[brennanb2025]

	Files	Added	Deleted	Net
Test	24	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​1341	$\color{#cf222e}{\Huge{\mathbf{−}}}$​6	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​1335
Prod	36	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​581	$\color{#cf222e}{\Huge{\mathbf{−}}}$​219	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​362

What this fixes

orca orchestration worker-start --agent omp times out at the agent_readiness stage on OMP 18.4.x, so the worker never receives its task. It was reported on Windows 11 with OMP 18.4.4 (#24068).

Anyone who starts or messages OMP workers through Orca hits it: OMP 18.4 changed its idle window title to π > <folder>, and Orca only recognised the older π - <folder>.

Simply accepting π > would be unsafe. OMP shows that same title before its first-run setup wizard (a provider picker) opens, so Orca would type the task into the wizard instead of OMP's message box.

Not affected:

• Other agents' readiness rules (only the shared title cleanup below touches them).
• The renderer's selected-target send path.
• Hook status mappings.
• Pi's legacy title syntax.

ELI5

Before Orca types a task into an agent, it asks "is the agent sitting at its message box?". For OMP it now answers yes when OMP has said "idle" in its window title for a few seconds and OMP is not showing a full-screen overlay such as its setup wizard.

Separately, Orca has a three-second timer that tidies up a spinner title when a pane goes quiet. That tidy-up used to also tell Orca "the agent is idle", which could make a busy agent look ready. Now the tidy-up only changes what is shown; it no longer counts as the agent saying it is done.

What Changed

OMP readiness

Before: Orca had no rule that accepted OMP 18.4's idle title, so worker-start waited until it timed out.

After: for a pane Orca knows is OMP, the idle title π > counts as ready once two things hold:

• the title has stood unchanged for the 3-second readiness window;
• Orca's own copy of the screen is not on the alternate screen.

Why the window is measured from when the title appeared, not from the last output: once a terminal answers OMP's capability probe, OMP re-sends a "bracketed paste on" sequence every second, so the output never goes quiet on a real pane.

Why the alternate screen: OMP draws its setup wizard, splash screen and every other full-screen overlay there, and keys typed there drive the overlay, not the message box.

The alternate-screen check applies to every way an OMP pane can become ready, not only to π >.

If Orca has no trustworthy copy of the screen (no screen model, or a size mismatch), π > never counts, because nothing can rule out the wizard.

Every other OMP title goes through the normal rules, as before this change:

• the π - <folder> title Orca's title extension paints after a turn;
• OMP 17's π: <folder> title;
• busy and needs-input titles.

Agent titles versus displayed titles (all agents)

Before: when the 3-second timer cleared a stale spinner title, Orca wrote the cleared title into its records as if the agent had sent it. Readiness waits, queued-message delivery and "is an agent running here" checks then treated a guess as the agent's own signal.

After: the records keep only what the agent itself sent. The timer's cleared title is stored separately on the terminal's in-memory record and lives as long as that record. It is retired by the agent's next real title, by a process restart, or when the record is dropped. Nothing new is persisted and nothing new goes over the wire.

Every display surface reads through one accessor that applies the cleared title, so it keeps showing what it showed before this change:

• orca worktree ps;
• the phone's status and titles;
• terminal list;
• orchestration pointers;
• adopted and revealed tabs;
• snapshot titles.

The timer path is now an explicit branch, and it sorts what the old path did into three groups:

• Skipped, because these are the agent's own evidence: the recorded title and status, prompt state, readiness waiters, the mailbox and message delivery.
• Still run, because the timer exists for "the agent may have exited": a re-check of which program is in the foreground, and the existing recovery that confirms a finished Codex behind a fresh "done" hook.
• Still published, using the cleared title: title facts, phone updates and display.

"Is an agent running here?" checks, and the terminal status query, read the title as it is displayed (the cleared title, ordered exactly as the display orders it), which is what they read before this change. Where that title is compared with the prompt state to spot a dialog waiting on the user, the prompt state is cleared the same way, so the comparison reaches the same verdict it did before this change. Prompt-delivery checks judge a prompt sent behind a cleared title against that same cleared prompt state, as before; readiness waits and message delivery never read it. A busy cwd spinner such as ⠋ repo clears to the plain folder name, which names no agent, so for an agent that exited behind it the foreground program decides. A spinner that clears to an agent name such as Claude Code still counts as that agent, as before, and the answer no longer depends on whether a renderer pane is open.

Why

Alternatives considered and rejected:

• Accept π > as an idle title for every agent. OMP's setup wizard then becomes ready, and so does any OMP pane whose spinner the timer cleared.
• Add a "came from the timer" flag to the records (the approach in fix(runtime): treat Pi/OMP state-marker idle title as idle for terminal wait #24111, fix(runtime): recognize Pi/OMP state-marker idle titles in tui-idle detection #24113 and fix: correct managed OMP status and readiness detection #24196). The slower fallback rule, waiters and delivery still accept the cleared status. A renderer update can also silently drop a flag stored on pane records.
• Require OMP's "⇧⇥ to change thinking effort" hint in the message box. OMP retires that hint after three uses, hides it once a conversation exists, and draws it differently per glyph style. Most users and every post-turn wait would never be ready.
• Require output silence before trusting π >. OMP's once-a-second bracketed-paste sequence means real panes never go quiet. Live QA confirmed worker-start still timed out under that design.

Bug class removed: a display-only guess can no longer satisfy a readiness wait or message delivery for any agent. Presence and terminal status keep reading the displayed title, as they always did, so they match what the user sees.

An earlier draft said the timer false-ready "reproduces on main for all agents". That needs qualifying: it requires at least 3 seconds of total silence behind a static busy title. On main it gave an immediate ready only for OpenCode; for Claude, Codex, Gemini and Pi it gave a slower fallback ready after quiet. For hook-managed Claude and Codex, a fresh "working" hook vetoes it.

Known weaknesses

• Startup window. OMP shows π > before the rest of its startup runs (stream setup and extension start hooks), and only then opens the setup wizard. If that tail ever reached 3 seconds on a launch that shows setup, Orca could type the task before the wizard opens. Measured on real Orca-launched first runs on macOS: the wizard opened 20–28 ms after the π > title (3 runs), far below the 3-second window. On Windows 11 the gap was 17–26 ms (3 runs).
• Windows SSH hosts (unverified). On a Windows SSH host the relay runs OMP through the system console host. If that console host drops the switch to the alternate screen, the setup wizard would look like a normal screen, and π > would count 3 seconds after it appeared.
• Wall-clock title age. The 3-second title age is measured with the wall clock, so a forward clock jump can shorten the window.
• Lost identity after the first delivery (temporary).
  • After the first message, an OMP started through bun loses its OMP identity. OMP's message bubbles end a shell-integration "command finished" zone, which retires Orca's launch identity, and Orca cannot yet recognise OMP running under bun from the foreground process.
  • From then on, later idle waits on that pane behave as on main. Live QA saw this on both macOS and Windows: a readiness wait on the idle message box after the first delivery timed out with no agent identity.
  • Worker-start's first readiness is unaffected.
  • Follow-up: teach foreground-process recognition to see OMP under bun. This is pre-existing and not caused by this PR.
• Limited live coverage. Windows 11 (the reported platform) was checked live for OMP readiness only, at c42c224; SSH hosts, WSL, real Claude Code or Codex agents and a signed-in OMP turn were not checked live.
• OMP inside tmux or screen. If Orca does identify such a pane as OMP, it never becomes ready, because the whole pane stays on the alternate screen. That fails safe.
• Pi or unknown identity. Panes Orca has not identified as OMP still become weakly ready on OMP's setup screen after it goes quiet, as on main.
• Codex name title. A Codex pane stuck on a name-shaped spinner (⠋ Codex working) after a fresh "done" hook is no longer released by the cleared title. This is intended. The cwd-shaped case is still released through the existing completion recovery.
• Relaunch or restored title. After an Orca relaunch, or when the title came only from a restore, an idle OMP at π > has no title age and waits for its next title. This is not worse than main.

Linked Issue

Fixes #24068. Closes STA-9000.

Supersedes the title-readiness part of #24111, #24113 and #24196, with co-author credit to drakeo338, Ahmed Nagy and Saurav M Hiremath (verified commit identities). Their branches are untouched, and the broader hook-lifecycle work in #24196 stays separate.

Visual Proof

Live QA ran the built app hidden on a macOS laptop (M4 MacBook Air) and a high-spec Windows 11 machine, with OMP 18.4.5 installed in an isolated home and no model provider signed in. Each result names the commit it was taken at.

OMP worker-start on macOS, at d7543e5. The provider picker ("Setup step 1 of 5") never became ready and no task text reached it:

Setup step 4 draws a preview of the message box inside the wizard; the worker correctly stayed not ready there:

After the last Esc the worker became ready 2.39 s later and the task was typed exactly once (OMP then reports it has no model, as expected without a provider):

A second launch on the finished home showed no wizard, became ready in 6.11 s, and received the task once:

Before the title-age change (at 995eabe) the same run timed out after setup, because OMP re-sends a bracketed-paste sequence every second (measured live: every 1.000 s, its only idle output), so the output never went quiet.

OMP worker-start on Windows 11 (high spec), at c42c224. Setup step 1 was not ready and no task text reached the picker; the worker became ready 1.87 s after the last Esc and the task was typed once; a second launch was ready in 6.27 s with the task once:

Gap between OMP's π > title and the setup screen opening, at c42c224: 20–28 ms on macOS and 17–26 ms on Windows 11 (3 first runs each), far below the 3-second window. The setup screen stayed on the alternate screen throughout and the worker never became ready during setup.

Claude Code and OpenCode stand-ins, at 995eabe and again at c42c224. These are scripted stubs, not the real agents. Each worker-start became ready once and received its task once. A busy stub that went silent behind its working title was not made ready and received no input. After the 3-second timer the display cleared as before this change: the tab title lost its spinner, the sidebar row went from working to idle, and orca worktree ps reported active.

Claude Code stub, busy and silent, after the timer (at c42c224): cleared tab, idle worker row:

OpenCode stub, busy and silent, after the timer (at 995eabe): cleared tab, idle worker row:

Not live-verified at the final head. The later commits aa76cbd through bc7d34e (presence, terminal status, blocked-dialog and prompt-verification behaviour matching main while a cleared title stands) are proven by unit tests against main's values, not by a live run.

Not covered live: real Claude Code or Codex agents, SSH hosts, WSL, and an OMP turn with a signed-in provider.

Testing

• I manually tested these changes locally (isolated OMP 18.4.5 captures, plus live QA on macOS and Windows 11 shown under Visual Proof)
• Automated tests added/updated

Captured fixtures (scrubbed real OMP 18.4.5 bytes, raw escapes kept):

• setup wizard at 120×40;
• message box after skipping setup, at 120×40;
• message box at 60×24.

Replaying the scrubbed and unscrubbed captures through the terminal serializer shows the same, already-known cursor-after-shrink divergence. It is recorded in the serializer test's known-baseline list; the serializer itself is unchanged.

New OMP tests:

• setup is never ready;
• the message box is ready with the 1-second bracketed-paste sequence running, and setup is not;
• the splash screen (alternate screen, no step heading) is refused;
• hintless and ASCII-glyph message boxes are ready (synthetic repaints, labelled as such);
• the post-turn π - title is ready;
• OMP 17's π: title takes the normal path;
• an answer that mentions "Setup step 2 of 4" on the normal screen does not block;
• with no trustworthy screen, π > is refused while π - still passes;
• Pi and unknown identities are not made ready immediately;
• a "waiting on the user" OMP stays pending so the screen check still runs.

New display and presence tests:

• worktree ps, the phone and terminal list show the cleared title and status;
• a restore re-seed does not bring "working" back;
• an SSH relay drop and same-session reattach keep the cleared display;
• an agent that exits loses its foreground identity;
• an agent that exited behind a cwd spinner cannot pass the send guard, including while the renderer still shows the spinner;
• a live agent is still found by its process, and terminal status reports what it reported before;
• a live, busy agent whose process Orca cannot recognize stays found through its cleared name title, with or without a renderer pane;
• a trust dialog painted after the clear still reads as a wait on the user: terminal status says permission, the guarded send refuses, and terminal show reports the dialog, with or without a renderer pane;
• a prompt sent behind a cleared spinner is judged as before: output alone after Enter still reads as stalled, and the agent's next spinner frame counts as the prompt's turn starting;
• a live Cursor spinner behaves as on main;
• a Codex completion behind a fresh "done" hook still delivers queued mail.

Each new test was run with its fix removed and failed. That covers 27 separate removals across six rounds, for example:

• restoring "output silence" fails the bracketed-paste test;
• restoring the heading check fails the "Setup step" answer test;
• dropping the foreground re-check fails both identity tests.

Named test files: 54 files, 2308 passed, 11 skipped, 0 failed, at the final commit.

Lint, formatting and the changed-code quality check pass. Typecheck was run separately.

Not run: a real SSH relay, WSL, and a full build. Windows 11 and Orca-launched OMP first runs with extensions loaded were checked live (see Visual Proof).

AI Disclosure

Review

Readiness is still decided on the execution host. There is no new wire field, message type or persisted field, and older clients receive the same display facts as before. Folder workspaces are covered: none of this depends on Git. A lost connection is never treated as proof that a process exited.

Agent skill upstream boundary

• Not applicable; no skill-installer source or assets are copied.

Notes

Author: @BrennanKB5. Draft; not ready for merge.

Co-authored-by: drakeo338 paranoyouz@gmail.com

Co-authored-by: Ahmed Nagy ahmednagy25t@gmail.com

Co-authored-by: Saurav M Hiremath sauravhiremath@gmail.com

Checklist

• This PR is small and focused (OMP readiness plus the shared title-evidence defect it depends on)
• 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 (live QA screenshots under Visual Proof)
• Self-reviewed for correctness, security, and performance
• Cross-platform, SSH/remote, and path/shortcut impact considered
• pnpm lint, pnpm typecheck, pnpm test, and pnpm build pass (named tests, lint, quality gate and typecheck run; CI covers the rest)

[brennanb2025]

Review status (draft; not marked ready, not merged)

The problem: orca orchestration worker-start --agent omp timed out at agent readiness on OMP 18.4.x (reported on Windows 11, OMP 18.4.4), so the worker never got its task. OMP 18.4 shows a new idle title, π > <folder>, that Orca did not accept. OMP also shows that same title while its first-run setup wizard is open, so simply accepting it would have typed the task into the provider picker.

What changes for users: an OMP worker now becomes ready once OMP's idle title has stood for 3 seconds and OMP is not showing a full-screen overlay (its setup wizard). The task is typed once into the message box, never into the wizard. Separately, for every agent, the 3-second timer that tidies a stale spinner title now only changes what is displayed. It can no longer make a busy agent look ready. Tab titles, the sidebar, orca worktree ps, the phone and terminal status still show exactly what they showed before.

Fixed during review (all in new commits on this branch, head bc7d34e):

• The first version required an onboarding hint that OMP stops showing (after three uses, after the first message, with other glyph styles), so most users and every later wait would never be ready. Replaced by refusing only while OMP is on its full-screen setup screen.
• The first version also required 3 s of output silence, but OMP re-sends a bracketed-paste sequence every second, so a real OMP pane never went quiet and worker-start still timed out (confirmed live). Readiness now counts from when the idle title appeared.
• π > was briefly accepted for every agent. It is now accepted only for panes identified as OMP, so other panes keep main's behaviour.
• Keeping the agent's own title on the records at first left the phone and orca worktree ps stuck on "working", let a reconnect bring "working" back, dropped the foreground-program re-check and the Codex completion recovery, and let a send reach the shell of an exited agent. All of these now match main: display surfaces go through one accessor, and the timer runs as an explicit branch that skips only the agent's own evidence.
• Follow-on mismatches (presence of a live agent with an unrecognised program, a trust dialog after the timer, prompt-delivery checks behind a cleared title) were fixed to match main's results.
• A CI failure from the new OMP recordings was the terminal serializer's existing cursor-after-shrink limitation, the same one already recorded for other recordings. It is recorded in that test's known list; the serializer is unchanged.

Deferred:

• After the first delivery, an OMP started through bun loses its OMP identity (OMP's message bubbles end a shell "command finished" zone, and Orca cannot recognise OMP under bun), so later idle waits on that pane time out. This is the same on main and not caused by this PR; it needs a follow-up ticket for recognising OMP under bun.
• Nits: a same-millisecond tie between a hook and the timer, the replay snapshot's raw title, a repeated status re-detection, and one display field the projection does not carry.

Verified:

• Live on macOS (M4 MacBook Air) and Windows 11 with OMP 18.4.5 in an isolated home: never ready during setup and nothing typed into the wizard; ready 2.39 s (macOS) / 1.87 s (Windows) after setup is skipped, task typed once; later launches ready in about 6 s. Screenshots are in the PR body.
• OMP's idle title comes 17–28 ms before its setup screen opens on both platforms, far inside the 3-second window.
• Claude Code and OpenCode stand-in scripts: ready once and task once; a busy, silent agent was not made ready; the display cleared after the timer exactly as before.
• Unit tests compare presence, terminal status, blocked dialogs and prompt-delivery checks against main's values across 465 scenario results; they are identical except when a hook and the timer land in the same millisecond (the deferred nit above). Every new test was run with its fix removed and failed. 54 named test files pass (2308 tests), the node typecheck passes, and lint and the changed-code quality check pass. CI on bc7d34e is green: all 5 test shards, typecheck, static analysis, relay integration, packaging and the Windows/Linux/macOS native smoke tests pass (18 pass, 18 skipped, 0 failed).

Not verified: the last three code commits (aa76cbd, bdb6e8f, bc7d34e: presence, status, blocked-dialog and prompt-delivery parity) were not re-run live. Real Claude Code or Codex agents, SSH hosts (including Windows SSH hosts, where the remote console host might drop the setup-screen switch), WSL, and an OMP turn with a signed-in provider were not checked live. A full build was not run locally (all changes are in the main process, which the node typecheck covers); CI covers the rest.

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Updates terminal readiness detection and title handling for OMP 18.4.

The reviewed changes appear safe to merge; no actionable regression was established.

Summary

This PR adds guarded readiness for OMP 18.4’s π > title and separates timer-cleared display titles from the agent’s own readiness evidence.

• OMP readiness requires title age and a trustworthy screen outside the alternate screen.
• Terminal, phone, worktree, and orchestration displays project the cleared title without rewriting native title records.
• Transcript and regression tests cover setup refusal, composer readiness, and display behavior.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Agent OSC title] --> B[Native PTY and leaf records]
  B --> C[Readiness and delivery]
  B --> D[Display projection]
  T[Stale-title timer] --> E[Display-only clear]
  E --> D
  S[OMP screen model] --> C
  D --> F[Terminal, phone, worktree and orchestration views]

Loading

_{Reviews (1) · Last reviewed commit: "fix(runtime): verify prompts behind a st..."}

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Important

Review skipped

We couldn't safely recover the incremental review. No full review was started, and the last reviewed checkpoint was preserved. Retry later, or explicitly request a full review by commenting @coderabbitai full review.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

The change adds OMP readiness checks based on ruled-screen input vetoes and quiescent idle titles. It also adds a display-only clear for stale working titles and applies that clear to terminal, worktree, mobile, lifecycle, and prompt-status projections. New tests and OMP transcript fixtures cover readiness, title clearing, agent presence, and related terminal interactions.

Priority: ➖ Normal

Severity of issue fixed: Medium

Merge Risk: 🟡 Moderate · up to bc7d3

OMP readiness can occasionally treat a setup screen as ready if the screen-mode change has not yet been applied, so a task may be sent into the setup wizard. Fix the write-ordering gap before merging. A new title-age test also depends on the real clock and could fail intermittently.

Security Architecture Review

Security architecture risk: 🔵 Low · up to bc7d3

The change prevents cosmetic status cleanup from being mistaken for readiness to receive work and adds checks before sending tasks. Unusual startup and recovery conditions remain incompletely verified.

Retained concerns
No architecture-level concerns identified.

Security review details

Security Blast Radius

• inferred — The inspected authority path affects task input to a PTY and its attached terminal leaves. The shared clear projection also propagates into worktree and mobile status surfaces. This establishes local task-delivery and status scope, not expanded tenant, credential, or service authority.

Trust Boundaries and Controls

• observed — Timer-generated clears cannot directly establish native idle evidence: their branch avoids native title/status mutation, idle-waiter resolution, and mailbox idle notification. This separates a cosmetic inference from observations used to authorize delivery.
• observed — Title-triggered waiters re-evaluate readiness, and queued delivery rechecks the same evaluator before injection. The new OMP title lane remains pending without readable screen evidence; ruled-screen reads reject pending hydration, partial models, unrepainted reflow, and grid-size mismatch.

Resilience and Maintainability Implications

• observed — Projection cleanup is tied to genuine title observations and provider-generation reset. These transitions contain stale-state inheritance across process replacement, which is important because readiness and delivery depend on process-specific evidence.

Hardening Proposals

• proposed — Consider validating readiness against an explicit composer signal and screen/output-sequence agreement before broadening supported title or overlay conventions. This would strengthen the current heuristic; it is not a verified PR-introduced vulnerability.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 36.67% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 30 functions across 41 files. (4 skipped:…	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes both primary changes: OMP 18.4 readiness at the composer and display-only stale-title clears.
Description check	✅ Passed	The description is complete and directly matches the template. It explains the user impact, implementation, rationale, linked issue, visual proof, testing, limitations, compatibility considerations, a…
Linked Issues check	✅ Passed	Issue [#24068] requires fresh idle OMP panes with the π > title to reach readiness. The PR limits this evidence to identified OMP panes, requires the title to remain observed for the quiescence inte…
Out of Scope Changes check	✅ Passed	The title-display-clear changes support the readiness objective. They prevent a stale spinner clear from becoming agent evidence while preserving the cleared title on display surfaces. The related pro…

Full details: Docstring Coverage

Explanation

Docstring coverage is 36.67% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 30 functions across 41 files. (4 skipped: 4 unsupported.)

✨ Finishing Touches 💡 1

📝 Generate docstrings 💡

• Commit to this branch
• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 2

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Advanced

Run ID: 9459e4f8-acac-4098-bed8-0279f51da44f

📥 Commits

Reviewing files that changed from the base of the PR and between 3135fbb and bc7d34e.

📒 Files selected for processing (47)

• src/main/daemon/serialize-grid-transcript-replay.test.ts
• src/main/runtime/__fixtures__/omp-18-composer-narrow.meta.json
• src/main/runtime/__fixtures__/omp-18-composer-narrow.txt
• src/main/runtime/__fixtures__/omp-18-composer.meta.json
• src/main/runtime/__fixtures__/omp-18-composer.txt
• src/main/runtime/__fixtures__/omp-18-setup.meta.json
• src/main/runtime/__fixtures__/omp-18-setup.txt
• src/main/runtime/agent-transcript-replay-test-harness.ts
• src/main/runtime/codex-quiet-ready-screen.test.ts
• src/main/runtime/omp-readiness-transcripts.test.ts
• src/main/runtime/omp-terminal-readiness.ts
• src/main/runtime/orca-runtime-adopt-terminal-orphans-from-inventory.ts
• src/main/runtime/orca-runtime-apply-tracked-pty-title.ts
• src/main/runtime/orca-runtime-build-pty-terminal-summary.ts
• src/main/runtime/orca-runtime-emit-daemon-pty-transient-fact.ts
• src/main/runtime/orca-runtime-focus-terminal.ts
• src/main/runtime/orca-runtime-get-terminal-interactive-wait.ts
• src/main/runtime/orca-runtime-get-unpersisted-tracked-title-for-pty.ts
• src/main/runtime/orca-runtime-get-worktree-ps.ts
• src/main/runtime/orca-runtime-has-exact-persisted-terminal-surface-identity.ts
• src/main/runtime/orca-runtime-maybe-hydrate-headless-from-renderer.ts
• src/main/runtime/orca-runtime-prune-mobile-session-tab-group-layout.ts
• src/main/runtime/orca-runtime-publish-pty-backed-mobile-session-terminal.ts
• src/main/runtime/orca-runtime-record-agent-prompt-lifecycle-state.ts
• src/main/runtime/orca-runtime-resolve-authoritative-terminal-wait-permission.ts
• src/main/runtime/orca-runtime-resolve-exit-waiters.ts
• src/main/runtime/orca-runtime-runtime-id.ts
• src/main/runtime/orca-runtime-serialize-agent-prompt-submission.ts
• src/main/runtime/orca-runtime-stop-requested-pty-ids.ts
• src/main/runtime/orca-runtime-visible-snapshot-preview.ts
• src/main/runtime/orca-runtime-write-orchestration-pointer-pty.ts
• src/main/runtime/runtime-mobile-session-projection-contract.ts
• src/main/runtime/runtime-mobile-session-projection.ts
• src/main/runtime/runtime-terminal-agent-presence.ts
• src/main/runtime/runtime-terminal-agent-status-query.ts
• src/main/runtime/runtime-terminal-orphan-adoption.ts
• src/main/runtime/runtime-terminal-orphan-session-adoption.ts
• src/main/runtime/runtime-terminal-state-records.ts
• src/main/runtime/runtime-worktree-ps-activity.ts
• src/main/runtime/runtime-worktree-status-projection.ts
• src/main/runtime/screen-ruled-agent-readiness.ts
• src/main/runtime/stale-title-completion-recovery.test.ts
• src/main/runtime/stale-title-display-clear.test.ts
• src/main/runtime/stale-title-display-surfaces.test.ts
• src/main/runtime/structured-worker-agent-presence.test.ts
• src/main/runtime/tui-idle-evidence.test.ts
• src/main/runtime/tui-idle-evidence.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 8 remain after this review.

[coderabbitai]

🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

🔎 Supported by static analysis

🏁 Script executed:

# Map emulator write handling before inspecting its implementation.
fd '^headless-emulator\.ts$' src |
while IFS= read -r file; do
  ast-grep outline "$file" --items all --match 'write|tryWriteSync'
  rg -n -C 10 'tryWriteSync|forwardQueryReplies|terminal\.write' "$file"
done

# Locate ingestion and readiness dispatch, including pending-write tracking.
rg -n -C 6 --glob '*.ts' \
  'onPtyData\s*\(|writeChain|resolveTuiIdleWaiters\s*\(|resolvePtyTuiIdleWaiters\s*\(|checkDeliverySettledAndArmRecheck\s*\(' \
  src/main/runtime

Repository: stablyai/orca

Length of output: 45651

---

🏁 Script executed:

set -eu
printf '%s\n' '--- headless emulator write and construction ---'
sed -n '120,230p' src/main/daemon/headless-emulator.ts
printf '%s\n' '--- runtime PTY ingestion definitions ---'
rg -n -C 18 '^\s*(public\s+)?onPtyData\s*\(|onPtyData\s*\(' src/main/runtime --glob '*.ts' | head -n 500
printf '%s\n' '--- emulator writes in runtime ---'
rg -n -C 12 'headless.*\.write|\.emulator\.write|terminal\.write|writeSync\(' src/main/runtime src/main/daemon --glob '*.ts' | head -n 500
printf '%s\n' '--- readiness and delivery definitions ---'
rg -n -C 20 'resolveTuiIdleWaiters|resolvePtyTuiIdleWaiters|checkDeliverySettledAndArmRecheck|deliverPendingMessagesForLeaf|readRuledScreen|alternateScreen' src/main/runtime --glob '*.ts' | head -n 700

Repository: stablyai/orca

Length of output: 42931

---

🏁 Script executed:

set -eu
printf '%s\n' '--- changed file ---'
sed -n '1,240p' src/main/runtime/orca-runtime-visible-snapshot-preview.ts
printf '%s\n' '--- PTY ingestion and runtime writes ---'
rg -n -C 18 'onPtyData|\.write\(|writeSync\(|headless|emulator' src/main/runtime/orca-runtime*.ts src/main/runtime/*pty*.ts src/main/runtime/*delivery*.ts src/main/runtime/*idle*.ts --glob '*.ts' | head -n 1200
printf '%s\n' '--- exact readiness symbols ---'
rg -n -C 30 'resolveTuiIdleWaiters|resolvePtyTuiIdleWaiters|checkDeliverySettledAndArmRecheck|deliverPendingMessagesForLeaf|readRuledScreen|alternateScreen' src/main/runtime --glob '*.ts' | head -n 1400

Repository: stablyai/orca

Length of output: 45625

---

🏁 Script executed:

set -eu
printf '%s\n' '--- production onPtyData declarations ---'
rg -n 'onPtyData' src/main/runtime --glob '*.ts' --glob '!*.test.ts'
printf '%s\n' '--- production writeChain references ---'
rg -n -C 12 'writeChain' src/main/runtime --glob '*.ts' --glob '!*.test.ts'
printf '%s\n' '--- production headless emulator writes ---'
rg -n -C 12 'emulator\.write|\.emulator\.write|HeadlessEmulator' src/main/runtime --glob '*.ts' --glob '!*.test.ts'
printf '%s\n' '--- delivery/readiness callers ---'
rg -n -C 16 'checkDeliverySettledAndArmRecheck|deliverPendingMessagesForLeaf|resolvePtyTuiIdleWaiters|resolveTuiIdleWaiters|readRuledScreen' src/main/runtime --glob '*.ts' --glob '!*.test.ts'

Repository: stablyai/orca

Length of output: 45599

---

🏁 Script executed:

set -eu
printf '%s\n' '--- exact onPtyData declarations ---'
rg -n 'onPtyData[[:space:]]*([:=]|\\()|onPtyData[[:space:]]*\\(' src --glob '*.ts' --glob '!*.test.ts' --glob '!**/tests/**' --glob '!**/*test*'
printf '%s\n' '--- exact writeChain assignments and reads ---'
rg -n 'writeChain[[:space:]]*=' src/main --glob '*.ts' --glob '!*.test.ts'
rg -n -C 8 'writeChain' src/main/runtime/runtime-terminal-state-records.ts src/main/runtime/orca-runtime-core.ts src/main/runtime --glob '*.ts' --glob '!*.test.ts' | grep -E 'writeChain|emulator\\.write|onPtyData|resolve|deliver' | head -n 300
printf '%s\n' '--- possible inherited method files ---'
rg -n 'class OrcaRuntime|extends .*OrcaRuntime|onPtyData' src/main --glob '*.ts' --glob '!*.test.ts' | grep -v '/tests/' | grep -v '\\.spec\\.ts'

Repository: stablyai/orca

Length of output: 328

---

🏁 Script executed:

set -eu
printf '%s\n' '--- onPtyData declarations ---'
rg -n -F 'onPtyData(' src/main --glob '*.ts' --glob '!*.test.ts' --glob '!**/*.spec.ts'
printf '%s\n' '--- writeChain assignments ---'
rg -n -C 12 'writeChain[[:space:]]*=' src/main --glob '*.ts' --glob '!*.test.ts' --glob '!**/*.spec.ts'
printf '%s\n' '--- runtime classes and inheritance ---'
rg -n -C 8 'class OrcaRuntime|extends .*OrcaRuntime|onPtyData' src/main --glob '*.ts' --glob '!*.test.ts' --glob '!**/*.spec.ts' | head -n 300

Repository: stablyai/orca

Length of output: 41516

---

🏁 Script executed:

set -eu
printf '%s\n' '--- onPtyData implementation ---'
cat -n src/main/runtime/orca-runtime-on-pty-data.ts
printf '%s\n' '--- direct headless tracking definitions ---'
rg -n -C 30 'trackHeadlessTerminalData|createPtyHeadlessTerminalState|outputSequence|emulator\.write' src/main/runtime --glob '*.ts' --glob '!*.test.ts' --glob '!**/*.spec.ts' | head -n 700
printf '%s\n' '--- PTY listener call ordering ---'
sed -n '45,95p' src/main/ipc/pty/provider/bind-listeners.ts
sed -n '105,140p' src/main/ipc/pty/provider/local-configure.ts

Repository: stablyai/orca

Length of output: 42271

---

🏁 Script executed:

set -eu
printf '%s\n' '--- trackHeadlessTerminalData definition ---'
rg -n -C 40 'trackHeadlessTerminalData' src/main/runtime --glob '*.ts' --glob '!*.test.ts' --glob '!**/*.spec.ts'
printf '%s\n' '--- headless state contract ---'
rg -n -C 25 'type RuntimeHeadlessTerminal|interface RuntimeHeadlessTerminal|writeChain:' src/main/runtime --glob '*.ts' --glob '!*.test.ts' --glob '!**/*.spec.ts' | head -n 500

Repository: stablyai/orca

Length of output: 41752

---

🏁 Script executed:

set -eu
printf '%s\n' '--- readRuledScreen consumers ---'
rg -n -C 24 'readRuledScreen|alternateScreen' src/main/runtime --glob '*.ts' --glob '!*.test.ts' --glob '!**/*.spec.ts'
printf '%s\n' '--- delivery gate implementation ---'
rg -n -C 36 'checkDeliverySettledAndArmRecheck' src/main/runtime --glob '*.ts' --glob '!*.test.ts' --glob '!**/*.spec.ts'

Repository: stablyai/orca

Length of output: 42148

---

Wait for the headless write before evaluating delivery.

onPtyData() queues the PTY chunk, then evaluates the title edge synchronously. The queued emulator write runs later. readRuledScreen() can therefore read alternateScreen: false before the transition applies. The overlay veto can be bypassed, and pending input can be delivered into the setup wizard.

Defer readiness and delivery evaluation until the corresponding model write completes. Preserve per-PTY write ordering.

[coderabbitai]

🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win

Freeze the clock for the title-age assertions.

This case places the title only 1,000 ms below the readiness threshold. If execution pauses for that interval before evaluation, the title becomes ready and the assertion on Line 352 fails.

Use fake timers and a fixed system time within this new describe block. Restore real timers after each test.

Based on learnings: “do not rely on actual system time” in time-dependent tests; use a fake clock.

Source: Learnings

[pullfrog]

ℹ️ No critical issues — minor suggestions inline.

Reviewed changes

Reviewed the full bc7d34e diff (47 files, 12 commits) plus the surrounding readiness, presence, and status-query code.

• OMP tui-idle lane — tui-idle-evidence.ts accepts OMP's π > idle title only after it ages a quiescence window on a readable, non-alternate screen, with a screen input veto (screen-ruled-agent-readiness.ts / omp-terminal-readiness.ts) that downgrades any ready verdict while OMP's setup overlay is on screen.
• Display-only stale-title clear — the 3 s timer's clear moves to titleDisplayClear on the PTY record and is projected at every display/status read via getPtyDisplayRecord / getLeafDisplayRecord / getDisplayPromptLifecycle, while readiness and delivery keep reading the native title.
• Native-first title precedence — hasExplicitIdleTitle and getAdoptedPtyExplicitIdleStatus now prefer the native OSC title over renderer/leaf titles, so a renderer-echoed cleared title cannot satisfy a wait.
• Tests and fixtures — scrubbed OMP 18.4.5 captures with replay suites, plus display/presence/prompt-verification coverage.

I verified the display projection reproduces main's in-place rewrite (the timer emits rawTitle === normalizedTitle === cleared, so clear.status recomputes identically; Math.max reproduces the seed re-stamp), and that the native-first precedence is safe for genuine flows because real OSC titles, hook synthetic terminal frames, and restore seeds all write both native records. The one inline note is about a test that is weaker than its name.

ℹ️ Nitpicks

• The new projection doc says records keep the agent's own title for "readiness, delivery and presence" (src/main/runtime/runtime-worktree-status-projection.ts:66-70), and the record field doc says "evidence never reads it" (src/main/runtime/runtime-terminal-state-records.ts:90-93) — but RuntimeTerminalAgentPresence.isRunning now projects through getPtyDisplayRecord. Worth updating so the evidence-vs-display split stays legible.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using DeepSeek Flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

This 1_000 ms timeout is below the 3 000 ms quiescence window every lane uses, so the pane stays pending for any agent identity. With the input.agent === 'omp' gate removed (tui-idle-evidence.ts:308), the case still passes: the OMP lane's title-age clause is unmet at 1 s, and readScreenInputVeto is null for non-OMP agents. Passing a timeout above quiescence, or asserting a behavior that distinguishes the gate, would make this case actually pin what its name claims.

[pullfrog]

ℹ️ No critical issues — one scope/coverage question inline.

Reviewed changes

Reviewed the commits added since the last pullfrog review (bc7d34e → ab9a4e3). The only new commit is a merge of origin/main, so the delta is main's agent-state-rules/tier-0-hook/readiness-census refactor reconciled with this PR rather than new feature work.

• Readiness evidence split — tui-idle-evidence.ts was split, moving per-site evidence gathering to a new tui-idle-evidence-source.ts and isolating the OMP screen veto in screen-input-veto.ts (renamed from screen-ruled-agent-readiness.ts).
• Ranking reconciled with main — the OMP π > title-age lane and the over-verdict screen veto now sit above main's new readAgentRuleVerdict and tier-0 hook lanes without changing their behavior; no dangling imports of the removed file or symbols remain.
• Census fixtures — OMP composer/setup/native-title recordings were added to the readiness census.
• Test input plumbing — the new readScreenInputVeto / titleObservedAtEpochMs fields were added to the affected test inputs.

I confirmed the refactor is behavior-preserving and that the display projection stays wired at every read site. The one inline note is a scope/coverage question about the veto on an unreadable grid.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}

[pullfrog]

readScreenInputVeto answers null — not true — when the agent has a veto rule but the ruled grid is unreadable (readRuledScreen is null on a restored/pending-hydration pane, a reflowed grid, or a PTY/grid size mismatch). The OMP π > lane guards against that locally by requiring === false, but only for the native π > title; a name-only OMP or legacy π: title falls through to the rule/sustained lanes, which can return ready-weak once the stream is quiet — with no wizard check. The new matrix pairs π: with a readable wizard and π >/π - with no screen, but never π:/name-only with no screen. If typing into the wizard is in scope for those titles, treat an unreadable grid as blocking for OMP titles other than the post-turn π -; otherwise a case pinning it as intentionally unguarded would make the scope explicit.

Technical details

# Unreadable ruled grid leaves non-`π >` OMP ready lanes unguarded

## Affected sites
- `src/main/runtime/screen-input-veto.ts:16-23` — returns `veto && screen ? veto(screen) : null`; a null screen yields null (no veto).
- `src/main/runtime/tui-idle-evidence.ts:258-265` — `evaluateTuiIdle` downgrades a ready verdict only on `readScreenInputVeto() === true`.
- `src/main/runtime/tui-idle-evidence.ts:305-315` — the OMP lane requires `=== false`, but only fires when `isOmpIdleStateTitle(lastOscTitle)` (native `π >`).
- `src/main/runtime/tui-idle-evidence.ts:334-342` — `readAgentRuleVerdict` (omp.json: title+idle → weak idle requiring quiet) and `hasSustainedTitleIdle` can return `ready-weak` for other OMP titles.
- `src/main/runtime/omp-readiness-transcripts.test.ts:188-213` — covers `π:` with a readable wizard and `π >`/`π -` with no screen, but not `π:`/name-only with no screen.

## Required outcome
- Decide whether an OMP pane whose current title is not the native `π >` (legacy `π:` or name-only `OMP`) and whose ruled grid is unreadable may still settle `ready-weak`, or whether it must be blocked the way `π >` is — while keeping the post-turn `π -` path ready with no screen (pinned at `omp-readiness-transcripts.test.ts:204-212`).

## Open questions for the human
- Reachability is narrow: it needs a non-native OMP title plus a quiet output clock (`hasQuietOutput`), which the 18.4.5 bracketed-paste keepalive normally prevents. Is this path in scope for the PR's wizard protection, or intentionally left to the normal rules?