feat(GTM-38): let the MozFest hosted agent run on a Fireworks model

[humantorch]

Why

#1333 seeded Fireworks prices and taught the ledger the fireworks provider; #1346 gave the backend a shared provider connection factory. The hosted-agent endpoint still hardcoded Anthropic. MozFest runs on GLM 5.3 via Fireworks, so this is the last piece of the agent path.

What

• Provider from the model table. AGENT_MODEL resolves to a provider through the same managed-model data /v1/chat uses, not by sniffing the id prefix. Startup refuses an id it cannot place (a typo, a model in another Fireworks account, a bare id) with an error naming the id and the two supported shapes, so a misconfiguration fails at boot instead of 503ing every request. claude-opus-5 is no longer accepted as AGENT_MODEL even though its price row still exists. No new setting.
• Model factory. createAgentModel builds Fireworks through @ai-sdk/openai-compatible with includeUsage: true (so the stream's final chunk carries usage the ledger can bill) over createManagedProviderConnection('fireworks', ..., { source: 'agent' }), so the key, base URL, instrumented fetch and the agent source label are all shared with /v1/chat.
• Admission and ledger use the Fireworks identity, so the 0033 prices apply.
• Startup refuses AGENT_ENABLED=true with a Fireworks model and no FIREWORKS_API_KEY, via the factory's existing check, same path as the Anthropic refusal.
• Reasoning is not forwarded (sendReasoning: false), so visitors never see raw chain-of-thought from GLM or MiniMax, matching the Anthropic path. Fireworks requests carry x-prompt-cache-isolation-key derived from a hash of the user id so one account's prompt cache cannot be probed from another.
• No file parts. The hosted agent takes text only for now, whatever the model: a request with a file part gets 400 AGENT_FILES_UNSUPPORTED before admission. feat(GTM-36): use the hosted agent in the web app for MozFest #1353 stops the client sending them.
• A usage chunk with null token counts logs agent_usage_missing instead of landing a silent zero-cost row.
• Dependency: @ai-sdk/openai-compatible@^2.0.81, the current ai-v6 line. It shares @ai-sdk/provider and provider-utils with the existing ai@6 stack; the only nested copy is tinfoil's pre-existing 3.x.
• Docs: .env.example and configuration.md (Hosted agent block, AGENT_MODEL and FIREWORKS_API_KEY rows), models.md (the "follow-up release" sentence now describes current behaviour, and the seed migration list includes 0035).

Tests

Derivation (both seeded ids, Anthropic ids, lookalikes), startup refusal, a Fireworks request landing on https://api.fireworks.ai/inference/v1/chat/completions with the bearer key and stream_options.include_usage, admission seeing { provider: 'fireworks', model }, and a ledger row against migrated PGlite costed from the 0033 GLM row (184,000 nano-USD for 100 in / 10 out). Six revert checks.

• bun test src/hosted-agent src/inference --randomize --rerun-each 5: 1740 pass, 0 fail
• tsc, eslint, prettier clean

Live smoke against Fireworks was not run (no key in the dev env). Steps: PORT=8011 AGENT_ENABLED=true AGENT_MODEL=accounts/fireworks/models/glm-5p3 bun run dev, one anonymous chat through /v1/agent/chat, check inference_usage for a fireworks row.

Size

Against main: 16 files changed, 469 insertions(+), 119 deletions(-), including the lockfile. Over the 400-line guideline after the review round (table-based provider resolution, the file-part rejection and the new tests).

[greptile-apps]

Confidence Score: 4/5

[High impact] Adds Fireworks model support to the hosted agent.

The required helper comments still need to be added before merging.

Findings

1.  New helpers lack required comments ▶

Summary

The hosted agent can now use supported Fireworks models alongside Anthropic.

• The hosted agent routes supported model IDs to Anthropic or Fireworks.
• The hosted agent refuses requests that include file attachments.
• Hosted-agent replies keep reasoning private and isolate Fireworks prompt caches by user.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[AGENT_MODEL] --> B[Managed model table]
  B --> C[Anthropic connection]
  B --> D[Fireworks connection]
  E[Text-only request] --> F[Admission check]
  F --> C
  F --> D
  C --> G[Reply without reasoning]
  D --> G
  C --> H[Usage ledger]
  D --> H

Loading

_{Reviews (3) · Last reviewed commit: "test(GTM-38): boot the wiring test with ..." · Reviewed by Greptile}

[github-actions]

Semgrep Security Scan

No security issues found.

[github-actions]

PR Metrics

Metric	Value
Lines changed (prod code)	+197 / -58
JS bundle size (gzipped)	🟢 653.1 KB → 652.9 KB (-184 B, -0.0%)
Test coverage	🟢 82.85% → 82.85% (+0.0%)
Performance (preview)	Preview not ready — Render deploy may have timed out
Accessibility	—
Best Practices	—
SEO	—

Updated Sun, 11 Oct 2026 19:57:57 GMT · run #3410

[github-actions]

🔭 thunder-deep-review (advisory)

Reviewed the diff — no issues to report. ✅ Never approves, never requests changes, never gates merge.
_{head: 40c796c5c571 · mode: single}

[github-actions]

Preview environment destroyed 🧹

Stack preview-pr-1359 and its Cloudflare subdomain have been cleaned up.

[darkbanjo]

did a deep pass, comments are tagged blocking / convention / nit. the real end-to-end billing test is great. a few things outside the diff:

• convention: fireworks caches prompts by default at about a fifth of the input price, but the ledger zeroes cache tokens for non-anthropic providers (usage-ledger.ts:79-82). the agent resends the whole conversation every turn, so from turn two most prompt tokens bill at ~5x real cost and the anonymous 10¢ window runs out well before real spend does. worth pricing cache reads before the event, or at least dropping "slightly" in models.md:94?
• nit: a usage chunk with null token counts becomes 0, so the run lands as a zero-cost row with no agent_usage_missing log. and history.test.ts:107-112 only covers the forged providerMetadata case with an anthropic key, an openaiCompatible one would pin the boundary this provider now relies on.
• follow-ups: a run cut at the 120s timeout is never billed (fireworks only reports usage in the last chunk), and with client system messages dropped and an empty AGENT_SYSTEM_PROMPT, reply language falls to the model's default, which changes going from claude to glm.

[humantorch]

Thanks for the review, and for the cache-pricing catch. Everything tagged is addressed in 2e39e2a and answered on its thread. The items outside the diff:

• Cache-read pricing in the ledger is ticketed in Linear and being picked up now; models.md no longer says "slightly".
• A usage chunk with null token counts logs agent_usage_missing instead of landing a silent zero-cost row, and history.test.ts has the openaiCompatible forged-providerMetadata case beside the Anthropic one.
• The 120 s timeout billing gap and reply language are ticketed in Linear as follow-ups.

One behaviour change worth knowing: with the table-based resolution, claude-opus-5 is no longer accepted as AGENT_MODEL even though migration 0028 still seeds its price. The live Fireworks smoke is still unrun for lack of a key in the dev env.

[greptile-apps]

New helpers lack required comments

containsFilePart has no JSDoc comment. This violates the repository directive: “Add JSDoc comments to new utility functions.” The new createFilesUnsupportedResponse helper in backend/src/hosted-agent/routes.ts also needs one. Add short comments explaining both helpers before merging.

Context Used: CLAUDE.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!