qa-cli-mcp-api

Contributors

GitHub-linked commit authors for this SKILL.md at the saved revision. Co-authors and history before file renames are not included.

File history ↗

Select and run explicitly requested, risk-based QA for Anarlog's CLI, webhooks, stdio MCP, hosted Cloud API, and remote MCP. Test only affected lanes unless comprehensive coverage is requested.

.agents/skills/qa-cli-mcp-api/SKILL.md

Download bundle ↓
main · 143e3d81 bundle fileScanned 2026-09-15

SKILL.md

2,944 tokens · o200k_base · 13,413 bytes

QA: CLI, MCP, and API

Default to the smallest set of programmatic-interface lanes that can prove the change. Automated tests are necessary but do not replace a live smoke test when the changed boundary is only exercised by a real client or deployment.

This QA workflow is independent from releasing. A release request alone does not invoke it, and its results do not approve or block a release.

Choose the scope first

Inspect the exact branch, commit, or diff and map changed code to its direct consumers before creating fixtures or credentials. In a GitButler workspace, use but status and but show <commit-or-branch>; do not use the synthetic workspace HEAD as the candidate or combine unrelated applied branches.

Find the original reproduction in the current or past Codex task, linked issue, PR, support report, or regression test. State the selected lanes and the reason for each before testing.

Targeted regression mode (default)

Select lanes by behavior, not by the existence of this checklist:

  • CLI parsing, output, or local DB access → CLI plus the closest contract tests.
  • Webhook endpoints, signing, or delivery retries → the webhook cases only.
  • Shared agent-access DTOs, exports, filtering, or pagination → every direct consumer, including hosted REST or remote MCP when affected, plus cross-surface parity only for the changed fields.
  • Hosted auth, snapshots, entitlements, isolation, purge, or Supabase policy → the affected hosted REST lifecycle and negative cases.
  • Local or remote MCP protocol/tool changes → that MCP lane and its direct transport/contract dependency.
  • Shared hosted REST/MCP behavior → both hosted consumers, but not unrelated local surfaces.

Run the closest affected automated tests, the original live reproduction, and only credible boundary cases. A Rust or shared-crate change does not trigger all lanes unless every lane consumes the changed behavior. Create only the minimum non-sensitive fixture required for the selected checks. If a required deployment, account, fixture, or client is unavailable, mark that check BLOCKED; do not substitute unrelated lanes. Stop when the mapped risks are covered.

Comprehensive Interface QA

Run every fixture, baseline, live lane, lifecycle case, privacy check, and cross-surface comparison below only when the user explicitly requests full or comprehensive programmatic-interface QA.

Safety and evidence

  • Use a dedicated QA account and non-sensitive fixture meetings.
  • Use a Pro or trialing account for the hosted lane and a separate free or expired test account for entitlement checks.
  • Never put API keys, JWTs, webhook secrets, database credentials, or personal meeting content in commands, screenshots, reports, or repository files. Load secrets into environment variables without echoing them.
  • Store transient response bodies in a directory created with mktemp -d. Remove that directory and revoke all generated keys after the run.
  • Record the exact candidate commit, app version, API deployment, Supabase migration version, operating system, and client versions.
  • In a GitButler workspace, identify the selected branch tip with but status and inspect it with but show <branch>; do not use the synthetic workspace HEAD as the candidate identity.
  • Mark a lane BLOCKED, not PASS, when its required deployment, account, fixture, or client is unavailable.

Comprehensive QA Fixture

Create two completed QA meetings through the desktop app:

  1. A standalone meeting whose title contains a unique run marker such as agent-access-2026-07-28T120000Z.
  2. Two meetings in the same recurring series.

The fixture must include:

  • a note and at least one generated summary;
  • two participants and one action item;
  • enough transcript words to require two pages when requested with a small limit;
  • punctuation and non-ASCII text;
  • a later edit to the title, note, or summary;
  • one meeting that will be deleted during lifecycle testing.

Record the meeting IDs and expected visible values. Do not seed SQLite or Postgres directly for the live happy-path tests.

Comprehensive Automated Baseline

Run from the repository root:

cargo test -p anarlog-cli
cargo test -p tauri-plugin-local-api
cargo test -p api-cloud
cargo test -p api openapi::tests
cargo check -p api-client
supabase test db
pnpm -F desktop exec vitest run \
  src/cloud-api/client.test.ts \
  src/settings/developers/index.test.tsx
pnpm -F desktop typecheck
pnpm exec dprint check

Require the CLI and MCP contract snapshots, OpenAPI composition tests, authentication tests, desktop account-switch tests, and database policy tests to pass without updating snapshots or generated clients during the run.

If the change touches a shared DTO or generated client, regenerate it using the repository command that owns the artifact, then require a clean second generation. A generated diff after the second run is a failure.

Local CLI

Build the candidate CLI with cargo build -p anarlog-cli, then use the built binary for every step.

  1. Run anarlog --json doctor.
    • PASS when it exits 0, reports schema_version: "1", uses command doctor, reports ready: true, and resolves the intended QA database.
  2. Run meetings list with JSON output.
    • Verify default ordering, the unique title query, exact series filtering, limit/offset pagination, and an empty result.
  3. For the recorded meeting ID, run:
    • meetings get ID
    • meetings note ID --kind note
    • meetings note ID --kind summary
    • meetings note ID --kind all
    • meetings transcript ID with at least two pages
    • meetings history ID with at least two pages
    • meetings export ID --format markdown
    • meetings export ID --format json
  4. Require every JSON response to have the correct schema_version, command, data, and pagination fields. Following next_offset must produce no duplicates or gaps.
  5. Export to a temporary file. A second export without --force must fail without changing the file; --force must replace it.
  6. A missing meeting ID and an unreadable or incompatible database must return machine-readable errors, a nonzero exit, and no panic or partial export.

Webhooks

Launch the exact desktop candidate and add a temporary receiver under Settings → Developers → Webhooks.

  1. Adding an endpoint must return a whsec_ secret exactly once, and a non-HTTP URL must be rejected.
  2. Trigger a test delivery, a meeting completion, and a note enhancement.
    • Verify the event name, delivery ID, timestamp, body, and HMAC-SHA256 signature over the exact raw body.
    • Verify retry behavior with one deliberate transient failure.
  3. Confirm the last-delivery status shown in settings matches the receiver.
  4. Delete the webhook and prove that no later delivery arrives.
  5. Quit the desktop app, complete no further work, and confirm no deliveries are queued or replayed on the next launch.

Local stdio MCP

Start anarlog mcp through a real MCP client over stdio. Do not validate this lane by invoking server handlers directly.

  1. Connect with the current MCP lifecycle and run tools/list. Repeat the handshake with a legacy 2025-11-25 client.
  2. Require exactly these tools:
    • list_meetings
    • get_meeting
    • get_meeting_transcript
    • get_recurring_meeting_history
    • export_meeting
    • propose_summary_edit
    • propose_memo_edit
    • list_proposals
    • get_proposal
    • decline_proposal
  3. Call every tool against the fixture. Verify query and series filters, two-page transcript/history traversal, missing IDs, and invalid arguments.
  4. Run resources/list, resources/templates/list, and resources/read for a meeting, transcript page, and recurring series.
  5. Require read-only, non-destructive, and idempotent tool annotations.
  6. Compare the returned values with the CLI lane.
  7. Capture stdout and stderr separately. Stdout must contain only MCP protocol frames; diagnostics belong on stderr. Client shutdown must terminate the server without leaving a process behind.

Hosted Cloud API

Use the deployed candidate API and Supabase migration with the Pro QA account. Begin with Cloud API & Connectors disabled.

  1. Before opt-in, confirm there are no server-readable snapshot rows for the account and a previously valid key returns 403 cloud_api_not_enabled.
  2. Enable the feature in the desktop UI and wait for backfill to finish.
  3. Create a short-lived cloud key and exercise the same read endpoints, filters, pagination, error cases, and exports as the CLI lane.
  4. Require:
    • no key, malformed key, and revoked key → 401;
    • free or expired account → 403 subscription_required;
    • disabled account → 403 cloud_api_not_enabled;
    • invalid input → 400 invalid_request;
    • missing meeting → 404 not_found;
    • request burst above the documented quota → 429 with retry-after.
  5. Edit the fixture locally and confirm the hosted result changes. Delete the lifecycle fixture and confirm the hosted endpoint returns 404.
  6. Sign the desktop into another account before a queued upload or retry can complete. No snapshot from the first account may appear in the second.
  7. Disable the feature.
    • PASS when all server-readable snapshots are purged, the cloud key returns cloud_api_not_enabled, local data remains, and normal encrypted sync data is unchanged.
  8. Re-enable and confirm a fresh backfill restores only currently existing meetings. Revoke the QA key when finished.

Remote MCP

Connect a real Streamable HTTP MCP client to the deployed /mcp endpoint.

  1. Use server/discover with MCP 2026-07-28, list tools without creating a session, and require exactly list_meetings, get_meeting, get_meeting_transcript, get_recurring_meeting_history, and export_meeting. Verify the required protocol metadata and standard HTTP headers on every modern request. Repeat discovery with a legacy 2025-11-25 stateless client to prove compatibility.
  2. Call every tool and traverse at least two transcript/history pages.
  3. Compare its structured results with the hosted REST responses.
  4. Repeat initialization or a tool call with no credential, a malformed key, a revoked key, an expired account, and after opt-out. Unauthenticated MCP requests must return WWW-Authenticate pointing at https://api.anarlog.so/.well-known/oauth-protected-resource/mcp.
  5. When OAuth is affected, complete MCP OAuth 2.1 discovery and consent from at least one documented host (Claude Code, Cursor, ChatGPT/Codex, or Copilot). Confirm the consent screen is Anarlog's /oauth/consent route, the issued token is bound to https://api.anarlog.so/mcp, and a tool call then reads the marked meeting. Repeat with a static anl_ key for hosts that cannot complete OAuth.
  6. Connect at least one supported agent client using the documented setup and ask it to identify the marked meeting, summarize it, and cite a transcript detail. Verify the answer against the fixture.
  7. Close the client and confirm the server releases the session cleanly.

Cross-surface parity

For the marked meeting, compare CLI, local MCP, hosted REST, and remote MCP:

FieldRequired parity
MeetingID, title, kind, status, timestamps, timezone, language, series
Documentscanonical note, summary titles and markdown
Peopleparticipants and organizations
Actionstext, assignee, completion state
Transcripttext, word order, timestamps, speakers, page boundaries
HistoryIDs, newest-first ordering, pagination
Errorsstable code, appropriate protocol status, no secret leakage

Local and hosted values must match after backfill settles. Hosted payloads must not contain local paths, audio paths, file paths, control characters, secrets, or fields outside the disclosed server-readable copy.

Reporting

For targeted mode, report the candidate branch/commit, base, selected lane and risk, PASS/FAIL/BLOCKED, and a one-line evidence note. List unrelated lanes once as NOT APPLICABLE, and say explicitly that the result is not comprehensive interface QA.

For comprehensive mode, produce one table with rows for:

  • automated baseline;
  • CLI;
  • webhooks;
  • local stdio MCP;
  • hosted REST;
  • remote MCP;
  • lifecycle and account isolation;
  • privacy purge;
  • cross-surface parity.

Use PASS, FAIL, or BLOCKED with a one-line evidence note. Include the candidate and deployment identifiers, fixture marker, clients tested, and redacted response artifact locations. List every skipped negative case.

Any required FAIL or BLOCKED result means comprehensive QA did not pass. Report that outcome without inferring release approval or blocking.

Discovery context

Discovered by repository scan. No exact path reference found in the snapshot’s root AGENTS.md.