Skip to content

Changelog

This file records notable changes to zdev.

  • Coding guidance now accompanies planning, implementation and independent review. Shared rules cover design, names, contracts, structure and completion; focused supplements cover Rust, TypeScript, MiniZinc, Java, Kotlin, Python, C++, web behavior, React and React Router framework mode.
  • Reviews distinguish required corrections from optional suggestions. Ordinary implementation uses one full review, one correction pass and one focused recheck. Recovery retains that review stage.
  • Explicit blanket approval for compatible parallel batches remains valid throughout an ongoing goal or loop, within the approved area and execution limits. Each batch still rechecks admission and verifies each task.
  • Research workflow tests use three representative local scenarios instead of a broad fault-injection matrix. Tests use scripted workers and never invoke an LLM or a live coding harness.
  • Codex and Claude Code support authorized serial and rolling research with measured baselines, correctness checks, retained candidates, bounded work and command allowances, and recovery from durable records.
  • Codex and Claude Code can coordinate sequential workers through T3 Code.
  • The companion area-report skill produces editable MDX and standalone HTML.
  • zdev skill install codex --to <directory> --portable generates a relocatable bundle for Codex plugin distribution. The corresponding skill check accepts the same flag.
  • zdev tasks list without an area prints the available areas as normal output and exits successfully, including when no areas exist.
  • The normal profile uses GPT-6.1 Sol for routine OpenAI implementation roles, matching standard and advanced work. Luna requires the explicit Codex simple profile or a worker override. Standard Sol implementation and verification use high reasoning.
  • Failed derived-task publication removes only files created by that operation, preserving existing task files when a generated filename collides.
  • Work-context capture and comparison reject dirty tracked submodules instead of accepting incomplete evidence of their working contents.
  • Rebaseline rejects untracked code created during measurement before accepting samples as measurements of the committed revision.
  • Parent changes and managed rebases acquire the existing state lock before reading and updating area metadata.
  • Rolling supporting work uses the ordinary workflow’s findings check once, preserving work allowances for implementation and verification.
  • Standalone Claude verification preserves task complexity when checking the resolved worker dispatch.
  • Research completion commits use the standard Zdev-Change-Id trailer.
  • Rolling integration accepts unchanged measured code after actual record-only changes without rerunning the evaluator.
  • Shared research records remain readable when clone-local experiment refs are absent. Admission checks the recorded candidate commit when available and reports how to fetch missing retained code.
  • Cargo source packages include the research probe needed to compile tests. Release checks compile packaged tests and install from the extracted package.
  • Existing 1.x records and worker configuration remain valid. The record schema remains version 1. See Upgrade to 2.0.
  • Refresh binary-installed integrations with zdev skill install <harness> --force to receive the new guidance and fixes. Use the same scope, destination and guidance selection as the existing install.
  • zdev tasks list without an area now reports the available areas, their lifecycle, and their open-task counts. Open areas appear before closed areas in both text and JSON output.
  • Verification no longer starts a corrective implementation round when its only findings are validation writes and the canonical work context remains unchanged. Semantic findings and retained checkout changes still require correction.
  • Existing records and worker configuration remain valid. Reinstall integrations with zdev skill install <harness> --force to receive the updated verification behavior.
  • Areas can select an execution profile independently for each harness. Area profiles take precedence over saved local and global defaults while leaving one-off and run-specific profile choices in control.
  • Installed harness skills can inspect and change area profiles from a plain configuration request when the area and harness are named.
  • zdev config profile show <profile> <harness> now prints every resolved role, including its model, effort, and any fallback source.
  • Command help now explains profile precedence, derived-task review and apply paths, and the safeguards and recovery steps for squash cleanup. Automated coverage includes every reachable help screen.
  • Existing areas without profile mappings keep their previous defaults. Reinstall integrations with zdev skill install <harness> --force to receive the updated configuration route and area-aware profile resolution.
  • Shared prose guidance now incorporates Cursor pstack’s unslop method. Zdev asks every supported harness to preserve meaning while using concrete facts, plain active language, readable sentences, and stable repository terms.
  • Briefs, slice briefs, tasks, and worker instructions now incorporate Matt Pocock’s writing-for-agents method. Durable records keep shared meaning in one place, use branch-specific reading pointers, avoid restating facts an agent can inspect cheaply, and define checkable completion conditions.
  • Durable planning records and transient worker payloads have separate writing rules, removing repeated instructions while preserving the context each worker actually loads.
  • Existing records, tasks, and worker configuration remain valid. Reinstall integrations with zdev skill install <harness> --force to receive the new writing guidance.
  • Google Antigravity (agy) is now a supported harness integration, with workspace and user-scoped Skills and custom subagents under Antigravity’s native .agents and ~/.gemini/config layouts.
  • Antigravity worker profiles use Gemini 3.8 Flash with role-specific zdev effort levels: low for routine work, medium for implementation and verification, and high for planning and advanced implementation.
  • Antigravity subagents are explicitly sandboxed and marked as subagent-only; they cannot be selected as the primary agent.
  • Antigravity coordinates zdev workflows through the installed Skill and invoke_subagent; durable goal state remains in .zdev.
  • Existing records and integrations remain valid. Install or refresh the new integration with zdev skill install agy --force.
  • An approved parallel-task route coordinates isolated worktrees, serial integration, independent verification, and one durable commit per task. The route is available in Codex, Claude Code, Pi, OpenCode, and Oh My Pi, with harness-specific fallbacks where native orchestration is unavailable.
  • Named execution profiles configure planner, implementer, and verifier model settings. Profiles can be selected globally, per harness, or for one command, and zdev plan exposes the same scoped routing for plan-only work.
  • Workflow and Git transitions that change the shape of the work now require explicit approval, while directly necessary technical adjustments can remain within an accepted task’s scope.
  • Task-writing and worker guidance use plainer, agent-focused language and keep actionable blocker context available for recovery.
  • zdev commit --root commits the selected repository even when inherited GIT_DIR or GIT_WORK_TREE variables point elsewhere.
  • Work-context snapshots detect content changes in existing untracked files, so validation cannot silently change a new implementation file while its snapshot still compares equal.
  • Existing records, tasks, and worker configuration remain valid. Reinstall integrations with zdev skill install <harness> --force to receive the new parallel, planning, profile-routing, and recovery workflows.
  • Task and planner paths describe expected changes rather than exhaustive lists of permitted files. Necessary files with clear task ownership no longer turn work within the approved scope into a blocker.
  • Goal loops distinguish external or safety blockers from unfinished work that can still proceed. Claude Code retries actionable worker blockers and stops repeated attempts when snapshot comparison shows no attributable progress.
  • Task drafting and focused selection keep unattended repository work ahead of work that requires physical devices, credentials, production access, or other user involvement.
  • Verified, coherent work proceeds through the normal completion and commit flow instead of remaining dirty solely because a later attended check exists.
  • Verifiers prefer check or dry-run forms for generators, avoiding dirty worktrees caused by verification when generation belongs to implementation.
  • Claude Code project integrations resolve their installed task-workflow contract from the project destination instead of a user-scoped path.
  • Existing records and task state remain valid. Reinstall integrations with zdev skill install <harness> --force to receive the updated loop, task-writing, worker, and verifier contracts.
  • Claude Code uses Haiku for deterministic work-context preflight, snapshot, and comparison agents instead of inheriting the session model.
  • Verifier instructions explicitly forbid extra semantic keys and require validation-written files as rework findings across supported harnesses.
  • Claude Code accepts completion fields written as blocks, brief prose before a status line, and commit IDs wrapped in backticks. These presentation differences no longer stop a valid workflow result.
  • Existing records and task state remain valid. Reinstall integrations with zdev skill install <harness> --force to receive the updated workflows and verifier instructions.
  • Claude Code accepts advanced-task plans that include supporting findings or normalized absolute checkout paths. These useful planner results no longer fail the workflow’s semantic envelope check.
  • Existing records and task state remain valid. Reinstall integrations with zdev skill install <harness> --force to receive the updated planner contract and Claude Code workflow.
  • Tasks may declare optional afk and priority metadata. With no focus, zdev selects ready work by AFK suitability, priority, then numeric task ID.
  • Area loops accept fuzzy focus text. The coordinating model inspects the full ready frontier, chooses the best-fitting task, and admits that exact task through zdev work-context <area> --task <task-id>.
  • Claude Code workflows pass compact stored work-context locators instead of relaying full task and Git context through model output.
  • Worker-result parsing accepts one valid JSON object inside brief prose or a Markdown fence while still rejecting malformed, truncated, ambiguous, or semantically invalid results.
  • Worker prompts are concise and role-specific across Claude Code, Codex, OpenCode, Pi, and Oh My Pi. Loop progress names task selection, worker stages, completed tasks, and commits.
  • Reopening a completed task preserves its earlier Result under History instead of discarding it or moving it into live Context.
  • Focused loops no longer silently fall back to the lowest ready task.
  • Long task context and harmless JSON wrapping no longer cause whole-workflow retries in Claude Code.
  • Compact snapshot handoffs remain readable throughout implementation, verification, and rework.
  • Installed worker contract paths resolve across user, project, and explicit integration destinations.
  • Existing tasks remain valid. Omitted afk and priority values behave as false and normal, preserving the earlier numeric ordering.
  • Existing pending task reviews remain usable when their bundles omit the new optional task metadata.
  • Reinstall integrations with zdev skill install <harness> --force to receive the 1.2.0 workflows and worker contracts.
  • Planner workers now return a small, harness-independent semantic result. Claude Code and Oh My Pi constrain that result at dispatch, while all harnesses share the same strict planner contract without a formatting retry.
  • Verifier workers report only semantic conclusions. Coordinators own snapshot comparison and public-envelope bookkeeping, and Claude Code avoids redundant full-context collection around verification.
  • Task-challenge follow-up is proportional to the change instead of requiring a new full review for every adjustment.
  • Claude Code workflows use the rendered inline contract when CLAUDE_PLUGIN_ROOT is unset or does not point to a readable installed contract.
  • Claude Code planner validation now rejects duplicate keys inside nested plan objects as well as malformed top-level results.
  • Existing records and task state remain valid. Reinstall integrations with zdev skill install <harness> --force to receive the updated worker contracts and harness workflows.
  • Committed initial task imports now publish complete project and pull-request records. Personal records remain local, and later imports into existing tracked areas retain their previous behavior.
  • Explicit trunk areas for personal and project records. Several areas can safely share the configured project trunk while task selection, lifecycle, and commits retain exact area and task ownership.
  • Lightweight slice briefs, a standing general area convention, explicit area closure, and zdev next --any for deterministic cross-area discovery.
  • Read-only zdev goal and zdev work-context projections, plus one installed skill that routes implement, verify, audit, and goal/loop work on each of the five harnesses.
  • Optional content-addressed work-context files with exact show and compact fresh-compare commands for large worker handoffs.
  • Layered worker configuration and routine, standard, and advanced task complexity with fresh independent verification and bounded escalation.
  • Atomic derived-task review and application for authorized investigation follow-ups and implementation splits.
  • Derived-task manual review now stores its proposal and Markdown in Git administrative state, so coordinators can show and apply an opaque review identity without replaying proposal content or carrying a fingerprint.
  • Harness integrations now render from one strict canonical template path and use typed worker-result envelopes. Route contracts live inside the single skill; native commands, prompts, and workflows remain harness adapters. Small audits use one verifier unless the user explicitly supplies bounded review lenses.
  • Task imports return their ready frontier, committed imports can include an approved area brief, and review identities are opaque coordinator handoffs.
  • Task review stores a presentable Markdown document and canonical bundle in repository-local Git state, so zdev can show and import the current review without reconstructing Markdown or asking users to transport a fingerprint.
  • Verifier PASS and completion exchange one stored work-context ID instead of serializing raw Git diffs through worker envelopes and prompts; zdev performs the fresh comparison before task mutation.
  • Claude task workflows load one rendered contract from the installed plugin instead of repeating it in every worker prompt, while keeping the detailed role instruction as a same-call fallback.
  • Safe-but-stale isolated branches remain workable with one advisory; explicit trunk areas do not use freshness checks or rebasing.
  • Existing project, area, task, and slice records remain valid. New optional lifecycle, complexity, slice, and area-mode fields are written only when the corresponding feature is used.
  • Reinstall integrations with zdev skill install <harness> --force to receive the 1.1.0 commands, workflows, worker profiles, and strict envelope contract.
  • Claude Code installs now include native plugin workflows. Codex and Oh My Pi use native goal tools when available; OpenCode and Pi stop after one task and report whether more work is ready.
  • A file-based task loop for planning, selecting, completing, and checking work.
  • Personal, project, and pull-request policies for storing .zdev planning records.
  • Stable Zdev-Change-Id trailers for commits created with zdev commit.
  • Managed area relationships, rebasing, recovery guidance, and squash cleanup.
  • Reviewed JSON task bundles with approval IDs that bind review to import.
  • Native integrations for Codex, Claude Code, OpenCode, Pi, and Oh My Pi.
  • Human-readable output and versioned JSON output for scripting.
  • Release archives for macOS and Linux on x86-64 and Arm64.