fix(context): batasi squash_log ke tool bash saja

Ditemukan reviewer whole-branch final: looks_log_shaped murni berbasis
konten (>=3 baris berpola error/warn/fail), jadi hasil grep/search yang
match ke kode error-handling ikut lolos ambang itu -- padahal
squash_log punya cap keras 20 error + 10 warning tanpa budget byte,
diam-diam membuang match yang sah di luar cap itu. Sekarang hanya tool
bash (penghasil log sungguhan) yang boleh lewat squash_log; tool lain
yang kebetulan konten-nya mirip log jatuh ke squash_generic yang lebih
longgar (head/tail + budget byte). Tambah test regresi yang membedakan
kedua jalur lewat retensi baris terakhir.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

docs/superpowers/specs/2026-07-16-context-compaction-overhaul-design.md

@@ -167,7 +167,14 @@ shape:
   lines, up to 10 highest-scored warning lines, all summary lines, plus a ±2-line context
   window around each kept line → single `[N lines omitted]` marker for drops (not
   comment-shaped, per rtk's own finding on LLM confusion). Applied when the output isn't
-  valid JSON and has ≥3 lines matching error/warn/stack-trace patterns.
+  valid JSON, the tool is `bash`, and the output has ≥3 lines matching error/warn/stack-trace
+  patterns. The tool restriction (added after the final whole-branch review) matters: a `grep`
+  result full of matches against error-handling code trips the same ≥3-line keyword threshold
+  as a real build log, but `squash_log`'s hard 20-error/10-warning cap has no byte budget and
+  would silently drop legitimate matches past it — the wrong compressor for search results.
+  Only `bash` (the actual log-producing tool) routes through `squash_log`; every other tool
+  whose output happens to look log-shaped falls through to the gentler, byte-budgeted
+  `squash_generic` instead.
 - `squash_generic(&str, budget) -> String` — importance-ranked truncation: keeps the first 10
   and last 10 lines plus any line matching a small "looks important" heuristic (non-blank,
   not a byte-for-byte repeat of the immediately preceding line), single `[N lines omitted]`

src/app/runtime/context/squash.rs

@@ -3,8 +3,8 @@
 //! ever enter conversation history, dispatching by content shape.
 //!
 //! Flow: `apply(tool_name, output)` -> `read` tool or under the size
-//! floor? pass through unchanged : valid JSON? `squash_json` : looks
-//! log-shaped? `squash_log` : `squash_generic`.
+//! floor? pass through unchanged : valid JSON? `squash_json` : tool is
+//! `bash` and looks log-shaped? `squash_log` : `squash_generic`.
 //!
 //! Why: a single large `bash`/`grep` result can dominate a
 //! conversation's token budget even on its first occurrence, long
@@ -30,6 +30,17 @@ const GENERIC_BUDGET_BYTES: usize = SQUASH_FLOOR_BYTES * 2;
 /// agent's view of real file content.
 const NEVER_SQUASH: &[&str] = &["read"];
 
+/// Tools whose output the log classifier is allowed to run on.
+/// `looks_log_shaped` keys purely on content (>=3 error/warn/fail-shaped
+/// lines), which a `grep`/`search` result full of matches against
+/// error-handling code would trip just as easily as a real build log —
+/// but `squash_log` caps at 20 error + 10 warning lines with no byte
+/// budget, silently dropping legitimate matches past that cap. Only
+/// `bash` (the actual log-producing tool) is allowed to route through
+/// it; everything else that looks log-shaped falls through to the
+/// gentler, byte-budgeted `squash_generic` instead.
+const LOG_SHAPED_TOOLS: &[&str] = &["bash"];
+
 /// Compress a tool's raw output before it's stored in conversation
 /// history.
 ///
@@ -43,7 +54,7 @@ pub fn apply(tool_name: &str, output: &str) -> String {
     if serde_json::from_str::<serde_json::Value>(output).is_ok() {
         return squash_json(output);
     }
-    if looks_log_shaped(output) {
+    if LOG_SHAPED_TOOLS.contains(&tool_name) && looks_log_shaped(output) {
         return squash_log(output);
     }
     squash_generic(output, GENERIC_BUDGET_BYTES)
@@ -379,6 +390,32 @@ mod tests {
         assert!(result.len() < text.len());
     }
 
+    #[test]
+    fn non_bash_tool_with_log_shaped_content_is_not_log_compressed() {
+        // A grep result whose matched lines all mention "error" would
+        // trip `looks_log_shaped`'s >=3-line keyword threshold just like
+        // a real build log — but `squash_log` caps at 20 highest-scored
+        // error lines with no guaranteed tail retention, silently
+        // dropping legitimate matches past that cap. Only `bash` is
+        // treated as log-shaped; `grep` must fall through to
+        // `squash_generic`, which always keeps the first and last 10
+        // lines regardless of score. With every line tied at the same
+        // score, a `squash_log` route would keep indices 0-19 (stable
+        // sort preserves original order on ties) and drop index 49 —
+        // so asserting the tail survives is a route-distinguishing
+        // check, not just a content check.
+        let lines: Vec<String> = (0..50)
+            .map(|i| format!("src/file{i}.rs:{i}: error handling for case {i}"))
+            .collect();
+        let text = lines.join("\n");
+        assert!(text.len() > SQUASH_FLOOR_BYTES);
+
+        let result = apply("grep", &text);
+
+        assert!(result.contains("src/file0.rs:0: error handling for case 0"), "generic keeps head");
+        assert!(result.contains("src/file49.rs:49: error handling for case 49"), "generic keeps tail — squash_log would have dropped this");
+    }
+
     #[test]
     fn generic_large_text_is_truncated_with_omission_marker() {
         let lines: Vec<String> = (0..500).map(|i| format!("line number {i} of plain output")).collect();