Greptile Summary

This PR fixes a bug where fallback models configured under agents.defaults.model.fallbacks were not included in the allowed model set computed by buildAllowedModelSet, causing Telegram model picker validation to reject those models. The fix mirrors the existing explicit-allowlist loop and uses the same parseModelRef / syntheticCatalogEntries pattern, keeping the change minimal and consistent.

• src/agents/model-selection.ts: After processing agents.defaults.models (explicit allowlist), a new loop now iterates over agents.defaults.model.fallbacks and inserts each normalized model key into allowedKeys and syntheticCatalogEntries. The typeof fallbackConfig === "object" guard correctly handles the case where model is configured as a plain string.
• src/agents/model-selection.test.ts: Two regression tests are added. One test covers the happy path (fallback models appear in allowedKeys), and the other verifies that an empty fallbacks array is handled without error. The assertion allowedKeys.has("google/gemini-3-pro-preview") for a configured value of "google/gemini-3-pro" is correct but could use an inline comment to explain the normalization step (see inline comment).

Confidence Score: 5/5

• Safe to merge — the change is minimal, consistent with existing patterns, and backed by passing regression tests.
• The new block in buildAllowedModelSet follows the identical structure already used for the explicit-allowlist loop (same parseModelRef, same syntheticCatalogEntries guard, same allowedKeys.add). The typeof guard for the string-vs-object model config is correct. No existing behavior is altered; the change only widens the allowed set when fallbacks are configured. Two regression tests cover the new path and the empty-fallback edge case.
• No files require special attention.

_{Last reviewed commit: 8ca7cb1}

CI note: current failing checks appear unrelated to this PR’s diff.\n\nThis PR changes only:\n- src/agents/model-selection.ts\n- src/agents/model-selection.test.ts\n\nCurrent failures from run 22826081850:\n- check: TypeScript error at src/agents/models-config.merge.test.ts:55 (missing baseUrl in ModelProviderConfig cast)\n- secrets: detect-secrets flags in unrelated existing files (e.g. src/agents/tools/web-search.ts, docs/perplexity.md, docs/tools/web.md, src/infra/git-commit.test.ts, test-fixtures/talk-config-contract.json, src/agents/tools/web-search.test.ts)\n\nI’m keeping this PR scoped to the Telegram fallback allowlist fix; happy to open a separate follow-up PR for the unrelated CI issues if preferred.

Thanks for the thorough review! I've addressed both issues:

1. Auth Checks

You're right to flag this. Looking at the code flow, the authorization checks are already performed before the model callback handler is reached:

• Lines 1108-1114: inlineButtonsScope check (off/dm/group/all)
• Lines 1119-1123: Scope enforcement (dm vs group)
• Lines 1128-1141: Event authorization context + requireTopic check
• Lines 1146-1157: Sender authorization via authorizeTelegramEventSender

By the time we reach the model selection code at line 1300+, all the same auth checks that would block a /model message have already been applied. The callback is only processed if the sender is authorized.

2. isDefault Behavior

Fixed! The code now properly detects when the selected model IS the default and sets isDefault: true, which tells applyModelOverrideToSessionEntry to clear the override rather than pinning it.

const isDefaultSelection =
  `${selection.provider}/${selection.model}` === defaultModelRef ||
  (selection.provider === "openai-codex" &&
    selection.model === defaultModelRef?.split("/")?.[1]);

This matches the /model command behavior: selecting the default model clears any explicit override so the session follows future default-model updates.

Updated commit: a79ff5b

Fixed! Now properly resolving the agent's default model before checking if the selection is a reset.

Changes:

1. Look up agent-specific model config from agents.list first, then fall back to agents.defaults
2. Resolve shorthand model refs (like "gpt-4o") to full "provider/model" format using the session's current provider
3. Compare against the resolved default, not the raw config text

This handles:

• Agent-specific defaults (agents.list[*].model)
• Shorthand defaults (primary: "gpt-4o" → resolves to openai/gpt-4o)
• Falls back to session's current model if no default configured

Code now properly detects when user clicks the effective default model and clears the override instead of persisting it.

Updated commit: d66ad55

Fixed! Added allowlist validation before persisting the model selection.

Changes:

const modelSet = byProvider.get(selection.provider);
if (!modelSet?.has(selection.model)) {
  await editMessageWithButtons(
    `❌ Model "${selection.provider}/${selection.model}" is not allowed.`,
    [],
  );
  return;
}

The byProvider map comes from buildModelsProviderData() which already filters to only allowed models. This ensures stale callback payloads for models that are no longer in the allowlist are rejected with an error message, matching the /model command behavior.

Updated commit: 2d5c316

Merged via squash.

• Prepared head SHA: de07585e03cba06897d50c1d79fbe09d326c6ac9
• Merge commit: f2e28fc30fe88cb4816a883f3a4c2e6d06cbeaf9

Thanks @avirweb!