Per-agent provider selection for custom agents
Dieses Issue hat noch niemand übernommen.
- Vorherrschende Sprache
- Shell
- Sterne
- 11.2k
- Forks
- 1.9k
- Ø Merge
- 14 Std. 16 Min.
- Gemergte PRs (30 T.)
- 6
Beschreibung
I know about BYOK (COPILOT_PROVIDER_BASE_URL plus COPILOT_MODEL), /model, and the model field in custom-agent frontmatter. We use all three. None of them lets two agents in one session talk to two different endpoints, because COPILOT_PROVIDER_BASE_URL is process-wide and nothing scopes it to an agent. That is the difference between this and #3624 and copilot-sdk#1545, both closed within hours pointing at the BYOK docs.
The ask
A provider field in custom-agent frontmatter. Or the per-agent env/env_file scoping that #2879 proposes for secrets, which would do the same job. The main agent keeps the session provider. The subagent gets its own base URL.
What the runtime already does
I put a recording HTTP server at COPILOT_PROVIDER_BASE_URL on 1.0.83-3 and dispatched a custom agent through the task tool. Three requests came out of one session:
1 model=local-worker-model tools=24 main agent
2 model=ROUTE-ME-LOCAL tools=1 subagent; model from its .agent.md frontmatter,
its own instructions in the system prompt
3 model=local-worker-model tools=24 main agent
The model override on the task tool arrives the same way. The per-agent model id already reaches the provider, and v1.0.62 ("Keep custom agents on their configured model when using BYOK providers") and v1.0.65 for subagents made that deliberate. What it cannot carry is an endpoint to send that id to.
Use case
A cloud main agent plans and reviews. It hands renames, comment edits, log lines and single-file changes to a subagent running a 27B model on the developer's Mac (MLX, 48 GB). Those tokens cost nothing. I am not claiming the local model matches the cloud one. On our own eight-task benchmark two local models came out indistinguishable from each other, and the only claim I will stand behind is free tokens for mechanical work. Judgement stays in the cloud.
What we tried
A routing proxy at COPILOT_PROVIDER_BASE_URL, keyed on the model id. Known local ids go to the MLX server and the proxy forwards the rest to GitHub. Two problems killed it. It puts the main agent's traffic through GitHub endpoints that are not documented for that use. And the docs say "A declared model or effort that can't be honored falls back to the session's value instead of failing the dispatch", so any proxy error on a local id sends that task to the session model instead. You find out from the bill rather than from an error.
Related
- #2879 asks for per-agent
env:/env_file:to isolate secrets. Same mechanism, different reason for wanting it. - #2728 asks for per-agent LSP config, so per-agent configuration beyond
modelis a shape this project already accepts. - #3709, #3282, #1354, #2554 and #2758 want the model version of this. A commenter on #3709 already asked for subagents to use the other models a provider offers.
- copilot-sdk#1544 named routing subagent tasks to a local endpoint. Its author closed it 14 minutes later while still investigating, and nobody from GitHub replied.
Beitragsleitfaden
Erste Schritte
- Lies das ganze Issue und danach den Beitragsleitfaden des Projekts.
- Schreib ins Issue, dass du es übernimmst — das erspart doppelte Arbeit.
- Forke das Repository und arbeite in einem Branch.
- Öffne einen Pull Request, der die Issue-Nummer nennt.
Rechercherichtung
Beginne mit dem Dispatch-Pfad des task tool und der Verarbeitung des custom-agent frontmatter und verfolge anschließend, wie COPILOT_PROVIDER_BASE_URL für die gesamte Sitzung und das agentenspezifische Modell den Provider erreichen. Lege fest, wie ein Subagent seinen eigenen Endpoint auswählt, während der Hauptagent den Sitzungs-Provider beibehält, und überprüfe, dass eine Sitzung mit beiden Agents Anfragen korrekt weiterleitet, ohne Verwirrung durch Fallbacks.
Vom Indexierungsmodell aus dem Issue-Text verfasst.
Bewertung
- Bereich
- ai, cli
- Issue-Typ
- Feature
- Schwierigkeit
- 5/5
- Geschätzter Aufwand
- Über eine Woche
- Aktivitätsstatus
- Aktiv
- Klarheit
- Größtenteils klar
- Anfängerfreundlichkeit
- 45/100