feat: show local plan usage for Claude, Cursor, Codex, and more signed-in CLIs

[marcelocecin]

What does this PR do?

Shows plan usage for signed-in local CLIs/editors inside Multica. The daemon on the host reads the session already present on that machine and uploads only derived fields: provider, window, percent used, reset time, plan name, and collected_at. The server does not call vendor APIs and does not store bearer tokens, cookies, or auth files.

A missing or unusable session (not signed in, API-key-only auth, CLI missing, expired local token) is an empty snapshot. It does not fail agent tasks. An expired token is not refreshed and does not replace the last good snapshot.

Related Issue

Closes #

Type of Change

• New feature (non-breaking change that adds functionality)

Changes Made

• Daemon collectors (host where the CLI/editor is signed in):
  • Claude Code: parse stdout of claude --print --no-session-persistence --strict-mcp-config /usage. No keychain read and no login refresh.
  • Cursor: read the local editor session and call GET https://cursor.com/api/usage-summary. Upload autoPercentUsed / apiPercentUsed, billing-cycle end, and membership type. Cookie stays in process memory.
  • Codex: GET https://chatgpt.com/backend-api/wham/usage using ~/.codex/auth.json. Map primary_window and secondary_window. API-key-only auth reports empty and does not call the API.
  • GitHub Copilot: gh hosts file or gh auth token, then the Copilot quota API. GH_TOKEN / GITHUB_TOKEN are ignored.
  • Antigravity: ~/.gemini/oauth_creds.json only, then Cloud Code quota. No keychain prompt and no language-server scan.
  • Grok: trusted https://auth.x.ai entry in ~/.grok/auth.json. Expired tokens are not refreshed (unlike CodeNotch fix(daemon): support Windows by extracting platform-specific SysProcAttr #393 in-memory refresh).
  • Kimi: ~/.kimi-code/credentials/kimi-code.json (KIMI_CODE_HOME).
  • Kiro: kiro-cli chat --no-interactive /usage stdout. The sqlite token store is not opened.
  • OpenCode: per data directory, in order: auth.json key opencode-go, then SQLite opencode.db table credential (read-only), then the mirrored opencode OAuth entry in auth.json. Go key → https://opencode.ai/zen/go/v1/usage; OAuth → https://opencode.ai/inference/go/v1/usage with x-opencode-org-id when metadata.orgID is present. Expired credentials upload credential_expired and are not refreshed (needed for OpenCode 1.18+ / 2.x, which stopped writing auth.json on sign-in — CodeNotch v1.20.0 / fix(inbox): archive at issue level instead of event level #398).
• Poll about every five minutes, back off on HTTP 429, keep the last good snapshot.
• New runtime_provider_usage_snapshot rows, separate from task_usage. No foreign keys. 564 creates the table; concurrent indexes are 565 and 566. Upstream main occupies 548–563; open PR feat(agent): add Prime Agent as a native ACP provider #6641 uses 567.
• POST /api/daemon/runtimes/{runtimeId}/provider-usage and GET /api/runtimes/{runtimeId}/provider-usage, plus GET /api/runtimes/provider-usage for the list. The handler rejects token-like fields.
• Runtimes list: a Usage column, one cell per row.
• Runtime detail: Plan usage shows only that runtime’s provider.

Sync with main

Merged current multica-ai/multica main through 2ea01ae4e (changelog v0.6.1, 2026-10-01) into feat/provider-plan-usage as 17d4f775a. Catch-up merges; no content conflicts on the latest sync.

Provider-usage migrations stay 564 / 565 / 566. Main still ends at 563. Sibling PR #6641 owns 567.

Contracts kept: derived upload fields only; token-like fields rejected; OpenCode order above; no token refresh.

CodeNotch follow-ups

• v1.19.0: compared v1.18.0...v1.19.0. No collector change (notch UI + Settings hang unrelated to our reads).
• v1.20.0: OpenCode credential path updated as above. Grok in-memory refresh and Abacus.AI left out of scope.

How to Test

1. cd server && go test -count=1 -timeout 180s ./internal/daemon/providerusage/
2. cd server && go test -count=1 -timeout 120s ./internal/handler/ -run 'TestRejectCredentialFields|TestNormalizeProviderUsageReport|TestProviderUsageErrorDoesNotEchoSecrets|TestWorkspaceDeletionManifestCoversPublicSchema|TestProviderUsageBatchGroupsByRuntime'
3. cd server && go test -count=1 -timeout 60s ./cmd/migrate/ -run TestEveryConcurrentUpBuildHasCleanup
4. pnpm --filter @multica/core exec vitest run api/schemas.test.ts
5. pnpm --filter @multica/views exec vitest run runtimes/components/provider-usage-block.test.tsx runtimes/components/runtime-plan-usage-cell.test.tsx locales/parity.test.ts
6. On a machine with any of the supported CLIs already signed in, run this daemon and open the runtimes list, then each runtime page. Confirm the list cell matches that provider only, and the detail page does not show other providers.

Default tests use fixture stdout and HTTP responses. They do not run a real CLI or call vendor APIs.

Checklist

• I have included a thinking path that traces from project context to this change
• I have run tests locally and they pass
• I have added or updated tests where applicable
• If this change affects the UI, I have included before/after screenshots
• I have updated relevant documentation to reflect my changes
• If I added a new runtime / coding tool / UI tab, I synced the change to landing copy (apps/web/features/landing/i18n/) and relevant docs (apps/docs/content/docs/)
• If this PR touches Chinese product copy, I checked it against apps/docs/content/docs/developers/conventions.zh.mdx (terminology, mixed-rule for task / issue / skill)
• I have considered and documented any risks above
• I will address all reviewer comments before requesting merge

AI Disclosure

AI tool used: Cursor

Prompt / approach:
Collect plan limits only on the local daemon from the already signed-in CLI or editor session. Extend collectors for additional Multica-relevant providers that expose a local session (inspired by CodeNotch’s provider surface). Upload derived percents, reset time, and plan name. Keep vendor tokens off the server. Show one usage cell per runtime in the list, and only that runtime on its detail page.

Screenshots (optional)

Risks

• Several collectors use undocumented HTTP endpoints and local session files already on disk. Credentials stay in process memory and are not written to the snapshot or logs. Those endpoints and file formats can change without notice. Parse failures keep the last good snapshot and do not fail tasks.
• Claude and Kiro are parsed from CLI /usage text, which can change format.
• A provider that cannot be read returns an empty snapshot (cli_unavailable, session_unavailable, credential_expired, or unsupported).
• Providers without a safe local plan-percent path (e.g. Qianwen browser cookie, Gemini CLI token logs, Amp/Devin/Ollama and similar) are intentionally not collected here.

[vercel]

@marcelocecin is attempting to deploy a commit to the IndexLabs Team on Vercel.

A member of the Team first needs to authorize it.