v2.0.0-beta.6 — Gemini 4 Argon on OpenCode Go

Add opencode-go-gemini-4-argon-contributor for Gemini 4 Argon Contributor through OpenCode Go.

• Default reasoning effort: high; OPENCODE_EFFORT still overrides it.
• Supports review and delegate, including steerable delegation.
• Select explicitly with -a opencode-go-gemini-4-argon-contributor.
• Existing user-local configs need the new profile copied from config.example.json.

OpenCode Go exposes the Contributor version. Its privacy terms allow Google to use prompts and completions for model training; it is not ZDR.

Validation: installed OpenCode 1.19.0 catalog checked; 58 focused profile checks and 55 orchestration tests passed with fake backends. No live inference call was run.

v2.0.0-beta.5 — Harden prompt input and quota reads

Prevent unintended input waits throughout Porch's prompt transports.

• Reject empty or whitespace-only explicit prompts and guidance instead of falling back to inherited stdin. Reject conflicting argument/file sources and empty file-option values.
• Stop shell prompt builders from reading ambient stdin. Close unused stdin for Grok/Devin file-based backend transports.
• Read caller-supplied paths literally, including files named -, in prompt transports, all code-review depths, and judge input.
• Replace costly shell whitespace substitutions so large file prompts remain responsive.
• Reuse the concurrent JSON-RPC client for Codex quota queries, enforcing one response deadline while draining stderr and handling partial response lines.

Compatibility: provide exactly one nonempty argument/file prompt, or supply the prompt through stdin when neither is selected. Review context alongside an explicit question still requires --context-stdin.

Validation: 36 focused input tests and 7 quota tests passed. The full offline harness reports 1001 passed, 0 failed; all backends were fakes. No live provider smoke test was run.

v2.0.0-beta.4 — Fix review stdin hangs and add explicit GPT-6.1 Sol

Fix review ask waiting indefinitely on inherited stdin when a question is already supplied as an argument or with --prompt-file.

• Add --context-stdin to explicitly append piped context. Stdin-only questions continue to work.
• Reject blank explicit questions immediately without waiting for EOF.
• Add the opt-in codex-gpt-6.1-sol profile (gpt-6.1-sol, high effort) for review and delegate; the enabled codex profile already selects GPT-6.1 Sol.
• Add 10 offline input transport regressions and model/effort argv checks.

Compatibility: commands that pipe context alongside an explicit question must now add --context-stdin, for example git diff HEAD | "$PORCH" review ask --context-stdin "Review this diff". User-local configs need the new profile copied from config.example.json.

Validation: full offline suite on updated main reports 999 passed, 0 failed, including input, platform, and steering checks. No live provider smoke test was run.

v2.0.0-beta.3 — Windows launcher fix

Fix npm-installed backend launchers under Windows Git Bash. Porch now selects the executable extensionless sibling of a resolved .CMD shim in the same directory, without substituting another installation from PATH. Native Python emits LF-only shell assignments so Windows line endings cannot corrupt backend paths and role names.

Includes regression coverage for Windows-style paths, spaces and Unicode, shell quoting, missing siblings, and non-.CMD binaries. Builds on av-frox's contribution in #2, with the completed fix in #3. Since beta.1, the default Codex profile also moves to GPT-6.1 Sol and the optional Claude Sonnet judge to Sonnet 5.5.

Validation: 993 offline checks and 401 steer checks passed, plus 22 platform tests (four native-Windows-only tests skipped on Linux) and 13 UI observer tests. Native Windows CI was not run: GitHub Actions jobs could not start because of an account billing lock. Live-provider smoke tests were not run.

v2.0.0-beta.1 — Porch live UI

Porch 2.0 beta adds an optional read-only dashboard for existing delegated runs. Open it with porch ui --desktop or use porch ui for a local URL. The view groups runs by launching agent, filters by launcher, executor and project, streams recorded output, and can focus on the current launching session or a specific run. Rows show Porch turn counts; run details show the executor's native session ID when available.

The macOS desktop app and helper processes are branded Porch. Closing the UI leaves workers running, and headless delegation does not need the UI dependencies.

Validation: 986 offline runtime checks, 13 observer tests, eight UI presentation tests, production build/lint, and Playwright desktop acceptance passed locally. Live-provider smoke tests were not run.

v1.5.0 — Grok 4.7 and GPT-6 Sol/Luna defaults

Enable GPT-6 Sol (codex, high effort) and GPT-6 Luna (codex-gpt-6-luna, low effort) in the default review pool. Keep GPT-6 Astra available explicitly as codex-gpt-6-astra.

Upgrade native Grok and remote visual QA defaults to Grok 4.7. Existing user-local configuration retains precedence.

Validation: offline suite completed with 971 passing checks and two incorrect new Grok flag assertions; the corrected profile/argv block passed all 36 checks. Native platform suite: 20 tests passed. Live provider calls were not exercised.

v1.3.4 — Focused delegation guidance

Delegation guidance is organized around task preparation, launch, supervision, correction, waiting, and acceptance. SKILL.md retains the mandatory delegation essentials so they are always loaded, including the first-minute check and the journal and verification protocol for less capable workers.

Detailed wait handling, backend delivery semantics, and Git/JJ observation now live in optional references. The guide clarifies completed-run recovery through durable Codex continuation and includes Devin in the supported backend list. Runtime code is unchanged.

Validation: skill structure, 97 links and anchors, release versions, workflow-example Bash parsing with the existing id placeholder substituted, and diff checks passed. Live-model behavior was not evaluated.

v1.3.3 — Task-relative review acceptance

Review findings are now judged against the task requirements and intended use, so optional improvements do not automatically trigger another correction round. A shared acceptance guide defines evidence-based triage, focused verification, and a stopping rule while preserving material blockers and disclosure of consciously accepted limitations.

The guidance applies to delegated work, self-review, external reviewers, and LLM judge verdicts. Runtime behavior is unchanged.

Validation: skill structure, reference links and anchors, release version consistency, and diff checks passed. Live-model behavior has not been evaluated.

v1.3.2 — remote-agents: lock-recovery docs + Xvfb/VNC pattern

Docs hardening from comparing with the client-side rdp-agent approach and verifying claims on a live Windows host.

remote-agents

• Locked sessions: disconnect locks the session; whether an NLA reconnect clears it is configuration-dependent (verified it does NOT on a default deployment). Documented the real recovery order — remove lock sources at provisioning on owned hosts, operator unlocks once via the manual client, and why no jiggler ships.
• Xvfb+VNC shared view: documented the xfreerdp→Xvfb→x11vnc pattern as an opt-in Linux alternative when a pixel-level live view + input injection into the holder's session is needed without stealing it — vs the watch-only observer stream, with the client-side-input caveats noted.
• windows-desktop.md de-macOS'd (title, observer, manual session now covers mstsc/Remmina).

Not adopted: client-side XTEST input, password typing (our in-session driver + auth-failure latch already cover these cases differently), human-in-the-loop escalation (agent product concern).

v1.3.1 — remote-agents: real-VM smoke fixes

Fixes found by smoke-testing onboard.py check and host.sh against live Linux and Windows hosts.

remote-agents

• Probe JSON parsing no longer broken by PowerShell CLIXML progress output on stderr (stdout/stderr now handled separately).
• host.ssm tolerates profiles without ssm:GetConnectionStatus — PingStatus fallback, WARN if neither is callable.
• host.sh wait_for_ssm accepts a working SSH probe as readiness proof (same IAM-restriction case).
• host.pwsh-shell warns only when pwsh is actually missing as DefaultShell.
• Firewall rule names are configurable: bridge.fwRuleWg/bridge.fwRuleSmb (default <fwRulePrefix>-*) plus -RulePrefix on setup-host-bridge.ps1.

Validated: Windows VM 23 pass / Linux VM 13 pass on onboard.py check; work-bridge status, desktop status/probe chain, and host.sh run verified on both OSes.