Skip to content

Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors

Rust

fix(hooks): exclude the model id by kind, not by identifier shape (#490)

Samir Hanna Verza0b32a81fde0f
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 92b3b2e5..869a3b04 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,28 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
## [Unreleased]
+### Fixed
+- Stopped the scaffolding filter discarding terse prompts. The identifier
+ branch added in #484 fired on any single token carrying a digit and a
+ hyphen, underscore or bracket, which is the shape of a bare model id and
+ equally the shape of `pr-477`, `issue-484` or `commit-a0bc43d`: measured
+ against the two populations, it matched 10 of 10 model ids and 7 of 11
+ plausible terse references, so a page could lose a real title. The class is
+ now excluded at its source instead — `derive_title` skips `SessionStart`
+ outright, the only kind whose title `best_title_hint` fills from the
+ harness's `model` field — and the branch is gone. That also excludes the
+ router's default `title_hint.unwrap_or(kind)`, the literal `session-start`
+ written whenever a harness sends neither `model` nor `title`. The trade is
+ narrower than a shape rule and stated as such: a model id a user *typed* is
+ now kept, because at that point it is what they wrote. Where the prompts
+ are unusable a page falls to the next observation title, which the router
+ defaults to the observation's own kind, so `Session {id}` is reached only
+ by a session carrying nothing else at all — measured over 332 real sessions
+ it never was, and 5 of them landed on `tool file` / `tool non-file` against
+ 327 keeping a real title. `looks_like_scaffolding` is re-exported from
+ `ai-memory-core`, so its narrowing is a public-API behaviour change even
+ though `derive_title` is its only caller in tree. (#484)
+
### Docs
- Documented why registering ai-memory through Kimi Code's own
`kimi mcp add` breaks every model turn: that command writes the plain
diff --git a/crates/ai-memory-core/src/scaffolding.rs b/crates/ai-memory-core/src/scaffolding.rs
index fab3a80d..82403881 100644
--- a/crates/ai-memory-core/src/scaffolding.rs
+++ b/crates/ai-memory-core/src/scaffolding.rs
@@ -14,6 +14,17 @@
//! silently, because a bad title just looks like a title. `transcript.rs`
//! already carries a narrower version of this idea for `<system-reminder>`
//! and `<user_info>` on the workstream path.
+//!
+//! **A bare model id is not tested for here, on purpose.** It has the same
+//! shape as a terse human reference — `gpt-5` and `pr-477` differ only in
+//! meaning — so any lexical rule that catches one catches the other, and this
+//! predicate errs toward keeping text. Every model id that reached a title
+//! in the surveyed corpus arrived on a `SessionStart` observation, which is
+//! the only kind whose title `best_title_hint` fills from the harness's
+//! `model` field, so `derive_title` skips that kind outright instead.
+//!
+//! That is narrower than a shape rule, deliberately: a model id somebody
+//! *typed* is now kept, because at that point it is what the user wrote.
/// Characters that open a decorated shell prompt rather than prose.
/// Box-drawing and the powerline separators commonly echoed into a paste.
@@ -61,19 +72,6 @@ pub fn looks_like_scaffolding(candidate: &str) -> bool {
return true;
}
- // A bare identifier rather than a sentence: no whitespace anywhere, and
- // carrying identifier punctuation. Real prompts essentially always
- // contain a space; requiring the punctuation too keeps a one-word prompt
- // like "help" or "continue" as a legitimate title.
- if !trimmed.contains(char::is_whitespace)
- && trimmed
- .chars()
- .any(|c| c.is_ascii_digit() || c == '[' || c == ']')
- && trimmed.chars().any(|c| c == '-' || c == '_' || c == '[')
- {
- return true;
- }
-
false
}
@@ -82,16 +80,40 @@ mod tests {
use super::*;
#[test]
- fn the_three_observed_classes_are_scaffolding() {
- // From the #484 corpus survey: 16 IDE blocks, 4 prompt echoes, 1
- // model id across 330 real session pages.
+ fn the_two_shape_classes_are_scaffolding() {
+ // From the #484 corpus survey: 16 IDE blocks and 4 prompt echoes
+ // across 330 real session pages. The survey's third class, a bare
+ // model id, is excluded at its source instead — see below.
assert!(looks_like_scaffolding(
"<ide_opened_file>The user opened the file /home/samir/x/main.rs"
));
assert!(looks_like_scaffolding(
"┌─[samir@samirb3 12:56:36] ~/x/ai-usagebar/gnome-extension"
));
- assert!(looks_like_scaffolding("claude-opus-5[1m]"));
+ }
+
+ #[test]
+ fn a_bare_identifier_is_not_judged_by_shape() {
+ // A model id and a terse human reference are the same string shape,
+ // so a rule catching `claude-opus-5[1m]` also catches every entry
+ // below. Measured against the corpus survey's two populations, one
+ // such rule matched 10 of 10 model ids and 7 of 11 references.
+ // Keeping text is the cheaper error: `derive_title` excludes model
+ // ids by observation kind, and nothing excludes a lost title.
+ for reference in [
+ "pr-477",
+ "issue-484",
+ "fix-484",
+ "commit-a0bc43d",
+ "main-2.rs",
+ "step-1",
+ "task_7",
+ ] {
+ assert!(
+ !looks_like_scaffolding(reference),
+ "{reference:?} is a plausible terse prompt and must survive"
+ );
+ }
}
#[test]
diff --git a/crates/ai-memory-hooks/src/synth.rs b/crates/ai-memory-hooks/src/synth.rs
index 65700532..c8b55aca 100644
--- a/crates/ai-memory-hooks/src/synth.rs
+++ b/crates/ai-memory-hooks/src/synth.rs
@@ -68,13 +68,33 @@ pub fn synthesize_session_page(
///
/// Prefers the first user prompt that reads as something a person wrote.
/// A prompt payload is whatever the harness put there, so IDE context
-/// blocks, an echoed shell prompt, or a bare model id can arrive in the
-/// same field (#484) — those are skipped rather than becoming the title the
-/// page is indexed and displayed under.
+/// blocks and an echoed shell prompt can arrive in the same field (#484) —
+/// those are skipped rather than becoming the title the page is indexed and
+/// displayed under.
+///
+/// A `SessionStart` is skipped as a source outright. `best_title_hint`
+/// fills its title from `model` before `title` (`payload.rs`), so a bare
+/// model id became the page title whenever the prompts were unusable — and
+/// by the time it is a string it is indistinguishable from a terse human
+/// reference like `pr-477`, so it is excluded here by kind instead.
+///
+/// The skipped field is `["model", "title"]`, so this drops more than a
+/// model id: OpenCode's `session.created` puts `info.title` in the same
+/// slot. Its plugin subscribes to `created`/`idle`/`deleted`/`compacted`
+/// and never `session.updated`, and starts a session once per id, so the
+/// only value that can arrive is the creation-time name — in practice the
+/// shared literal "New session", never a rename. **That last part is an
+/// assumption about an external component, not an invariant:** it is read
+/// off OpenCode's subscription list as of v1.32.1, and if a future release
+/// surfaces a renamed session here, this would discard a good title with
+/// nothing in the tree failing. What a `SessionStart` carries is whatever
+/// the harness *named the session*, and no harness names it with something
+/// a person typed.
///
/// Falls through in decreasing order of confidence: a usable prompt, then
-/// any usable observation title, then the session's own path. The literal
-/// "session" remains only for the case where there is nothing else at all.
+/// any other usable observation title, then the session's own path. The
+/// literal "session" remains only for the case where there is nothing else
+/// at all.
fn derive_title(observations: &[Observation], session_id: SessionId) -> String {
for obs in observations {
if obs.kind == ObservationKind::UserPrompt
@@ -85,6 +105,9 @@ fn derive_title(observations: &[Observation], session_id: SessionId) -> String {
}
}
for obs in observations {
+ if obs.kind == ObservationKind::SessionStart {
+ continue;
+ }
if !obs.title.is_empty() && !looks_like_scaffolding(&obs.title) {
return obs.title.clone();
}
@@ -541,9 +564,8 @@ mod tests {
for scaffold in [
"<ide_opened_file>The user opened the file /home/samir/x/main.rs",
"\u{250c}\u{2500}[samir@samirb3 12:56:36] ~/x/ai-usagebar",
- "claude-opus-5[1m]",
] {
- let obs = vec![
+ let candidates = vec![
obs(ObservationKind::UserPrompt, scaffold),
obs(
ObservationKind::UserPrompt,
@@ -551,7 +573,7 @@ mod tests {
),
];
assert_eq!(
- derive_title(&obs, test_session_id()),
+ derive_title(&candidates, test_session_id()),
"Make the backpressure test deterministic",
"{scaffold:?} must be skipped in favour of the next real prompt"
);
@@ -563,15 +585,79 @@ mod tests {
#[test]
fn an_all_scaffolding_session_falls_back_to_its_identity() {
let sid = test_session_id();
- let obs = vec![
+ // The model id arrives on the `SessionStart`, which is where
+ // `best_title_hint` puts the harness's `model` field — not as a
+ // prompt somebody typed.
+ let candidates = vec![
+ obs(ObservationKind::SessionStart, "claude-opus-5[1m]"),
obs(ObservationKind::UserPrompt, "<ide_opened_file>x"),
- obs(ObservationKind::UserPrompt, "claude-opus-5[1m]"),
];
- let title = derive_title(&obs, sid);
+ let title = derive_title(&candidates, sid);
assert_eq!(title, format!("Session {sid}"));
assert_ne!(title, "session", "must not collapse to a shared literal");
}
+ /// The shape measured in production on 2026-08-21: a 17-observation
+ /// session whose `UserPrompt` carried an **empty** title, so the fallback
+ /// loop reached the `SessionStart` and the harness's model id became the
+ /// page title. The prompt was never scaffolding — it was absent, which is
+ /// why no filter on the string could have caught this.
+ #[test]
+ fn a_model_id_on_session_start_never_becomes_the_title() {
+ let sid = test_session_id();
+ let with_a_later_title = vec![
+ obs(ObservationKind::SessionStart, "claude-opus-5[1m]"),
+ obs(ObservationKind::UserPrompt, ""),
+ obs(ObservationKind::PostToolUse, "tool non-file"),
+ ];
+ assert_eq!(derive_title(&with_a_later_title, sid), "tool non-file");
+
+ let nothing_else = vec![
+ obs(ObservationKind::SessionStart, "claude-opus-5[1m]"),
+ obs(ObservationKind::UserPrompt, ""),
+ ];
+ assert_eq!(
+ derive_title(&nothing_else, sid),
+ format!("Session {sid}"),
+ "the session's identity, never the model it ran on"
+ );
+ }
+
+ /// Skipping the kind also excludes what the router writes when a harness
+ /// sends neither `model` nor `title`: `title_hint.unwrap_or(kind)`, i.e.
+ /// the literal "session-start". The shape predicate never caught that,
+ /// and it is a string every such page would share.
+ #[test]
+ fn the_routers_default_session_start_title_is_not_a_page_title() {
+ let sid = test_session_id();
+ for stored in ["session-start", "New session"] {
+ let candidates = vec![
+ obs(ObservationKind::SessionStart, stored),
+ obs(ObservationKind::UserPrompt, ""),
+ ];
+ assert_eq!(
+ derive_title(&candidates, sid),
+ format!("Session {sid}"),
+ "{stored:?} is what the harness named the session, not a title"
+ );
+ }
+ }
+
+ /// The cost of the rule this replaces: a terse reference is the same
+ /// string shape as a model id, so filtering one filtered the other.
+ #[test]
+ fn a_terse_reference_is_still_a_usable_title() {
+ let sid = test_session_id();
+ for prompt in ["pr-477", "issue-484", "commit-a0bc43d"] {
+ let candidates = vec![obs(ObservationKind::UserPrompt, prompt)];
+ assert_eq!(
+ derive_title(&candidates, sid),
+ prompt,
+ "{prompt:?} is a plausible prompt and must survive as a title"
+ );
+ }
+ }
+
/// The title is also what `page_descriptor` filters against — a body line
/// equal to it is dropped as a repeat. Changing which string becomes the
/// title therefore changes which body lines survive, so a filtered page