[nwparker]

ELI5

Reasonix becomes a selectable, launchable terminal agent with its own identity, saved history, hooks, and same-session resume. This draft consolidates the original proposals and the reviewed integration onto pinned main.

What Changed

• Repairs the hosted CI integration: registers the existing literal Reasonix captures with the census, adds only their missing evidence entries, and corrects legacy search/preload expectations to match the verified independent capability contracts.
• Fixes restored terminal text gaining bold or dim from the live pen: the shared desktop/daemon snapshot wrapper resets intensity when moving from normal to alternate-buffer paint, then retains the final pen for subsequent output. The focused control covers empty and nonempty normal buffers with bold, dim, and both together.
• Registers the third-party Reasonix CLI in the existing discovery, installation, picker, configuration, prompt delivery, and manual-process recognition paths.
• Reads bounded native RX4F history through the execution host's existing scanners, with integrity checks, committed-tail rules, regular-file admission, and cancellation cleanup.
• Routes native lifecycle hooks into the canonical status store and resumes the original saved session ID through the existing host-owned launch path.
• Preserves main's newer search scope handling and terminal-only Build identity. Adapts Reasonix model overrides to the current option catalog and keeps unsupported history agents out of ranking and paging.

Folder workspaces, git worktrees, native/WSL/SSH host ownership, mixed-version capability negotiation, and conservative unsupported operations retain their existing architecture.

Why

The literal setup capture exposed six real restore errors. An intensity-only reset corrects that cause while leaving unrelated known color limitations visible; the original failed captures and exact reduction audit are retained.

Users requested a distinct Reasonix agent instead of a generic terminal workaround. Reusing the current catalog, launch, history, and status owners keeps provider identity and host authority consistent; a separate adapter subsystem would duplicate those decisions.

Linked Issue

Addresses the Reasonix requests in #12200, #23216, and #23774.

Builds on @BobcGn's #12330 and @makiisthenes's #23217. Both proposals remain open and unmerged; this consolidation preserves their verified contributor credit without claiming either original was merged. Their original commit identities are credited in the consolidation commit.

Visual Proof

The retained Reasonix 1.39.7 PTY captures distinguish setup from a ready composer. The native history/hook capture records HTTP 401, three trusted submitted users, zero assistant messages, and zero tokens. These are captured CLI evidence, not rendered Orca acceptance; new rendered before/after proof remains outstanding.

Testing

Current repair is committed at 221fb444332fdf818da88ca6285dbbd07a68fb3f, tree 13c278bdaaf0ae3af9ed8732adb7ef68f06ff544, on the ordinary additive composition with pinned main d9173ffbdbfa10ed7a24af893d15f44f5baf7c35.

Normal macOS/Node commands use ORCA_BACKGROUND_LAUNCH=1. Cohorts overlap and their counts are not added:

• Directly failed census/coverage/bundle, service search, actual old-parser and paired web-preload suites: 289 passing tests in 12 files.
• All 100 literal captured serialize tests pass, including Reasonix with zero accepted failures. Snapshot/fuzz cohort: 145 passing tests in seven files and two existing optional skips.
• Differential serializer tests with the unchanged addon materialized from entry c7: 103 passing tests in two files, with the optional differential controls enabled.
• Snapshot caller regressions: 62 passing tests in eight files; focused bold/dim/combined control: six passing cases, also included above.
• Reasonix/capability/current-main-shell regressions: 235 passing tests in 30 files. DSH/Jcode/hooks/search/resume regressions: 448 passing tests in 39 files, with four existing platform/opt-in skips.
• Full pnpm tc, intended-file formatting, and check:code-quality:changed against pinned d917 pass. Normal commit hooks run oxlint, React Doctor and oxfmt. All 17 repair files in the commit match the recorded tested bytes; tracked worktree is clean.

No literal transcript changed. Twelve old failure counts only decrease; Hermes/Build bold signatures disappear. Two Freebuff lifecycle jitter hashes change because the unwanted intensity is removed from the row descriptor; its existing background mismatch remains pinned. No new replay failure allowance, timeout weakening or test skip was added.

The original hosted run used synthetic 4dd271d0910fc74e44ce2eef90ab5b7049d4f0d5 (c7+c4); its failures were reproduced before the repair and all original failed attempts remain recorded. Fresh hosted success is not claimed. No app/window was launched: the required Electron skill is unavailable here, so rendered before/after acceptance remains outstanding. Authenticated provider generation, real Windows/Linux/WSL/SSH and paired/mobile/native packaging acceptance remain separate gaps; historical captured HTTP401 is not provider success.

Historical validation before this repair (retained verbatim):

Fresh macOS checks against the composed tracked bytes:

• Full pnpm tc passes.
• Focused integration/regression: 1,403 passing tests across 106 files, with three existing platform/opt-in skips.
• Current option-catalog regression: 14 passing tests; argument-removal/startup regression: 23 passing tests across two files (one startup file overlaps the focused run).
• pnpm exec oxlint on intended files and intended-file formatting pass. The changed-code quality gate is checked against pinned main b407d06c1ecd9d237010984ac02d9cd3abe37df0; normal commit hooks remain enabled.

The initial composition's duplicate import, obsolete option override property, and stale search expectations failed checks and were corrected. Installed pnpm 12.8.1 reports verifyDepsBeforeRun=warn, retired managePackageManagerVersions=undefined, and an existing dependency/patch metadata mismatch; no dependency install or manifest/lock change was needed for the successful checks.

No installed reasonix command was available on this terminal's PATH for fresh isolated version/help proof. Authenticated generation, successful tools/approval/cancel, Studio2, rendered UI, real Windows/Linux/WSL/SSH, paired clients, mobile, and native packaging remain unverified. Historical whole-run failures remain failures; the previous 33-check independent SOURCE CLEAR applies only to 1ba2, not this new composition. Fresh hosted results are not claimed.

• I manually tested these changes locally
• Automated tests added/updated, or explained why not below

AI Disclosure

Review

Draft for independent review of the final source/main composition. The coordinator owns acceptance and any future user-authorized lifecycle decision; no merge or predecessor closure is requested here.

Agent skill upstream boundary

• Not applicable, or this change follows docs/reference/agent-skill-sharing-upstream-boundary.md and copies or mechanically translates no upstream skill-installer source, tests, fixtures, registry entries, path tables, comments, or documentation.

Notes

The experimental content-lifetime gate remains disabled. Reasonix is third-party and distinct from official DSH. Source/capture tests establish bounded integration behavior, not authenticated provider or platform acceptance.

Checklist

• This PR is small and focused
• 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)