v0.3.0a7 — Foreground tasks and handoff automation

Install project guidance and register a window together, then process ready tasks and explicit handoff receipts from one bounded foreground command.

• start --apply-setup --dry-run previews setup and registration. Conflicts prevent registration; custom rules remain preserved.
• worker watch continues ready work, supports finite peer-change waits and explicit delivered-message receipts, emits compact progress and stops on decisions/command/scope failures. No command retries, lease takeover, automatic commits or GUI launches.
• Native OpenClaw verification binds the generated isolated profile before discovery. --native-executable selects a trusted local CLI while keeping managed-service launchers intact.
• Updated bundled setup Skill and bilingual foreground guide.

Validated: 420 source and 420 independently installed-package tests; Linux/Windows Python 3.10/3.13; three real worktrees completing four tasks and three explicit receipts; recorded a6→a7 Skill updates; native local OpenClaw discovery. Fixtures are deterministic and invoke no models. WorkBuddy GUI testing remains deferred.

python -m pip install --upgrade --pre coprogrammer
coprogrammer start --client codex --session codex-api --apply-setup --dry-run
coprogrammer start --client codex --session codex-api --apply-setup

Foreground tasks and messages · Installation maintenance

v0.3.0a6 — Refresh installed Skills and MCP safely

Package upgrades can now refresh copied Skills, templates and MCP launchers without overwriting project customization.

• Preview with coprogrammer setup --client codex --upgrade; add --apply to update files still matching their local setup receipt.
• Preserve project instructions and customized/unrecorded files; keep independent receipts per Git worktree and serialize cooperating installers.
• Include maintenance/recovery guidance and implementation choices informed by gstack, Beads and MCP Agent Mail.

Validated: 395 source tests and 395 independently installed-wheel tests; 7 client profiles × 3 editions; two real Python environments; native OpenClaw profile/tool discovery; actual a5→a6 in-place adoption. Model access remains separately checked and WorkBuddy GUI work remains deferred. Multi-file updates are not filesystem transactions.

Installation maintenance · Research decisions

python -m pip install --upgrade --pre coprogrammer
coprogrammer setup --client codex --upgrade
coprogrammer setup --client codex --upgrade --apply

v0.3.0a5 · Native activation checks and agent harness workflows

CoProgrammer 0.3.0a5 adds native activation checks and a foreground agent harness workflow.

• OpenClaw: isolated embedded profile, installed-client configuration validation and native MCP discovery.
• WorkBuddy: documented project .codebuddy/skills and .workbuddy/mcp.json paths, preserving older import kits and existing settings.
• Explicit worker run: claim/guard, invoke a trusted command, check changes, complete work and send selected local handoffs. Failure retains work for inspection.
• Opt-in inbox delivery: acknowledge only the harness’s explicit delivered-message receipts; refuse undelivered IDs or project edits.
• Problem-based README, use cases and package search metadata; no claim of adoption or measured productivity gains.

Install in a virtual environment: python -m pip install --pre --upgrade coprogrammer==0.3.0a5.

Native activation and harness guide · When this is useful

Native OpenClaw 2026.6.8 task/handoff calls were tested through its actual SessionMcpRuntime with a deterministic controller and an external harness fixture. No model was invoked. WorkBuddy 5.7.7 signature passed; GUI testing was deferred at the user’s request after UI reads timed out. No live model credentials were available in the inspected clients. Scope guards remain cooperative, not a sandbox; no background daemon, automatic retries, account setup or arbitrary-command MCP tool.

Validation: 380 source tests and 380 independently installed-wheel tests passed; Ubuntu/Windows Python 3.10/3.13 CI passed on PR and merged main. Fresh PyPI installation matches all 44 package files and passed real installed OpenClaw discovery, the installed runner receipt/edit/finish/handoff loop, 11 demo steps, four runtime profiles in three editions and stdio discovery of 34 tools/12 resources. Live model/account success and WorkBuddy GUI discovery are not claimed.

v0.3.0a4 · More AI runtimes, one collaboration workflow

CoProgrammer 0.3.0a4 extends the existing local collaboration workflow to more AI runtimes.

• OpenClaw: workspace Skills and an importable nodeHost.mcp.servers recipe.
• WorkBuddy: six Skills and a stdio MCP import kit, without guessed native config paths.
• Agent harnesses and other hosts: generic Skill/MCP kits or CLI work-boundary calls.
• Grok: explicit advisory reviews using the Responses API, account-selected models, bounded output and store disabled.
• Additive activation reports distinguish generated configuration from client connection.

All profiles reuse task ownership, scope guards, explicit inbox acknowledgement and durable handoffs. Installation offers full, skills and mcp editions, preserving existing configuration. No runtime dependency or event/schema migration.

Install in a virtual environment: python -m pip install --pre --upgrade coprogrammer==0.3.0a4. Preview with coprogrammer setup --client openclaw --edition full --include-content; review the result before adding --apply.

Installation and activation guide

Local CLI/MCP and provider-fixture validation does not establish live model access, OpenClaw node pairing or WorkBuddy GUI discovery. Generic Harness support assumes no particular product API. Other computers and independent clones continue to use portable handoffs rather than distributed live leases.

Validation: 364 source tests and 364 independently installed-wheel tests passed. Ubuntu/Windows Python 3.10/3.13 CI passed on both the PR and merged main. The published PyPI wheel matches all 41 package files and passed fresh installation, 11 demo steps, four new runtime profiles in three editions, real stdio MCP discovery, task worker cycles and Grok offline preview.

CoProgrammer 0.3.0a3 — Automatic teamwork, Skill and MCP editions

CoProgrammer now reduces manual coordination between Codex, Claude Code and GLM/DeepSeek-backed coding windows, with Skill, MCP and full installation editions.

• worker step / MCP worker_step: preview by default; explicitly apply to pulse, inspect inbox/decisions, select work and atomically claim or guard/renew existing ownership. Actual Git changes and planned paths are checked. Pending messages and decisions pause work.
• worker finish / MCP worker_finish: guard changes, finish/release the claim and send handoffs to explicit local recipients in one transaction. Identical retries reuse completion and messages; invalid recipients roll back the operation.
• start --provider ... --model ...: record declared provider/model identity. Task eligibility consistently matches the host or provider label during dispatch, claim and replay. Resume cannot silently change identity.
• setup --edition skills|mcp|full: choose six Skills plus Markdown, only native MCP configuration, or both. Existing instructions and conflicting files remain protected.
• Bilingual READMEs, the automatic collaboration guide and mirrored active-sync Skill explain first use, inbox handling, provider routing, scope boundaries and retries.

In an active Python environment:

python -m pip install --pre --upgrade "coprogrammer==0.3.0a3"
coprogrammer demo
coprogrammer setup --client codex --edition full --include-content

After reviewing the preview, repeat setup with --apply, reload the coding client and complete its normal trust flow. Each concurrent task uses its own committed Git worktree:

coprogrammer start --client codex --session codex-api
coprogrammer worker step --session codex-api --apply

A Claude Code window already configured for GLM can register --client claude --provider glm; use --provider deepseek for DeepSeek. These are declared routing labels, not credentials or proof of a model connection.

Validation: 352 source tests, 352 independently installed-wheel tests and all Ubuntu/Windows Python 3.10/3.13 CI jobs passed. Twenty-four new regressions cover real worktrees/concurrency, guard rollback, provider routing, inbox/decision pauses, notification retries and setup editions. The published PyPI wheel was checksum-verified and compared with the tested package, installed fresh, and exercised through the 11-step demo, real stdio MCP (34 tools, 12 resources), all three editions and a three-worktree provider-routed worker/notification/ACK cycle.

This remains an alpha preview. No event/config migration or runtime dependency was added. Participating agents call these boundaries; there is no background launcher, client wakeup, live model request, automatic ACK or silent takeover. Independent clones still have separate local Managers and reconcile portable handoffs.

Automatic collaboration guide · 中文 README · Plugin guide

CoProgrammer 0.3.0a2 — Teamwork for coding agents

CoProgrammer makes teamwork between coding windows easier to start and understand: shared tasks, file boundaries and durable handoffs for Codex, Claude Code and Copilot.

• coprogrammer start registers a window or explicitly resumes it, with dry-run and client setup preview.
• coprogrammer status shows eligible work, blockers and next steps. The read-only workspace_status MCP tool provides the same view without refreshing sessions, claiming tasks or acknowledging messages.
• English and Chinese READMEs now explain the problem, show a real demo excerpt and link a complete two-worktree walkthrough.
• Unicode Git branches/worktrees are decoded consistently across platforms.
• Atomic Manager commits retain their transaction lock across bounded Windows file-contention retries. Persistent errors preserve the previous log and fail clearly.

Install the preview in an active virtual environment:

python -m pip install --pre --upgrade "coprogrammer==0.3.0a2"
coprogrammer demo

In a committed project worktree on a named branch:

coprogrammer start --client codex --session codex-api
coprogrammer status --session codex-api

Validation: 328 source tests, 328 independently installed-wheel tests and the Ubuntu/Windows Python 3.10/3.13 CI matrix passed. The published PyPI wheel was downloaded, checksum-verified, compared with the tested package files, installed in a fresh environment and exercised through the 11-step demo and real stdio MCP calls (32 tools, 12 Markdown resources).

This remains an alpha preview. Existing event/config formats remain compatible; no migration is needed. Startup does not launch clients or install configuration. Coordination is local to cooperating worktrees; independent clones do not share a global lock. These checks do not establish native client GUI connections or live model/provider access.

English README · 中文说明 · Two-window walkthrough

CoProgrammer v0.3.0a1 — workspace dispatch and portable handoffs

v0.3 adds portable context checks for teams developing in multiple coding windows and moving work between computers. This is an alpha release.

• Preview tasks eligible for the current session, client and worktree before the existing atomic claim.
• Run an explicit local check, then detect stale results using file-content fingerprints, command hashes, runtime identity and expiry.
• Export JSON or Markdown handoffs and compare the receiving clone's package version, exact HEAD/base, repository lineage and content.
• Inspect the same workflow through five read-only MCP tools and matching packaged Skills.
• Fix Windows Manager lock contention, concurrent setup path checks and UTF-8 workspace artifact output.

Validation: 310 source tests and 310 tests against an independently installed wheel, including 28 added regressions and real two-clone/stdio MCP journeys. The final PR checks cover Ubuntu and Windows with Python 3.10 and 3.13. Config, manifests, integration plan, JSON/YAML and repository Markdown links were checked; the native Claude plugin validator passed.

python -m pip install --pre --upgrade "coprogrammer==0.3.0a1"
coprogrammer --version

Read dispatch and handoff and check receipts.

Handoffs compare context; they do not transfer ownership between independent Managers. Receiving computers still need local validation and claims. Check records are unauthenticated local evidence and exclude ignored files and external dependencies. Existing events and client configuration need no migration. Native client GUI connections and live model-provider requests were not exercised.

Publication readback: the existing Trusted Publishing workflow succeeded. PyPI exposes version 0.3.0a1 with a wheel and source archive. The downloaded wheel matched its published SHA-256 and all 38 package files matched the independently tested wheel. A fresh installation passed the 11-step offline demo and stdio MCP checks (31 tools, including the five additions, and 12 resources). The published MCP server verified the current source check as fresh.