1.2.0: what real sessions say about retirement, and memory for subagents

.claude-plugin/marketplace.json

@@ -6,14 +6,14 @@
   },
   "metadata": {
     "description": "Persistent memory for Claude Code and Codex",
-    "version": "1.1.0"
+    "version": "1.2.0"
   },
   "plugins": [
     {
       "name": "muninn",
       "source": "./plugin",
       "description": "Persistent memory for Claude Code and Codex that never serves a replaced decision, and turns CLAUDE.md / AGENTS.md rules into permission rules and deny hooks. Local: one binary, SQLite, no cloud, no model in the read path.",
-      "version": "1.1.0",
+      "version": "1.2.0",
       "author": {
         "name": "Muninn contributors"
       },

CHANGELOG.md

@@ -4,6 +4,61 @@ Names like v17 or v38 in these entries are test runs, numbered in the order they
 they are not releases. Releases are the numbered headings, such as 0.2.0. See
 [how to read the test names](docs/technical-overview.md#how-to-read-the-test-names).
 
+## 1.2.0 — 2026-11-02
+
+These come from 773 sessions and 1,200 subagent transcripts from the ten stores one person
+had used Muninn in. The measurements are usage-v1 to usage-v3 in
+[muninn-experiments](https://github.com/ilien-dev/muninn-experiments), which publishes counts
+and record ids only. The projects stay private.
+
+- **Subagents get memory.** Before this, 1 of the 1,200 subagents had been given any, and 667 of
+  them wrote files. A new `SubagentStart` hook hands a subagent the invariants and corrections a
+  session start shows, the catalogue, and one line on how to read them. In usage-v3 there were six
+  tasks, each with an obvious default that one recorded rule forbids. The subagent followed the
+  rule in 5 of 6 with the hook and in 1 of 6 without it. That is six tasks on one model
+  (`claude-sonnet-5-5`). Over seven `perf` runs its p95 was 6.7 to 9.7 ms, and SessionStart's 7.6 to 9.8 ms. `muninn config subagent-delivery off`
+  turns it off.
+- **Fewer records retired for no reason.** The 95 typed records that had been retired with a
+  replacement were labelled without seeing the code: 2 retirements were valid, and 25 lost
+  something the user had really said. Replaying every session with each build, 1.1.0 retired 20 of those
+  25 and this build retires 9. Retirements that pointed at an unrelated record went from
+  21 to 8. Both valid ones still happen. Four causes were fixed:
+  - A correction's subject came from a turn index that two turns could share, so a new correction
+    replaced an unrelated old one. That happened 19 times in the replay and now never does. A
+    correction now replaces only one from the same session that shares at least 10 % of its
+    content words.
+  - Episodes took part in replacement by text. They are literal excerpts and no longer do.
+  - An answer picked in AskUserQuestion is recorded but retires nothing by its wording.
+  - A commit could retire a statement made after it, and hand it to an older record. Only a commit
+    made after a record can retire it now, and only a newer record that names a word the commit
+    added inherits it. Of the records the turn produced, only those that name the removed value go
+    with it. Retirements handed to an older record went from 64 to 9.
+- **`maintain` brings back what those paths retired wrongly, once.** On copies of the ten stores it
+  restores 56 records, most of them episodes. 21 are among the 95 labelled: 13 real losses, 7
+  old statements that were never facts, and 1 that had been handed to an unrelated record. It restores neither valid retirement. Of 14 losses checked
+  by hand, 7 come back.
+- **Messages typed while the agent was working were never read.** Claude Code delivers them as
+  `queued_command` attachments, not as user turns. 28 of the 31 such messages the review found
+  missing are now stored, and the change adds no record built from a task notification.
+- **Sessions the Agent SDK runs no longer produce decisions.** Their prompts are written by
+  a program, so they are kept as episodes and nothing in them is typed.
+- **A long sentence was listed twice in the catalogue.** The decision keeps the whole line and the
+  correction from the same message cuts it at 160 characters, so the two did not match. The comparison now
+  cuts both at that length.
+- **`muninn show` called a live record retired** when the budget left it out of the block. It now
+  says the record is on record, not shown in this block, and to ask for it on its own.
+- **The startup text says to ask `muninn why` in the records' language.** The search matches words.
+  On 30 Spanish records the question reached the record 27 times in Spanish and 5 in English
+  (usage-v2). A model wrote those questions from the records, so they may be closer to the records'
+  wording than a real question would be.
+- **Dependencies:** `rusqlite` 0.41.1, `serde_json` 1.0.154, `sha2` 0.11.1, `tree-sitter` 0.27.1.
+- **Held out:** on the loop7 and loop8 scenarios, which none of these fixes was written from, every
+  figure is the same as on 1.1.0, and nothing is retired on the noise arm.
+- **Not fixed:** a user statement read as a change of an earlier one on the same subject still
+  retires it. That is most of what is still retired wrongly in the replay. `perf --strict` still fails on Windows on the full
+  UserPromptSubmit: 8.9 to 11.5 ms p95 against a 10 ms limit over seven runs of each build, on 1.1.0 and on
+  this build alike. The medians are 10.6 ms on 1.1.0 and 9.7 ms on this build.
+
 ## 1.1.0 — 2026-10-18
 
 These come from reading the stores of several projects one person had used Muninn in.

Cargo.toml

@@ -12,14 +12,14 @@ members = [
 ]
 
 [workspace.package]
-version = "1.1.0"
+version = "1.2.0"
 edition = "2021"
 license = "AGPL-3.0-only"
 repository = "https://github.com/ilien-dev/muninn"
 rust-version = "1.85"
 
 [workspace.dependencies]
-rusqlite = { version = "0.40", features = ["bundled"] }
+rusqlite = { version = "0.41", features = ["bundled"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 thiserror = "2"

crates/muninn-bench/src/main.rs

@@ -824,6 +824,13 @@ fn main() -> Result<()> {
                 &serde_json::json!({"session_id":"b","cwd":cwd}),
                 runs,
             )?;
+            let subagent = run_hook(
+                &bin,
+                root,
+                "SubagentStart",
+                &serde_json::json!({"session_id":"b","cwd":cwd}),
+                runs,
+            )?;
             let ups_gated = run_hook(
                 &bin,
                 root,
@@ -848,6 +855,13 @@ fn main() -> Result<()> {
                     unit: "ms",
                     floor: Some(start.1),
                 },
+                Contract {
+                    name: "hook SubagentStart p95",
+                    value: subagent.1,
+                    limit: 10.0,
+                    unit: "ms",
+                    floor: Some(start.1),
+                },
                 Contract {
                     name: "hook UserPromptSubmit gated p95",
                     value: ups_gated.1,
@@ -897,7 +911,7 @@ fn main() -> Result<()> {
                     serde_json::json!({
                         "records": records, "cues": cues, "runs": runs, "populate_ms": populate_ms,
                         "process_start": start,
-                        "hook": { "SessionStart": ss, "UserPromptSubmit_gated": ups_gated, "UserPromptSubmit_full": ups },
+                        "hook": { "SessionStart": ss, "SubagentStart": subagent, "UserPromptSubmit_gated": ups_gated, "UserPromptSubmit_full": ups },
                         "cue_eval": { "glob": glob, "ancestor": anc },
                         "ingest_200_ms": ingest_ms,
                         "limit_scale": scale,
@@ -914,6 +928,7 @@ fn main() -> Result<()> {
                 for (n, v) in [
                     ("process start (`--version`, no work)", start),
                     ("hook SessionStart", ss),
+                    ("hook SubagentStart", subagent),
                     ("hook UserPromptSubmit (gated)", ups_gated),
                     ("hook UserPromptSubmit (full)", ups),
                     ("cue eval GLOB per file", glob),