mirror of
https://github.com/stablyai/orca.git
synced 2026-10-07 08:02:21 +00:00
Brennan Benson
761d63a4e5
* feat(agent-launch): host-side prompt delivery for agent.launch The host's agent.launch typed any launch prompt into the shell as part of the launch command. A long or multi-line prompt then ran line by line in the shell, and an agent that never showed readiness or crashed at startup had nothing guarding where its text went. agent.launch now carries a prompt on the typed line only when the line stays one line, control-free and at most 512 bytes; otherwise the agent starts clean and the host pastes the prompt once the agent's own ready signal fires (bracketed paste plus its composer marker or a quiet render, read only after the shell's last hand-off, never while the pane's own shell is proven in front), with main's draft-paste bytes and an Enter 50 ms later. Orchestration worker starts wait on tui-idle as before. A replay-safe launch admits and claims its ledger row in one write, Qwen Code gets a second Enter, the desktop and phone share one launch-refusal classifier, and hosts advertise agent.launch.prompt-carry.v1. Split out of #23748, which moves the desktop source-control buttons onto this path. * fix(agent-launch): keep a short-lined multi-line prompt on a local zsh launch line, as main did #24257 moved every multi-line or over-512-byte prompt off the typed launch line and pasted it after readiness. The phone's AI buttons and review notes, whose multi-line prompts main typed whole into zsh, then reached Claude 0.5-3 s later and their RPC reply waited for the paste. The host now names the shell a local macOS or Linux line is typed into, the way the spawn picks it, and a multi-line prompt rides a zsh line when every line is at most 512 bytes and the whole line at most 8 KB. A real-zsh test types such a line through Orca's own ready barrier and startup write, including when a slow user config makes the write land early. Elsewhere the measured unsafe cases keep the paste: bash 3.2 runs multi-line lines piecemeal, fish drops an early multi-line write, and any shell loses a line over 1 KB written early. * test(agent-launch): keep the real-zsh launch-line test out of the Windows lane's gate scan The Windows lane registration check read `const ZSH_PATH = process.platform === 'win32'` (the head of a multi-line ternary) as a Windows-true flag, so `describe.skipIf(!ZSH_PATH)` looked like a Windows-only suite. The file is POSIX-only; the zsh lookup is now a function. * refactor(protocol): move the agent.launch capabilities into their own module Main's protocol-version.ts sits at the 300-line cap, so the prompt-carry capability pushed it over. The four agent.launch capabilities and their doc move to agent-launch-runtime-capability.ts, re-exported by name and spread into RUNTIME_CAPABILITIES at the same position; the advertised lists and every export are unchanged. * refactor(protocol): import the agent.launch capabilities from their own module `export *` from protocol-version left the four names undefined under the mobile recording loader, which resolves a relative import through a Proxy with no own keys, so 37 phone recordings lost agent.launchReplay. Importers now name agent-launch-runtime-capability directly; protocol-version only spreads its list. * refactor(agent-launch): drop the unshipped viewMode field and trusted local caller id Both were inert in step 1 and existed only for step 2. agent.launch will become a public plugin API, so every wire field is permanent once shipped; a top-level viewMode reads as "choose terminal vs chat", which the host decides. Step 2 introduces placement and view intent under a placement object instead. * fix(agent-launch): read Codex's provisional startup from the rule files' hold anchor Main (#24375) moved Codex's provisional-header check into codex.json's provisional_startup hold anchor and deleted codex-terminal-readiness.ts, so the launch readiness hold now asks showsHoldAnchor, as main's own settled check does. * fix(agent-launch): hold rule-file name titles to quiet for a launch, and census the zsh fixture Main (#24375) answers a name-only title from each agent's rule file ahead of the sustained-title lane, so gemini.json's name_title settled a launch readiness wait on the shell's auto-title while Gemini was still booting. A launch now asks quiet of every weak idle verdict, as that lane did. Main's readiness census requires a recorder for every runtime fixture; the zsh prompt recording is a non-agent control. Gemini's synthetic baseline is regenerated for this PR's stated change: a bare gemini title is no longer its rest mark, so name-only rows settle weak, and a fresh working or blocked status is no longer overridden. * fix(agent-launch): paste a launch prompt only when the launched agent is proven in front A launch pasted its prompt unless a shell was proven in the terminal's foreground, so any read that could not prove one let the prompt through. After an agent exited at startup, its shell turned bracketed paste on at the next prompt, readiness fired on it, and the prompt was typed into the shell: - macOS: a pane runs its shell under login, so the process-group fence's root was never the shell's group and never proved it; the cached foreground name could also still name the exited process. - Windows Git Bash and WSL: the shell-alone-in-its-job check never answers. Now one fresh read of the terminal's foreground decides: agent, shell or unknown. Only 'agent' lets a write through (paste, Enter, second Enter, reused panes too); 'shell' still drops a ready signal. A Windows host never proves the agent, so there the launch line carries the prompt at any size, as on main. * test(agent-launch): cover the Windows QA stub, a grok override that exits at once * fix(agent-launch): keep the local socket alive while a prompted launch waits for its agent A launch with a prompt now waits up to 60 s for the terminal agent to be ready before it writes the prompt, and reports not-delivered when the agent never is. The local runtime socket closes a connection idle for 30 s unless the request is a long poll, so a launch whose agent exited at startup lost its reply and the caller saw 'runtime closed the connection' instead of not-delivered. Classify a prompted agent.launch and agent.launchReplay as a long poll, as orchestration.workerStart already is for the same wait. * refactor(agent-launch): narrow the launch params by 'in' instead of a cast * fix(agent-launch): find a launched agent behind a wrapper that leads its process group A tcsh or nu launch line runs the agent from /bin/sh '<script>', and a wrapper script that does not exec its agent does the same: the wrapper leads the terminal's foreground process group and the agent is a member of it. The fresh foreground read names the group's leader, sh, so a prompted launch was refused or pasted late (M4Air tcsh: 2 of 4 not delivered, 2 pasted ~9 s late). Before that read, take the host's process-group observation as positive proof when it names the launched agent among the foreground group's members and is younger than a ready signal's quiet window. It never proves a shell. * fix(agent-launch): judge the foreground-group proof by when its capture began, not how long ps took The age the host stamps on a process-group observation runs from the start of its whole-machine ps, so on a loaded Mac a capture begun after the read was asked for still read as older than 1 s and the proof was dropped. Count an observation whose capture began after the read was asked for, less the window a shared capture is reused across. * test(agent-launch): keep the crash-guard live test out of the Windows lane's gate scan The Windows-lane registration scan read the const assigned from a platform check as a Windows-only gate, though the suite runs everywhere but Windows; find zsh in a function instead, as the real-zsh typed-line test does. Under load the fresh foreground scan can fail to answer, which lets the shell's prompt settle readiness (2 of 4 paired runs). The guard still refuses that write, so assert the refused write, the property that must always hold. * perf(agent-launch): read a local pane's foreground from its own terminal, not the whole process table The foreground read that gates every launch paste ran the daemon's inspectProcess capture and then a fresh scan, each a whole-machine ps; the fresh one also waits for any capture already running before it starts its own. Measured here at load 5: 1.2 s a read (M4Air QA: 3.4-5.0 s, and worker starts 17.6-32 s against main's 9-12 s at load 25-84). On a local macOS or Linux host, take the pane's root pid from the provider's session inventory and run one ps limited to that pane's terminal. Its foreground process group decides: the launched agent or any non-shell member is the agent (a wrapper that did not exec its agent leads the group), a group of shells alone is the shell. Same pane, same verdict: 2.7 ms a read. SSH hosts keep the relay's observation and name. * test(mobile): re-measure the web app's script sweep after agent.launch's capabilities moved out of protocol-version The mobile web bundle check failed at 124 assets against a ceiling of 123. Main already sat exactly on that ceiling: its sweep table read 69 scripts at 16 routes while the tree builds 73, the whole margin of 4. This branch imports the agent.launch capabilities from their own module, so protocol-version is no longer pulled into the root layout and four other routes. That moves which routes share which modules, and the Qoder capability module, imported by protocol-version and the AI-vault resume path, no longer shares an importer set with anything, so it gets a chunk of its own: 74 scripts. The fence says to re-derive the bound rather than raise it, so the sweep is re-measured on this head (every prefix of the sorted route list). The worst route now adds 10 scripts (session), not 9, which moves the pinned shell crossing from 32 to 30 routes; main re-measured on its own lands on the same crossing. * fix(agent-launch): a worker's brief needs its agent found in front, and Grok's start answers on its composer A paired-server worker start whose agent exited at startup typed its brief into the server's shell, which ran it: the idle wait can settle on a shell back at its prompt, and the brief was written with no foreground read. Both worker-start paths now check before each brief write, as a launch prompt is checked: on a host that can find the agent in front it must be there; on one that cannot (Windows) a shell proven in front still refuses, and anything else writes as before. A Grok worker start waited ~10 s more than main: its only rest signal is its bare name, which a launch holds to quiet output, and Grok animates its logo for ten seconds after its composer glyph. A worker start for an agent whose rest signal is its bare name and whose composer draws a marker (Grok, DSH, mimo-code) now also answers on that marker, whichever comes first.