Summary

Markdown docs drifted after the open-source completion path moved from instruction-tuned to base models (+ a default-off constrained decoder) and the visual-context summarizer was removed. This brings them back in line with main. Each change was verified against the current code by a per-file audit.

• README.md — model table now lists the tabby-2-* base GGUFs (Qwen3.5 0.8B/2B/4B + Gemma E2B/E4B) with real sizes and mradermacher sources; the Open Source engine is reframed as base-model continuation (conditioned, not instructed).
• ARCHITECTURE.md / AGENTS.md / .claude/CLAUDE.md — drop the deleted LlamaVisualContextSummarizer and instruct LlamaPromptRenderer; describe the OCRTextHygiene cleanup (no summarization step), BaseCompletionPromptRenderer, and the default-off constrained decoder; point the llama integration at the CotabbyInference SwiftPM package instead of a nonexistent LlamaRuntime/ dir / llama.swift.
• docs/POLLING_AND_DELAYS.md — remove the deleted summarizer timeout row.

CONTRIBUTING.md, RELEASING.md, the issue/PR templates, CODE_OF_CONDUCT.md, and the ship skill were audited and needed no changes.

Validation

Docs-only change (no build). Audited by per-file subagents that cross-checked every factual claim against the current code; a repo-wide sweep confirms no residual references to the deleted LlamaVisualContextSummarizer, LlamaPromptRenderer, VisualContextSummaryPromptRenderer, llama.swift, or the old tabby-1-* / instruct model names.

Linked issues

None.

Risk / rollout notes

• Documentation only; no code or behavior change.
• The project wiki (separate repo, no PR mechanism) was refreshed in the same pass and pushed directly: How-Cotabby-Works, Debugging-Guide, How-To-Add-a-Feature, Privacy-and-Safety-Model.

Greptile Summary

This is a documentation-only PR that refreshes five markdown files to align with code that has already shipped: the OSS path migrated from instruction-tuned to base GGUFs, LlamaVisualContextSummarizer was removed, and LlamaPromptRenderer was replaced by BaseCompletionPromptRenderer.

• README.md: Model table updated from four tabby-1-* instruct entries to five tabby-2-* base-model entries (filenames, sizes, and mradermacher HuggingFace links); all values confirmed exact matches against LlamaRuntimeModels.swift.
• AGENTS.md / ARCHITECTURE.md / .claude/CLAUDE.md: Remove LlamaVisualContextSummarizer, describe OCRTextHygiene as the cleanup step, document BaseCompletionPromptRenderer and the default-off cotabbyConstrainedDecoderEnabled/runConstrainedDecode path; point the llama integration at the CotabbyInference SwiftPM package.
• docs/POLLING_AND_DELAYS.md: Drops the now-deleted LlamaVisualContextSummarizer 3 s timeout row.

Confidence Score: 5/5

Documentation-only change; no code or runtime behavior is modified.

Every factual claim in the updated docs was cross-checked against the live source: model filenames, sizes, and HuggingFace URLs match LlamaRuntimeModels.swift exactly; OCRTextHygiene, BaseCompletionPromptRenderer, runConstrainedDecode, and cotabbyConstrainedDecoderEnabled all exist in the codebase; LlamaVisualContextSummarizer and LlamaPromptRenderer are confirmed absent; the CotabbyInference SwiftPM package reference matches the Xcode project. No stale claims remain.

No files require special attention.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    subgraph Visual Context
        VC[VisualContextCoordinator] --> SCG[ScreenshotContextGenerator]
        SCG --> WSS[WindowScreenshotService]
        SCG --> STE[ScreenTextExtractor\nVision OCR]
        STE --> OCR[OCRTextHygiene\npure cleanup]
        OCR --> EXCERPT[Bounded Excerpt]
    end

    subgraph Runtime Generation
        SER[SuggestionEngineRouter] -->|Open Source| LSE[LlamaSuggestionEngine]
        SER -->|Apple Intelligence| FMSE[FoundationModelSuggestionEngine]
        LSE --> LRM[LlamaRuntimeManager]
        LRM --> LRC[LlamaRuntimeCore\nserialised actor]
        LRC -->|default| NORM[Normal Decode]
        LRC -->|cotabbyConstrainedDecoderEnabled default-off| CD[runConstrainedDecode]
    end

    subgraph Prompt Rendering
        LSE --> BCPR[BaseCompletionPromptRenderer\nbase-model continuation]
        FMSE --> FMPR[FoundationModelPromptRenderer\ninstruct-shaped]
    end

    subgraph Models tabby-2
        M1[tabby-2-mini Qwen3.5-0.8B ~0.8 GB]
        M2[tabby-2-base default Qwen3.5-2B ~1.4 GB]
        M3[tabby-2-pro Qwen3.5-4B ~2.6 GB]
        M4[tabby-2-gemma-mini gemma-4-E2B ~4.5 GB]
        M5[tabby-2-gemma-pro gemma-4-E4B ~5.0 GB]
    end

    LRC --> M1
    LRC --> M2
    LRC --> M3
    LRC --> M4
    LRC --> M5

_{Reviews (1): Last reviewed commit: "Refresh docs for the base-model migratio..." | Re-trigger Greptile}