[brennanb2025]

	Files	Added	Deleted	Net
Test	12	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​1205	$\color{#cf222e}{\Huge{\mathbf{−}}}$​124	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​1081
Prod	27	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​791	$\color{#cf222e}{\Huge{\mathbf{−}}}$​209	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​582

ELI5

When you talk to OpenCode, Grok, OMP or Pi in Orca's chat view, the conversation still shows up in Session History as a plain terminal conversation, not as your chat. Clicking resume on it runs the agent in a terminal (for example opencode --session <id>) on the same conversation your chat is still using, so two programs write to one conversation at once. Claude and Codex chats already avoid this. This PR gives the other four agents the same behaviour.

What Changed

User-facing:

• Your chat's conversation opens your chat. In Session History (the desktop list and search), opening an OpenCode, Grok, OMP or Pi conversation that a chat owns opens the chat. Copy-resume and delete are hidden, as for Claude and Codex.
  • The row keeps the title the agent's CLI saved, the same title its terminal conversations show (Pi saves none, so it shows the first prompt). The chat tab keeps its own name, and renaming the tab doesn't change Session History, just as for terminals.
  • Claude and Codex rows are unchanged: they still show the chat's name.
• A terminal resume of a chat's conversation is refused. A resume Orca launches for such a conversation (from Session History, a session tab, mobile, or orca terminal send) gets the same "this conversation is open in a chat" refusal Claude and Codex already get.
  • This includes resumes that name no conversation: opencode -c, grok -c, a bare grok --resume, omp -c, a bare omp --resume, pi -c and a bare pi -r.
  • They are refused while any saved chat of that agent exists, because they could pick the chat's conversation. That is the existing rule for claude -c.
• Mobile builds that can't open these agents' chats no longer see these conversations. Tapping one there would have reopened it in a terminal, as a second writer. Mobile hides them, as clients without chat support already do.
• Search results no longer offer a terminal resume command for a conversation a chat owns, for any agent, Claude and Codex included.
  • For example, orca CLI search no longer prints resume: claude --resume <id> for a chat's conversation.
  • Opening the result opens the chat.
• The resume check reads a command line the way a shell would, for every agent, Claude and Codex included.
  • It splits text into commands at &&, ||, ;, |, & and line breaks, also when they touch a word (opencode -s <id>; echo, cd /x&&claude -r <id>), but never inside quotes. A line ending in \, a backtick or ^ continues onto the next line.
  • It removes quotes the way a shell does, so --session="<id>" reads as the id.
  • An agent counts only as a command's program: the first word after VAR=x prefixes, a path, or env, exec, command, sudo, nohup, time, npx, bunx, pnpm dlx or wsl --.
  • A name anywhere else is an argument or prompt text. So "Fix the Pi --resume flag", sent while a Pi chat exists, is no longer refused.
  • Resume ids compare ignoring case.
  • What this changes for Claude and Codex:
    • Newly refused: a resume of the owned conversation with a separator glued on (claude -r <owned>;), a quoted value after =, or the id in another case.
    • Also newly refused: claude -r <other> ; claude -r <owned>. Main read only the first mention.
    • Newly allowed, because the agent is not the program: echo claude -r <owned>, an agent run through a wrapper not in the list above (xargs, an option value such as sudo -u bob), or nested in bash -c '…'.
    • Still refused: cd ~/src/codex && claude --resume <owned>.
    • Still allowed: claude ; echo --resume <id>.
• "Resume in New CLI" now works for Grok and Pi chats too. It opens a copy of the conversation in a terminal and leaves the chat alone, as for Claude and Codex (grok --resume <id> --fork-session, pi --fork <session file>). It shows in the Session History row and the chat tab's right-click menu.
  • OpenCode's interactive CLI and OMP have no way to open a copy, so their chats keep only Resume (which opens the chat). A follow-up will get the copy from the running chat instead.
• Unchanged here: "Resume in New Native Chat" (turning a terminal conversation into a chat) is still Claude and Codex only. Grok and OMP get it in Resume in New Native Chat for Grok and OMP terminal conversations #28049, stacked on this PR; OpenCode and Pi follow.

Mechanism:

• Each agent's registration says how its chats own Session History rows. Every agent registered for native chat carries a required sessionHistory entry:

  • which history rows it owns;
  • the row id for a chat's conversation;
  • how to read a terminal resume of it.

  Shared code reads these entries instead of checking agent names.

  Agent	Rows it owns	Row id	Resume forms it reads
  Claude	claude	the session id	--resume/-r, --continue/-c; --fork-session is allowed
  Codex	codex	the thread id	resume; codex fork is allowed
  OpenCode	opencode and opencode2	the ses_ id	--session/-s
  Grok	grok	the id	--resume/-r/--load; a value that isn't an id is a title lookup, so it's refused
  OMP	omp	the id in the session file's header	--resume/-r/--session, which also accept an id prefix or a file path; a bare --resume opens a picker, so it's refused
  Pi	pi	the id in the session file's name (Pi's chat stores the file path)	--session with an id, prefix or path, and --session-id (reopens that id when it exists); a bare -r/--resume opens a picker, so it's refused

  For the four new agents, --continue/-c names no conversation and fork flags are allowed.

• Adoption reads the registration too. The registration's transcriptImport says how to find a conversation the agent's own CLI recorded.

  • Claude and Codex fill it in; the others leave it empty.
  • The host refuses to adopt for an agent without one, as before.
  • This moves the Claude/Codex branches out of the shared adoption code.

• One ownership index. The duplicate-adoption check and Session History both read one list of which chat owns which conversation. Each row id comes from the owning agent's registration. A test pins that no two registrations claim the same kind of history row.

• Per-client row visibility. Session History decides per row whether the requesting client can open that agent's chat, reusing the existing "does this client render this agent" check. Claude and Codex rows behave as before for every client.

• Nothing new crosses the wire. Clients and remote hosts exchange the same messages as before; ownership is worked out on the host that runs the chat.

Why

This is the third bug of one kind: a feature written for Claude and Codex never reached the newer agents. The earlier two were orchestration workers (#26776) and the floating-workspace folder (#26981).

Each agent's registration now answers "which history rows does this agent's chat own". A future agent gets this behaviour by filling in its own entry, instead of someone hunting down agent-name checks spread through shared code.

Matching the command text is still needed because mobile, including older mobile builds, sends resumes as plain text, and orca terminal send carries typed commands.

Alternatives considered:

• Publishing per-agent "adopts transcripts" and "owns history" flags to clients. This was the first version (Native chat: answer 'can this agent do X' from each agent's registration #27078) and is dropped. Each flag's only reader could already answer from what it had, and a new field must be reconciled between old and new clients and hosts.
• Refusing only where Orca builds the command. That misses older mobile clients and typed commands.
• Leaving these agents unprotected. That leaves the second-writer bug in place.

Linked Issue

N/A: maintainer follow-up. It is a deferred item from #26981 (floating-workspace chats of every agent). It replaces #27078, whose registration changes it now includes in slimmer form.

Visual Proof

QA on a hidden, isolated macOS test build (window code at cdd7221, background process at 216ba66), clicked by a separate test agent. OpenCode, Grok, OMP and Pi ran with isolated homes and a local fake model; Claude and Codex were the real signed-in CLIs.

OpenCode: the chat tab is renamed "Renamed OC tab", but its Session History row keeps OpenCode's own title ("QA chat", from the fake model). Opening the row opens the chat, and there is no "Resume in New CLI".

Pi: the row keeps Pi's title (its first prompt) and offers "Resume in New CLI", which ran pi --fork <session file> in a new terminal.

After the copy, the chat still answers into its own conversation; the copy is a second Pi row (same title, since both start with the same prompt).

Grok: "Resume in New CLI" is offered and opened grok --resume <id> --fork-session.

OMP: no "Resume in New CLI" (its CLI can't open a copy).

CLI copies, checked outside the app with throwaway homes: pi --fork and grok --resume <id> --fork-session each created a new session and left the original byte-identical.

Claude and Codex (control): rows still show the chat's name and offer "Resume in New CLI".

Seen in QA, not changed here: new rows and message counts appear only after Refresh (existing indexing behaviour), and a CLI copy of a Pi chat shows the same title as the chat's row.

Testing

• Unit tests per agent:
  • Owned rows and search hits open the chat; Claude and Codex rows show the chat's name, the other four keep their CLI's title.
  • "Resume in New CLI" is offered for Claude, Codex, Grok and Pi chats and not for OpenCode or OMP, and every copy command Orca builds passes the resume check.
  • Each resume form is refused or allowed as intended.
  • Every command Orca's own builders produce for an owned row is refused, in posix, PowerShell and cmd, with or without a custom agent command, for all seven history row kinds including Claude and Codex.
  • Shell-reading tests for every agent cover glued and quoted separators, quoted = values, command position, and prompt text.
  • The check reads text in one pass. 16 MB takes about 0.5–3 s on this machine, comparable to main.
  • Rows are visible per client kind.
  • No two agents claim the same history row kind.
  • An agent with no transcript importer is refused before any adoption work.
• Claude/Codex:
  • Every existing test assertion is kept.
  • The only commands main refused that this branch allows are the "newly allowed" forms listed under What Changed.
• Checks:
  • The full local static-analysis and typecheck run (24 checks) passes on cdd7221.
  • The targeted ai-vault, runtime, tab-menu and right-sidebar suites pass.
  • The only local failures are tests that need node:sqlite, which this machine's test runtime lacks. They fail the same way on main.
• I manually tested these changes locally
• Automated tests added/updated

Agent skill upstream boundary

• Not applicable, or this change copies or mechanically translates no upstream skill-installer source, tests, fixtures, registry entries, path tables, comments, or documentation.

Notes

Known limits, deferred:

• The check doesn't catch hand-typed command forms it can't parse:
  • clustered or repeated short flags (opencode -cs <id>);
  • an agent run through a wrapper not in the list, or nested in bash -c '…', $( … ) or backticks;
  • a backslash-escaped space in the program's path (/My\ Apps/claude);
  • cmd.exe's ;, which is an argument there but a separator here;
  • the stored cmd /d /s /c "cd /d ""C:/x"" && opencode --session ""<id>""" form. This one predates this PR, and Claude and Codex have the same gap.
• Pre-existing, and the same for Claude and Codex:
  • An owned row's details panel still labels its action "Resume in Worktree", though clicking it focuses the chat.
  • The CLI prints the raw agent_session_conflict when it refuses.
• A chat created on OpenCode 1.x and migrated to 2.x shows as two rows with the same name.
• After a Pi fork or a failed restore, the chat still owns the old conversation, as Claude and Codex chats do today.

Other notes:

• Remote/SSH: ownership is decided on the host that runs the chat, and no new message crosses between client and host.
  • An older desktop paired with a newer host also gets these agents' chat-owned rows marked as belonging to the chat. That's because the older desktop reports that it can show these agents' chats.
  • Main's row code acts on that mark for any agent, so it should open the chat. This is read from code, not run.
• Mobile: current mobile can't open these agents' chats from history, so their owned rows are hidden there. Making mobile open chats from Session History (for Claude and Codex too) is a separate follow-up.

Checklist

• 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)

[brennanb2025]

Review status: ready for your review (draft, not merged; stacked on #27078)

The problem. A conversation you have with OpenCode, Grok, OMP or Pi in Orca's chat view showed up in Session History as a plain terminal row. "Resume" on it ran the agent in a terminal on the same conversation the chat was still using, so two programs wrote to one conversation. Claude and Codex chats were already protected, but that protection was written only for those two. This is the item deferred from #26981.

What changes for users:

• Session History shows the chat. An OpenCode/Grok/OMP/Pi conversation a chat owns shows with the chat's name. Opening it opens the chat, and copy-resume and delete are hidden.
• Terminal resumes of it are refused. Any terminal resume Orca launches for an owned conversation is refused, including "most recent" resumes like opencode -c while a chat of that agent exists. Claude/Codex work the same way.
• Older mobile hides these rows. Mobile builds that can't open these chats no longer see such rows, rather than starting a second writer.
• Search hits drop the resume command. Search hits for an owned conversation no longer carry a terminal resume command, for any agent.
• The resume check reads command lines shell-style. It splits at &&, ;, | and line breaks. For Claude/Codex:
  • A resume within one command that main refused is still refused.
  • A resume flag sitting in a different command from the program is no longer read as one. For example, claude ; echo --resume X was refused and is now allowed.

Fixed during review:

• Round 1:
  • Hid owned rows from clients that can't open that agent's chat; the first version showed them to current mobile, where a tap failed with a raw error.
  • Dropped the resume command from owned search hits.
  • Made the new agents' flag matching case-sensitive, so rg … src/pi -C 3 is no longer refused.
  • Added tests that feed Orca's own resume-command builders into the check.
• Readiness:
  • The check took seconds on large terminal send text; it is now linear (16 MB in about 1–2 s).
  • Right-clicking these agents' chat tabs ran a pointless history scan; that is gone.
• Narrow rounds 3–4: replaced a span rule with "split into shell commands, then each agent reads from its first mention". That fixed a Claude command main refused that the span version allowed, false refusals of real forks, and a bare carriage return not ending a command.
• End: merged Native chat: answer 'can this agent do X' from each agent's registration #27078, which includes current main, with no conflicts.

Deferred:

• Hand-typed command forms not parsed:
  • Glued or clustered short flags such as opencode -cs <id>, and repeated flags.
  • A resume nested in $( … ) after another resume, or a continuation mark used for a non-matching shell. Main allows both too.
  • cmd.exe's ;, which is an argument there.
• Possible false refusal: claude --model pi -r X while a Pi chat exists.
• Duplicate rows: an OpenCode 1.x chat migrated to 2.x shows two rows with the same name.
• Old conversations stay owned: after a Pi fork or a failed restore, the old conversation stays owned. This matches Claude/Codex.
• Pre-existing for every agent, seen in QA:
  • The details panel says "Resume in Worktree" on owned rows (it opens the chat).
  • The CLI prints the raw agent_session_conflict.
  • Naming a chat creates an extra OpenCode conversation in the app folder.
• Follow-up: mobile opening chats from Session History.

Verified:

• QA on a hidden macOS build (207f0d6) with real OpenCode 2.0.21 and a local fake model:
  • The owned row and search hit open the chat.
  • opencode --session <owned> and opencode -c are refused, and no OpenCode process starts.
  • A plain terminal conversation still resumes in a terminal.
  • Screenshots are in the PR body.
• Fuzz against main's parser (300,000 Claude/Codex commands): no resume inside a single command that main refused is allowed now.
• Every command Orca builds for an owned row of each agent is refused, on posix, PowerShell and cmd.
• Local checks: pnpm tc:node, pnpm tc:web, all 24 orca-ci-checks, and the targeted tests pass at a357a6b.

Not verified:

• Grok/OMP/Pi live. Running real Grok writes the real Grok home, and the QA machine's Pi is too old for chat. They are covered by unit tests. Grok's id match and Pi's grammar are from code and docs, not observed.
• Windows/Linux/SSH live.
• Mobile on a device.
• This PR's full CI suite only runs on PRs into main, so it runs once Native chat: answer 'can this agent do X' from each agent's registration #27078 merges and this PR retargets main.

[brennanb2025]

Live Electron QA at f4ce1d9: every agent, all rows pass

Hidden macOS build of this PR, with the real OpenCode (2.0.14), Grok, OMP, Pi, Claude and Codex CLIs, each on its own existing sign-in. A separate agent drove every click.

Check	Result
A chat per agent, one exchange, tab renamed to "<Agent> QA chat"	Pass, all 6
Session History row titled with the chat's name; menu has no Copy Resume Command or Delete	Pass, all 6
Opening the row switches to the chat tab; no terminal or new tab is created	Pass, all 6
Searching for the chat's word finds the chat, with no resume command (resumeCommand: null)	Pass, all 6
Terminal resume of the chat's conversation is refused and starts nothing	Pass: opencode --session, opencode2 --standalone --session, grok --resume, omp --resume, pi --session and each bare -c. Agent process count stayed the same.
Forms this PR now catches	Refused: opencode -s <id>;echo done, --session="<id>", nohup pi --session <id>
Prompt-like text while a Pi chat exists	echo Fix the Pi --resume flag is accepted and only echoes
Plain terminal OpenCode conversation (control)	A normal row with Resume, Copy Resume Command and Delete; a terminal resume is allowed
Claude and Codex unchanged	Chat-owned rows as above, terminal resume refused. "Resume in New Native Chat" on a plain Claude conversation opens a chat that remembers it.
History id vs the chat's own id	Equal for Grok, OMP, OpenCode, Claude and Codex. Pi's row uses the id from the session file's name, as designed.

The agents' real config files were hashed before and after. Only Grok's and Pi's own sign-in files changed: each agent refreshed its own token.

Found along the way, not from this PR (the same on main):

• "Resume in Worktree" on a plain OpenCode 2 row types opencode2 --standalone …. A Homebrew OpenCode 2.0.14 install only has opencode, so that fails with command not found. The command comes from tui-agent-config.ts, which this PR doesn't touch.
• The refusal shows the raw agent_session_conflict text.
• Each chat's automatic naming adds a "Name this chat from the user's first request…" Claude row to Session History.

Grok chat's row: titled with the chat's name; no Copy Resume Command or Delete:

Opening the OMP chat's row switches to the chat tab:

Searching a Pi chat's word finds the chat, with no resume command:

Control: a plain terminal OpenCode conversation keeps Resume, Copy Resume Command and Delete:

Claude unchanged: "Resume in New Native Chat" on a plain Claude conversation opens a chat that remembers it:

[coderabbitai]

📝 Walkthrough

Walkthrough

The change adds shared CLI conversation contracts and shell parsing utilities. Agent runtime registrations now provide session-history behavior, and selected agents also provide transcript import. AI Vault ownership checks use registered agent behavior to match history rows and detect terminal resume commands. Session and search projections now apply client chat-opening policies. The renderer updates fork eligibility, history switching, and chat-title handling for supported agents.

---

Priority: ➖ Normal

Merge Risk: 🟡 Moderate · up to cdd72

A terminal command that runs through a shell, such as bash -c "pi --session ...", can still resume a conversation that a chat already owns. Two writers could then edit the same conversation. Close this gap, or explicitly accept it, before merging.

Pre-merge checks · 4 · 1

Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage		Docstring coverage is 46.67% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 60 functions across 38 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.
Title check		The title clearly summarizes the primary change: Session History treats OpenCode, Grok, OMP, and Pi chats as chat-owned and blocks conflicting terminal writers.
Description check		The description is detailed and follows the required structure. It explains the user impact, implementation, rationale, testing, visual evidence, compatibility considerations, and known limitations. T…

---

• Fix all pre-merge checks with AI

✨ Finishing Touches 💡 1

📝 Generate docstrings 💡

• Commit to this branch
• Create a new PR

---

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

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

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

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to GitHub limitations.

⚠️ Outside diff range comments (1)

🟠 Major · Handle resume commands hidden behind shell… · structured-session-ownership.ts:254-259

src/main/ai-vault/structured-session-ownership.ts:254-259
🗄️ Data Integrity & Integration | 🟠 Major | 🏗️ Heavy lift

Handle resume commands hidden behind shell interpreters.

The terminal entrypoints pass caller-supplied commands to this guard. For bash -lc 'pi --session /path/to/owned.jsonl', shellCommands keeps the quoted script as one word. commandProgramIndex therefore selects bash, not pi. No invocation is found, so the guard returns before checking ownership. A shell can then resume the owned Pi session as a second terminal writer.

Parse supported interpreter payloads such as -c and -lc recursively. If the payload cannot be inspected safely, refuse the command instead of treating it as having no resume invocation.

Source: Learnings

---

ℹ️ Review info

⚙️ Run configuration

• Configuration used: Repository UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: c581a18e-771e-4c0e-b931-7f803a749484

📥 Commits

Reviewing files that changed from the base of the PR and between f4ce1d9 and cdd7221.

📒 Files selected for processing (15)

• src/main/ai-vault/structured-session-ownership-acp-and-pi.test.ts
• src/main/ai-vault/structured-session-ownership-client-audience.test.ts
• src/main/ai-vault/structured-session-ownership.ts
• src/main/pi/pi-session-history.ts
• src/renderer/src/components/right-sidebar/AiVaultSessionRow.structured-title.test.tsx
• src/renderer/src/components/right-sidebar/ai-vault-session-cli-fork.test.ts
• src/renderer/src/components/right-sidebar/ai-vault-session-cli-fork.ts
• src/renderer/src/components/right-sidebar/ai-vault-session-surface-switch.test.ts
• src/renderer/src/components/right-sidebar/ai-vault-session-surface-switch.ts
• src/renderer/src/components/right-sidebar/use-ai-vault-session-display-title.ts
• src/renderer/src/components/tab-bar/tab-session-history-switch.test.ts
• src/renderer/src/components/tab-bar/tab-session-history-switch.ts
• src/shared/agent-resume-argv.ts
• src/shared/structured-chat-row-name.ts
• src/shared/tui-agent-fork-startup.test.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 7 remain after this review.