — conversation loop with failure context (error details, failed movement, last message), run session data, and piece structure injected into the system prompt
• Dedicated instruct system prompt for completed/failed task re-instruction — injects task name, content, branch changes, and retry notes directly into the prompt instead of using the generic interactive prompt
• Direct re-execution from takt list — "execute" action now runs the task immediately in the existing worktree instead of only requeuing to pending
• startReExecution atomic task transition — moves a completed/failed task directly to running status, avoiding the requeue → claim race condition
• Worktree reuse in task execution — reuses existing clone directory when it's still on disk, skipping branch name generation and clone creation
• Task history injection into interactive and summary system prompts — completed/failed/interrupted task summaries are included for context
• Previous run reference support in interactive and instruct system prompts — users can reference logs and reports from prior runs
• findRunForTask and getRunPaths helpers for automatic run session lookup by task content
• isStaleRunningTask process helper extracted from TaskLifecycleService for reuse

Changed

• Interactive module split: interactive.ts refactored into interactive-summary.ts, runSelector.ts, runSessionReader.ts, and selectorUtils.ts for better cohesion
• requeueTask now accepts generic allowedStatuses parameter instead of only accepting failed tasks
• Instruct/retry actions in takt list use the worktree path for conversation and run data lookup instead of the project root
• save_task action now requeues the task (saves for later execution), while execute action runs immediately

Internal

• Removed DebugConfig from models, schemas, and global config — simplified to verbose mode only
• Added stdin simulation test helpers (stdinSimulator.ts) for E2E conversation loop testing
• Added comprehensive E2E tests for retry mode, interactive routes, and run session injection
• Added check:release npm script for pre-release validation

[0.18.2] - 2026-02-18

Added

• Added codex_cli_path global config option and TAKT_CODEX_CLI_PATH environment variable to override the Codex CLI binary path used by the Codex SDK (#292)
  • Supports strict validation: absolute path, file existence, executable permission, no control characters
  • Priority: TAKT_CODEX_CLI_PATH env var > codex_cli_path in config.yaml > SDK vendored binary

[0.18.1] - 2026-02-18

Added

• Added multi-tenant data isolation section and authorization-resolver consistency code examples to security knowledge
• Added "prefer project scripts" rule to coding policy — detects direct tool invocation (e.g., npx vitest) when equivalent npm scripts exist

[0.18.0] - 2026-02-17

Added

• deep-research builtin piece: Multi-angle research workflow with four steps — plan, deep-dive, analyze, and synthesize
• Project-level .takt/ facets (pieces, personas, policies, knowledge, instructions, output-contracts) are now version-controllable (#286)
• New research facets added: research policy, knowledge, comparative-analysis knowledge, dedicated persona, and instructions

Changed

• Refactored the research piece — separated rules and knowledge embedded in the persona into policy, knowledge, and instruction files, conforming to the faceted design
• Added knowledge/policy references to existing pieces (expert, expert-cqrs, backend, backend-cqrs, frontend)

Fixed

• Fixed a bug where facet directories were not tracked because .takt/ path prefix was written with .takt/ prefix in the .takt/.gitignore template (dotgitignore)

Internal

• Created knowledge facet style guide (KNOWLEDGE_STYLE_GUIDE.md)
• Added regression tests for dotgitignore patterns

[0.17.3] - 2026-02-16

Added

• Added API client generation consistency rules to builtin AI anti-pattern policy and frontend knowledge — detects handwritten clients mixed into projects where generation tools (e.g., Orval) exist

Fixed

• Fixed EPERM crash when releasing task store locks — replaced file-based locking with in-memory guard

Internal

• Unified vitest configuration for e2e tests and added forceExit option to prevent zombie workers

[0.17.2] - 2026-02-15

Added

• expert-mini and expert-cqrs-mini pieces: Lightweight variants of Expert pieces — plan → implement → parallel review (AI anti-pattern + supervisor) → fix workflow
• Added new pieces to "Mini" and "Expert" piece categories

Fixed

• Fixed an error being thrown when permission mode could not be resolved — now falls back to readonly

[0.17.1] - 2026-02-15

Changed

• Changed .takt/.gitignore template to allowlist approach — ignores all files by default and tracks only config.yaml. Prevents ignore gaps when new files are added

[0.17.0] - 2026-02-15

Added

• Mini piece series: Added default-mini, frontend-mini, backend-mini, backend-cqrs-mini — lightweight development pieces with parallel review (AI anti-pattern + supervisor) as successors to coding/minimal
• Added "Mini" category to piece categories
• supervisor-validation output contract: Requirements Fulfillment Check table format that presents code evidence per requirement
• getJudgmentReportFiles(): Phase 3 status judgment target reports can now be filtered via use_judge flag
• Added finding_id tracking to output contracts (new/persists/resolved sections for tracking findings across iterations)

Changed

• BREAKING: Removed coding and minimal pieces — replaced by the mini piece series. Migration: coding → default-mini, minimal → default-mini
• BREAKING: Unified output contract to item format — use_judge (boolean) and format (string) fields are now required; OutputContractLabelPath (label:path format) is removed
• Moved runtime environment directory from .runtime to .takt/.runtime
• Enhanced supervisor requirements verification: extracts requirements individually and verifies one-by-one against code (file:line) — "roughly complete" is no longer valid grounds for APPROVE

Fixed

• Added retry mechanism for deleting clone/worktree directories (maxRetries: 3, retryDelay: 200) — reduces transient deletion failures caused by file locks

Internal

• Removed review-summary output contract (consolidated into supervisor-validation)
• Updated all builtin pieces, e2e fixtures, and tests to the new output contract format

[0.16.0] - 2026-02-15

Added

• Provider-specific permission profiles (provider_profiles): Define default permission modes per provider and per-movement overrides in global (~/.takt/config.yaml) and project (.takt/config.yaml) config — 5-level priority resolution (project override → global override → project default → global default → required_permission_mode floor)

Changed

• BREAKING: permission_mode → required_permission_mode: Renamed movement's permission_mode field to required_permission_mode — acts as a floor value; the actual permission mode is resolved via provider_profiles. Old permission_mode is rejected by z.never(), no backward compatibility
• Rewrote builtin config.yaml template: reorganized comments, added provider_profiles description and examples, added OpenCode-related settings

Internal

• Added tests for provider profile resolution (global-provider-profiles, project-provider-profiles, permission-profile-resolution, options-builder)
• Added missing loadProjectConfig mock to parallel execution tests

[0.15.0] - 2026-02-15

Added

• Runtime environment presets: piece_config.runtime.prepare and global config runtime.prepare allow environment preparation scripts to run automatically before piece execution — builtin presets (gradle, node) isolate dependency resolution and cache setup to the .runtime/ directory
• Loop monitor judge instruction: loop_monitors judge config now supports instruction_template field — externalizes loop judgment instructions as an instruction facet, applied to builtin pieces (expert, expert-cqrs)

Internal

• Added runtime environment tests (runtime-environment, globalConfig-defaults, models, provider-options-piece-parser)
• Added provider e2e test (runtime-config-provider)

[0.14.0] - 2026-02-14

Added

• takt list instruct mode (#267): Added instruct mode for issuing additional instructions to existing branches — refine requirements through a conversation loop before piece execution
• takt list completed task actions (#271): Added diff view and branch operations (merge, delete) for completed tasks
• Claude sandbox configuration: provider_options.claude.sandbox supports excluded_commands and allow_unsandboxed_commands
• provider_options global/project config: provider_options can now be set in ~/.takt/config.yaml (global) and .takt/config.yaml (project) — acts as lowest-priority fallback for piece-level settings

Changed

• Consolidated provider/model resolution into AgentRunner: Fixed provider resolution to prioritize project config over custom agent config. Added step-level stepModel/stepProvider overrides
• Unified post-execution flow: Shared postExecution.ts for interactive mode and instruct mode (auto-commit, push, PR creation)
• Added scope-narrowing prevention to instructions: plan, ai-review, and supervise instructions now require detecting missed requirements — plan mandates per-requirement "change needed/not needed" judgments with rationale, supervise prohibits blindly trusting plan reports

Fixed

• Fixed a bug where interactive mode options were displayed during async execution (#266)
• Fixed OpenCode session ID not being carried over during parallel execution — server singleton prevents race conditions in parallel runs
• Extended OpenCode SDK server startup timeout from 30 seconds to 60 seconds

Internal

• Large-scale task management refactor: split TaskRunner responsibilities into TaskLifecycleService, TaskDeletionService, and TaskQueryService
• Split taskActions.ts by feature: taskBranchLifecycleActions.ts, taskDiffActions.ts, taskInstructionActions.ts, taskDeleteActions.ts
• Added postExecution.ts, taskResultHandler.ts, instructMode.ts, taskActionTarget.ts
• Consolidated piece selection logic into pieceSelection/index.ts (extracted from selectAndExecute.ts)
• Added/expanded tests: instructMode, listNonInteractive-completedActions, listTasksInteractiveStatusActions, option-resolution-order, taskInstructionActions, selectAndExecute-autoPr, etc.
• Added Claude Code sandbox option (dangerouslyDisableSandbox) to E2E tests
• Added OPENCODE_CONFIG_CONTENT to .gitignore

[0.13.0] - 2026-02-13

Added

• Team Leader movement: New movement type where a team leader agent dynamically decomposes a task into sub-tasks (Parts) and executes multiple part agents in parallel — supports team_leader config (persona, maxParts, timeoutMs, partPersona, partEdit, partPermissionMode) (#244)
• Structured Output: Introduced JSON Schema-based structured output for agent calls — three schemas for task decomposition, rule evaluation, and status judgment added to builtins/schemas/. Supported by both Claude and Codex providers (#257)
• provider_options piece-level config: Provider-specific options (codex.network_access, opencode.network_access) can now be set at piece level (piece_config.provider_options) and individual movements — Codex/OpenCode network access enabled in all builtin pieces
• backend builtin piece: New backend development piece — parallel specialist review by backend, security, and QA reviewers
• backend-cqrs builtin piece: New CQRS+ES backend development piece — parallel specialist review by CQRS+ES, security, and QA reviewers
• AbortSignal for part timeouts: Added timeout control and parent signal propagation via AbortSignal for Team Leader part execution
• Agent usecase layer: agent-usecases.ts consolidates agent call usecases (decomposeTask, executeAgent, evaluateRules) and centralizes structured output injection

Changed

• BREAKING: Public API cleanup: Significantly narrowed the public API in src/index.ts — internal implementation details (session management, Claude/Codex client internals, utility functions, etc.) are no longer exported, reducing the API surface to a stable minimum (#257)
• Revamped Phase 3 judgment logic: Removed JudgmentDetector/FallbackStrategy and consolidated into status-judgment-phase.ts with structured output-based judgment. Improves stability and maintainability (#257)
• Report phase retry improvement: Report Phase (Phase 2) now automatically retries with a new session when it fails (#245)
• Unified Ctrl+C shutdown: Removed sigintHandler.ts and consolidated into ShutdownManager — graceful shutdown → timeout → force-kill in three stages, unified across all providers (#237)
• Scope-deletion guardrails: Added rules to coder persona prohibiting deletions and structural changes outside the task instruction scope. Added scope discipline and reference material priority rules to planner persona
• Added design token and theme scope guidance to frontend knowledge
• Improved architecture knowledge (both en/ja)

Fixed

• Fixed checkout failure for existing branches during clone — now passes --branch to git clone --shared then removes the remote
• Removed # from issue-referenced branch names (takt/#N/slug → takt/N/slug)
• Resolved deprecated tool dependency in OpenCode report phase; migrated to permission-based control (#246)
• Removed unnecessary exports to ensure public API consistency

Internal

• Added Team Leader tests (engine-team-leader, team-leader-schema-loader, task-decomposer)
• Added structured output tests (parseStructuredOutput, claude-executor-structured-output, codex-structured-output, provider-structured-output, structured-output E2E)
• Added unit tests for ShutdownManager
• Added unit tests for AbortSignal (abort-signal, claude-executor-abort-signal, claude-provider-abort-signal)
• Added unit tests for Report Phase retry (report-phase-retry)
• Added unit tests for public API exports (public-api-exports)
• Added tests for provider_options (provider-options-piece-parser, models, opencode-types)
• Significantly expanded E2E tests: cycle-detection, model-override, multi-step-sequential, pipeline-local-repo, report-file-output, run-sigint-graceful, session-log, structured-output, task-status-persistence
• Refactored E2E test helpers (extracted shared setup functions)
• Removed judgment/ directory (JudgmentDetector, FallbackStrategy)
• Added ruleIndex.ts utility (1-based → 0-based index conversion)

[0.12.1] - 2026-02-11

Fixed

• Fixed silent fallthrough to a new session when the session was not found — now shows an info message when no session is detected

Internal

• Set OpenCode provider report phase to deny (prevents unnecessary writes in Phase 2)
• Skip copying tasks/ directory during project initialization (TASK-FORMAT is no longer needed)
• Added stream diagnostics utility (streamDiagnostics.ts)

[0.12.0] - 2026-02-11

Added

• OpenCode provider: Native support for OpenCode as a third provider — SDK integration via @opencode-ai/sdk/v2, permission mapping (readonly/edit/full → reject/once/always), SSE stream handling, retry mechanism (up to 3 times), and hang detection with 10-minute timeout (#236, #238)
• Arpeggio movement: New movement type for data-driven batch processing — CSV data source with batch splitting, template expansion ({line:N}, {col:N:name}, {batch_index}), concurrent LLM calls (Semaphore-controlled), and concat/custom merge strategies (#200)
• frontend builtin piece: Frontend development piece — React/Next.js knowledge injection, coding/testing policy, parallel architecture review
• Slack Webhook notifications: Automatic Slack notification on piece completion — configured via TAKT_NOTIFY_WEBHOOK env var, 10-second timeout, non-blocking on failure (#234)
• Session selector UI: On interactive mode startup, select a resumable session from past Claude Code sessions — shows latest 10 sessions with initial input and last response preview (#180)
• Provider event logs: Claude/Codex/OpenCode execution events written to NDJSON files — .takt/logs/{sessionId}-provider-events.jsonl, with automatic compression of large text (#236)
• Provider/model name display: Active provider and model name shown in console output at each movement execution

Changed

• Revamped takt add: Auto-add to task on issue selection, removed interactive mode, added task stacking confirmation on issue creation (#193, #194)
• max_iteration → max_movement unification: Unified terminology for iteration limits; added ostinato for unlimited execution (#212)
• Improved previous_response injection: Implemented length control and always-inject Source Path (#207)
• Task management improvements: Redefined .takt/tasks/ as storage for long-form task specs; completeTask() removes completed records from tasks.yaml (#201, #204)
• Improved review output: Updated review output format; moved past reports to history log (#209)
• Simplified builtin pieces: Further streamlined top-level declarations across all builtin pieces

Fixed

• Fixed Report Phase blocked behavior: Report Phase (Phase 2) now retries with a new session when blocked (#163)
• Fixed OpenCode hang and termination detection: Suppressed prompt echo, suppressed question prompts, fixed hang issues, corrected termination detection (#238)
• Fixed OpenCode permission and tool wiring: Corrected permission and tool wiring during edit execution
• Worktree task spec copy: Fixed task spec not being correctly copied during worktree execution
• Fixed lint errors (merge/resolveTask/confirm)

Internal

• Comprehensive OpenCode provider tests added (client-cleanup, config, provider, stream-handler, types)
• Comprehensive Arpeggio tests added (csv, data-source-factory, merge, schema, template, engine-arpeggio)
• Significantly expanded E2E tests: cli-catalog, cli-clear, cli-config, cli-export-cc, cli-help, cli-prompt, cli-reset-categories, cli-switch, error-handling, piece-error-handling, provider-error, quiet-mode, run-multiple-tasks, task-content-file (#192, #198)
• Added providerEventLogger.ts, providerModel.ts, slackWebhook.ts, session-reader.ts, sessionSelector.ts, provider-resolution.ts, run-paths.ts
• Added ArpeggioRunner.ts (data-driven batch processing engine)
• AI Judge now routes through provider system (Codex/OpenCode support)
• Added/expanded tests: report-phase-blocked, phase-runner-report-history, judgment-fallback, pieceExecution-session-loading, globalConfig-defaults, session-reader, sessionSelector, slackWebhook, providerEventLogger, provider-model, interactive, run-paths, engine-test-helpers

[0.11.1] - 2026-02-10

Fixed

• Fixed AI Judge to route through provider system — changed callAiJudge from a Claude-only implementation to provider-based (runAgent), enabling correct AI judgment with the Codex provider
• Reduced instruction bloat — set pass_previous_response: false in implement/fix movements, prioritizing reports in the Report Directory as primary information source (en/ja)

Internal

• Improved CI workflow to automatically sync npm next dist-tag to latest on stable releases (with retry)

[0.11.0] - 2026-02-10

Added

• e2e-test builtin piece: E2E test focused piece — E2E analysis → E2E implementation → review → fix flow (for Vitest-based E2E tests)
• error status: Separated provider errors from blocked, enabling clear distinction of error states. Added retry mechanism to Codex
• Centralized task YAML management: Unified task file management into tasks.yaml. Structured task lifecycle management (pending/running/completed/failed) via TaskRecordSchema
• Task spec documentation: Documented the structure and purpose of task specs (#174)
• Review policy: Added shared review policy facet (builtins/{lang}/policies/review.md)
• SIGINT graceful shutdown E2E test: E2E test to verify Ctrl+C behavior during parallel execution

Changed

• Simplified builtin pieces: Removed top-level policies/personas/knowledge/instructions/report_formats declarations from all builtin pieces, migrating to implicit name-based resolution. Piece YAML is now simpler
• Updated piece category spec: Improved category configuration and display logic. Enhanced category management in global config (#184)
• Improved takt list priority and resolution: Optimized branch resolution performance. Introduced base commit cache (#186, #195, #196)
• Improved Ctrl+C signal handling: Stabilized SIGINT handling during parallel execution
• Strengthened loop prevention policy: Enhanced policy to prevent agent infinite loops

Fixed

• Fixed original instruction diff processing not working correctly (#181)
• Fixed task spec goal being inappropriately scope-expanded — goal is now always fixed to implementation and execution

Internal

• Large-scale task management refactor: removed parser.ts and split into store.ts/mapper.ts/schema.ts/naming.ts. Split branch resolution into branchGitResolver.ts/branchBaseCandidateResolver.ts/branchBaseRefCache.ts/branchEntryPointResolver.ts
• Significantly expanded and refactored tests: added aggregate-evaluator, blocked-handler, branchGitResolver-performance, branchList-regression, buildListItems-performance, error-utils, escape, facet-resolution, getFilesChanged, global-pieceCategories, instruction-context, instruction-helpers, judgment-strategies, listTasksInteractivePendingLabel, loop-detector, naming, reportDir, resetCategories, rule-evaluator, rule-utils, slug, state-manager, switchPiece, task-schema, text, transitions, watchTasks, etc.
• Refactored Codex client
• Improved facet resolution logic in piece parser

[0.10.0] - 2026-02-09

Added

• structural-reform builtin piece: Full project review and structural reform — iterative codebase restructuring with staged file splits, powered by loop_monitors
• unit-test builtin piece: Unit test focused piece — test analysis → test implementation → review → fix, with loop_monitors for cycle control
• test-planner persona: Specialized persona for analyzing codebase and planning comprehensive test strategies
• Interactive mode variants: Four selectable modes after piece selection — assistant (default: AI-guided requirement refinement), persona (conversation with first movement's persona), quiet (generate instructions without questions), passthrough (user input used as-is)
• persona_providers config: Per-persona provider overrides (e.g., { coder: 'codex' }) — route specific personas to different providers without creating hybrid pieces
• task_poll_interval_ms config: Configurable polling interval for takt run to detect new tasks during execution (default: 500ms, range: 100–5000ms)
• interactive_mode piece field: Piece-level default interactive mode override (e.g., set passthrough for pieces that don't benefit from AI planning)
• Task-level output prefixing: Colored [taskName] prefix on all output lines during parallel takt run execution, preventing mid-line interleaving between concurrent tasks
• Review policy facet: Shared review policy (builtins/{lang}/policies/review.md) for consistent review criteria across pieces

Changed

• BREAKING: Removed all Hybrid Codex pieces (*-hybrid-codex) — replaced by persona_providers config which achieves the same result without duplicating piece files
• Removed tools/generate-hybrid-codex.mjs (no longer needed with persona_providers)
• Improved parallel execution output: movement-level prefix now includes task context and iteration info in concurrent runs
• Codex client now detects stream hangs (10-minute idle timeout) and distinguishes timeout vs external abort in error messages
• Parallel task execution (takt run) now polls for newly added tasks during execution instead of only checking between task completions
• Parallel task execution no longer enforces per-task time limits (previously had a timeout)
• Issue references now routed through interactive mode (as initial input) instead of skipping interactive mode entirely
• Builtin config.yaml updated to document all GlobalConfig fields
• Extracted conversationLoop.ts for shared conversation logic across interactive mode variants
• Line editor improvements: additional key bindings and edge case fixes

Fixed

• Codex processes hanging indefinitely when stream becomes idle — now aborted after 10 minutes of inactivity, releasing worker pool slots

Internal

• New test coverage: engine-persona-providers, interactive-mode (532 lines), task-prefix-writer, workerPool expansion, pieceResolver expansion, lineEditor expansion, parallel-logger expansion, globalConfig-defaults expansion, pieceExecution-debug-prompts expansion, it-piece-loader expansion, runAllTasks-concurrency expansion, engine-parallel
• Extracted TaskPrefixWriter for task-level parallel output management
• Extracted modeSelection.ts, passthroughMode.ts, personaMode.ts, quietMode.ts from interactive module
• InteractiveMode type model added (src/core/models/interactive-mode.ts)
• PieceEngine validates taskPrefix/taskColorIndex pair consistency at construction
• Implementation notes document added (docs/implements/retry-and-session.ja.md)

[0.9.0] - 2026-02-08

Added

• takt catalog command: List available facets (personas, policies, knowledge, instructions, output-contracts) across layers (builtin/user/project)
• compound-eye builtin piece: Multi-model review — sends the same instruction to Claude and Codex simultaneously, then synthesizes both responses
• Parallel task execution: takt run now uses a worker pool for concurrent task execution (controlled by concurrency config, default: 1)
• Rich line editor in interactive mode: Shift+Enter for multiline input, cursor movement (arrow keys, Home/End), Option+Arrow word movement, Ctrl+A/E/K/U/W editing, paste bracket mode support
• Movement preview in interactive mode: Injects piece movement structure (persona + instruction) into the AI planner for improved task analysis (interactive_preview_movements config, default: 3)
• MCP server configuration: Per-movement MCP (Model Context Protocol) server settings with stdio/SSE/HTTP transport support
• Facet-level eject: takt eject persona coder — eject individual facets by type and name for customization
• 3-layer facet resolution: Personas, policies, and other facets resolved via project → user → builtin lookup (name-based references supported)
• pr-commenter persona: Specialized persona for posting review findings as GitHub PR comments
• notification_sound config: Enable/disable notification sounds (default: true)
• Prompt log viewer: tools/prompt-log-viewer.html for visualizing prompt-response pairs during debugging
• Auto-PR base branch now set to the current branch before branch creation

Changed

• Unified planner and architect-planner: extracted design knowledge into knowledge facets, merged into planner. Removed architect movement from default/coding pieces (plan → implement direct transition)
• Replaced readline with raw-mode line editor in interactive mode (cursor management, inter-line movement, Kitty keyboard protocol)
• Unified interactive mode save_task with takt add worktree setup flow
• Added -d flag to caffeinate to prevent App Nap process freezing during display sleep
• Issue references now routed through interactive mode (previously executed directly, now used as initial input)
• SDK update: @anthropic-ai/claude-agent-sdk v0.2.34 → v0.2.37
• Enhanced interactive session scoring prompts with piece structure information

Internal

• Extracted resource-resolver.ts for facet resolution logic (separated from pieceParser.ts)
• Extracted parallelExecution.ts (worker pool), resolveTask.ts (task resolution), sigintHandler.ts (shared SIGINT handler)
• Unified session key generation via session-key.ts
• New lineEditor.ts (raw-mode terminal input, escape sequence parsing, cursor management)
• Extensive test additions: catalog, facet-resolution, eject-facet, lineEditor, formatMovementPreviews, models, debug, strip-ansi, workerPool, runAllTasks-concurrency, session-key, interactive (major expansion), cli-routing-issue-resolve, parallel-logger, engine-parallel-failure, StreamDisplay, getCurrentBranch, globalConfig-defaults, pieceExecution-debug-prompts, selectAndExecute-autoPr, it-notification-sound, it-piece-loader, permission-mode (expansion)

[0.8.0] - 2026-02-08

Formal release of 0.8.0-alpha.1 content. No functional changes.

[0.8.0-alpha.1] - 2026-02-07

Added

• Faceted Prompting architecture: Prompt components are managed as independent files and can be freely combined across pieces
  • personas/ — persona prompts defining agent role and expertise
  • policies/ — policies defining coding standards, quality criteria, and prohibitions
  • knowledge/ — knowledge defining domain knowledge and architecture information
  • instructions/ — instructions defining movement-specific procedures
  • output-contracts/ — output contracts defining report output formats
  • Piece YAML section maps (personas:, policies:, knowledge:) associate keys with file paths; movements reference by key
• Output Contracts and Quality Gates: Structured definitions for report output and AI directives for quality criteria
  • output_contracts field defines reports (replaces report field)
  • quality_gates field specifies AI directives for movement completion requirements
• Knowledge system: Separates domain knowledge from personas, managed and injected at piece level
  • knowledge: section map in piece YAML defines knowledge files
  • Movements reference by key via knowledge: field
• Faceted Prompting documentation: Design philosophy and practical guide added to docs/faceted-prompting.md (en/ja)
• Hybrid Codex piece generation tool: tools/generate-hybrid-codex.mjs auto-generates Codex variants from Claude pieces
• Failed task re-queue: select failed task branches from takt list and re-execute (#110)
• Branch name generation strategy is now configurable (branch_name_strategy config)
• Added auto-PR feature and unified PR creation logic (#98)
• Piece selection now also applies for issue references (#97)
• Optional macOS idle sleep prevention during piece execution (#100)

Changed

• BREAKING: Renamed resources/global/ directory to builtins/
  • resources/global/{lang}/ → builtins/{lang}/
  • Changed files field in package.json from resources/ to builtins/
• BREAKING: Renamed agent field to persona
  • Piece YAML: agent: → persona:, agent_name: → persona_name:
  • Internal types: agentPath → personaPath, agentDisplayName → personaDisplayName, agentSessions → personaSessions
  • Directories: agents/ → personas/ (global, project, and builtin)
• BREAKING: Changed report field to output_contracts
  • Unified legacy report: 00-plan.md / report: [{Scope: ...}] / report: {name, order, format} formats to output_contracts: {report: [...]} format
• BREAKING: Renamed stances → policies, report_formats → output_contracts
• Migrated all builtin pieces to Faceted Prompting architecture (separated domain knowledge from old agent prompts into knowledge facets)
• SDK updates: @anthropic-ai/claude-agent-sdk v0.2.19 → v0.2.34, @openai/codex-sdk v0.91.0 → v0.98.0
• Added policy/knowledge fields to movements (referenced by section map keys)
• Added policy-based evaluation to interactive mode scoring
• Refreshed README: agent → persona, added section map description, clarified control/management classification
• Refreshed builtin skill (SKILL.md) for Faceted Prompting

Fixed

• Fixed report directory path resolution bug
• Fixed PR issue number link not being set correctly
• Fixed gitignored files being committed in stageAndCommit (removed git add -f .takt/reports/)

Internal

• Large-scale builtin resource restructuring: removed old agents/ directory structure (default/, expert/, expert-cqrs/, magi/, research/, templates/) and migrated to flat personas/, policies/, knowledge/, instructions/, output-contracts/ structure
• Added Faceted Prompting style guides and templates (PERSONA_STYLE_GUIDE.md, POLICY_STYLE_GUIDE.md, INSTRUCTION_STYLE_GUIDE.md, OUTPUT_CONTRACT_STYLE_GUIDE.md, etc. in builtins/ja/)
• Added policy, knowledge, and instruction resolution logic to pieceParser.ts
• Added/expanded tests: knowledge, policy-persona, deploySkill, StreamDisplay, globalConfig-defaults, sleep, task, taskExecution, taskRetryActions, addTask, saveTaskFile, parallel-logger, summarize
• Added policy and knowledge content injection to InstructionBuilder
• Added taskRetryActions.ts (failed task re-queue logic)
• Added sleep.ts utility
• Removed old prompt files (interactive-summary.md, interactive-system.md)
• Removed old agent templates (templates/coder.md, templates/planner.md, etc.)

[0.7.1] - 2026-02-06

Fixed

• Fixed Ctrl+C not working during piece execution: SIGINT handler now calls interruptAllQueries() to stop active SDK queries
• Fixed EPIPE crash after Ctrl+C: dual protection for EPIPE errors when SDK writes to stdin of a stopped child process (uncaughtException handler + Promise.resolve().catch())
• Fixed terminal raw mode leaking when an exception occurs in the select menu's onKeypress handler

Internal

• Added integration tests for SIGINT handler and EPIPE suppression (it-sigint-interrupt.test.ts)
• Added key input safety tests for select menu (select-rawmode-safety.test.ts)

[0.7.0] - 2026-02-06

Added

• Hybrid Codex pieces: Added Codex variants for all major pieces (default, minimal, expert, expert-cqrs, passthrough, review-fix-minimal, coding)
  • Hybrid configuration running the coder agent on the Codex provider
  • en/ja support
• passthrough piece: Minimal piece that passes the task directly to the coder
• takt export-cc command: Deploy builtin pieces and agents as Claude Code Skills
• Added delete action to takt list, separated non-interactive mode
• AI consultation action: takt add / interactive mode can now create GitHub Issues and save task files
• Cycle detection: Added CycleDetector to detect infinite loops between ai_review and ai_fix (#102)
  • Added arbitration step (ai_no_fix) to the default piece for when no fix is needed
• CI: Added workflow to auto-delete skipped TAKT Action runs weekly
• Added Hybrid Codex subcategory to piece categories (en/ja)

Changed

• Simplified category configuration: merged default-categories.yaml into piece-categories.yaml, changed to auto-copy to user directory
• Fixed subcategory navigation in piece selection UI (recursive hierarchical display now works correctly)
• Refreshed Claude Code Skill to Agent Team-based design
• Unified console.log to info() (list command)

Fixed

• Fixed YAML parse error caused by colons in Hybrid Codex piece descriptions
• Fixed invalid arguments passed to selectPieceFromCategoryTree on subcategory selection

Internal

• Refactored list command: separated listNonInteractive.ts, taskDeleteActions.ts
• Added cycle-detector.ts, integrated cycle detection into PieceEngine
• Refactored piece category loader (pieceCategories.ts, pieceSelection/index.ts)
• Added tests: cycle-detector, engine-loop-monitors, piece-selection, listNonInteractive, taskDeleteActions, createIssue, saveTaskFile

[0.6.0] - 2026-02-05

Formal release of RC1/RC2 content. No functional changes.

[0.6.0-rc1] - 2026-02-05

Fixed

• Fixed infinite loop between ai_review and ai_fix: resolved issue where ai_fix judging "no fix needed" caused a return to plan and restarted the full pipeline
  • Added ai_no_fix arbitration step (architecture-reviewer judges the ai_review vs ai_fix conflict)
  • Changed ai_fix "no fix needed" route from plan to ai_no_fix
  • Affected pieces: default, expert, expert-cqrs (en/ja)

Changed

• Changed default piece parallel reviewer from security-review to qa-review (optimized for TAKT development)
• Moved qa-reviewer agent from expert/ to default/ and rewrote with focus on test coverage
• Added iteration awareness to ai_review instruction (first iteration: comprehensive review; subsequent: prioritize fix verification)

Internal

• Restricted auto-tag workflow to merges from release/ branches only, unified publish job (resolves chained trigger failure due to GITHUB_TOKEN limitations)
• Removed postversion hook (conflicts with release branch flow)
• Updated tests: adapted to security-reviewer → qa-reviewer change

[0.6.0-rc] - 2026-02-05

Added

• coding builtin piece: Lightweight development piece — design → implement → parallel review → fix (fast feedback loop without plan/supervise steps)
• conductor agent: Dedicated agent for Phase 3 judgment. Reads reports and responses to output judgment tags
• Phase 3 judgment fallback strategy: 4-stage fallback (AutoSelect → ReportBased → ResponseBased → AgentConsult) to improve judgment accuracy (src/core/piece/judgment/)
• Session state management: Saves task execution results (success/error/interrupted) and displays previous result on next interactive mode startup (#89)
• TAKT meta information (piece structure, progress) injection mechanism for agents
• /play command: Immediately executes task in interactive mode
• E2E test infrastructure: mock/provider-compatible test infrastructure, 10 E2E test specs, test helpers (isolated-env, takt-runner, test-repo)
• Added detection rule for "logically unreachable defensive code" to review agents

Changed

• Changed Phase 3 judgment logic from session-resume approach to conductor agent + fallback strategy (improved judgment stability)
• Refactored CLI routing as executeDefaultAction() function, reusable as fallback from slash commands (#32)
• Input starting with / or # is now accepted as task instruction when no command/issue is found (#32)
• Simplified isDirectTask(): only issue references execute directly, all others go to interactive mode
• Removed pass_previous_response: true from all builtin pieces (redundant as it is the default behavior)

Internal

• Added E2E test config files (vitest.config.e2e.ts, vitest.config.e2e.mock.ts, vitest.config.e2e.provider.ts)
• Added getReportFiles(), hasOnlyOneBranch(), getAutoSelectedTag() to rule-utils.ts
• Added report content and response-based judgment instruction generation to StatusJudgmentBuilder
• Added piece meta information (structure, iteration counts) injection to InstructionBuilder
• Added tests: judgment-detector, judgment-fallback, sessionState, pieceResolver, cli-slash-hash, e2e-helpers

[0.5.1] - 2026-02-04

Fixed

• Windows environment file path handling and encoding issues (#90, #91)
  • Improved .git detection for Windows
  • Added mandatory .git check for Codex (error if not found)
  • Fixed character encoding issues
• Codex branch name summary processing bug

Internal

• Test memory leak and hanging issues resolved
  • Added cleanup handlers for PieceEngine and TaskWatcher
  • Changed vitest to single-threaded execution for improved test stability

[0.5.0] - 2026-02-04

Changed

• BREAKING: Complete terminology migration from "workflow" to "piece" across entire codebase
  • All CLI commands, configuration files, and documentation now use "piece" terminology
  • WorkflowEngine → PieceEngine
  • workflow_categories → piece_categories in config files
  • builtin_workflows_enabled → builtin_pieces_enabled
  • ~/.takt/workflows/ → ~/.takt/pieces/ (user piece directory)
  • .takt/workflows/ → .takt/pieces/ (project piece directory)
  • All workflow-related file names and types renamed to piece-equivalents
  • Updated all documentation (README.md, CLAUDE.md, docs/*)

Internal

• Complete directory structure refactoring:
  • src/core/workflow/ → src/core/piece/
  • src/features/workflowSelection/ → src/features/pieceSelection/
• File renames:
  • workflow-types.ts → piece-types.ts
  • workflowExecution.ts → pieceExecution.ts
  • workflowLoader.ts → pieceLoader.ts
  • workflowParser.ts → pieceParser.ts
  • workflowResolver.ts → pieceResolver.ts
  • workflowCategories.ts → pieceCategories.ts
  • switchWorkflow.ts → switchPiece.ts
• All test files updated to reflect new terminology (194 files changed, ~3,400 insertions, ~3,400 deletions)
• Resources directory updated:
  • resources/global/*/pieces/*.yaml updated with new terminology
  • All prompt files (*.md) updated
  • Configuration files (config.yaml, default-categories.yaml) updated

[0.4.1] - 2026-02-04

Fixed

• Workflow execution bug where previous step's response was incorrectly bound to subsequent steps
  • Fixed MovementExecutor, ParallelRunner, and state-manager to properly isolate step responses
  • Updated interactive summary prompts to prevent response leakage

[0.4.0] - 2026-02-04

Added

• Externalized prompt system: all internal prompts moved to versioned, translatable files (src/shared/prompts/en/, src/shared/prompts/ja/)
• i18n label system: UI labels extracted to separate YAML files (labels_en.yaml, labels_ja.yaml) with src/shared/i18n/ module
• Prompt preview functionality (src/features/prompt/preview.ts)
• Phase system injection into agents for improved workflow phase awareness
• Enhanced debug capabilities with new debug log viewer (tools/debug-log-viewer.html)
• Comprehensive test coverage:
  • i18n system tests (i18n.test.ts)
  • Prompt system tests (prompts.test.ts)
  • Session management tests (session.test.ts)
  • Worktree integration tests (it-worktree-delete.test.ts, it-worktree-sessions.test.ts)

Changed

• BREAKING: Internal terminology renamed: WorkflowStep → WorkflowMovement, StepExecutor → MovementExecutor, ParallelSubStepRawSchema → ParallelSubMovementRawSchema, WorkflowStepRawSchema → WorkflowMovementRawSchema
• BREAKING: Removed unnecessary backward compatibility code
• BREAKING: Disabled interactive prompt override feature
• Workflow resource directory renamed: resources/global/*/workflows/ → resources/global/*/pieces/
• Prompts restructured for better readability and maintainability
• Removed unnecessary task requirement summarization from conversation flow
• Suppressed unnecessary report output during workflow execution

Fixed

• takt worktree bug fix for worktree operations

Internal

• Extracted prompt management into src/shared/prompts/index.ts with language-aware file loading
• Created src/shared/i18n/index.ts for centralized label management
• Enhanced tools/jsonl-viewer.html with additional features
• Major refactoring across 162 files (~5,800 insertions, ~2,900 deletions)

[0.3.9] - 2026-02-03

Added

• Workflow categorization support (#85)
  • Default category configuration in resources/global/{lang}/default-categories.yaml
  • User-defined categories via workflow_categories in ~/.takt/config.yaml
  • Nested category support with unlimited depth
  • Category-based workflow filtering in workflow selection UI
  • show_others_category and others_category_name configuration options
  • Builtin workflow filtering via builtin_workflows_enabled and disabled_builtins
• Agent-less step execution: agent field is now optional (#71)
  • Steps can execute with instruction_template only (no system prompt)
  • Inline system prompts supported (agent string used as prompt if file doesn't exist)
• takt add #N automatically reflects issue number in branch name (#78)
  • Issue number embedded in branch name (e.g., takt/issue-28-...)

Changed

• BREAKING: Permission mode values unified to provider-independent format (#87)
  • New values: readonly, edit, full (replaces default, acceptEdits, bypassPermissions)
  • TAKT translates to provider-specific flags (Claude: default/acceptEdits/bypassPermissions, Codex: read-only/workspace-write/danger-full-access)
  • All builtin workflows updated to use new values
• Workflow naming changes:
  • simple workflow replaced with minimal and review-fix-minimal
  • Added review-only workflow for read-only code review
• Agent prompts updated with legacy対応禁止ルール (no backward compatibility hacks)
• Documentation updates:
  • README.md and docs/README.ja.md updated with v0.3.8+ features
  • CLAUDE.md significantly expanded with architectural details and implementation notes

Internal

• Created src/infra/config/loaders/workflowCategories.ts for category management
• Created src/features/workflowSelection/index.ts for workflow selection UI
• Enhanced src/shared/prompt/select.ts with category display support
• Added comprehensive tests for workflow categories (workflow-categories.test.ts, workflow-category-config.test.ts)

[0.3.8] - 2026-02-02

Added

• CLI option to specify workflow/config file paths: --workflow <path> and --config <path> (#81)
• CI-friendly quiet mode for minimal log output (#70)
• Mock scenario support for testing workflow execution
• Comprehensive integration tests (7 test files, ~3000 lines of test coverage)

Changed

• Rule evaluation improved: detectRuleIndex now uses last match instead of first match (#25)
• ai_fix step significantly improved:
  • Added {step_iteration} counter to show retry attempt number
  • Explicit fix procedure defined (Read → Grep → Edit → Test → Report)
  • Coder agent now prioritizes reviewer feedback over assumptions
• README and docs updated with clearer CLI usage and CI/CD examples

Fixed

• Workflow loading priority corrected (user workflows now take precedence over builtins)
• Test stability improvements (flaky tests skipped, ai_fix test updated)
• Slack notification configuration fixed

Internal

• Refactored instruction builder: extracted context assembly and status rules logic (#44)
• Introduced src/infra/task/git.ts for DRY git commit operations
• Unified error handling with getErrorMessage()
• Made projectCwd required throughout codebase
• Removed deprecated sacrificeMode
• 35 files updated for consistency (console.log → blankLine(), etc.)

[0.3.7] - 2026-02-01

Added

• --pipeline flag for explicit pipeline/non-interactive mode execution (#28)
• Pipeline mode can be used with both --task and --issue options

Changed

• Log file naming changed from base36 to human-readable YYYYMMDD-HHmmss-random format (#28)
• --task option description updated to clarify it's an alternative to GitHub issue

[0.3.6] - 2026-01-31

Fixed

• ai_review workflow step now correctly includes pass_previous_request setting

[0.3.5] - 2026-01-31

Added

• --create-worktree <yes|no> option to skip worktree confirmation prompt

Fixed

• Various CI/CD improvements and fixes (#66, #67, #68, #69)

[0.3.4] - 2026-01-31

Added

• Review-only workflow for code review without modifications (#60)
• Various bug fixes and improvements (#14, #23, #35, #38, #45, #50, #51, #52, #59)

[0.3.3] - 2026-01-31

Fixed

• Fixed takt add #N passing issue content through AI summarization and corrupting task content (#46)
  • Changed to use resolveIssueTask result directly as the task when referencing issues

[0.3.1] - 2026-01-31

Added

• Interactive task planning mode: takt (no args) starts AI conversation to refine task requirements before execution (#47, #5)
  • Session persistence across takt restarts
  • Read-only tools (Read, Glob, Grep, Bash, WebSearch, WebFetch) for codebase investigation
  • Planning-only system prompt prevents code changes during conversation
  • /go to confirm and execute, /cancel to exit
• Boy Scout Rule enforcement in reviewer/supervisor agent templates

Changed

• CLI migrated from slash commands (takt /run-tasks) to subcommands (takt run) (#47)
• /help and /refresh-builtin commands removed; eject simplified
• SDK options builder only includes defined values to prevent hangs

Fixed

• Claude Agent SDK hanging when model: undefined or other undefined options were passed as keys

[0.3.0] - 2026-01-30

Added

• Rule-based workflow transitions with 5-stage fallback evaluation (#30)
  • Tag-based conditions: agent outputs [STEP:N] tags matched by index
  • ai() conditions: AI evaluates free-text conditions against agent output (#9)
  • all()/any() aggregate conditions for parallel step results (#20)
  • 5-stage evaluation order: aggregate → Phase 3 tag → Phase 1 tag → AI judge → AI fallback
• 3-phase step execution model (#33)
  • Phase 1: Main work (coding, review, etc.)
  • Phase 2: Report output (when step.report defined)
  • Phase 3: Status judgment (when tag-based rules exist)
  • Session resumed across phases for context continuity
• Parallel step execution with concurrent sub-steps via Promise.all() (#20)
• GitHub Issue integration: execute/add tasks by issue number, e.g. takt #6 (#10, #34)
• NDJSON session logging with real-time streaming writes (#27, #36)
• Builtin resources embedded in npm package with /eject command for customization (#4, #40)
• edit property for per-step file edit control
• Rule match method visualization and logging
• Report output auto-generation from YAML report.format
• Parallel review support in builtin workflows with spec compliance checking (#31)
• WorkflowEngine mock integration tests (#17, #41)

Changed

• Report format unified to auto-generation; manual order/instruction_template for reports removed
• gitdiff report type removed in favor of format-based reports

Fixed

• Report directory correctly includes .takt/reports/ prefix (#37, #42)
• Unused import in eject.ts (#43)

[0.2.3] - 2026-01-29

Added

• /list-tasks command for branch management (try merge, merge & cleanup, delete)

Changed

• Isolated execution migrated from git worktree to git clone --shared to prevent Claude Code SDK from traversing back to main repository
• Clone lifecycle: auto-deletion after task completion removed; use /list-tasks for cleanup
• worktree.ts split into clone.ts + branchReview.ts
• Origin remote removed from clones to block SDK traversal
• All workflow report steps granted Write permission
• git clone --shared changed to --reference --dissociate

Fixed

• Version read from package.json instead of hardcoded 0.1.0 (#3)

[0.2.2] - 2026-01-29

Added

• /review instruct action for executing instructions on task branches
• AI-powered task name summarization to English slugs for branch names
• Worktree session inheritance
• Execution Rules metadata (git commit prohibition, cd prohibition)

Changed

• Status output rule headers auto-generated
• Instructions auto-include worktree change context
• Try Merge changed to squash merge
• expert-review renamed to expert-cqrs; common reviewers consolidated under expert/

Fixed

• Tasks incorrectly progressing to completed on abnormal termination

[0.2.1] - 2026-01-28

Added

• Language setting (ja/en)
• Multiline input support for /add-task
• /review-tasks command
• Cursor-based (arrow key) menu selection replacing numeric input
• answer status, autoCommit, permission_mode, verbose logging options

Fixed

• Multiple worktree-related bugs (directory resolution, session handling, creation flow)
• ESC key cancels workflow/task selection

[0.2.0] - 2026-01-27

Added

• /watch command for file system polling and auto-executing tasks from .takt/tasks/
• /refresh-builtin command for updating builtin resources
• /add-task command for interactive task creation
• Enhanced default workflows

[0.1.7] - 2026-01-27

Added

• Schema permission support for workflow validation

[0.1.6] - 2026-01-27

Added

• Mock execution mode for testing

Changed

• -r option omitted; default changed to conversation continuation mode

[0.1.5] - 2026-01-27

Added

• Total execution time output

Fixed

• Workflow unintentionally stopping during execution

[0.1.4] - 2026-01-27

Changed

• Workflow prompts strengthened
• Transition prompts consolidated into workflow definitions

[0.1.3] - 2026-01-26

Fixed

• Iteration stalling issue

[0.1.2] - 2026-01-26

Added

• Codex provider support
• Model selection per step/agent
• Permission mode configuration
• Worktree support for isolated task execution
• Project .gitignore initialization

Changed

• Agent prompts refined

[0.1.1] - 2026-01-25

Added

• GitHub Actions workflow for npm publish

Changed

• Interactive mode removed; CLI simplified