git: git_log description doesn't say results are newest-first or that timestamp filters are optional; small models filter to an empty window

Aperta Adatta ai principianti
#5,552 2 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
1/5
Tempo stimato
Meno di un'ora
Idoneità per principianti
88/100
Tipo di issue
Documentazione
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
git, typescript
Ambito
devtools

Direzione di ricerca

Trova la definizione dello strumento git_log e leggi la descrizione attuale. Aggiorna la descrizione specificando che i risultati sono ordinati dal più recente al meno recente e che start_timestamp e end_timestamp sono filtri facoltativi per l’intervallo di date, usando la formulazione suggerita nell’issue. Il lavoro è completato quando la descrizione comunica chiaramente entrambi i punti; l’issue non indica alcun test da eseguire.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

enhancement server-git v2
Which server?

git (mcp-server-git)

Protocol era

Both, or not era-specific

MCP client

Custom SDK client: AgentGrade (open-source eval harness on the MCP Python SDK 2.3.0 client); model qwen3:4b-instruct via Ollama, temperature 0.7

The problem

Server: mcp-server-git==2026.10.10 (checked: 2026.11.11 has the same git_log description and the same empty output). Run with uvx mcp-server-git --repository <repo>, stdio, macOS.

git_log's description is "Shows the commit logs". It doesn't say results come newest first, or that start_timestamp / end_timestamp are optional. Asked "Who made the most recent commit on the main branch?", a small model repeatedly called:

{"repo_path": "<repo>", "max_count": 1, "start_timestamp": "now", "end_timestamp": "now"}

The server returned an empty Commit history: and the agent concluded "the commit history is empty… unknown author". This happened in 5 of 6 runs of that question (two independent runs), even when the agent knew the correct repo path.

Measured on a deterministic repository (identical commit hashes every run), 5 read-only questions x 3 runs, answers checked in code, no LLM judge.

Case study (method, all tables, caveats): https://github.com/Tanishq67m/AgentGrade/blob/main/docs/CASE_STUDY.md
Raw reports: https://github.com/Tanishq67m/AgentGrade/tree/main/results

Solution you have in mind (optional)

Change only git_log's description to something like:

Shows the commit log, newest commit first. To see the latest commits, call it with just repo_path (and optionally max_count). start_timestamp and end_timestamp are OPTIONAL filters for a specific date range - leave them out unless the user asks about particular dates, because a narrow range returns no commits.

With only that change, the suite went from 13/15 to 15/15, and the "most recent commit" question from 1/6 to 3/3.

Two smaller ideas in the same spirit:

  • When no commits match, say so with a hint ("No commits in this range; omit the timestamps to see the latest commits") instead of an empty Commit history:.
  • The repo_path error already names the allowed repository, which is great, but small models usually stopped there instead of retrying (all 9 failures in one run), and a 120B model mis-copied the long path twice. A default repo_path (the configured --repository) would remove that failure class.

(We also hit git_diff rejecting main..feature in 2026.10.10. Thanks for already fixing that in #5194.)

Which MCP feature does it demonstrate? (optional)

No response

Alternatives or workarounds you have tried
  • Agent-side prompt advice ("If a tool returns an error, read it and retry"): 8/15 vs 7/15 baseline, so no real effect, and it made another task worse.
  • Telling the agent the repo path in its system prompt: 7/15 → 13/15 overall, but the "most recent commit" question stayed at 1/6. The timestamp trap is independent of the path.
  • A larger model (gpt-oss-120b) never used the timestamp filters and scored 14/15, so this mainly affects small, cheap models.
Already built it locally?

No server code was changed. The suggested description was tested by having the client show the model only a different git_log description, with the server binary unchanged and the same tasks and repeats. Verified by deterministic, code-checked answers over 15 runs per variant. The exact text is in "Solution" above.

Before you submit
  • I searched existing issues and this is not a duplicate.
  • This is a request for this repository or one of its servers, not a new server, the MCP specification or an SDK.
Lingua principale
TypeScript
Stelle
92k
Fork
12k
Merge medio
3h 12m
PR unite (30g)
174

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di modelcontextprotocol/servers

Tutte le issue di modelcontextprotocol/servers

Issue simili

Altre issue su TypeScript

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.