aiti-models-import · Models Config Import

❝ Don't copy the config — just move in. ❞

Scans the provider·model configs of other coding agents on this machine into drafts; a first-launch dialog guides a one-time import into pi's models.json. Key plaintext never leaves the server, changes are detected via source-level fingerprints, and saving always goes through the 🧙 aiti-model-wizard chain.

Source adapters

• codex: ~/.codex/config.toml (parsed with smol-toml) — [model_providers.*] → wire_api mapping (chat→openai-completions / responses→openai-responses); top-level model+model_provider becomes the default model; auth.json OAuth is not imported, keys go through apiKeyEnv + a warning
• zcode: top-level provider sections of ~/.zcode/v2/config.json — strips the builtin: prefix, excludes *-start-plan channels entirely, filters coding-plan-cache by plan authorization; keys are encrypted in the desktop credentials.json → drafts carry no key
• opencode: multi-file merge of ~/.config/opencode/opencode.json + auth.json (api key); the npm SDK package name maps to the protocol, and when models is empty a skeleton is imported + a hint
• Non-importable, grayed out: cursor (login state encrypted), qoder (credentials WASM-encrypted) — the reason is honestly labeled

Mechanism

• Auto-pops on first launch: GET /status (only checks file existence, zero polling) decides firstRun, and the dialog pops only when a usable source exists; once dismissed or imported, it never pops again
• GET /scan: scan each source → normalize → source-level sha256 fingerprint (key plaintext stripped, with hasKey semantics) → state machine available-new / changed / unchanged / source-gone + merge/overwrite conflict detection; drafts are output key-stripped (keyHint shows only the last 4 characters), and the model list carries preview metadata (context/output/reasoning/new|existing)
• POST /import: dedupe selections → fresh re-scan per source (not the /scan cache) → keepKey:true (existing keys of same-named providers are not cleared) → persisted via the aiti-model-wizard.saveBatch service → per-item results backfill the markers (failed items don't get marked; total failure doesn't count as done, and the first-launch dialog pops again)
• Stores <stateDir>/import-state.json (0600 atomic write); on successful import broadcasts pi-web:models-changed so the model selector refreshes immediately

Dependencies

• Hard dependency aiti-model-wizard.saveBatch (stashUndo before the batch rolls the whole batch back in one step / when absent, /import returns 503 prompting a wizard upgrade); soft dependency aiti-quickstart (the onboarding card probes the aiti-models-import.open service to open this dialog with zero re-implementation)
• npm smol-toml is installed into the shared root via extman reconcile; when missing, the codex adapter degrades without crashing

Sources

Planned M2 sources (openclaw / claude archives etc.) not implemented.