[nplez1]

	Files	Added	Deleted	Net
Test	12	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​931	$\color{#cf222e}{\Huge{\mathbf{−}}}$​19	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​912
Prod	32	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​1354	$\color{#cf222e}{\Huge{\mathbf{−}}}$​427	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​927

ELI5

Orca can write a short brief about what an agent session did: what it is working on, what it finished, what is blocking it, and a timeline of highlights. That feature was merged, but it never appeared anywhere — the panel only existed inside a page the app can no longer open. This puts it in the right sidebar, next to Agents, and gives it its own agent/model/effort setting so its cost can be tuned on its own.

What Changed

Before: there was no way to see a session summary. The panel was mounted only inside the Activity page's detail pane, and nothing has opened that page since the agent list became a sidebar tab (upstream stablyai#18222), so no build ever rendered it. Its model settings were also borrowed from the commit-message Source Control AI recipe: editing commit-message AI silently changed summary cost, and editing the commit prompt looked like it should change summaries when it could not.

After:

• A Summary tab in the right sidebar (info icon), beside Agents. It briefs the agent session you have focused in the active workspace and nothing else; with nothing focused it asks you to focus an agent session.
• Settings → Agents → Session summaries owns its agent, model, thinking effort, and custom command. Unset means your default agent with its default model and effort, so summaries work without configuring anything.
• Agents whose model list only exists in their CLI (Pi) now list real models with their effort levels — 63 of them on this machine, from pi --list-models. Copilot's hosted Claude models expose effort like its GPT models already did.
• The effort dropdown is always visible: enabled with the model's levels, or disabled with This model has no effort setting.
• A finished or hibernated session still summarizes when you focus its pane.

Mechanism:

• A new static summary tab renders the panel in RightSidebarPanelContent. The ledger view and header state moved out of the Activity strip (activity/session-summary-ledger-view.tsx, activity/session-summary-presentation.ts) so both hosts render one component instead of two copies.
• The subject is getFocusedAgentPaneKeyForWorktree (right-sidebar/session-summary-subject.ts). The pane keys on the pane and its provider session, because a new session can start in the pane you are already focused on.
• The fold reads the new sessionSummaryAi block (shared/session-summary-ai.ts) instead of Source Control AI's commitMessage operation. Resolution and the picker share one rule for the effective model and effort, so the dropdown always names what the CLI is actually given.
• Dynamic agents are probed once through the existing git:discoverCommitMessageModels path; only a real probe result is cached, keyed by agent, with a Refresh models control.
• Laziness, incrementality and caching are unchanged from feat(session-summary): on-view session summaries for agent sessions #51: a fold runs only while the pane is open, only messages past the ledger's cursor are extracted, and the cached brief is shown immediately.

Why

• A right-sidebar tab rather than resurrecting the Activity page. That page is unreachable, and the other thing it held (the agent list) has already moved to the sidebar. A tab puts the panel where agents now are.
• Its own setting. A summary is not a source-control action: it feeds no command template and no PR/branch flow. Borrowing the commit-message recipe made its cost untunable on its own. The shape is reused (agent, per-model effort, custom command) — including the selection code, which the AI-author card now shares instead of duplicating.
• Local-only discovery, host-keyed storage. The fold spawns the agent CLI locally and reads the transcript from the local filesystem, so a remote host's model list would offer models that spawn cannot run. The list is stored per host so a future host-aware fold needs no migration, but resolution reads the local slot.
• Per-model effort, always visible. Effort is stored per model (selectedThinkingByModel), so a model that advertises none cannot be given one; hiding the control made that read as a missing feature.

Linked Issue

None — fork-local feature work (no external issue exists; happy to attach one if the maintainer wants tracking).

Fixes #

Visual Proof

A "before" capture is impossible: the panel rendered in no build, so there was no screen to photograph. Evidence instead, both from a hidden dev instance over CDP:

• After — the new tab and pane (info icon, focus prompt when nothing is focused):

  

• After — the new setting with a probed dynamic agent (Pi, Deepseek Deepseek Flash, effort Low, Refresh models):

  

The setting's before/after is real and visible: Pi previously offered a single Config default entry and no effort control at all, because its model list exists only in its CLI output.

Testing

• pnpm tc clean. Changed-code quality gate, localization coverage/catalogs/extraction, and the fork's newer gates (check:upstream-identity-literals, check:max-lines-ratchet, check:dead-classes) all pass.

• 10,438 tests pass across 1,140 files in the touched areas (settings, right-sidebar, lib, shared, main/session-summary). New tests cover: refusing an unusable agent instead of substituting one, creating the fold's scratch directory, honouring a probed model's effort, keeping the spec default in the picker list, the session-identity remount, the missing-bridge pane state, and the fields component's three effort states.

• Rendered checks (hidden dev instance, no window shown): the tab and pane render; the setting drives sessionSummaryAi while commitMessageAi stays untouched; Pi's probe caches 63 models in both the agent slot and the host-keyed slot; a thinking: no Pi model shows the disabled effort control with its explanation; the value persists in profile state.

• macOS only. No Windows/Linux/SSH run, and no real end-to-end fold against a live agent CLI — the fold's own behaviour is unchanged from feat(session-summary): on-view session summaries for agent sessions #51 and covered by its unit tests.

• I manually tested these changes locally

• Automated tests added/updated, or explained why not below

AI Disclosure

Authored with the pi coding agent (xiaomi/mimo workhorse models) under human direction, then reviewed by github-copilot/gpt-6.1-sol at high effort. The review raised two blockers and several should-fix findings, all fixed in the last three commits:

1. The fold's scratch directory (<tmpdir>/orca-session-summary) was never created, so every fold would have failed with ENOENT — created before spawning.
2. An unusable agent setting silently ran the default agent (transcript to a provider the user did not choose) — now fails with a message naming the setting.
3. A picked effort was dropped for probed models, and the displayed model could differ from the executed one — one rule now serves both sides.
4. A new session in the same pane kept the previous session's brief — the pane keys on the provider session too.
5. Discovery was cached forever — added Refresh models.
6. A paired web client offered a tab it cannot run — the tab is omitted when the bridge is missing.

Agent skill upstream boundary

• Not applicable, or this change follows docs/reference/agent-skill-sharing-upstream-boundary.md and copies or mechanically translates no upstream skill-installer source, tests, fixtures, registry entries, path tables, comments, or documentation.

Notes

• Security: a fold runs only the agent configured for summaries (or your default agent); an unusable configuration fails rather than substituting a provider, so a transcript is never sent somewhere the user did not choose.
• Cross-platform / SSH: the spawn path is the existing cross-platform spawnSourceControlAgent; no bare rg, no shell-string spawning, no new path handling. The right sidebar and settings are desktop surfaces.
• Backwards compatibility: sessionSummaryAi is an optional settings key that older builds ignore, and the new summary tab is dropped by an older host's tolerant ui.set path (tolerateUnknownValues) rather than failing the batch.
• Performance: no polling and no background folds; one probe per dynamic agent, cached, refreshable on demand.
• Known and deliberately left open: while the pane stays open, new messages do not trigger another fold (re-folding on status changes would spend tokens continuously for a pane the user may be leaving); the feature-wall AI-author card does not probe dynamic agents, so Pi still shows its stub there; and parsePiModels double-labels Pi models ("Deepseek Deepseek Flash"), which Native Chat's picker shares.

Checklist

• This PR is small and focused
• I explained what changed and why (ELI5, the user-facing before/after, the mechanism, and why over the alternatives)
• Before/after screenshots or videos attached for UI changes, or N/A with reason
• Self-reviewed for correctness, security, and performance
• Cross-platform, SSH/remote, and path/shortcut impact considered (or N/A)
• pnpm lint, pnpm typecheck, pnpm test, and pnpm build pass (or CI will cover; local preferred)