feat: tree summarizer uses local AI exclusively (#518)

* refactor: update module visibility and enhance summarization logic

- Changed the visibility of the `ollama_api` module to `pub(crate)` to restrict its access to the current crate, improving encapsulation.
- Refactored the `run_summarization` and `propagate_node` functions to pass the chat model ID from the configuration, ensuring consistent usage of the model across summarization tasks.
- Enhanced the `summarize_to_limit` function to accept the model parameter, allowing for more flexible and dynamic summarization based on the specified model.
- Updated the `create_provider` function to enforce the requirement of enabling local AI for the tree summarizer, ensuring proper configuration before execution.

* fix: update Ollama provider credential handling

- Changed the credential parameter in the `create_local_ai_provider` function from `None` to `Some("ollama")`, clarifying that while Ollama does not require authentication, a non-None credential is necessary for the provider's configuration.

* feat: add tree summarization script for memory namespaces

- Introduced a new script `tree-summarizer-run-all.sh` that automates the process of running tree summarization across all memory namespaces.
- The script discovers namespaces from the specified workspace and supports various commands including `run`, `status`, `query`, and `rebuild`.
- Enhanced error handling and logging for improved usability and debugging during summarization tasks.

* style: apply cargo fmt to tree summarizer

* refactor: improve argument handling and output summary in tree summarizer script

- Refactored the argument construction in `tree-summarizer-run-all.sh` to use an array for better handling of command-line arguments.
- Enhanced the output summary to provide a clearer report of succeeded and failed namespace processing.
- Updated the while loop to read from `NAMESPACES` directly, improving readability and efficiency.

* fix: address PR review — array args, subshell loop, retry wrapper, narrow export

- Script: use bash array for args instead of string to prevent word-splitting
- Script: use here-string loop so SUCCEEDED/FAILED propagate to parent shell
- ops.rs: wrap Ollama provider in ReliableProvider for retry/backoff
- local_ai/mod.rs: keep ollama_api private, re-export only OLLAMA_BASE_URL

scripts/tree-summarizer-run-all.sh

@@ -0,0 +1,189 @@
+#!/usr/bin/env bash
+# tree-summarizer-run-all.sh — Run tree summarization for every memory namespace.
+#
+# Discovers namespaces by listing directories under the workspace's
+# memory/namespaces/ folder, then runs the tree-summarizer for each one.
+#
+# Usage:
+#   bash scripts/tree-summarizer-run-all.sh                 # run (drain buffer + summarize)
+#   bash scripts/tree-summarizer-run-all.sh status           # show status for all trees
+#   bash scripts/tree-summarizer-run-all.sh query [node_id]  # query all trees
+#   bash scripts/tree-summarizer-run-all.sh rebuild          # rebuild all trees from leaves
+#
+# Options:
+#   -v, --verbose    Enable debug logging
+#   --workspace DIR  Override OPENHUMAN_WORKSPACE
+#   --binary PATH    Override the openhuman-core binary path
+
+set -euo pipefail
+
+# ── Defaults ───────────────────────────────────────────────────────────
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+
+VERBOSE=""
+SUBCOMMAND="run"
+NODE_ID=""
+
+# Resolve binary: staged sidecar → debug build → release build
+resolve_binary() {
+    local arch
+    arch="$(uname -m)"
+    case "$arch" in
+        arm64|aarch64) arch="aarch64-apple-darwin" ;;
+        x86_64)        arch="x86_64-apple-darwin"  ;;
+        *)             arch="$arch-unknown-linux-gnu" ;;
+    esac
+
+    for bin in \
+        "$REPO_ROOT/app/src-tauri/binaries/openhuman-core-$arch" \
+        "$REPO_ROOT/target/debug/openhuman-core" \
+        "$REPO_ROOT/target/release/openhuman-core"; do
+        if [ -x "$bin" ]; then
+            echo "$bin"
+            return
+        fi
+    done
+
+    echo >&2 "error: could not find openhuman-core binary. Build with: cargo build --bin openhuman-core"
+    exit 1
+}
+
+OPENHUMAN_BIN="${OPENHUMAN_BIN:-$(resolve_binary)}"
+
+# Resolve workspace: env var → active user → first user dir
+resolve_workspace() {
+    if [ -n "${OPENHUMAN_WORKSPACE:-}" ]; then
+        echo "$OPENHUMAN_WORKSPACE"
+        return
+    fi
+
+    # Try the active user workspace
+    local active_user_file="$HOME/.openhuman/active_user.toml"
+    if [ -f "$active_user_file" ]; then
+        local user_id
+        user_id=$(sed -n 's/^user_id *= *"\([^"]*\)".*/\1/p' "$active_user_file" 2>/dev/null || true)
+        if [ -n "$user_id" ] && [ -d "$HOME/.openhuman/users/$user_id/workspace" ]; then
+            echo "$HOME/.openhuman/users/$user_id/workspace"
+            return
+        fi
+    fi
+
+    # Fallback: first user directory with a workspace
+    for user_dir in "$HOME"/.openhuman/users/*/; do
+        if [ -d "${user_dir}workspace" ]; then
+            echo "${user_dir}workspace"
+            return
+        fi
+    done
+
+    echo >&2 "error: could not resolve OPENHUMAN_WORKSPACE. Set it explicitly."
+    exit 1
+}
+
+export OPENHUMAN_WORKSPACE="${OPENHUMAN_WORKSPACE:-$(resolve_workspace)}"
+
+# ── Parse args ─────────────────────────────────────────────────────────
+
+while [ $# -gt 0 ]; do
+    case "$1" in
+        -v|--verbose)
+            VERBOSE="-v"
+            shift
+            ;;
+        --workspace)
+            export OPENHUMAN_WORKSPACE="$2"
+            shift 2
+            ;;
+        --binary)
+            OPENHUMAN_BIN="$2"
+            shift 2
+            ;;
+        run|status|query|rebuild)
+            SUBCOMMAND="$1"
+            shift
+            # For query, grab optional node_id
+            if [ "$SUBCOMMAND" = "query" ] && [ $# -gt 0 ]; then
+                case "$1" in
+                    -*) ;;  # skip flags
+                    *)  NODE_ID="$1"; shift ;;
+                esac
+            fi
+            ;;
+        -h|--help)
+            sed -n '2,/^$/{ s/^# //; s/^#$//; p }' "$0"
+            exit 0
+            ;;
+        *)
+            echo >&2 "unknown argument: $1"
+            exit 1
+            ;;
+    esac
+done
+
+# ── Discover namespaces ────────────────────────────────────────────────
+
+NAMESPACES_DIR="$OPENHUMAN_WORKSPACE/memory/namespaces"
+
+if [ ! -d "$NAMESPACES_DIR" ]; then
+    echo "No namespaces directory found at $NAMESPACES_DIR"
+    exit 0
+fi
+
+NAMESPACES=$(find "$NAMESPACES_DIR" -mindepth 1 -maxdepth 1 -type d | while read -r d; do basename "$d"; done | sort)
+
+if [ -z "$NAMESPACES" ]; then
+    echo "No memory namespaces found."
+    exit 0
+fi
+
+NS_COUNT=$(echo "$NAMESPACES" | wc -l | tr -d ' ')
+NS_LIST=$(echo "$NAMESPACES" | tr '\n' ' ')
+
+echo "Found $NS_COUNT namespace(s): $NS_LIST"
+echo "Workspace: $OPENHUMAN_WORKSPACE"
+echo "Binary:    $OPENHUMAN_BIN"
+echo "Command:   tree-summarizer $SUBCOMMAND"
+echo "---"
+
+# ── Strip ASCII art banner from output ─────────────────────────────────
+
+strip_banner() {
+    grep -v '▗\|▐\|▝\|▀\|█\|Contribute\|OpenHuman core' | grep -v '^[[:space:]]*$'
+}
+
+# ── Run for each namespace ─────────────────────────────────────────────
+
+FAILED=0
+SUCCEEDED=0
+
+while IFS= read -r ns; do
+    echo ""
+    echo "=== [$ns] ==="
+
+    args=("$SUBCOMMAND" "$ns")
+    if [ "$SUBCOMMAND" = "query" ] && [ -n "$NODE_ID" ]; then
+        args+=("$NODE_ID")
+    fi
+    if [ -n "$VERBOSE" ]; then
+        args+=("$VERBOSE")
+    fi
+
+    if output=$("$OPENHUMAN_BIN" tree-summarizer "${args[@]}" 2>&1); then
+        echo "$output" | strip_banner | head -40
+        SUCCEEDED=$((SUCCEEDED + 1))
+    else
+        echo "$output" | strip_banner | tail -5
+        echo "  ^^^ FAILED"
+        FAILED=$((FAILED + 1))
+    fi
+done <<< "$NAMESPACES"
+
+echo ""
+echo "---"
+echo "Done. $SUCCEEDED succeeded, $FAILED failed out of $NS_COUNT namespace(s)."
+
+if [ "$FAILED" -gt 0 ]; then
+    exit 1
+fi

src/openhuman/local_ai/mod.rs

@@ -15,6 +15,7 @@ pub mod sentiment;
 mod install;
 pub(crate) mod model_ids;
 mod ollama_api;
+pub(crate) use ollama_api::OLLAMA_BASE_URL;
 mod parse;
 pub(crate) mod paths;
 mod service;

src/openhuman/tree_summarizer/engine.rs

@@ -14,8 +14,6 @@ use crate::openhuman::tree_summarizer::types::{
     TreeStatus,
 };
 
-/// The model hint passed to the provider for summarization tasks.
-const SUMMARIZATION_MODEL: &str = "hint:fast";
 const SUMMARIZATION_TEMP: f64 = 0.3;
 
 /// Maximum characters for a summary response (hard limit enforced after LLM call).
@@ -81,12 +79,14 @@ pub async fn run_summarization(
             combined
         };
 
+        let model = &config.local_ai.chat_model_id;
         let hour_summary = summarize_to_limit(
             provider,
             &to_summarize,
             NodeLevel::Hour.max_tokens(),
             "hour",
             hour_id,
+            model,
         )
         .await
         .context("summarize hour leaf")?;
@@ -138,9 +138,16 @@ pub async fn run_summarization(
     ] {
         for (node_id, node_level) in &all_propagation_ids {
             if *node_level == level && seen.insert(node_id.clone()) {
-                propagate_node(config, provider, namespace, node_id, level)
-                    .await
-                    .with_context(|| format!("propagate {node_id}"))?;
+                propagate_node(
+                    config,
+                    provider,
+                    namespace,
+                    node_id,
+                    level,
+                    &config.local_ai.chat_model_id,
+                )
+                .await
+                .with_context(|| format!("propagate {node_id}"))?;
             }
         }
     }
@@ -233,16 +240,25 @@ pub async fn rebuild_tree(
     }
 
     // Propagate bottom-up: days, then months, then years, then root
+    let model = &config.local_ai.chat_model_id;
     for day_id in &day_ids {
-        propagate_node(config, provider, namespace, day_id, NodeLevel::Day).await?;
+        propagate_node(config, provider, namespace, day_id, NodeLevel::Day, model).await?;
     }
     for month_id in &month_ids {
-        propagate_node(config, provider, namespace, month_id, NodeLevel::Month).await?;
+        propagate_node(
+            config,
+            provider,
+            namespace,
+            month_id,
+            NodeLevel::Month,
+            model,
+        )
+        .await?;
     }
     for year_id in &year_ids {
-        propagate_node(config, provider, namespace, year_id, NodeLevel::Year).await?;
+        propagate_node(config, provider, namespace, year_id, NodeLevel::Year, model).await?;
     }
-    propagate_node(config, provider, namespace, "root", NodeLevel::Root).await?;
+    propagate_node(config, provider, namespace, "root", NodeLevel::Root, model).await?;
 
     let final_status = store::get_tree_status(config, namespace)?;
 
@@ -268,6 +284,7 @@ async fn propagate_node(
     namespace: &str,
     node_id: &str,
     level: NodeLevel,
+    model: &str,
 ) -> Result<()> {
     let children = store::read_children(config, namespace, node_id)?;
     if children.is_empty() {
@@ -305,7 +322,15 @@ async fn propagate_node(
             combined_tokens,
             max_tokens
         );
-        summarize_to_limit(provider, &combined, max_tokens, level.as_str(), node_id).await?
+        summarize_to_limit(
+            provider,
+            &combined,
+            max_tokens,
+            level.as_str(),
+            node_id,
+            model,
+        )
+        .await?
     };
 
     let now = Utc::now();
@@ -351,6 +376,7 @@ async fn summarize_to_limit(
     max_tokens: u32,
     level_name: &str,
     node_id: &str,
+    model: &str,
 ) -> Result<String> {
     let max_chars = (max_tokens as usize) * 4;
     let system_prompt = format!(
@@ -365,12 +391,7 @@ async fn summarize_to_limit(
     );
 
     let response = provider
-        .chat_with_system(
-            Some(&system_prompt),
-            content,
-            SUMMARIZATION_MODEL,
-            SUMMARIZATION_TEMP,
-        )
+        .chat_with_system(Some(&system_prompt), content, model, SUMMARIZATION_TEMP)
         .await
         .with_context(|| {
             format!("LLM summarization failed for node {node_id} (level={level_name})")

src/openhuman/tree_summarizer/ops.rs

@@ -4,7 +4,6 @@ use chrono::{DateTime, Utc};
 use serde_json::{json, Value};
 
 use crate::openhuman::config::Config;
-use crate::openhuman::providers;
 use crate::openhuman::tree_summarizer::{engine, store, types::*};
 use crate::rpc::RpcOutcome;
 
@@ -148,10 +147,46 @@ pub async fn tree_summarizer_rebuild(
 fn create_provider(
     config: &Config,
 ) -> Result<Box<dyn crate::openhuman::providers::traits::Provider>, String> {
-    providers::create_resilient_provider(
-        config.api_key.as_deref(),
-        config.api_url.as_deref(),
-        &config.reliability,
-    )
-    .map_err(|e| format!("failed to create provider: {e}"))
+    // Tree summarization runs exclusively on local AI to keep memory
+    // processing private and offline — no backend calls.
+    if !config.local_ai.enabled {
+        return Err("tree summarizer requires local_ai to be enabled in config".to_string());
+    }
+    create_local_ai_provider(config)
+}
+
+/// Create a provider backed by the local Ollama instance for summarization,
+/// wrapped in `ReliableProvider` for retry/backoff on transient failures.
+fn create_local_ai_provider(
+    config: &Config,
+) -> Result<Box<dyn crate::openhuman::providers::traits::Provider>, String> {
+    use crate::openhuman::local_ai::OLLAMA_BASE_URL;
+    use crate::openhuman::providers::compatible::{AuthStyle, OpenAiCompatibleProvider};
+    use crate::openhuman::providers::reliable::ReliableProvider;
+
+    let base_url = format!("{}/v1", OLLAMA_BASE_URL);
+    let inner = OpenAiCompatibleProvider::new_no_responses_fallback(
+        "ollama-local",
+        &base_url,
+        Some("ollama"), // Ollama ignores auth but the provider requires a non-None credential
+        AuthStyle::Bearer,
+    );
+
+    let providers: Vec<(
+        String,
+        Box<dyn crate::openhuman::providers::traits::Provider>,
+    )> = vec![("ollama-local".to_string(), Box::new(inner))];
+    let reliable = ReliableProvider::new(
+        providers,
+        config.reliability.provider_retries,
+        config.reliability.provider_backoff_ms,
+    );
+
+    tracing::debug!(
+        "[tree_summarizer] using local Ollama provider at {} with model '{}'",
+        base_url,
+        config.local_ai.chat_model_id
+    );
+
+    Ok(Box::new(reliable))
 }