feat: edit managed-model context windows in Settings

[22zyfeng-collab]

Summary

Add context-window inputs to Settings → Models for managed providers, including account/OAuth models such as ChatGPT Plus/Pro. Custom models already expose this setting; managed models currently require editing models.json by hand.

• Use Pi's native providers[provider].modelOverrides[modelId].contextWindow; no SDK patches or new dependencies.
• Keep these edits in the Models panel's existing draft and shared Save flow, rather than introducing a second writer that could overwrite custom-model edits.
• Show the SDK-resolved budget, validate positive safe-integer token counts on both client and server, and retain invalid/failed-save drafts.
• Clearing the input or choosing Default removes only the context override, preserving credentials, other overrides and custom models.
• Keep override-only entries under their managed provider instead of duplicating them in the custom-provider sidebar.
• Include English, Simplified Chinese and Traditional Chinese strings, accessible labels/errors, and a narrow-screen layout.

Application semantics

This changes Pi's context accounting, not the upstream provider's actual capacity. New session runtimes read the saved value. Loaded sessions need a Pi Web restart: an integration check confirmed that Pi 1.2's /reload retains their old model metadata, so the help text deliberately does not promise that /reload applies this edit.

Verification

• 92 targeted tests passed: context helper/component, ModelsConfig helpers, models-config HTTP handlers/store (including real SDK resolution), enabled-models runtime/edit engine and i18n registry.
• npm run lint passed.
• Production webpack build and TypeScript checks passed; the existing dynamic-import warning in the session export route remains.
• Ran checkModelContext against the packed production build at 1280px and 390px: edit, shared Save, reopen, invalid draft, server refusal/retry, Default, preservation of unrelated configuration, and no duplicate provider row.
• Additional isolated HTTP/browser integration with fixture-only credentials: real UI → models.json → newly created session contextUsage.contextWindow changed 272000 → 500000 → 272000, without changing auth.json or sending inference requests.

Existing test limitations

The full suite was attempted on Windows/Node 24 and did not finish cleanly. Unrelated MCP component/transport failures also reproduce on an untouched v0.12.1 worktree (53 failures among 62 tests in components/McpConfig.test.mjs and app/api/mcp/test/route.test.mjs). These baseline issues are not changed by this PR; I am not claiming a green full suite.

The E2E runner blocks service workers so its API route fixtures also intercept requests in production mode.

[agegr]

Thanks @22zyfeng-collab! I checked this on a merge with main first: pi 1.2.0 really applies modelOverrides[id].contextWindow (in a scratch agent dir openai-codex/gpt-5.5 resolves to 500000 instead of 272000, and a new AgentSession reports that window), the field goes through the shared Save without clobbering custom-model edits, and you're right that /reload keeps the old value, so the restart note is accurate.

The one thing I wanted changed was the weight of each row: the field took every model row from 43.5 to 78 px (117.5 px on a phone), so OpenRouter's 393 rows grew from about 17.5k to 31k px. I pushed five commits on top of your branch so it can go in:

• 7e5d7df fix(models): let a context override pi accepts through a save. pi accepts any positive contextWindow (400000.5 loads), but the normalizer refused any non-integer already in models.json, so one hand-edited value blocked every later Save, even for an unrelated custom provider. What you type in the field still has to be a whole token count; a value already in the file only needs to pass pi's own check, and values pi refuses are still refused, named by provider/id. A new route test returns 400 on your head and 200 now.
• 20d5faa refactor(models): reduce the context-window editor to one input per row. Each managed row now has one compact field. Its placeholder is the window pi resolves for that model, and blank means no override. The label, Default button and "Currently" line are gone (the aria-label names the model), the note above the list is one line on desktop, and the unused strings are removed in all three locales. Rows are 43.5px at 1280px and 390px, the same as main. The unit tests, e2e/model-context.mjs and docs/agents/models.md are updated.
• 1f7bb16, 8ee706c keep an overrides-only entry out of the custom list and form while loading: until /api/auth/providers answered, such an entry still flashed in the custom-provider list, and on a fresh browser the detail pane could open it as a custom-provider form.
• 7bc6fa3 docs: lib/model-context.ts and components/ModelContextWindow.tsx are in AGENTS.md's File Map and the models.md file list.

Checked on a merge with main: tsc, lint and the full unit suite pass, your e2e check passes in headless Chrome at 1280x800 and 390x844, and a real save round trip against next dev with a scratch agent dir writes, keeps and removes the override as expected.