diff --git a/README.md b/README.md
index fc62bbf..a0b7aaf 100644
--- a/README.md
+++ b/README.md
@@ -21,7 +21,7 @@ uv sync --extra dev
 uv run pytest
 ```
 
-## Usage (Phases 0–2)
+## Usage (Phases 0–3)
 
 ```bash
 uv run gitmate --help
@@ -32,12 +32,16 @@ uv run gitmate config set ignore_globs '*.gen.py,*.tmp'  # comma-separated, empt
 uv run gitmate config set max_context_tokens 8000        # override model context limit
 uv run gitmate config set reserved_output_tokens 1500    # reserve headroom for model reply
 uv run gitmate config set template_overhead 600          # reserve headroom for prompt template
+uv run gitmate config set cache_dir /path/to/cache       # custom prompt cache directory
 uv run gitmate debug-diff             # staged diff: per-file table + cleaned patch
 uv run gitmate debug-diff --summary   # per-file stats table only, no patch text
 uv run gitmate debug-diff --base main # branch-vs-main diff (PR path)
 ```
 
-`commit`, `pr-summary`, `changelog`, and `doc` are stubbed until their phases
+Phase 3 introduces the `LLMProvider` abstraction, backed by `GeminiProvider`
+(targeting `gemini-3.5-flash-lite` with tenacity exponential backoff retries)
+and wrapped by `CachedProvider` (hash-keyed disk caching via `diskcache`).
+`commit`, `pr-summary`, `changelog`, and `doc` remain stubbed until their phases
 land — see `AGENTS.md` and the phased roadmap for details.
 
 ## Diagrams
@@ -48,6 +52,6 @@ files in the draw.io desktop app to edit):
 | Diagram | Source | Export |
 |---|---|---|
 | System architecture (Phases 0–8, gemini-3.5-flash-lite) | [`docs/architecture.drawio`](docs/architecture.drawio) | [`docs/architecture.png`](docs/architecture.png) |
-| Class diagram (Phases 0–2: config, diff extraction, token budgeting) | [`docs/class-diagram.drawio`](docs/class-diagram.drawio) | [`docs/class-diagram.png`](docs/class-diagram.png) |
+| Class diagram (Phases 0–3: config, diff extraction, token budgeting, providers & caching) | [`docs/class-diagram.drawio`](docs/class-diagram.drawio) | [`docs/class-diagram.png`](docs/class-diagram.png) |
 | Use cases (implemented + planned) | [`docs/use-case.drawio`](docs/use-case.drawio) | [`docs/use-case.png`](docs/use-case.png) |
 | `config` command sequences | [`docs/sequence.drawio`](docs/sequence.drawio) | [`docs/sequence.png`](docs/sequence.png) |
diff --git a/docs/class-diagram.drawio b/docs/class-diagram.drawio
index 7812b43..c0498e4 100644
--- a/docs/class-diagram.drawio
+++ b/docs/class-diagram.drawio
@@ -1,16 +1,16 @@
 <?xml version="1.0" encoding="UTF-8"?>
 <mxfile host="drawio" version="26.0.0">
   <diagram name="class-diagram">
-    <mxGraphModel dx="1200" dy="800" grid="1" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="1500" pageHeight="1020" math="0" shadow="0">
+    <mxGraphModel dx="1200" dy="900" grid="1" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="1500" pageHeight="1150" math="0" shadow="0">
       <root>
         <mxCell id="0"/>
         <mxCell id="1" parent="0"/>
 
         <!-- Header -->
-        <mxCell id="title" value="gitmate — Architecture &amp; Class Diagram (Phases 0–2)" style="text;html=1;align=left;verticalAlign=middle;fontStyle=1;fontSize=22;" vertex="1" parent="1">
+        <mxCell id="title" value="gitmate — Architecture &amp; Class Diagram (Phases 0–3)" style="text;html=1;align=left;verticalAlign=middle;fontStyle=1;fontSize=22;" vertex="1" parent="1">
           <mxGeometry x="40" y="20" width="800" height="34" as="geometry"/>
         </mxCell>
-        <mxCell id="subtitle" value="config.py + diff_extractor.py + token_budget.py + cli.py · Strict typing, I/O behind protocols, offline-first verification" style="text;html=1;align=left;verticalAlign=middle;fontSize=12;fontColor=#555555;" vertex="1" parent="1">
+        <mxCell id="subtitle" value="config.py + diff_extractor.py + token_budget.py + providers/ + cli.py · Strict typing, I/O behind protocols, offline-first verification" style="text;html=1;align=left;verticalAlign=middle;fontSize=12;fontColor=#555555;" vertex="1" parent="1">
           <mxGeometry x="40" y="54" width="1100" height="22" as="geometry"/>
         </mxCell>
 
@@ -18,8 +18,8 @@
         <mxCell id="cfg_t" value="&lt;b&gt;GitmateConfig&lt;/b&gt;&lt;br/&gt;«dataclass»" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=13;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
           <mxGeometry x="50" y="90" width="260" height="40" as="geometry"/>
         </mxCell>
-        <mxCell id="cfg_b" value="+ model: str = DEFAULT_MODEL&lt;br/&gt;+ commit_style: str = &quot;conventional&quot;&lt;br/&gt;+ budget_cap_usd: float | None = None&lt;br/&gt;+ ignore_globs: list[str]&lt;br/&gt;+ max_context_tokens: int | None = None&lt;br/&gt;+ reserved_output_tokens: int = 2048&lt;br/&gt;+ template_overhead: int = 500" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=8;spacingTop=6;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
-          <mxGeometry x="50" y="130" width="260" height="120" as="geometry"/>
+        <mxCell id="cfg_b" value="+ model: str = DEFAULT_MODEL&lt;br/&gt;+ commit_style: str = &quot;conventional&quot;&lt;br/&gt;+ budget_cap_usd: float | None = None&lt;br/&gt;+ ignore_globs: list[str]&lt;br/&gt;+ max_context_tokens: int | None = None&lt;br/&gt;+ reserved_output_tokens: int = 2048&lt;br/&gt;+ template_overhead: int = 500&lt;br/&gt;+ cache_dir: str | None = None" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=8;spacingTop=6;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="50" y="130" width="260" height="135" as="geometry"/>
         </mxCell>
 
         <mxCell id="ss_t" value="&lt;b&gt;SecretStore&lt;/b&gt;&lt;br/&gt;«protocol»" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=13;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1">
@@ -83,8 +83,8 @@
         <mxCell id="tbm_t" value="&lt;b&gt;TokenBudgetManager&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=13;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
           <mxGeometry x="50" y="520" width="260" height="35" as="geometry"/>
         </mxCell>
-        <mxCell id="tbm_b" value="+ counter: TokenCounter&lt;br/&gt;+ model: str = DEFAULT_MODEL&lt;br/&gt;+ context_window: int = 1_048_576&lt;br/&gt;+ reserved_output_tokens: int = 2048&lt;br/&gt;+ default_template_overhead: int = 500&lt;br/&gt;+ count_diff_tokens(diffs) -&amp;gt; int&lt;br/&gt;+ assess(diffs, overhead) -&amp;gt; BudgetDecision" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=8;spacingTop=6;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
-          <mxGeometry x="50" y="555" width="260" height="130" as="geometry"/>
+        <mxCell id="tbm_b" value="+ counter: TokenCounter&lt;br/&gt;+ model: str = DEFAULT_MODEL&lt;br/&gt;+ context_window: int = 1_048_576&lt;br/&gt;+ reserved_output_tokens: int = 2048&lt;br/&gt;+ default_template_overhead: int = 500&lt;br/&gt;+ count_diff_tokens(diffs) -&amp;gt; int&lt;br/&gt;+ assess(diffs, overhead) -&amp;gt; BudgetDecision&lt;br/&gt;+ from_config(cfg, counter) -&amp;gt; TokenBudgetManager" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=8;spacingTop=6;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="50" y="555" width="260" height="145" as="geometry"/>
         </mxCell>
 
         <mxCell id="tc_t" value="&lt;b&gt;TokenCounter&lt;/b&gt;&lt;br/&gt;«protocol»" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=13;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1">
@@ -122,17 +122,67 @@
           <mxGeometry x="930" y="680" width="180" height="35" as="geometry"/>
         </mxCell>
 
+        <!-- Phase 3: Provider Abstraction & Cache Layer -->
+        <mxCell id="llmp_t" value="&lt;b&gt;LLMProvider&lt;/b&gt;&lt;br/&gt;«protocol»" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=13;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1">
+          <mxGeometry x="50" y="740" width="260" height="35" as="geometry"/>
+        </mxCell>
+        <mxCell id="llmp_b" value="+ generate(prompt: str, model: str) -&amp;gt; LLMResponse" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=8;spacingTop=6;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1">
+          <mxGeometry x="50" y="775" width="260" height="40" as="geometry"/>
+        </mxCell>
+
+        <mxCell id="gem_t" value="&lt;b&gt;GeminiProvider&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=12;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="50" y="850" width="260" height="30" as="geometry"/>
+        </mxCell>
+        <mxCell id="gem_b" value="- _client: genai.Client | None&lt;br/&gt;- _api_key: str | None&lt;br/&gt;+ generate(prompt, model) -&amp;gt; LLMResponse&lt;br/&gt;- _generate_with_retry() via tenacity&lt;br/&gt;- _is_transient(exc) -&amp;gt; bool" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=8;spacingTop=4;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="50" y="880" width="260" height="85" as="geometry"/>
+        </mxCell>
+
+        <mxCell id="cp_t" value="&lt;b&gt;CachedProvider&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=12;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="360" y="740" width="250" height="30" as="geometry"/>
+        </mxCell>
+        <mxCell id="cp_b" value="- _provider: LLMProvider&lt;br/&gt;- _cache: Cache (diskcache)&lt;br/&gt;- _ttl: float | None&lt;br/&gt;- _template_version: str = &quot;1&quot;&lt;br/&gt;+ generate(prompt, model) -&amp;gt; LLMResponse&lt;br/&gt;+ close() -&amp;gt; None" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=8;spacingTop=4;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="360" y="770" width="250" height="100" as="geometry"/>
+        </mxCell>
+
+        <mxCell id="ck_t" value="&lt;b&gt;cache_key()&lt;/b&gt;" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=11;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="360" y="900" width="250" height="25" as="geometry"/>
+        </mxCell>
+        <mxCell id="ck_b" value="sha256(version + \x00 + model + \x00 + prompt)" style="rounded=0;whiteSpace=wrap;html=1;align=center;verticalAlign=middle;fontSize=10;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="360" y="925" width="250" height="30" as="geometry"/>
+        </mxCell>
+
+        <mxCell id="llmr_t" value="&lt;b&gt;LLMResponse&lt;/b&gt;&lt;br/&gt;«dataclass»" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=13;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="660" y="740" width="230" height="35" as="geometry"/>
+        </mxCell>
+        <mxCell id="llmr_b" value="+ text: str&lt;br/&gt;+ model: str&lt;br/&gt;+ input_tokens: int | None = None&lt;br/&gt;+ output_tokens: int | None = None" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=8;spacingTop=6;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+          <mxGeometry x="660" y="775" width="230" height="75" as="geometry"/>
+        </mxCell>
+
+        <mxCell id="pu_t" value="&lt;b&gt;ProviderUnavailable&lt;/b&gt;&lt;br/&gt;«extends Exception»" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=12;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
+          <mxGeometry x="930" y="740" width="190" height="30" as="geometry"/>
+        </mxCell>
+        <mxCell id="pu_b" value="raised when backend fails retries" style="rounded=0;whiteSpace=wrap;html=1;align=center;verticalAlign=middle;fontSize=10;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
+          <mxGeometry x="930" y="770" width="190" height="25" as="geometry"/>
+        </mxCell>
+
+        <mxCell id="rpe_t" value="&lt;b&gt;RetryableProviderError&lt;/b&gt;&lt;br/&gt;«extends Exception»" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=12;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
+          <mxGeometry x="930" y="820" width="190" height="30" as="geometry"/>
+        </mxCell>
+        <mxCell id="rpe_b" value="internal transient error marker" style="rounded=0;whiteSpace=wrap;html=1;align=center;verticalAlign=middle;fontSize=10;fillColor=#f8cecc;strokeColor=#b85450;" vertex="1" parent="1">
+          <mxGeometry x="930" y="850" width="190" height="25" as="geometry"/>
+        </mxCell>
+
         <!-- CLI Box -->
         <mxCell id="cli_t" value="&lt;b&gt;gitmate CLI&lt;/b&gt;&lt;br/&gt;Typer app · commands only, no business logic" style="rounded=1;whiteSpace=wrap;html=1;align=center;fontSize=13;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
           <mxGeometry x="1100" y="160" width="340" height="40" as="geometry"/>
         </mxCell>
-        <mxCell id="cli_b" value="&lt;b&gt;Commands implemented:&lt;/b&gt;&lt;br/&gt;• config set-key()&lt;br/&gt;• config show()&lt;br/&gt;• config set(field, value)&lt;br/&gt;• debug-diff(--base, --from-ref, --to-ref, --summary)&lt;br/&gt;&lt;br/&gt;&lt;b&gt;Stubs for future phases:&lt;/b&gt;&lt;br/&gt;• commit() (Phase 5)&lt;br/&gt;• pr-summary() (Phase 7)&lt;br/&gt;• changelog() (Phase 7)&lt;br/&gt;• doc()" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=10;spacingTop=6;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
+        <mxCell id="cli_b" value="&lt;b&gt;Commands implemented:&lt;/b&gt;&lt;br/&gt;• config set-key()&lt;br/&gt;• config show()&lt;br/&gt;• config set(field, value) [incl. cache_dir]&lt;br/&gt;• debug-diff(--base, --from-ref, --to-ref, --summary)&lt;br/&gt;&lt;br/&gt;&lt;b&gt;Stubs for future phases:&lt;/b&gt;&lt;br/&gt;• commit() (Phase 5)&lt;br/&gt;• pr-summary() (Phase 7)&lt;br/&gt;• changelog() (Phase 7)&lt;br/&gt;• doc()" style="rounded=0;whiteSpace=wrap;html=1;align=left;verticalAlign=top;fontSize=11;spacingLeft=10;spacingTop=6;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1">
           <mxGeometry x="1100" y="200" width="340" height="190" as="geometry"/>
         </mxCell>
 
         <!-- Legend -->
         <mxCell id="legend" value="Legend" style="rounded=0;whiteSpace=wrap;html=1;fillColor=none;strokeColor=#666666;verticalAlign=top;fontStyle=1;fontSize=13;" vertex="1" parent="1">
-          <mxGeometry x="50" y="760" width="1060" height="60" as="geometry"/>
+          <mxGeometry x="50" y="1010" width="1060" height="60" as="geometry"/>
         </mxCell>
         <mxCell id="leg1" value="" style="rounded=0;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="legend">
           <mxGeometry x="20" y="30" width="30" height="16" as="geometry"/>
@@ -181,7 +231,7 @@
         <mxCell id="e_tbm_bd" value="produces" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;fontSize=11;endArrow=open;endFill=0;dashed=1;exitX=1;exitY=0.8;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="tbm_b" target="bd_b">
           <mxGeometry relative="1" as="geometry">
             <Array as="points">
-              <mxPoint x="340" y="659"/>
+              <mxPoint x="340" y="671"/>
               <mxPoint x="340" y="630"/>
             </Array>
           </mxGeometry>
@@ -193,6 +243,27 @@
           <mxGeometry relative="1" as="geometry"/>
         </mxCell>
 
+        <!-- Phase 3 Connectors -->
+        <mxCell id="e_gem_llmp" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;endArrow=block;endFill=0;dashed=1;exitX=0.5;exitY=0;exitDx=0;exitDy=0;entryX=0.5;entryY=1;entryDx=0;entryDy=0;" edge="1" parent="1" source="gem_t" target="llmp_b">
+          <mxGeometry relative="1" as="geometry"/>
+        </mxCell>
+        <mxCell id="e_cp_llmp" value="wraps" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;fontSize=11;endArrow=open;endFill=0;dashed=1;exitX=0;exitY=0.5;exitDx=0;exitDy=0;entryX=1;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="cp_t" target="llmp_t">
+          <mxGeometry relative="1" as="geometry"/>
+        </mxCell>
+        <mxCell id="e_llmp_llmr" value="produces" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;fontSize=11;endArrow=open;endFill=0;dashed=1;exitX=1;exitY=0.5;exitDx=0;exitDy=0;entryX=0;entryY=0.5;entryDx=0;entryDy=0;" edge="1" parent="1" source="llmp_b" target="llmr_t">
+          <mxGeometry relative="1" as="geometry">
+            <Array as="points">
+              <mxPoint x="330" y="795"/>
+              <mxPoint x="330" y="725"/>
+              <mxPoint x="640" y="725"/>
+              <mxPoint x="640" y="757"/>
+            </Array>
+          </mxGeometry>
+        </mxCell>
+        <mxCell id="e_ck_cp" value="" style="edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1;endArrow=open;endFill=0;dashed=1;exitX=0.5;exitY=0;exitDx=0;exitDy=0;entryX=0.5;entryY=1;entryDx=0;entryDy=0;" edge="1" parent="1" source="ck_t" target="cp_b">
+          <mxGeometry relative="1" as="geometry"/>
+        </mxCell>
+
       </root>
     </mxGraphModel>
   </diagram>
diff --git a/docs/class-diagram.png b/docs/class-diagram.png
index 1ff7799..64a5968 100644
Binary files a/docs/class-diagram.png and b/docs/class-diagram.png differ