Task: Fix Robot agent configuration loading

Identity

Reproduce the Robot page state where a selected agent remains indefinitely in configuration loading and the edit action stays disabled after device binding.
Trace the Renderer typed API, page effects, and Main-owned Works Square route without weakening authentication, ETag, idempotency, or response validation.
Fix the smallest confirmed client-side cause and add a regression test at the user-visible page seam.

The edit button may be enabled only after a real, versioned agent configuration has loaded; do not bypass the loading/error state.
Preserve Main ownership of Works Square credentials and upstream headers, and retain strict numeric revision plus ETag equality checks.
Keep activation codes and upstream response details out of Renderer state, logs, task records, and test output.
Work only in the isolated linked worktree; do not alter or release the existing main integration owner.

Task context:

Task ID: 20260815-robot-config-loading-4f8c2d
Mode: feature
Branch: codex/20260815-robot-config-loading-4f8c2d-robot-config-loading
Worktree: D:\Datas\OthersProjects\makelore-robot-config-loading-4f8c2d
Base commit: a4050f0a6567e8203630bf5d16b57f00c45caab8
Other active local tasks: the existing main integration owner, completed Robot source task, and isolated release, AI Design, AI Programming, updater, and packaging tasks returned by the registry.
Overlap or semantic-conflict assessment: the completed Robot task establishes the relevant trust and state boundaries; no active peer owns the Robot page or typed hardware API. The same-base AI Design points task is confined to AI Design. Placeholder peer records have unrelated titles and isolated worktrees, so no semantic conflict blocks this diagnosis.

Read:

Relevant understanding:

Project goal: provide a non-technical desktop workflow while keeping credentials and sensitive network behavior in Electron Main.
Current integrated focus: Robot is the single enabled hardware module and its typed Renderer → Host API → Main → Works Square chain is integrated.
Active task scope: restore the configuration-loading terminal state so the user can edit the selected agent after binding a device.
Active constraints: do not expose credentials/raw upstream errors, do not bypass revisions, and do not initialize Programming state on Robot routes.
Decisions affecting this task: one local account maps to one Xiaozhi binding; agents/devices are resources beneath it; Main owns auth/idempotency/ETags.
Evidence, reflections, or commitments affecting this task: the screenshot proves overview and device binding render successfully while the configuration pane remains loading; real production Robot smoke remains the decisive deployment check.
Files or modules likely involved: src/pages/AiHardware/index.tsx, src/lib/ai-hardware.ts, electron/api/routes/ai-hardware.ts, and their focused tests.
Unknowns, stale docs, or conflicts: the screenshot alone does not distinguish a never-settling promise from a rejected response followed by stale loading state; a red-capable page test is required before selecting a fix.

Gate result:

Reproduced the user-visible failure with a deterministic page test: after the configuration request rejected, the card still rendered 正在读取配置, exposed no retry action, and kept editing disabled.
Split configuration loading from its terminal failure state. A failed read now renders a safe local error state with 重试读取配置; a successful retry restores the prompt and enables editing only after a real configuration and revision have loaded.
Clear loading/error state when no agent is selected, preventing an abandoned request from leaving a stale spinner.
Extended the Main-owned upstream deadline through bounded response-body reading and JSON parsing. A response that returns headers and an initial chunk but never completes now terminates as the safe retryable 504 / AI_HARDWARE_TIMEOUT envelope instead of leaving Renderer pending forever.
Preserved strict response schemas, ETag/revision equality, Main-owned authentication, idempotency, response-size limits, and upstream error redaction.
Cross-repository inspection found no current source-contract mismatch for agent configuration fields, revision numbers, or ETags. Production still needs to be checked for the current Xiaozhi image and Liquibase migrations if a retry reports a server error.

Red page test on the pre-fix implementation: pnpm exec vitest run tests/unit/ai-hardware-page.test.tsx -t "stops loading and lets the user retry after configuration loading fails" — 1 failed, proving the stale loading state.
Red Main route test on the pre-fix implementation: stalled configuration response body exceeded the 100 ms test deadline without returning an envelope.
corepack pnpm exec vitest run tests/unit/ai-hardware-page.test.tsx — 25 passed.
corepack pnpm exec vitest run tests/unit/ai-hardware-routes.test.ts — 18 passed.
Focused Robot/Main regression suite across page, route, typed API, Host proxy, module navigation, layout gate, and provider gate — 7 files / 74 tests passed.
corepack pnpm run typecheck — passed.
ESLint on the four changed TypeScript/TSX files — passed.
git diff --check — passed; only Git line-ending conversion warnings were emitted.
Independent final Sol review — PASS; the reviewer reproduced the 7-file / 74-test result and found no code or documentation blocker.

After deployment, verify the Xiaozhi manager image includes numeric External API revision serialization and that Liquibase changesets 202608131500 and 202608131501 are applied if configuration reads still fail.
Perform a signed-in staging smoke: bind a device, load the selected agent configuration, edit systemPrompt, save with the returned revision, and confirm the updated prompt after refresh.