Remove the dead history_loader module (#686)

* Remove the dead history_loader module

`history_loader::load_all` was deleted when host history moved to ACP
`session/list`, but the module stayed behind: half of it was disk-scanner
dead code kept compiling only by `#[allow(dead_code)]`, and its header
still documented a per-CLI on-disk scan that no longer runs. The name
alone implied WTA cold-scans four CLI transcript stores at startup.

Dissolve it into the code that actually uses the survivors:

- `parse_iso_to_system_time` and `short_id` move to `session_history.rs`
  next to their only callers (ACP `SessionInfo::updated_at` and the
  no-title fallback label), with the Codex-specific doc rewritten.
- `codex_record_is_subagent_meta` / `codex_payload_is_subagent` move to
  `session_watcher/classify_codex.rs` as `record_is_subagent_meta` /
  `payload_is_subagent`, beside the classifier that consumes them.
- Delete the YAML block-scalar parser, the JSONL title extractors and
  their helpers, all unused.

Also refresh comments that still pointed at `load_all` or the removed
`wsl_acp` module, and mark the two spec sections describing the deleted
disk scanner as superseded.

No behavior change; tests move with their code.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 741e2eec-21f3-4286-8e2e-500eb06d553d

* Parse timestamps with the time crate instead of by hand

`parse_iso_to_system_time` was ~125 lines of hand-rolled ISO 8601: manual
offset scanning, a leap-year table, day-of-month validation, and its own
days-since-epoch arithmetic. Replace it with `time::OffsetDateTime` +
`Rfc3339`.

`time` v0.3.47 is already linked into wta through ratatui-widgets and
tracing-appender and is already listed in cgmanifest.json and NOTICE.md, so
promoting it to a direct dependency adds no third-party code; only the
`parsing` feature is new. Re-running Generate-WtaThirdPartyNotices.ps1
produces no diff (276 crates before and after).

Behavior is preserved, including the two non-RFC-3339 quirks the old parser
had, now explicit and tested:

- offset-less `YYYY-MM-DDTHH:MM:SS` is still read as UTC, via a retry with
  `Z` appended rather than silently dropping the value;
- pre-epoch timestamps still return `None`, so callers can keep treating
  `SystemTime` as an offset from `UNIX_EPOCH`.

All existing parse tests pass unchanged; added cases for the offset-less
fallback and for malformed input.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 741e2eec-21f3-4286-8e2e-500eb06d553d

* Address review: spelling and helper visibility

- Rename `parse_iso_reads_offsetless_timestamps_as_utc` to
  `..._without_offset_as_utc`. check-spelling flagged `offsetless`;
  rewording keeps the allowlist untouched for a single test name.
- Narrow the Codex subagent helpers moved in this PR:
  `record_is_subagent_meta` is `pub(super)` (its only caller is
  `session_watcher::process_change`) and `payload_is_subagent` is now
  private, since nothing outside `classify_codex` calls it. `classify`
  keeps its existing visibility; it predates this PR.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 741e2eec-21f3-4286-8e2e-500eb06d553d

* Convert to SystemTime without a panicking Add

`SystemTime::from(OffsetDateTime)` reaches `impl Add<Duration> for
SystemTime`, which panics on overflow, so the function's documented
"never panics" contract held only because `time` caps years at 9999.
That cap is the `large-dates` feature, and feature unification puts it
outside this crate's control, so fail closed instead: take
`unix_timestamp()` through `u64::try_from` (which also subsumes the
pre-epoch rejection) and add it with `SystemTime::checked_add`.

Adds `parse_iso_preserves_fractional_seconds`, since the `Duration` is
now assembled by hand and sub-second precision has to be carried across
explicitly.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 741e2eec-21f3-4286-8e2e-500eb06d553d

* Fix comments that misdescribed the agent-pane origin join

`classify_and_map` does not stamp `AgentPane` origin onto historical
rows — it consults the agent-pane index to subtract those sessions from
history outright. Repointing the old `history_loader::load_all` comments
at it carried over a claim that was true of the deleted disk scanner but
is not true of the ACP mapper.

- app.rs and the `SessionOrigin` doc now say the flag is stamped on the
  live row and kept after it ends, and that rows rebuilt from ACP
  `session/list` never carry it because they are filtered out.
- session_registry.rs: `session_history` maps `session/list` rows; the
  listing itself happens in its caller. Dropped the "ACP history scan"
  wording that implied it owns I/O.
- hybrid-agent-session-tracking.md: the superseded note now names the
  removed symbols the section still references, so readers do not go
  looking for deleted code.

Comment-only; no behavior change.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 741e2eec-21f3-4286-8e2e-500eb06d553d

* Attribute the overflow panic to the right impl

`SystemTime::from(OffsetDateTime)` has no `Add` of its own; it is
implemented as `SystemTime::UNIX_EPOCH + duration`, and that `Add` impl
is what panics on overflow. Reword so the comment names the impl the
`checked_add` is actually avoiding.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 741e2eec-21f3-4286-8e2e-500eb06d553d

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: 741e2eec-21f3-4286-8e2e-500eb06d553d

Author: DDKinger

doc/specs/hybrid-agent-session-tracking.md

@@ -242,6 +242,18 @@ helper-side and unrelated).
 
 ### Title resolution (and the codex AGENTS.md fix)
 
+> **Superseded.** The on-disk title scan described below was removed with the
+> rest of the disk scanner: titles now come from ACP `session/list`
+> (`SessionInfo::title`, mapped in `session_history.rs`, with
+> `session_history::short_id` as the fallback label). Only the codex
+> subagent-fork check survives, as
+> `session_watcher::classify_codex::record_is_subagent_meta`. The rationale
+> below is kept for history — every other file, function, and helper it
+> names (`history_loader.rs`, `try_refresh_title_from_disk`,
+> `lookup_title_for_session`, `codex_title_from_file`,
+> `codex_user_text_is_synthetic`, `codex_session_has_real_content`) no
+> longer exists in the tree, so do not go looking for it.
+
 A watcher row is created with a **synthetic** title (cwd basename, or empty),
 then upgraded from the CLI's on-disk artefacts by `try_refresh_title_from_disk`
 → `lookup_title_for_session`, the **same** disk-title path the hook and
@@ -283,7 +295,7 @@ title).
 | Apply / dedup / gate / reaper | `tools/wta/src/master/mod.rs` (`apply_watcher_event`, `handle_session_hook`, `ensure_watched_session_row`, `watcher_row_allowed`, `live_it_pane_guids`, `reap_dead_class_b_sessions`, `hook_owned` + `born_bound` sets) |
 | Born-bound registration | `session_registry.rs` (`build_born_bound_request`, `INTELLTERM_METHOD_SESSION_BORN_BOUND`), `main.rs` (`register_launched_session_with_master`) |
 | Row `bound_pid` field | `tools/wta/src/session_registry.rs` |
-| Codex title / subagent / phantom | `tools/wta/src/history_loader.rs` |
+| Codex subagent fork detection | `session_watcher/classify_codex.rs` (`record_is_subagent_meta`) |
 | User-input tool heuristic | `agent_sessions.rs` (`is_user_input_tool`) |
 
 ### Status detection (per-CLI)
@@ -403,7 +415,7 @@ of the live states.
   `session_hook_marks_*`; born-bound: `session_born_bound_marks_born_bound_not_hook_owned`,
   `born_bound_session_gets_watcher_activity_without_rebinding`,
   `real_hook_takes_over_born_bound_session`, `resume_binding_events_are_born_bound_not_hook_owned`),
-  `history_loader::tests` (codex title / subagent / phantom), and
+  `classify_codex::tests` (codex subagent fork detection), and
   `session_watcher` discovery/classify tests (incl. `classify_claude` turn-based:
   user→Working, `stop_reason` end_turn→Idle / tool_use→Working, streaming-partial
   stays Working, AskUserQuestion→Attention).

doc/specs/wsl-session-management.md

@@ -1,5 +1,11 @@
 # WSL Agent Session Management (Historical MVP)
 
+> **Superseded in part.** This spec is written against the on-disk history
+> scanner (`history_loader::load_all`) and the `wsl_acp` scan module, both of
+> which have since been removed. Host history now comes from the agent's own
+> ACP `session/list` (`session_history.rs`); there is no WSL history scan in
+> the tree today. The design rationale below is kept for history.
+
 ## Abstract
 
 Intelligent Terminal (IT) surfaces a list of agent-CLI sessions (Copilot /

tools/wta/Cargo.lock

@@ -61,3 +61,8 @@ sys-locale = "0.3"
 # (issue #287). Already in the lockfile as a transitive dep of clap, so
 # promoting it to a direct dep does not change the resolved graph.
 strsim = "0.11"
+# RFC 3339 parsing for ACP `SessionInfo::updated_at`. Already in the lockfile
+# as a transitive dep of ratatui and tracing-appender, so promoting it to a
+# direct dep does not change the resolved graph; only the `parsing` feature
+# is added on top of what those crates already enable.
+time = { version = "0.3", features = ["parsing"] }

tools/wta/Cargo.toml

@@ -27,9 +27,8 @@
 //   ENTER-routing work that will consume this field):
 //     {"v":2,"session_id":"<uuid>","origin":"agent_pane","pane_session_id":"<WT pane GUID>","started_at":"<RFC3339-ish>"}
 //
-// We deliberately do NOT record `cli_source` — `history_loader` already
-// derives it from which per-CLI on-disk artefact directory the session was
-// found in, so duplicating it here would create a second source of truth
+// We deliberately do NOT record `cli_source` — the session's own agent
+// reports it, so duplicating it here would create a second source of truth
 // that could drift. Same rationale for `owner_tab_id`: no caller needs it
 // yet, and we can always recover it via WT itself.
 //
@@ -44,8 +43,8 @@
 // The file is append-only; it is never read-then-written from this module.
 // Old entries become orphans naturally when the corresponding CLI session
 // directory is deleted by the user or the agent CLI itself — orphan entries
-// in the index are harmless because `history_loader` only consults the
-// index when constructing rows for sessions that *still exist on disk*.
+// in the index are harmless because the index is only ever consulted as a
+// filter against session ids the agent itself still reports.
 
 use std::collections::{HashMap, HashSet};
 use std::fs::{File, OpenOptions};
@@ -133,8 +132,8 @@ pub fn append_to(
 
 /// Load the default index into a `HashSet<String>` of session ids. Empty
 /// set if the file does not exist, cannot be opened, or is empty — never
-/// errors out to the caller, which lets `history_loader` proceed even on
-/// a fresh install or after a manual delete.
+/// errors out to the caller, so the history scan still proceeds on a fresh
+/// install or after a manual delete.
 ///
 /// Unpackaged dev binaries also merge the installed Intelligent Terminal
 /// package's LocalState index when present. That keeps diagnostics such as

tools/wta/src/agent_pane_origin.rs

@@ -89,7 +89,7 @@ pub(crate) fn title_is_placeholder(cli: &CliSource, title: &str) -> bool {
             23 => *byte == b'Z',
             _ => byte.is_ascii_digit(),
         })
-        && crate::history_loader::parse_iso_to_system_time(timestamp).is_some()
+        && crate::session_history::parse_iso_to_system_time(timestamp).is_some()
 }
 
 #[derive(Clone, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -153,10 +153,12 @@ pub enum LivenessState {
 ///
 /// Populated authoritatively by `agent_pane_origin`: WTA appends a record
 /// to the on-disk index whenever it creates an ACP session for an agent
-/// pane (i.e. `--owner-tab-id` was supplied), and `history_loader` joins
-/// that index when reconstructing historical rows. Live rows default to
-/// `Unknown` because the UI only surfaces this badge for ended/historical
-/// sessions, where it is most useful.
+/// pane (i.e. `--owner-tab-id` was supplied). A row is stamped while it is
+/// live, when a routed event's session id turns up in that index, and it
+/// keeps the flag once the session ends — which is where the UI surfaces
+/// it. Rows rebuilt from ACP `session/list` never carry it:
+/// `session_history::classify_and_map` uses the same index to drop
+/// agent-pane sessions from history entirely rather than to badge them.
 #[derive(Clone, Debug, Default, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub enum SessionOrigin {
     /// Origin not recorded — either the session pre-dates the index, was
@@ -1147,8 +1149,8 @@ impl AgentSessionRegistry {
     /// is `Historical` or `Ended`, upgrade it to `Live` (`AgentStatus::Idle`)
     /// and bind the pane.
     ///
-    /// Motivation: at startup the on-disk history scan
-    /// (`history_loader::load_all`) and the helper's `list_sessions`
+    /// Motivation: at startup the ACP `session/list` history scan
+    /// (`session_history::classify_and_map`) and the helper's `list_sessions`
     /// bootstrap can land in either order, and a WTA process attached
     /// to an existing master in another WT window may never see the
     /// originating `SessionStarted` hook event. Without this join, a

tools/wta/src/agent_sessions.rs

@@ -732,8 +732,10 @@ where
 
     // Stamp `AgentPane` origin on the live session if the agent-pane
     // origin index recorded its session id. This is what flips the
-    // "agent pane" prefix on for *live* rows — historical rows pick up
-    // the same flag through `history_loader::load_all`'s join. We
+    // "agent pane" prefix on for *live* rows. Rows rebuilt from ACP
+    // `session/list` never get it: `session_history::classify_and_map`
+    // consults the same index to subtract agent-pane sessions from
+    // history outright rather than to badge them. We
     // re-read the index on every routed event (small file, infrequent
     // event) rather than caching, to stay correct after a new session
     // is created while wta is already running.