Skip to content

docs: make Agent Quickstart CLI-first with browser login and OAuth MCP - #100

Merged
eddiearc merged 18 commits into
mainfrom
docs/agent-quickstart-cli-first-20260909
Sep 10, 2026
Merged

eddiearc merged 18 commits into
mainfrom
docs/agent-quickstart-cli-first-20260909

Conversation

@eddiearc

@eddiearc eddiearc commented Sep 9, 2026

Copy link
Copy Markdown
Contributor

The Agent Quickstart is now a shorter bootstrap: install/read one official Skill, use CLI command help for login, and use MCP/API fallbacks when appropriate. It no longer embeds the headless-login manual or numbered sections that suggest installing both CLI and MCP. Read the Agent-facing English Markdown.

Against the earlier12df325 candidate, English whole-page bytes fall14,525→8,122 (44.1%); bootstrap before direct HTTP falls10,528→4,125 (60.8%). Whitespace word counts fall2,132→1,116 and1,575→559. English/Chinese keep the entire direct-HTTP tail byte-for-byte, including API-key fallback and the exact language-matched paid-api-approval-first XML.

Responsibilities:

  • Quickstart retains connection reuse, npx installation, complete three-file curl fallback when the installer cannot run, honest host-local/native Skill persistence, CLI/MCP choice and necessary API fallback.
  • Skill PR51 adds a short Sign-in route to actual help; it does not duplicate the manual.
  • CLI PR23 owns actual URL/persistent-terminal/one-time-result/reuse/expiry/balance mechanics. Runtime supplies the current URL/status. Login implementation is unchanged.

Working connections and configured credentials are reused. Setup is not spending approval. CLI/MCP preserve matching quotes and valid scope/cost approval; direct paid HTTP preserves endpoint/charging-contract review, exact scope/rate/maximum, a separate later user approval, no paid probes and refusal when provider bounds are missing.

Review map: Agent pages are agent-quickstart.mdx and its Chinese counterpart; earlier human introduction/copy prompt stays in agent-skills/quickstart.mdx and its Chinese counterpart. Current overall PR against main347835f:4 files,+197/-503. This latest consolidation changes only two Agent pages,+32/-182 against12df325. No new route, service or installer package.

Validation:

  • Current two-page Mintlify preview HTTP200; exact public Raw bytes match the reviewed guide. Candidate Skill0332d85 real curl download works with no Node on PATH and matches all three files; native Codex loader reports aisa enabled/displayAIsa. Preservation/failure controls are retained for the unchanged download script, with only its immutable source pin updated.
  • Actual packed CLI help retrieval plus five controlled login/PTY/balance states: final full5/5, removed-help2/5 semantic result (mechanical5/5 vs4/5). All three complete rounds are retained; no selective resampling. Current results, source pins and independent audit. No native Work OAuth, production key mint, real balance or paid-data call was performed.
  • Earlier budget evaluation retains its original55f017a inputs and6/6 calibrated result; it is not relabeled as a new run. The direct-HTTP policy tail remains unchanged in this consolidation.

Merge and production verification authorized. The user explicitly requested: verify the CLI release, merge Docs and confirm the real remote content, then merge evals only after both pass. Official npm0.5.2 has been freshly downloaded/installed; its archive integrity and compiled login-help hash match the reviewed release. Skill PR51 and CLI PR23 are merged. The reviewed Docs head remains5587b3b. After merge, verify the official English/Chinese Agent Markdown and human Quickstart pages; GitHub merge status alone is not publication proof. Evals PR14 stays unmerged until that check passes.

Production verification (2026-09-10): merged as 0a3b0fcd65b6bc1c0fd4f45ae6e7801da3837440; Mintlify main check 102771389555 succeeded. All four files on GitHub main match reviewed head 5587b3b. Actual English and Chinese Agent Markdown bodies match after only Mintlify formatting normalization; both paid-approval XML blocks are byte-identical. Both rendered Agent pages expose the new Skill pin and CLI help reference, and both human Quickstart pages expose the one-line setup prompt and spending boundary. CLI 0.5.2 was freshly downloaded and installed from official npm before this merge.

Replace key-first navigation with an English copy prompt, one official
aisa skill via npx skills, browser login, and a no-terminal MCP OAuth path.
Keep Chinese pages and Agent Skills Quickstart on the same installer.
Raw <id> was parsed as JSX and blocked Mintlify preview. Present on origin/main.
Unpinned skills 1.5.25 engines require Node.js >=22.20.0; keep a working
CLI on its Node 18+ floor. Restore INSTALL-hermes.mdx to origin/main
baseline. Full-site raw <id> parse failures stay documented, not cleaned.
Shorten English-first onboarding, point Skills Quickstart at the canonical
flow, and keep Claude/Codex MCP examples as help-sourced copyable steps.
Keep one MCP tool table, cite Claude/Codex help without verification
disclaimers, and treat remote-read errors as cause-specific.
Move the public copy prompt onto Agent Skills Quickstart. Keep
agent-quickstart.mdx as operating instructions without a self URL.
Keep CLI/MCP on quote-and-scope rules. Use direct HTTP only when both
are unavailable and an existing key is present; do not switch transport
to bypass auth, credit, or outage.
Limit the XML contract to costly data endpoints, allow public discovery
without a key, keep one fallback ladder, and use aisa.one docs links in
the raw Agent page.
Reuse AISA_API_KEY, ~/.aisa/key, or the client secret store; browser
login first; console key creation only when login is unavailable.
Canonical skill path moved to the repo root aisa/ package.
Canonical skill path is the platform category package after the root
Vercel discovery limitation.
HTTP must read the selected operation, not a catalog overview. If quote
or documented max exceeds remaining approval, stop instead of asking
for an under-cap retry.
Require the OpenAPI/operation read before presenting a paid request
for approval, not only after the user has already approved.
@eddiearc
eddiearc marked this pull request as ready for review September 10, 2026 06:39
@eddiearc
eddiearc merged commit 0a3b0fc into main Sep 10, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant