Rebrand and harden Independent QA as Crosscheck

[jtcchan]

Released September 7, 2026

Crosscheck 0.3.0-rc.1 is published and PR #2 is merged. Independent QA passed all required criteria under the explicit Codex-first scope clarification. Authenticated Claude execution remains unverified and is not a required release gate. All source-issue acceptance and proof items below are complete for this release candidate; automatic fleet rollout is separate. The original human-intent quote is retained.

Objective

Rebrand and harden the open-source Independent QA Agent as Crosscheck: the fresh-context, read-only verification gate that must return PASS before completed work can move to merge or closure.

Human intent

Audit it for improvements now that we’re on Astra; make it more robust depending on the type of task or decision being QA’d; improve completion evidence such as screenshots and videos; and rebrand it to Final Boss. It should run after the assignment and/or No Mistakes, use an ephemeral independent agent with fresh context, remain read-only, publish a plain-language PASS/FAIL report to the worker thread and GitHub, and remain an open-source project maintained under 2x.

Current-state audit

There are two divergent public implementations:

• 2xgrowthagency/independent-qa-agent is the intended organizational home but contains the older v0.1 implementation.
• jtcchan/workboard-qa-agent contains the stronger v0.2.1 implementation: deliverable/decision/process modes, target fingerprinting, execution-separation proof, stale-target invalidation, structured findings, plain-language summaries, and idempotent GitHub publication.

The organizational repo should become canonical. The personal repo should be treated as migration input and later archived or redirected only after compatibility and installation paths are proven.

Scope

In scope

1. Brand and ownership

   • Rename the product, plugin, marketplace entry, documentation, prompts, and report headings to Crosscheck.
   • Keep 2x Growth Agency as the canonical maintainer and repository owner.
   • Preserve a documented compatibility path for existing independent-verification, qa-agent, and workboard-qa-agent users during migration.

2. Astra-era verifier contract

   • Define a fresh, ephemeral verifier session with no inherited builder conversation or target write ownership.
   • Default 2x routing to Astra, with reasoning effort selected by risk/ambiguity rather than one global setting.
   • Keep the open-source instructions harness-agnostic; Astra-specific routing belongs in an explicit adapter/example.
   • Treat builder summaries, screenshots, logs, and claimed test results as leads, never as sufficient proof.

3. Task- and decision-aware QA profiles

   • Support at least: code/PR, UI/website, interactive workflow, document/content, data/report, operational/configuration, decision/recommendation, and QA-process evaluation.
   • For each profile, define required inputs, default checks, mandatory evidence, common failure modes, stop conditions, and what a PASS permits.
   • Allow more than one evidence lane when the acceptance contract crosses boundaries, while keeping one primary verdict target.
   • Separate required criteria from advisory observations before checks begin.

4. Evidence bundle and completion proof

   • Define one versioned evidence manifest that binds the exact target, base, environment, verifier session, capture time, tool/version, criteria, check results, and artifact hashes.
   • Require current screenshots for visual claims, including named viewport and exact URL/revision.
   • Require a video or trace for material multi-step interactions where still images cannot prove the sequence; do not require video for static or non-visual work.
   • Capture console/network/HTTP evidence when relevant, rendered-document proof for documents, reconciliations for data, and persisted-state readback for operations.
   • Mark every artifact local-only, safe-to-share, or redacted; never publish unreviewed screenshots, secrets, client data, cookies, or absolute local paths.
   • Make stale or changed targets invalidate the verdict automatically.

5. Gate semantics and rework loop

   • PASS is the only verdict that clears Crosscheck for the exact bound target.
   • FAIL means a required criterion is violated and routes to bounded rework.
   • BLOCKED means the verdict cannot be established safely; it is a non-pass and names the exact unblock owner/action.
   • Any later target change, push, rebase, deployment, evidence refresh, or material configuration change requires a fresh Crosscheck run.
   • Crosscheck never edits product work, fixes findings, merges, closes, deploys, changes account state, or expands authority.

6. Human-readable reporting and publication

   • Write a durable technical report and a small machine-readable gate receipt.
   • Publish an idempotent comment to every explicitly associated PR and issue, and an informational result to the original worker thread.
   • Human-facing output must explain: what the ticket was for, what was checked in ordinary language, the verdict and practical meaning, any blocker/failure, what happens next, and whether a person needs to act.
   • Follow the active communication guidance (for 2x, TASTE.md) at response time rather than copying private/local style rules into the public skill.
   • Keep publication status separate from the QA verdict so a comment failure cannot rewrite a conclusive product result.

7. Open-source quality and migration

   • Migrate the stronger public v0.2.1 behavior into the 2x repo without losing its safety boundaries.
   • Add deterministic schema/contract validation for the evidence manifest and gate receipt.
   • Add behavior-oriented fixtures for correct work, required defects, missing evidence, stale targets, publication failure, interactive-media requirements, and redaction/privacy failures.
   • Document Codex and Claude Code installation, generic agent usage, No Mistakes handoff, and Workboard integration without making Workboard mandatory.
   • Publish migration notes and a versioned release after fresh independent QA.

Out of scope

• Automatic merge, issue closure, deployment, production mutation, or account changes.
• Replacing code review, CI, No Mistakes, or human product judgment.
• Uploading local/private evidence without an explicit safe-sharing decision.
• Treating model identity alone as independence.
• Archiving or deleting the personal repository before the new canonical install and migration path are verified.

Suggested starting points

• Organizational baseline: 2xgrowthagency/independent-qa-agent
• Migration source: jtcchan/workboard-qa-agent at v0.2.1
• Primary skill: plugins/**/skills/independent-verification/SKILL.md
• Existing mode references: decision-analysis.md, process-evaluation.md
• Existing publication contract: result-publication.md
• Add task profiles, evidence schema/examples, behavior corpus, and migration documentation beside the skill.

Acceptance criteria

• The canonical public project and visible brand are Crosscheck, maintained under 2xgrowthagency.
• Existing users have a tested, documented compatibility/migration path.
• A fresh-context, read-only verifier contract is explicit and testable.
• Astra routing is defined for 2x without making the public skill Astra-only.
• Every supported QA profile has concrete checks, evidence expectations, and verdict meaning.
• Visual work requires current screenshot proof; multi-step interactions require video/trace proof when sequence matters.
• A versioned evidence manifest and machine-readable gate receipt validate deterministically.
• Only PASS clears the gate; FAIL, BLOCKED, or stale evidence cannot be treated as approval.
• PR, issue, and worker-thread summaries are plain-language, useful without prior context, and idempotently published.
• Publication failure is tracked separately from the verdict.
• Crosscheck remains read-only except for its own reports and explicitly authorized result comments.
• Behavior fixtures cover pass, fail, blocked, stale target, evidence-media, privacy/redaction, and publication scenarios.
• Repository validation, focused tests, and plugin/skill validation pass on the final diff.
• A separate fresh Crosscheck process evaluation passes the exact release candidate before release.

Required proof

• Audit/migration matrix showing which v0.2.1 behaviors were preserved, changed, or intentionally retired.
• Deterministic tests and schema validation output.
• Example evidence bundle with safe screenshots and a short interaction recording or trace.
• Example human-facing PASS, FAIL, and BLOCKED comments for a PR/issue and worker thread.
• Exact-head plugin and skill validation.
• Fresh independent process-verification report against the release candidate.
• Install smoke from the 2x repository plus compatibility smoke for the previous public name.

Stop and ask if

• A compatibility choice would silently break existing installations.
• The repository rename/redirect cannot preserve GitHub clone and marketplace behavior.
• Required media evidence would expose private/client data or credentials.
• A requested integration would give Crosscheck mutation authority beyond reports/comments.
• Release, archive, deletion, or other irreversible publication is required.

[jtcchan]

Execution and proof are tracked in https://github.com/2xgrowthagency/donna-workboard/issues/175. That record owns the 2x routing prerequisite; this issue remains the canonical product scope.

[jtcchan]

Independent source review: changes required

What this work is for: make the independent QA agent a reliable final check that proves completed work, explains the result clearly, and prevents an outdated or incomplete result from clearing the gate.

What was checked: a fresh read-only reviewer examined every changed file and the screenshot fixture at this PR revision. All 20 tests, metadata/schema validation, and four isolated compatibility-installation paths passed. These checks found three problems not covered by the existing tests; the coordinating reviewer reproduced all three independently.

What needs fixing:

1. A target can change while the publisher searches existing comments, yet the publisher still posts and returns the old PASS. Add freshness checks around comment delivery and tests for a change during lookup or delivery.
2. Replacing the human-readable QA report does not invalidate its receipt. Bind and verify the exact report bytes, including after publication-state updates.
3. Verifying a valid BLOCKED receipt returns the FAIL exit code. Preserve the distinct missing-proof versus observed-defect outcomes.

What happens next: the original implementation worker is fixing these three bounded findings on this same PR. The revised work needs tests and another source review, canonical skill integration, then fresh independent process QA before release. No action is needed from the owner for these fixes.

Review evidence

• Reviewed head: 98a53424fe944db0dfdc2fe8f96f646bc031afdf
• Base: 5587ecd5028a680b506068c1c374419f74a9795a
• Severity: three P2 findings; no P0/P1 identified in the reviewed source.
• A single-destination target-change reproduction returned PASS/published after its revision changed.
• A replaced-report reproduction still returned successful receipt verification.
• A valid BLOCKED receipt evaluated with exit 2 but verified with exit 1.
• The synthetic 960×640 screenshot was visually inspected; its matching ordered trace was read. This is fixture inspection, not a fresh browser capture or process-QA pass.
• The source checkout remained unchanged during review. Full review and reproduction evidence is retained in task artifacts.

This is a binary-capable independent source review, not an Autoreview PASS or the final process-QA verdict. Autoreview stopped before invoking a reviewer because of the PNG fixture. The PR remains draft; no merge or release has occurred.

[jtcchan]

Crosscheck: BLOCKED

What this was for: Upgrade the independent QA agent so completed work must be backed by current evidence before it can move forward, with clearer checks for different kinds of work and a tested migration to Crosscheck.

What I checked: A fresh independent reviewer checked the completed candidate, including 40 repository tests, 51 additional failure cases, six installation paths, the packaged instructions, evidence and privacy rules, stale-result rejection, and reporting retries. Codex loaded and used the installed instructions successfully. Nineteen of twenty required criteria passed; no product defect was found in the checks performed.

Why this is blocked: Claude discovers the instructions, but a real authenticated Claude run is still missing. Installation success alone does not prove that compatibility check passed. This result does not clear the candidate for merge or release.

What happens next: The release owner completes Claude sign-in, then runs the isolated canonical and legacy instruction checks and obtains fresh final verification. No source repair is requested, and this informational report does not start another worker.

Evidence: Full report and checked evidence are retained in the task artifacts. Repository checks are available in the exact candidate's CI run. Private logs and local evidence have not been uploaded.

[jtcchan]

Release scope clarification

This is a Codex-first Crosscheck release. Authenticated Claude model execution is not required for this release and remains unverified; Claude installation/discovery and legacy compatibility are preserved. This explicitly supersedes the earlier Claude-login release blocker, not its historical evidence. Every other required correctness, independence, privacy, evidence, review and publication gate remains in force.

The intended operating order is implementation, No Mistakes where applicable (otherwise the repository's normal tests/review), fresh read-only Crosscheck, then owner-controlled merge or closure only after a current PASS. A changed target invalidates the earlier PASS. Releasing this package does not itself install a universal hook, enable branch protection, or prove fleet-wide automatic enforcement.

The existing candidate is 0.3.0-rc.1. Release ownership has authorized review, merge and a version-matching GitHub prerelease after the fresh verdict and delivery checks. No stable-version or package-registry publication is implied. The old personal repository remains untouched.

[jtcchan]

Crosscheck: PASS

What this was for: turn Independent QA into Crosscheck, an independent final check that requires evidence before completed work can move forward.

What was checked: all 20 required checks passed. Fresh testing covered the shipped tests, the independent failure-case suite and new checks for misleading evidence, privacy, expired results and changed work. Installation, legacy compatibility, Codex instruction loading, reports and delivery behavior were also verified against the exact candidate.

Result: the Codex-first release candidate passes independent verification, with no required defects found. Authenticated Claude execution remains unverified and was explicitly excluded from required release acceptance by the owner before this run; it is not reported as passed. This supersedes the earlier blocked release assessment under the clarified scope, while preserving that report as history.

What happens next: the owner completes publication checks, merges the reviewed PR and publishes version 0.3.0-rc.1. No action is needed from a person. This is a package release, not proof that automatic Crosscheck enforcement is already enabled across every project. The original implementation worker should not start new work from this informational notice.

Evidence: exact candidate 76ce4f4c752c5c8d0c7edfe1861263e35714683a; passing CI; owner-approved scope. Full evidence is retained in the task artifacts. Only this reviewed summary is published; raw evidence stays private.

[jtcchan]

Crosscheck released

Crosscheck 0.3.0-rc.1 is published: release and installable files.

PR #2 is merged. Source review found no required fixes, all three candidate CI checks passed, fresh independent QA passed all 20 required criteria, and the PASS report was delivered and read back at the source issue, PR and original worker thread before merge.

The merge tree exactly matches the verified candidate. Main-branch CI also passed. The release tag identifies the exact tested candidate, and both published asset hashes match the reviewed wheel/plugin archive.

This is a Codex-first release candidate. Claude installation/discovery and legacy preservation were checked; authenticated Claude execution remains unverified and was explicitly excluded from required release acceptance by the owner. The older blocked result remains as historical evidence.

This completes the package/rebrand release scope. No further action is needed for this release. Automatic fleet-wide invocation or merge protection is not enabled merely by publishing this package; the intended default remains implementation → No Mistakes where applicable → fresh Crosscheck → owner-controlled merge or completed-issue closure.

Independent PASS. Raw evidence remains private. No production/client state, personal repository or live user plugin activation was changed.