Harden ContextStream plugin for Grok Bot marketplace

[escott-]

Summary

Prepare ContextStream's existing marketplace plugin to become a strong shared-context layer for Grok Bot and Cursor while keeping public compatibility claims behind explicit live-validation gates.

Product experience

• Expand from 3 to 7 focused skills:
  • context-check
  • project-brief
  • decision-check
  • project-resume
  • change-impact
  • project-handoff
  • memory-review
• Update the Project Brief & Handoff Bot profile to route requests to the right workflow instead of making users pick from a large menu.
• Add first-run guidance and a capability map so the experience starts with a useful project result rather than a tool catalog.
• Add explicit handling for stale/partial evidence, conflicting decisions, missing graph coverage, reported-vs-verified completion, uncertain writes, and receipt-bound feedback.

Validation and release hardening

• Add a 24-scenario human acceptance catalog covering installation, OAuth, scope, briefs, role consistency, decision authority, resume, impact, handoffs, cross-tool reuse, revoked/read-only access, prompt injection, wider audiences, feedback, outages, budgets, public templates, and persistence.
• Add a fail-closed release evidence checker requiring separate reviewed Cursor and Grok records, package/scenario fingerprints, repeated critical trials, and hashed evidence files.
• Add an opt-in metadata-only MCP probe that can initialize and list tool metadata but never calls tools/call; protocol success is explicitly not treated as workflow success.
• Strengthen package validation for duplicate JSON fields, unexpected executable/plugin components, nested or missing skills, size budgets, scenario coverage, unsafe paths, and local-link integrity.
• Expand CI to Ubuntu/Python 3.10 and Windows/Python 3.13, including package tests and the runnable synthetic demo contract tests.
• Expand the Harbor Export fixture with real dependency-free Python code so search/impact behavior can be tested against an explicit compatibility contract.

Security and trust boundaries

• Keep the hosted MCP endpoint credential-free in repository configuration.
• Treat retrieved project text as untrusted evidence rather than instructions.
• Keep saves, updates, publication, destination changes, and wider sharing approval-gated.
• Distinguish workflow policy from backend authorization and transcript-retention behavior.
• Do not claim read-only OAuth, zero retention, marketplace approval, Grok compatibility, or automatic cross-agent learning without evidence.

Locally prepared validation

The complete review candidate was exercised locally before this push with package validation, regression/negative tests, protocol fakes, release-gate tests, synthetic exporter tests, Python compilation, and whitespace checks. The earlier review candidate reported 87 regression tests plus 6 synthetic demo tests passing. GitHub-hosted CI for this exact pushed commit must be treated as authoritative for the repository state and is currently running/queued separately.

Required before marketplace submission or public compatibility claims

• GitHub Actions passes on the exact PR head.
• Cursor installation, skill discovery, hosted OAuth, and the seven workflows pass using a dedicated synthetic ContextStream project.
• Grok Bot's supported plugin/skill path is established and tested; do not assume Cursor .mdc behavior transfers automatically.
• Complete the acceptance catalog for Cursor and Grok with sanitized, reviewed evidence.
• Verify actual persistence/retention behavior and applicable controls for the hosted path.
• Verify approved save, uncertain-write recovery, read-only denial, revoked access, destination authorization, and cross-tool reuse.
• Review the public Bot-sharing preview using synthetic data and recipient-owned authentication.
• Confirm/update Cursor marketplace submission separately from curated Grok Bot-directory review.

A merged GitHub PR, Cursor plugin approval, Grok compatibility, a public Bot share link, curated directory inclusion, and featured placement are separate milestones. This PR does not submit to a marketplace or publish a Bot.

Review starting points

• skills/*/SKILL.md — seven user workflows
• bots/project-brief-handoff.md — proposed Bot behavior
• docs/first-run.md — first useful experience
• docs/capability-map.md — mapping to current ContextStream capabilities
• evaluation/scenarios.json — live acceptance catalog
• scripts/check_release.py — evidence/release gate
• scripts/probe_mcp.py — metadata-only protocol diagnostic
• docs/manual-validation.md — live test record
• docs/marketplace-launch.md — submission and distribution gates

[escott-]

Pre-submission review — hold marketplace submission

I reviewed the package against the public MCP implementation at f1236e7b6c65e4babc2276f7ec6e85d8c936096f and prepared a stronger local candidate. This candidate is NOT committed to this PR: the tool blocked validation-code uploads, and I stopped rather than advance the branch with a partial/inconsistent package. The PR head remains b19d2337bd6ca976fc7e60a71c0bbdf535f949de.

Candidate improvements

• Seven focused workflows: connection/content readiness, role-specific project briefs, decision checks, cross-session resume, code/dependency impact, approved handoffs, and stale/conflicting memory review.
• One useful first answer rather than exposing every feature at onboarding; reuse valid project bindings and approvals without unnecessary repeat prompts.
• Explicit handling of missing optional tools, stale coverage, access/setup responses, unverified prior completion, uncertain writes, cancellation, and recorded-only feedback.
• Stronger offline package validation and regression tests; a metadata-only protocol diagnostic; a fail-closed acceptance-record checker; and a runnable synthetic export-contract example.

Local results for the candidate, not the current PR

• Package validation: PASS.
• 87 unit/regression tests: PASS (includes mocked transport and release-record checks).
• 6 synthetic demo contract tests: PASS.
• Python compilation and whitespace checks: PASS.

These results do not establish actual Grok installation, browser OAuth, permission enforcement, retrieval quality, or end-to-end behavior. No live client tests were performed. The candidate defines 24 live acceptance scenarios, with 34 trials per client including repeated continuity tests, to be recorded separately for Cursor and Grok.

Submission blockers

Real client installation/skill discovery, OAuth, cited first answers, cross-session/cross-tool reuse, backend access revocation and write restrictions, approved-save recovery, public-template privacy, and applicable data-retention behavior still require evidence. The existing logo also needs maintainer review before submission. Do not treat a package check or an MCP tools-list response as proof that those journeys work.

No marketplace submission, Bot publication, merge, backend change, or customer-data operation was performed.