Repository navigation
Steer agents toward actionable platform feedback content - #789
Conversation
Co-authored-by: Kent C. Dodds <me+github@kentcdodds.com>
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (4)
💤 Files with no reviewable changes (1)
📝 WalkthroughWalkthroughPlatform feedback guidance now defines actionable issue content, expands submission metadata descriptions, updates approval and disclosure instructions, and removes the corresponding approval-flow bullet from the base MCP server instructions. ChangesPlatform feedback workflow
Estimated code review effort: 2 (Simple) | ~10 minutes Possibly related PRs
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
… description Co-authored-by: Kent C. Dodds <me+github@kentcdodds.com>
|
🔎 Preview deployed: https://kody-pr-789.kody-a99.workers.dev Worker: Mocks:
|
Why
The platform-feedback steering (the
platform_frictionguide,meta_platform_feedback_submitschema descriptions, and MCP server instructions) thoroughly covers consent, attribution, and privacy, but says nothing about what makes the feedback content itself useful to triage. Research on high-quality product feedback (Mozilla bug-writing guidelines, OpenTelemetry good-issues guide, Atlassian product-discovery guidance, NN/g task-context research) converges on: one issue per submission, a specific scannable summary, goal context, minimal repro steps, expected vs actual, verbatim errors, frequency, impact/workaround cost, and problem-first framing for suggestions. Agents are uniquely positioned to capture that firsthand context while it is still in the conversation — but nothing told them to.What changed
docs/guides/platform-friction.md: new## Write feedback admins can act onsection covering one issue per submission, summary quality (with a good-vs-vague example, noting admins triage from the summary-only list view), a details checklist (goal context, exact capability names, repro steps, expected vs actual, verbatim errors, frequency, impact/workaround), problem-first framing forsuggestionfeedback, fact-vs-suspicion labeling, and excerpt-over-transcript quoting.meta-platform-feedback-submit.ts: thecategory,summary, anddetailsfield descriptions now steer content quality (when to pick each category; specific scannable summary; per-issue details checklist), and the capability description points atcoding_guide_get({ guide: "platform_friction" })for the approval flow and content guidance. No handler or shape changes.kody-official-guide.ts: theplatform_frictioncatalog summary andguideenum description now also advertise the one-actionable-issue content bar.server-instructions.ts: the platform-feedback bullet is removed from the base server instructions. Server instructions hold a high bar (workflows and cross-tool conventions, perdocs/contributing/mcp-server-patterns.md); feedback discovery now lives on the capability surfaces themselves — themetadomain description, the submit capability's description/keywords, and thecoding_guide_getguide catalog — which agents hit throughsearchand tool listings.No runtime behavior, schema shape, validation, or storage changes — text-only steering.
System recap — extends existing primitives (medium risk, prompt-surface text only)
Mode: recap · Base:
main@84952ee3· Head:8dbc4962Classification: extends — the change edits agent-facing contract text (schema descriptions, guide body) on existing primitives and removes one bullet from the base server instructions. No runtime behavior, schema shape, or storage changes.
Primitives touched
platform-feedbackcategory/summary/detailsdescribe() textcapability-registrycoding_guide_getplatform_friction summary/describe textmcp-serverSystem map
An agent discovers the feedback flow from the submit capability and guide catalog descriptions (no longer from server instructions), loads the guide body from
docs/guides/platform-friction.mdon main, and then calls the unchanged submit handler.Legend: green = composes (wiring only) · amber = extended by this PR · red = new primitive · gray = context (unchanged, included only when an edge crosses it).
Testing
npm run validate(format:check, lint, typecheck, unit tests, Playwright E2E, MCP E2E) locally.Summary by CodeRabbit