[nplez1]

ELI5

The Summary tab next to the Agents list is supposed to be a quick written brief of what an agent session is doing, so you don't have to scroll back through a terminal. It never showed a brief — not once, for any session — and it often claimed it was "Up to date" while the pane was empty. Separately, picking GitHub Copilot as the agent that writes those briefs offered a model list that stopped somewhere in the past. This PR fixes all three, plus the tab's icon.

What Changed

Before: opening the Summary tab on an agent session showed a loading skeleton forever, then settled on "Up to date" with nothing under it. Hiding and re-showing the panel changed nothing. Choosing Copilot in Settings → Agents → Session summaries listed models no longer offered and never listed current ones.

After: the tab produces a real brief for a session, using the transcript of the session itself — goal, what the agent is doing now, what it has finished, blockers, decisions, and a timeline. Copilot's model list is read from the installed CLI, so it always matches what that CLI actually accepts.

There are three independent causes, and each is fixed at its own layer.

1. The brief could not find the session's transcript

A summary is built from the agent CLI's own transcript file, not by re-reading the terminal. The code that located that file only knew the "native chat" agents — Claude, Codex, Grok, OMP, OpenCode. For anything else it returned nothing, so the pane settled on "No readable transcript for this session yet." and stayed there permanently.

On the machine this was reported from, the sessions are Pi (462), Copilot (16), OpenCode (11) and Cursor (2) — so almost every session was one the brief could not read. That is why it never worked.

It now also resolves:

• Pi / Prime Agent — the hook already reports the exact session file, and Orca only exposes their session identity at all when that file exists, so there is nothing to search for.
• Copilot — reports no file path, but its session id is the name of its session directory, so the transcript is $COPILOT_HOME/session-state/<id>/events.jsonl. The id is validated before being joined onto that path, so a malformed one cannot escape the sessions root.
• A hook-reported path that no longer exists, or that points at a directory, is no longer mistaken for a transcript.

The decoding side already understood every one of these formats, so no parser changed.

2. "Up to date" was reported for folds that were running, or had failed

Two bugs in the header line:

• The panel's "a fold is running" flag was only set once the first chunk of text came back from the model. So for the whole of a first fold — the entire window a user actually watches — the header said "Up to date" over a skeleton, and a fold that could not start never set the flag at all. It is now set when the fold starts.
• The header text ignored the fold's own status, so "no readable transcript" and "the summarizer failed" both fell through to "Up to date". They now say what happened. A failure over a session that already has a good brief keeps counting how many events are unfolded instead, because the old brief is still on screen.

3. Copilot's model list was a hand-written snapshot

An array pinned 16 model ids, ending at gpt-5.5. The installed CLI offers 31, including claude-sonnet-5.5, claude-opus-5.5, gpt-6.1-sol, gpt-6-luna, grok-4.7 and kimi-k3. Anything released after that array was written was unreachable, and retired ids stayed on offer.

The CLI has no "list models" command and its interactive picker is filtered per account, but copilot help config documents the model setting as the list of ids it accepts — the CLI's own catalog. The probe now reads that list. An unparseable help format parses to nothing, which falls back to a refreshed pinned copy and is not cached, so the next attempt probes again rather than retiring the probe forever.

4. The icon

The tab and the pane header used a generic information circle — the same icon the app uses for notices, so the tab read as a warning rather than a place to read a brief. Both now use a sticky note.

Why

Why probe help config rather than refresh the array. A refreshed array is stale again the next time GitHub ships a model; the CLI's own catalog is the thing that tracks that, and we already have a probe pipeline for every other dynamic agent.

Why not fix only Copilot's list. The reporting user's own settings select Pi as the summarizing agent, and the sessions were Pi. The model list and the missing brief are genuinely separate defects that happened to be reported together.

Why the header changes are in the same commit as the transcript coverage. They are the two halves of one symptom: a user cannot tell "nothing is here yet" from "everything is done" if the header lies. Fixing only the first would have replaced a silent empty pane with a pane that still claimed to be up to date whenever a fold failed.

Linked Issue

N/A — this is a fork PR against nplez1/main, reported directly by the fork owner rather than filed as an issue first.

Visual Proof

Not attached: this work was done headlessly in a worktree, and the rendered check was not run. Recording it needs the app driven by CDP on a focused agent session plus a completed fold, which is more than the change warrants; say the word and I'll capture it before this goes anywhere near upstream.

The visual deltas are exact and pinned by tests, so a reviewer can see them without a screenshot:

• Icon — the Summary tab and the pane header change from lucide-info to lucide-sticky-note (SessionSummaryPane.test.tsx asserts the class).
• Header text — new strings for a session with no readable transcript ("No readable transcript") and for a failed fold ("Could not summarize this session"), replacing "Up to date" in both cases.
• No layout, spacing, colour or typography change.

Testing

• Automated tests added/updated
• Manually exercised in the running app — the feature path was driven end to end in the main process (below), but no rendered-UI check was run

Full runs, against the pushed head

Every command below was run in this worktree against f6adc4ca02 — the head of this PR — after the commits were made, so the evidence describes exactly what is being reviewed. An earlier, narrower run during development produced 80 failures; 42 of them were phantom, because a bare pnpm exec vitest run does not export ORCA_BACKGROUND_LAUNCH=1 the way the repo's pnpm test wrapper does, which turned 41 green browser-manager cases red. Running the documented commands is what separated the real signal from that noise.

Command	Result
pnpm tc	clean
pnpm lint	exit 0 — all 19 steps, including repo-wide oxlint with its max-lines rule
pnpm test	113,874 tests — 112,887 passed, 57 failed across 14 files, 1 expected failure, 929 skipped

All 57 failures are pre-existing and unrelated to this change. None of the 14 failing files is one this PR touches, and none of them has been modified since the upstream merge (git log 7c0a53a148..HEAD -- <paths> is empty, so they arrived red and unchanged). They fall into three classes:

Class	Files	Why they are red
Upstream-red logic — agent-session-*, agent-launch-*, structured-agent-session-*	9	Recorded in LOCAL-PATCHES.md as "proved not ours"; they reproduce on pre-merge 807aa6062c
Relay region correction	1	Needs the uninstalled cloud/ workspace (pg)
PTY / subprocess suites that only run red under the bun project	4	daemon/repro-13767-shell-ready-marker-lost-to-exec, daemon/shell-ready, shell-startup-feature-channel, orcad-bundled-runtime.integration produce no child output under the bun runtime and pass under the node project — a test-project assignment gap, the same class upstream re-seated in stablyai#26028

The third class only becomes visible with bun installed (pinned at config/.bun-version, 1.4.2); it is a property of this test environment, not of this diff — none of those four files is touched here.

Focused tests added or updated for this change

• New src/main/session-summary/session-summary-transcript-source.test.ts (5 cases): a Pi session read from its hook-reported file, a Copilot session read from its session-state directory, a traversal-shaped Copilot id refused, a directory refused as a transcript, and an agent the fold cannot parse returning nothing.
• session-summary-service.test.ts: the fold reports itself in flight before its first chunk returns.
• session-summary-presentation.test.ts: "no readable transcript" and the failure case are named instead of "Up to date"; a failed fold over an existing brief still counts its backlog.
• commit-message-agent-spec.test.ts: Copilot probes help config; the id parser reads only the model setting's block, adds auto back, and parses to nothing on a format change.
• Changed-lines gate pnpm run check:code-quality:changed and the localization verifiers pass.

End to end

On the reporting machine, with the reporter's own settings (sessionSummaryAi = Pi / deepseek/deepseek-flash) and a real Pi transcript: the service reaches ready in 7.4s with a populated brief. Before this change the same call returned "no readable transcript". The Copilot probe was likewise run through the real main-process discovery path, returning 31 models with catalogOrigin: 'probe'.

Platforms actually exercised: macOS only, local worktree, no SSH. The Copilot path is macOS-verified (/opt/homebrew/bin/copilot); the Pi path depends only on the hook's reported path, which is platform-neutral, but Windows and Linux were not run.

Notes

• Remote SSH / mobile: unaffected in the sense that nothing new crosses the wire — sessionSummary is local-only, and a paired web client still gets no tab. Transcript resolution deliberately stays local because the fold spawns the agent CLI locally.
• Backwards compatibility: an unchanged or older host behaviour is preserved for every agent type that resolved before; agents still unresolvable return nothing and the pane says so, which is now an honest message rather than "Up to date".
• Performance: one stat per open, plus one spawn of copilot help config when the probe runs (it reads local documentation, makes no network call, and is cached).
• Known gaps, unchanged here: Pi transcripts decode only their text blocks for this consumer — Pi's thinking and tool-call blocks use shapes the shared decoder does not read, so a Pi brief is text-only. That decoder is shared with Claude, Codex and OpenCode, so it is deliberately out of scope. OpenCode sessions still need a hook-supplied transcript path.

AI Disclosure

Written by an AI agent (Pi, running a github-copilot model) working directly in this worktree, under the fork owner's direction and review.

Review

Agent skill upstream boundary

• Not applicable — no skill installer source, tests, fixtures, registry entries, path tables, comments or documentation touched.