C2C 2.5.0 — Developer QA and clearer evidence

C2C now supports developer QA with evidence from authorized local tests. The coordinating chat executes requested checks; peers challenge the supplied results, missing cases and regressions.

• QA runs require an exact-plan assurance record linking selected options, accepted fixes and proposed, observed or blocked checks to supplied evidence. Invalid hashes, ambiguous JSON and future observations are rejected.
• A shared spending preference applies in both apps. New launches honor the stricter current or sealed policy; included-only workers require current route-matched evidence before an attempt is reserved. No paid recovery or account changes.
• Shorter discussion excerpts and a compact completion record preserve findings, source links and review limits without repeating the full plan. All three deliverables remain: revised plan, discussion and result.
• Updated guidance checks contradictions, source holds, negative cases, retests and execution entry points. Retained runs keep their original runtime; current user restrictions must still be honored when resuming older versions.

Use $C2C in Codex or /C2C in Claude Code. Ask it to review a change and run the relevant authorized local tests, or request test planning only.

Validation uses offline regressions, native platform CI and isolated instruction walkthroughs. These checks do not establish live provider quality, production certification or a token-saving percentage. QA guide · Setup and updates.

C2C 2.4.0 — Explicit effort and clearer revision checks

C2C 2.4.0 gives planners clearer effort settings and reviewers a focused view of plan changes while keeping the complete evidence available.

• Select and record compatible worker effort explicitly. Requested effort stays separate from effective effort, which remains unknown when the provider does not report it.
• Include bounded revision diffs alongside the full current plan, findings, decisions and security review. Oversized diffs fall back safely without dropping authoritative context.
• Apply consistent outcome, dependency, acceptance and severity guidance to software, content and mixed projects.
• Reduce unnecessary status polling where the host supports longer waits, and preserve reviewed plans when adding navigation or later amendments.
• Register the frozen 2.3.0 evaluation baseline without inventing measured outcomes or token-saving percentages.

The three-file handoff, usage limits, existing runtime pins and managed update/rollback behavior remain supported. Start a new chat after updating to use the new instructions for a new plan.

Validated with 525 passing local tests and nine platform skips, documentation checks, independent source review and fictional instruction walkthroughs. All 15 native CI jobs passed on the exact release candidate across Linux, macOS and Windows with Node 18, 22 and 24 (CI results). These checks are not a live provider trial or a measured planning-quality/token-savings comparison.

C2C v1.2.1 — Explain execution choices

Execution recommendations now explain why before showing a percentage.

• Connect the actual task and its acceptance checks to supported model capabilities or observed evidence.
• Explain why the choice fits better than its alternative for this scope, what the tradeoff is, and what would change the recommendation.
• Explain any supported saving: lower published rates or a measured usage difference. Price alone cannot establish fewer reasoning tokens, fewer retries, model size or equivalent results.

For a small feature, the explanation stays a few decision-focused sentences. Advice remains advisory; no new worker call, coding action, model switch or billing change is introduced. Runtime, schemas and research scope are unchanged.

All 192 documentation targets and affected metadata passed validation. Independent source review found no actionable defects. A fictional advice exercise differentiated routine UI work from unresolved admin-upload work, explained task fit before the price comparison, and rejected unsupported claims about model size/retries. It does not establish real execution quality or savings.

Update with git pull --ff-only, then node scripts/install.mjs --update, and start a new chat. Use $C2C in Codex or /C2C in Claude Code.

Recommendation policy · Uninstall steps.

C2C v1.2.0 — Execution model recommendations

C2C now recommends which model should execute the plan's first work item, with a justified alternative, dated evidence and a clear reason to reconsider the choice as the work changes.

What the plan includes

• A task-specific execution model and supported effort, distinct from the models that planned or reviewed the task.
• Official model/configuration research and relevant firsthand community evidence, with freshness, eligibility and quality limits visible.
• Separate labels for task-token measurements, equal-token API price estimates and subscription allowance. Unsupported savings stay unknown; price differences never become token-saving claims.
• Advice placed beside Start here before plan verification and reused in the optional implementation handoff.

Execution recommendations are advisory. They do not switch your model, launch coding, change settings or enable extra paid usage. Existing planning stages, provider routing, security controls and budgets remain intact. There is no permanent model ranking or guaranteed performance/usage improvement.

Validation

All 192 documentation targets and affected skill metadata passed. Independent source review resolved an instruction conflict between keeping full research local and supplying concise execution evidence for verification. Reviewers can evaluate supplied rates and assumptions; source links do not grant them browsing access.

Two independent offline instruction cases used fictional data. One selected a task-fit execution option and calculated the weighted API estimate while keeping unsupported token/allowance savings unknown. The other retained an unresolved authorization requirement, stale-research and identity gaps, and the included-only spending boundary. These are instruction checks, not real provider calls or a measured execution-quality comparison.

All 15 cross-platform CI jobs passed, running the existing suite on Windows, Linux and macOS.

Use it

Ask $C2C in Codex or /C2C in Claude Code to plan the task. For more detail, add:

Recommend an execution model for each milestone within my included allowance. Show sourced savings where supported and mark unknowns.

Update from your clone:

git pull --ff-only
node scripts/install.mjs --update

Start a new chat. Execution recommendation policy · Uninstall steps.

C2C v1.1.1 — Focused single-provider planning

C2C's single-provider workflow now gives planning and implementation review complementary questions within the existing stages. The goal is fewer material mistakes and clearer next steps for the resources used.

Changes

• The planning author connects user outcomes, constraints and decisive assumptions to a minimal viable design.
• In independent plan mode, the other model develops its own implementation proposal from interfaces, state transitions and failure paths.
• Critiques test consequential decisions and remedies against evidence, prioritize blockers and avoid unnecessary scope. Verification can challenge accepted fixes as well as rejected concerns.
• A solo fallback traces material decisions through evidence or assumptions, a failure case or alternative, a response and a proposed check. It stays honestly labeled as one-model self-critique.

No extra worker stage or call is required. Cross-provider prompts and workflow remain unchanged, as do identity, security, usage and no-paid-recovery boundaries. Reviews focus on corrections and unresolved risks instead of repeating the plan. One provider does not acquire another provider's capabilities; distinct models may share blind spots.

Validation and evidence limits

The full local suite passed 273 tests with seven platform-specific skips. Focused tests cover eight background-author and four host-authored route combinations, including independent inputs, real role attribution and exact call counts. Independent source review found no actionable defects, and 183 documentation targets passed.

Native-agent acceptance checks against captured fictional prompts identified an account-authorization defect, rejected an unsupported login-only resolution, and accepted a sound scoped plan without invented findings. These are instruction checks, not a live CLI exchange or a comparison with ordinary planning.

All 15 native CI jobs passed across Windows, Linux and macOS. Accuracy, speed and token-saving improvements remain unproven until a matched whole-task comparison supports them. More discussion is not the success metric.

Update

git pull --ff-only
node scripts/install.mjs --update

Start a new chat, then use $C2C in Codex or /C2C in Claude Code. Changelog.

C2C v1.1.0 — Start from either app

C2C can start equally from Codex or Claude Code. The initiating chat coordinates; when the other provider is unavailable, the skill selects a planner and a distinct coding-focused critic within the available provider, provided both are eligible within your limits.

What changed

• Symmetric README examples and setup routes, with Planner / Reviewer workflow diagrams.
• Automatic same-provider selection before worker calls, using real separate contributions. Explicit participant/model choices and waiting instructions take precedence.
• Evidence-based challenges, coordinator responses and reviewer verification, without forced agreement or invented disagreements.
• One bounded linked route after an opposite-provider block, preserving prior work and remaining limits. The coordinator accounts for aggregate usage; the runner enforces each run's supplied caps.
• Honest solo self-critique when no eligible distinct pair can run, retaining security, proposed tests, unresolved risks and a clear Start here.

No forced installation or login, paid-credit recovery, billing/account change, sealed participant replacement or fresh budget reset. Different models can still share blind spots; requested identities and reported identities remain distinct.

Update and use

git pull --ff-only
node scripts/install.mjs --update

Start a new chat in either app:

Use $C2C to plan this feature. Check the project first, include security
and tests, and show me the plan before coding.

In Claude Code, start the same request with /C2C.

Validation

Nine independent offline instruction scenarios checked both entry points, available-provider selection and the limits that must prevent calls. Existing runtime tests cover both directions, distinct author/critic workflows, identity checks and preserved failures. Independent review corrected reply attribution before release. Documentation checks cover 177 targets.

Native CI covers Windows, Linux and macOS. These checks do not establish a new live model exchange, model independence or measured token savings. Setup · Changelog.

C2C v1.0.1 — Compatible workers and focused discovery

C2C now finds a compatible installed worker when an earlier CLI on PATH lacks required capabilities, and guides project assessment toward relevant current evidence.

What changed

• Checks distinct installed CLI candidates in order. Explicit binary overrides remain authoritative; execution, login and quota failures do not trigger another selection.
• Shows checked executable paths, available versions and rejection reasons in local setup diagnostics.
• Uses current scoped records, source-drift checks, bounded reading and selected JSON fields to reduce unnecessary discovery output. Security evidence and review depth remain required.

Required worker restrictions, no-paid-limit-recovery behavior and provisional planning remain intact. Discovery makes no model call, installs nothing and changes no global configuration.

Update

git pull --ff-only
node scripts/install.mjs --update

Start a new chat afterward. Use $C2C in Codex or /C2C in Claude Code. Setup help.

Validation

The local suite passed 273 tests with seven expected platform skips, including ten new adapter regressions. Independent source review found no actionable issue. The final isolated discovery walkthrough preserved important source, security and test limitations without dumping the large irrelevant record field. Native CI covers Windows, Linux and macOS.

Local CLI readiness is not proof of fresh authentication, model access or included usage. This patch does not claim measured token savings or a new live peer exchange. Full changelog.

C2C v1.0.0 — Stable plan delivery

C2C 1.0 helps Codex and Claude Code turn features and project briefs into practical plans with independent proposals, critique, security considerations, proposed tests and a clear execution entry point. You can also explicitly choose different models within one provider.

Stable plan delivery

Review progress and completion now live in generated DISCUSSION.md and RESULT.md. Announcing success leaves the verified plan unchanged, avoiding a repeat review caused only by rewriting its status paragraph.

All substantive revision checks remain intact. Changes to reviewed plan or security content, prior decisions or supplied evidence still need the bounded final check or an explicitly provisional result. Headings and status text have no integrity exemption.

Included workflows

• Project assessment, proportional feature review and project roadmaps.
• Real worker reports, attributed discussion and recorded finding decisions.
• Fresh model research for eligible worker routes, within user limits.
• Security, proposed acceptance checks and a concrete Start here.
• Provisional planning when required tools or allowance are unavailable, without forced setup or paid limit recovery.
• Resumable runs and managed updates with backups.

Update

git pull --ff-only
node scripts/install.mjs --update

Start a new chat afterward. Use $C2C in Codex or /C2C in Claude Code. Installation and first use.

Validation

The local release suite passed 264 tests with six expected platform skips. Two focused regressions cover stable completion and strict revision checks. An independent offline walkthrough finished a verified synthetic plan without changing its reviewed content or adding a worker call. Cross-platform CI covers Linux, macOS and Windows.

These checks do not guarantee plan correctness, runtime model identity, account billing or comparative token savings. Full changelog.

C2C v0.7.0 — Codex + Claude planning with less repetition

C2C lets Codex and Claude Code assess a project, review each other’s plans, and record decisions before implementation. Invoke $C2C in Codex or /C2C in Claude Code.

What is new in 0.7.0

• Shorter entry instructions, focused review prose, and reuse of returned reports.
• Less bookkeeping in peer prompts, with complete selected evidence and sealed local records retained.
• Optional compact JSON output preserves every field. Models, security requirements, and review stages stay unchanged.

Planning features included

Project and production-status assessment; clarification of goals and scope; small-feature reviews and project roadmaps; independent proposals and mutual review; visible stage updates and DISCUSSION.md; security review and proposed acceptance tests; advisory task sizing and model recommendations; saved Markdown handoffs and recovery state.

In this historical release, focused review uses 2 successful peer calls; independent planning uses 3. Failed attempts also consume the attempt allowance. See the current README for later features and routes.

Validation and limits

135 automated tests passed locally. All six CI jobs passed for this release commit on Windows and Ubuntu with Node 18, 22, and 24. These test the controller, not a user’s future implementation. They do not establish lower total token usage, faster responses or better planning quality.

Install and learn

• Quick start and invocation examples
• 12-page historical setup and usage guide (PDF)
• FAQ · Update instructions

This release uses the existing v0.7.0 tag at a67657d4787c08bf89c22eb2ff6872797bcf061e. The linked documentation is maintained on main.