[devin-ai-integration]

Summary

Before this change, requesty configure for Pi wrote only {"id": "<model>"} into ~/.pi/agent/models.json. Pi then used its custom-model defaults (contextWindow ?? 128000, maxTokens ?? 16384 in modelFromJson), so a 1M-context model run through plain pi auto-compacted at 128k.

The shared modelpicker already holds the full client.Model from /v1/models / /v1/models/managed. modelpicker.ChosenMsg now carries that Model instead of just the ID, and the integrations wizard copies its ContextWindow and MaxOutputTokens into ConfigureOptions. In both configure paths (merge and overwrite), Pi builds its entry from those with piCatalogModelFrom, the same function the requesty pi launch catalog uses:

{"id": "anthropic/claude-sonnet-5-5", "contextWindow": 1000000, "maxTokens": 128000}

If the list has no value for one of these, the launch defaults are written (piCatalogContextWindow 200k, piCatalogMaxTokens 8192). Other harnesses ignore the new options.

The values are a snapshot from configure time; the pi-requesty extension, if installed, replaces them with live limits on startup.

Link to Devin session: https://app.devin.ai/sessions/ed8e81b7f4e246f79baeef970ae8f691
Open in Devin Desktop: https://app.devin.ai/desktop/session/ed8e81b7f4e246f79baeef970ae8f691?variant=devin
Requested by: @requesty-JohnCosta27

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[github-actions]

AI Cost

No spend detected for branch devin/1791370894-pi-configure-context-window in the last 30 days.

_{Updated 2026-10-27T12:46Z · last 30 days · data from Requesty}

[devin-ai-integration]

Tested 70043b8 end to end with the real Requesty catalog, an isolated HOME, and plain Pi v1.4.0 with extensions disabled (--no-extensions).

• Picking Sonnet on the Models tab and choosing merge wrote contextWindow: 1000000 and maxTokens: 128000 to models.json. Pi replied ok and its footer showed 0.2%/1.0M (auto), not /128k.
• Overwrite mode and a managed-policy pick both wrote 1000000/128000.
• Picking Sonnet, pressing Esc on the mode step, then picking Haiku wrote 200000/64000, so the Sonnet values were not kept.
• A policy with no max_output_tokens fell back to maxTokens: 8192.
• Merge kept an unrelated provider and its models.

Plain Pi response and footer	Pi model registry

Not exercised: actual compaction near 1M tokens, and the 200k fallback for a model with no context_window. Merge still replaces the existing providers.requesty.models array; that behaviour predates this PR.

Written by Devin

[devin-ai-integration]

Replaced by #40 (branch renamed to fix/eng-2213-..., squashed).