From d9abed6b9b6e4346adc07305e059a81c817a841b Mon Sep 17 00:00:00 2001 From: Aaron Ogle Date: Tue, 4 Aug 2026 16:08:36 -0500 Subject: [PATCH] fix(agent): keep working copy on agent view at session end MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Session-end used to switch the working copy back to the session's parent view. That made the review flow clunky: exit the harness and you land back on dev, forced to hunt 'atomic view list' for the draft view holding the agent's single change before you can insert it. Leave the working copy on the session's agent view instead so the user lands where the work happened and can review (log/diff) and insert from there. parent_view is still recorded — it remains the fork baseline and the 'attest --pending' delta reference. --- .../src/turn/orchestrator/session_end.rs | 38 ++++++------------- .../src/turn/orchestrator/session_start.rs | 5 ++- atomic-agent/src/turn/orchestrator/tests.rs | 9 +++-- .../vault/skills/atomic-vault.md | 3 +- docs/adding-agents.md | 2 +- 5 files changed, 22 insertions(+), 35 deletions(-) diff --git a/atomic-agent/src/turn/orchestrator/session_end.rs b/atomic-agent/src/turn/orchestrator/session_end.rs index 4f0701ce..bddc64f9 100644 --- a/atomic-agent/src/turn/orchestrator/session_end.rs +++ b/atomic-agent/src/turn/orchestrator/session_end.rs @@ -13,6 +13,10 @@ impl TurnOrchestrator { /// Transitions the session to Ended, saves it, and creates an /// attestation covering all changes recorded during the session. /// + /// The working copy stays on the session's agent view so the user + /// lands where the work happened and can review it before inserting + /// it into a shared view. + /// /// The attestation is a graph-level audit node — it captures agent /// identity, timing, and which changes were recorded. Cost and token /// data are left at zero (they're not available from the hook) and @@ -46,9 +50,7 @@ impl TurnOrchestrator { // Flush any unrecorded turn BEFORE finalizing. Headless agents such as // Cursor's CLI fire sessionStart/postToolUse/sessionEnd but no per-turn // `stop` (TurnEnd), so the turn's file changes are still uncommitted at - // session end. Record them now — crucially *before* switching back to - // the parent view below, which restores the working copy and would - // otherwise discard the agent's work. Idempotent: agents that already + // session end. Record them now. Idempotent: agents that already // recorded each turn on `stop` leave a clean working copy here, so // `record_turn` returns `EmptyTurn` and this is a no-op for them. { @@ -157,30 +159,12 @@ impl TurnOrchestrator { self.create_session_attestation(&session); } - // Switch back to the user's original view. session-start switched - // to the agent view so that all file writes happened there. Now - // that the session is over, restore the user's view. - if let Some(ref parent) = session.parent_view { - match atomic_repository::Repository::open_existing(&self.repo_root) { - Ok(mut repo) => { - if let Err(e) = repo.switch_view(parent) { - log::warn!( - "Could not switch back to user view '{}': {} (non-fatal)", - parent, - e, - ); - } else { - log::info!("Restored user view '{}'", parent); - } - } - Err(e) => { - log::warn!( - "Could not open repository to restore user view: {} (non-fatal)", - e, - ); - } - } - } + // Deliberately do NOT switch back to `session.parent_view`: the + // working copy stays on the session's agent view so the user lands + // where the work happened and can review it (`atomic log`/`diff`) + // before deciding to insert it into a shared view. Switching back + // forced users to hunt through `atomic view list` for the view + // holding their agent's changes. Ok(DispatchResult::new(session_id, session.phase)) } diff --git a/atomic-agent/src/turn/orchestrator/session_start.rs b/atomic-agent/src/turn/orchestrator/session_start.rs index b614bde8..a2d89483 100644 --- a/atomic-agent/src/turn/orchestrator/session_start.rs +++ b/atomic-agent/src/turn/orchestrator/session_start.rs @@ -240,8 +240,9 @@ impl TurnOrchestrator { // Switch to the agent view so that all file writes during // the session (tool calls, npm install, builds, etc.) happen // while current_view points to the agent view. This ensures - // status/add/record see the right view. session-end will - // switch back to the user's view. + // status/add/record see the right view. The working copy + // stays on the agent view after session-end so the user + // lands where the agent's work happened. if let Err(e) = repo.align_to_view(&session.view_name) { log::warn!( "Could not align to agent view '{}': {} (non-fatal)", diff --git a/atomic-agent/src/turn/orchestrator/tests.rs b/atomic-agent/src/turn/orchestrator/tests.rs index 74fd21d2..2881bf53 100644 --- a/atomic-agent/src/turn/orchestrator/tests.rs +++ b/atomic-agent/src/turn/orchestrator/tests.rs @@ -961,7 +961,7 @@ async fn test_session_start_adopts_declared_managed_view() { assert_eq!( session.parent_view.as_deref(), Some("dev"), - "parent view is remembered so session-end can restore it" + "parent view is remembered as the fork/insert baseline" ); let stamp = session .managed_run @@ -987,15 +987,16 @@ async fn test_session_start_adopts_declared_managed_view() { ); assert_eq!(result.view.as_deref(), Some("sherpa-run-view")); - // Session end restores the parent view. + // Session end leaves the working copy on the agent's view: the user + // lands where the work happened and can review/insert from there. orch.dispatch(session_end_event("sess-managed")) .await .unwrap(); let repo = Repository::open_existing(dir.path()).unwrap(); assert_eq!( repo.current_view(), - "dev", - "session-end restores the user's view" + "sherpa-run-view", + "session-end leaves the working copy on the agent's view" ); } diff --git a/atomic-repository/vault/skills/atomic-vault.md b/atomic-repository/vault/skills/atomic-vault.md index b9e9925a..dc15bb2e 100644 --- a/atomic-repository/vault/skills/atomic-vault.md +++ b/atomic-repository/vault/skills/atomic-vault.md @@ -141,7 +141,8 @@ workflow orchestrator own all of that: view — depending on how you were launched). - **Turn end** records automatically with full AI provenance (model, session, decision graph). -- **Session end** restores the original view. +- **Session end** leaves the working copy on the session's view, where + the recorded work lives — review it there. - **Merging into a shared view** (`atomic insert`) is a human/orchestrator decision made at review time — never run it yourself. diff --git a/docs/adding-agents.md b/docs/adding-agents.md index 61dafce4..2e09902b 100644 --- a/docs/adding-agents.md +++ b/docs/adding-agents.md @@ -213,7 +213,7 @@ invocation is a fresh, stateless CLI process keyed by `session_id`. | `PreToolUse` | `before-tool`, `pre-tool` | before a tool runs | Bookkeeping only (no output yet) | | `PostToolUse` | `after-tool`, `post-tool` | after a tool runs | Appends a classified node (`Exploration` / `Commitment` / `Verification` / `Execution` / `Error`) plus inferred causal edges to the session's **decision graph** | | `TurnEnd` | `stop`, `after-agent`, `turn-end` | agent finishes a turn | **Records a change** on the draft view with full AI provenance embedded | -| `SessionEnd` | `session-end` | agent session closes | Flushes any unrecorded work, writes the **AI attestation**, restores the parent view | +| `SessionEnd` | `session-end` | agent session closes | Flushes any unrecorded work, writes the **AI attestation**; the working copy stays on the session's view so the user lands where the work happened | **The golden rule of responsibility:** the agent side (plugin or hook command) only does two things — fire the event at the right moment, and attach a JSON