fix(agent): clamp provider context images

[roboomp]

Repro

A focused regression test that imports clampProviderContextImages and builds a 31-image umans context failed before the fix because @oh-my-pi/pi-coding-agent/session/provider-image-budget did not exist: bun test packages/coding-agent/test/session/provider-image-budget.test.ts exited 1 with Cannot find module '@oh-my-pi/pi-coding-agent/session/provider-image-budget'.

Cause

packages/snapcompact/src/snapcompact.ts shipped providerImageBudget("umans") === 10, but packages/coding-agent/src/sdk.ts only installed transformProviderContext when secret obfuscation or snapcompact inline imaging was enabled. No coding-agent request path clamped existing image blocks, so live contexts with more than 10 images could still be sent unchanged to umans.

Fix

• Added packages/coding-agent/src/session/provider-image-budget.ts with clampProviderContextImages(), which preserves text and the newest images while dropping oldest image blocks above the active provider cap.
• Updated packages/coding-agent/src/sdk.ts to always run the provider-context transform and apply the clamp after obfuscation and snapcompact inline transforms.
• Added packages/coding-agent/test/session/provider-image-budget.test.ts coverage for the 31-image umans regression and the below-budget no-op case.
• Added the coding-agent changelog entry for Regression: #3227 not actually fixed — clamp logic from PR #3228 never merged; 31-image 400 still fires on umans #3230.

Verification

bun test packages/coding-agent/test/session/provider-image-budget.test.ts passed with 2 tests; bun run --cwd packages/coding-agent check passed. Fixes #3230

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: eb88ecd7bb

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Preserve a placeholder for fully dropped tool-result images

When an older successful tool result contains only images, dropping all image parts here returns content: []. I checked the Anthropic/umans conversion path: buildToolResultBlock() only fills empty content for isError results, so once the provider budget is exceeded an image-only successful tool result is serialized as an empty tool_result rather than a textual omission marker. Add a placeholder text block when the clamp removes the last tool-result block so the request stays meaningful and protocol-safe.

Useful? React with 👍 / 👎.

[roboomp]

Added a [image omitted: provider image limit] text block when clamping removes every block from a successful tool result.
Added regression coverage for image-only tool results so they stay non-empty after provider image clamping.