fix: skip Claude's <synthetic> placeholder when reading transcript models

[MetalZealot]

Problem

After an API error (e.g. 529 Overloaded), Claude Code appends locally-fabricated assistant rows to the session transcript stamped with the placeholder model "<synthetic>" — the API-error notice, "No response requested.", and session-limit messages. extractClaudeEventModel accepts whatever model the last assistant row carries, so the backward transcript scan in getCurrentActiveModel adopts the placeholder and reports <synthetic> as the session's active model until a genuine turn lands.

Repro

1. Have a Claude session hit an API error (529 Overloaded, session limit, etc.) — the CLI writes assistant rows with "model": "<synthetic>" to the session JSONL.
2. Switch to / reopen that session → the active model resolves to <synthetic> instead of the model the session was actually running.

Fix

extractClaudeEventModel now rejects angle-bracketed placeholder values for both event.model and message.model, so the scan keeps walking back to the last genuine turn and recovers the session's real model. A transcript containing only synthetic rows falls back to the catalog default, same as before when no model was found.

Tests

New server/modules/providers/tests/claude-models.test.ts (node:test, same isolated-database harness as opencode-sessions.test.ts):

• control: the last assistant turn's model is read from the transcript
• synthetic error rows are skipped and the real model is recovered
• an all-synthetic transcript falls back to the catalog default

Run with npx tsx --tsconfig server/tsconfig.json --test server/modules/providers/tests/claude-models.test.ts — 3/3 pass with the fix; the two guard tests fail on current main. tsc --noEmit -p server/tsconfig.json and eslint are clean.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Improved Claude model detection by ignoring synthetic placeholder model values.
  • Active model selection now correctly retains the last genuine model after errors or synthetic transcript entries.
  • Added/verified fallback behavior when no valid model is available.
• Tests
  • Added coverage to validate active model extraction across transcripts containing synthetic placeholder entries.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 454abaee-1aba-4daa-beb7-69a4b1b3eed7

📥 Commits

Reviewing files that changed from the base of the PR and between ce7b3bc and 351f987.

📒 Files selected for processing (1)

• server/modules/providers/tests/claude-models.test.ts

🚧 Files skipped from review as they are similar to previous changes (1)

• server/modules/providers/tests/claude-models.test.ts

---

📝 Walkthrough

Walkthrough

Claude model extraction now ignores synthetic <...> placeholders from session events. New isolated-database tests verify selection of the latest real model, skipping synthetic rows, and fallback behavior when no real model exists.

Changes

Claude model filtering

Layer / File(s)	Summary
Placeholder model extraction server/modules/providers/list/claude/claude-models.provider.ts	Recognizes <...> placeholder values and excludes them from direct and nested Claude event model fields.
Model selection validation server/modules/providers/tests/claude-models.test.ts	Sets up isolated transcript data and tests real-model selection, skipping synthetic rows, and returning the default model when all rows are synthetic.

Possibly related PRs

• siteboon/claudecodeui#1047: Modifies the Claude model selection and JSONL transcript resolution path.

Poem

I nibbled placeholders away,
So real Claude models could stay.
Through transcript rows I hop and roam,
And find the last true model home.
If none appear, the default takes flight.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes the main change: ignoring Claude synthetic placeholder models when reading transcript models.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧹 Nitpick comments (1)

server/modules/providers/tests/claude-models.test.ts (1)

90-105: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Cover placeholders in event.model too.

The test currently validates only message.model; add model: '<synthetic>' to one synthetic row so both changed extraction paths are protected.

Proposed test update

         {
           type: 'assistant',
           sessionId: PROVIDER_SESSION_ID,
+          model: '<synthetic>',
           message: {
             model: '<synthetic>',
             content: [{ type: 'text', text: 'API Error: 529 Overloaded.' }],

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@server/modules/providers/tests/claude-models.test.ts` around lines 90 - 105,
Update the synthetic assistant event fixtures in the test to include model:
'<synthetic>' on one row, while retaining message.model coverage on another row,
so extraction from both event.model and message.model is validated.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@server/modules/providers/tests/claude-models.test.ts`:
- Around line 90-105: Update the synthetic assistant event fixtures in the test
to include model: '<synthetic>' on one row, while retaining message.model
coverage on another row, so extraction from both event.model and message.model
is validated.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 37ebf5f8-de5e-45c0-a5de-72cd3a8fd748

📥 Commits

Reviewing files that changed from the base of the PR and between 27eaf01 and ce7b3bc.

📒 Files selected for processing (2)

• server/modules/providers/list/claude/claude-models.provider.ts
• server/modules/providers/tests/claude-models.test.ts

[blackmammoth]

Hello @MetalZealot, thanks for the PR. I tried replicating the issue but it doesn't seem to occur. Can you check and verify again? If so, can you share a more clear reproduction step?

[MetalZealot]

Thanks for checking. I re-verified this against current main (c2408f0) and the issue is still present.

The trigger is narrower than “any 529”: it occurs only when Claude Code appends a locally generated assistant row whose message.model is "<synthetic>". If the failed request does not write one of those rows, the bug will not reproduce.

A deterministic reproduction is:

1. Create/open a Claude session and let one genuine assistant turn land.
2. In that session's JSONL, append a synthetic assistant row after the real turn (replace <session-id> with the transcript's session id):

{"type":"assistant","sessionId":"<session-id>","message":{"model":"claude-sonnet-5","content":[]}}
{"type":"assistant","sessionId":"<session-id>","message":{"model":"<synthetic>","content":[{"type":"text","text":"You've hit your session limit"}]}}

3. Switch away and reopen the session, then open /model.

Actual on current main: the active model resolves to <synthetic>.

Expected: the scan skips that placeholder row and keeps walking back to claude-sonnet-5.

The same rows are produced naturally by session-limit notices, some API/auth errors, and "No response requested.". As a real-world check, I found 173 synthetic rows across 75 local Claude transcript files, including recent rows from July 30. One sanitized adjacent pair is:

2026-07-17T14:21:54.526Z  claude-fable-5
2026-07-17T14:21:55.603Z  <synthetic>  You've hit your session limit

I also merged this PR locally into current upstream main and reran:

npx tsx --tsconfig server/tsconfig.json --test server/modules/providers/tests/claude-models.test.ts

All 3 tests pass on the current merge result.