feat: add icm wake-up command + icm_wake_up MCP tool (#84)

* feat: add `icm wake-up` command + `icm_wake_up` MCP tool

Adds a compact critical-facts pack generator for LLM system-prompt
injection, inspired by MemPalace's wake-up command but grounded in ICM's
importance/decay model.

The command selects critical/high memories (plus global preferences)
optionally scoped by project, ranks by importance × recency × weight,
categorizes into Identity / Decisions / Constraints / Errors / Milestones
/ Context sections, and truncates to fit a token budget.

- `icm-core`: new `wake_up` module with pure, testable `build_wake_up()`
  and `build_wake_up_from_memories()` APIs. 14 unit tests cover filtering,
  project scoping, budget truncation, formatting, and categorization.
- `icm-cli`: new `wake-up` subcommand with `--project`, `--max-tokens`,
  `--format markdown|plain`, `--no-preferences`. Auto-detects project
  from PWD/git remote; use `--project -` to disable.
- `icm-mcp`: new `icm_wake_up` tool with the same parameters, clamped
  token budget [20, 4000]. 7 MCP integration tests cover empty store,
  filtering, project scoping, plain format, budget clamping, preference
  opt-out, and tool listing.

Total: 21 new tests, all green. Clippy clean.

* refactor(wake-up): address review feedback

- project_matches is now segment-aware: the topic and the project filter
  are both split on `-./_:/` and every project segment must appear as a
  complete topic segment. Rejects false positives like "icm" matching
  "icmp-notes" while still allowing "icm" -> "decisions-icm-core" and
  "icm-core" -> "decisions-icm-core".
- compute_score now uses max(created_at, last_accessed) so frequently
  recalled memories stay fresh under decay.
- truncate_by_budget and approx_tokens now count chars (not bytes) so
  multibyte summaries are not over-penalized and the token count printed
  in the header is honest for non-ASCII.
- sanitize_summary flattens newlines and collapses runs of spaces, so a
  user-stored summary containing "\n## Fake header" cannot break the
  rendered section structure.
- render_body trailing-newline loop fixed (was "\n\n\n" check that never
  triggered — now "\n\n").
- cmd_wake_up prints "Project: <detected>" to stderr when auto-detecting,
  so users understand why specific topics appear or not.
- tool_wake_up accepts "" and "-" as project-disable sentinels, matching
  the CLI.
- build_wake_up_from_memories marked #[must_use].
- +6 new unit tests covering segment matching, include_preferences=false,
  multiline sanitization, multibyte budget, access-aware recency, and
  plain format header suppression (20 core tests total).

Author: pszymkowiak

crates/icm-cli/src/main.rs

@@ -17,8 +17,9 @@ use clap::{Parser, Subcommand, ValueEnum};
 use serde_json::Value;
 
 use icm_core::{
-    keyword_matches, topic_matches, Concept, ConceptLink, Feedback, FeedbackStore, Importance,
-    Label, Memoir, MemoirStore, Memory, MemoryStore, Relation, MSG_NO_MEMORIES,
+    build_wake_up, keyword_matches, topic_matches, Concept, ConceptLink, Feedback, FeedbackStore,
+    Importance, Label, Memoir, MemoirStore, Memory, MemoryStore, Relation, WakeUpFormat,
+    WakeUpOptions, MSG_NO_MEMORIES,
 };
 use icm_store::SqliteStore;
 
@@ -279,6 +280,29 @@ enum Commands {
         limit: usize,
     },
 
+    /// Print a compact critical-facts pack for LLM system-prompt injection
+    ///
+    /// Selects critical/high memories (and preferences) optionally scoped by
+    /// project, ranks them by importance × recency × weight, then truncates
+    /// to fit the token budget. Inspired by MemPalace's `wake-up` command.
+    WakeUp {
+        /// Project filter (default: auto-detect from PWD/git remote; use "-" to disable)
+        #[arg(short, long)]
+        project: Option<String>,
+
+        /// Approximate token budget (1 token ≈ 4 characters)
+        #[arg(short = 't', long, default_value = "200")]
+        max_tokens: usize,
+
+        /// Output format
+        #[arg(short, long, default_value = "markdown")]
+        format: CliWakeUpFormat,
+
+        /// Exclude global preferences/identity memories
+        #[arg(long)]
+        no_preferences: bool,
+    },
+
     /// Auto-save context for the current project (detects from PWD / git remote)
     SaveProject {
         /// Summary of what was done in this session
@@ -633,6 +657,21 @@ impl From<CliImportance> for Importance {
     }
 }
 
+#[derive(Clone, Copy, ValueEnum)]
+enum CliWakeUpFormat {
+    Markdown,
+    Plain,
+}
+
+impl From<CliWakeUpFormat> for WakeUpFormat {
+    fn from(val: CliWakeUpFormat) -> Self {
+        match val {
+            CliWakeUpFormat::Markdown => WakeUpFormat::Markdown,
+            CliWakeUpFormat::Plain => WakeUpFormat::Plain,
+        }
+    }
+}
+
 #[derive(Clone, ValueEnum)]
 enum CliImportFormat {
     Auto,
@@ -923,6 +962,12 @@ fn main() -> Result<()> {
         }
         Commands::RecallContext { query, limit } => cmd_recall_context(&store, &query, limit),
         Commands::RecallProject { limit } => cmd_recall_project(&store, limit),
+        Commands::WakeUp {
+            project,
+            max_tokens,
+            format,
+            no_preferences,
+        } => cmd_wake_up(&store, project, max_tokens, format, no_preferences),
         Commands::SaveProject {
             content,
             importance,
@@ -2578,6 +2623,48 @@ fn cmd_recall_project(store: &SqliteStore, limit: usize) -> Result<()> {
     Ok(())
 }
 
+/// Build and print a wake-up pack for LLM system-prompt injection.
+///
+/// Selects critical/high memories (plus preferences) optionally scoped to a
+/// project, ranks by importance × recency × weight, and truncates to fit the
+/// token budget.
+fn cmd_wake_up(
+    store: &SqliteStore,
+    project: Option<String>,
+    max_tokens: usize,
+    format: CliWakeUpFormat,
+    no_preferences: bool,
+) -> Result<()> {
+    // Resolve project: explicit "-" disables, None auto-detects, Some(name) uses it.
+    let detected;
+    let project_ref: Option<&str> = match project.as_deref() {
+        Some("-") => None,
+        Some(p) => Some(p),
+        None => {
+            detected = detect_project();
+            if detected.is_empty() || detected == "unknown" {
+                None
+            } else {
+                // Make auto-detection visible so users understand why
+                // specific topics show (or don't).
+                eprintln!("Project: {detected} (auto-detected; use --project - to disable)");
+                Some(detected.as_str())
+            }
+        }
+    };
+
+    let opts = WakeUpOptions {
+        project: project_ref,
+        max_tokens,
+        format: format.into(),
+        include_preferences: !no_preferences,
+    };
+
+    let pack = build_wake_up(store, &opts)?;
+    print!("{pack}");
+    Ok(())
+}
+
 fn cmd_save_project(
     store: &SqliteStore,
     embedder: Option<&dyn icm_core::Embedder>,

crates/icm-core/src/lib.rs

@@ -9,6 +9,7 @@ pub mod memoir;
 pub mod memoir_store;
 pub mod memory;
 pub mod store;
+pub mod wake_up;
 
 /// Default embedding vector dimensions (used when no embedder is configured).
 pub const DEFAULT_EMBEDDING_DIMS: usize = 384;
@@ -25,6 +26,7 @@ pub use memory::{
     Importance, Memory, MemorySource, PatternCluster, Scope, StoreStats, TopicHealth,
 };
 pub use store::MemoryStore;
+pub use wake_up::{build_wake_up, build_wake_up_from_memories, WakeUpFormat, WakeUpOptions};
 
 pub use learn::{learn_project, LearnResult};