Skip to content

Prepare Autohand Code 0.9.10 for release - #740

Closed
igorcosta wants to merge 86 commits into
mainfrom
cto/release-0.9.10-final-20261104
Closed

igorcosta wants to merge 86 commits into
mainfrom
cto/release-0.9.10-final-20261104

Conversation

@igorcosta

Copy link
Copy Markdown
Contributor

Supersedes #739, whose recorded head stopped advancing during the current GitHub Pull Requests/Git Operations incident. This fresh validation branch points directly to verified commit 06ebff5b6effcdccf53aa52203fc5607428ede22; the release code is unchanged. The old validation PR will be closed only after this PR tracks that exact commit.

Prepare Autohand Code 0.9.10 with the integrated runtime, skills, provider, and terminal improvements.

Restored sessions accept the first turn and connect MCP in the background without resetting history. Adaptive skills load pinned instructions only when needed, retain registry discovery and local authoring, and keep verified learning and decision requests separately controlled. Queue steering, resize, command navigation, settings/sync panels, stale-write protection, safe imports, and provider approval boundaries are covered by the integrated changes.

The package version, changelog, curated notes at docs/changelog/whats-new-0.9.10.md, and legacy formula URL are aligned to 0.9.10. Previously released goal defaults and the continue alias remain in the 0.9.9 changelog section. The stable release workflow retains mandatory Developer ID signing, Apple notarization, native artifact verification, and npm/Homebrew publication gates.

Validation:

  • The preceding runtime revision 9252d13 passed all 12 CI jobs: 13,775 unit tests, all four Linux terminal shards, Windows terminal checks, and Linux/macOS/Windows builds. All four independently published stable/alpha installer checks passed.
  • Final version preparation passed 65 focused tests, three real-terminal version scenarios, build, and frozen-lock verification. The built CLI reported 0.9.10.
  • The initial complete proof passed all 13,775 unit tests and completed 76 of 109 terminal files without recorded test failures before the 60-minute job deadline canceled it. This is not treated as a completed proof.
  • The opt-in complete proof now uses two isolated terminal workers, keeps cases within a file serial, and has a 90-minute job ceiling. Ordinary terminal runs remain serial; individual deadlines and assertions are unchanged. Worker configuration passed 29 focused tests and 27 real-terminal adaptive-skill/settings/resume scenarios with two workers, plus lint, typecheck, and build.
  • Cleanup safety coverage deliberately uses project paths containing -D and --force, verifies the safe argument sequences, and executes the printed commands. This removes random temporary-directory false positives while preserving safety assertions. The targeted fixture and worker settings passed 20 tests plus lint.
  • Composer focus coverage holds the mocked response until an actual nonzero working-status refresh is observed, then checks cursor visibility and draft retention through completion. It no longer assumes a two-second wall-clock sample. Status-frame classification passed 17 focused tests, and all 34 composer/adaptive-skill terminal scenarios passed with two workers, plus lint, typecheck, and build.
  • Final CI and the complete proof are requested against the final 0.9.10 source and metadata. Publication waits for their results. The repository maintainer has confirmed authority to release; the administrator override will be used transparently for the author-review restriction while preserving linear history.

The release workflow is currently disabled and no stable tag, public release, npm publication, or branch deletion has been performed. The managed CTO branch avoids duplicate push and PR CI matrices.

igorcosta and others added 30 commits October 16, 2026 12:59
Chrome only lets extensions listed in the native host's allowed_origins
connect, and the installer wrote an empty list when no --extension-id was
given, so store installs could not reach the CLI. The install command and
the /browser auto-install now default to the published store id; an
explicit --extension-id or a configured id still wins for unpacked builds.
Folder skills (skills/<name>/SKILL.md) are loaded in place from
~/.claude/skills, so they are reported and skipped rather than copied
into duplicates. The global --dry-run option claimed the flag before the
import subcommand saw it, so dry runs wrote data; the subcommand now reads
both.
Rewrite the Multi-Provider Support bullet (free Fantail model with an
Autohand account, or 15+ providers) and add a Coming from Claude Code
bullet pointing at autohand import claude.
install.sh and install.ps1 now seed the optional first run with
"What can you do in this folder? Give me 3 quick ideas." instead of
"hello world". Both scripts only interpolate the variable.
…ions

Consume a versioned private settings envelope before provider construction, preserve saved configuration, and resolve SDK session references before creating a session. Return the actual persisted session identity and reject invalid or disabled forks.

chore(release): hold SDK startup protocol publication pending the coordinated CLI and SDK release.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Retain both public continue/resume selectors and private SDK session references while incorporating current main. The combined settings/auth regressions pass 85 tests; public and SDK startup terminal scenarios pass all 17 checks. Required aggregate proof remains recorded separately with baseline failures.

chore(release): keep the paired startup protocol unpublished until a compatible CLI release is approved.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Keep third-party, custom and local inference independent of Autohand account quotas. Select first-party gateway behavior explicitly, suppress unrelated quota UI and clear cached plan notices when providers change. Cover Tatch requests, exhausted accounts and terminal usage flows.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Retry transient cleanup errors with bounded backoff and keep cleanup failures from replacing captured diffs or original Git errors. Add regression coverage and document Windows diagnosis and validation limits.

Validation: 13 focused tests, scoped TypeScript, and scoped ESLint passed. Full repository proof remains blocked by unreadable cloud-placeholder files; native Windows execution is unverified.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Make compact scripted dispatch the default across providers, keep discovered
schemas out of the native tool envelope, and serialize runtime context without
rewriting the leading policy. Respect Retry-After, configured context windows,
optional background-call controls, and provider thread limits.

Preserve nested permissions, hooks, images, cancellation, goal completion
delivery, and Moa plan/commit/failure reviews. Route scripted context cropping
through the live conversation handler.

Validation: lint, typecheck, build, 13,000 unit tests, and 86 changed-file terminal
tests passed. Aggregate terminal proof remains non-green due to the reproduced
baseline OpenAI /usage expectation and host disk-exhaustion setup failures.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Limit the session decision cache to 1,000 entries with the shared bounded-map helper. Preserve explicit approvals and denials when older cached decisions are evicted or cleared, and cover pressure, re-recording, and disabled-memory behavior.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Configure the signed-in /usage scenario for first-party hosted inference so its quota assertions agree with the provider isolation policy.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Support Ctrl+S alongside Ctrl+Enter while preserving draft queueing, selected message identity, and unrelated composer text. Reserve the chord against extension bindings and document the portable shortcut.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Initialize the shared Ink composer before manager and MCP startup can perform synchronous work. Preserve single renderer ownership and cover first paint, silent MCP handshakes, first responses, and exit in both screen modes.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Pin CLI CI, installers, builds, and transported-artifact checks to macOS 26 for Apple Silicon and Intel. Preserve the Computer Use host deployment target, apply Bun runtime entitlements, and validate stapled tickets and Gatekeeper acceptance before publication. Document the Apple credential setup and add native signing coverage.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Defer the Git baseline scan until the first instruction, where the existing finalizer still protects pre-existing user changes. Reset stale baselines on fresh and restored sessions without scanning before the composer renders. Cover silent MCP startup and user-edit isolation with regression tests.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Prepare local search settings without holding the composer on the optional
3.5-second account capability request. Resolve shared, cancellable discovery
before advertising web_search, retaining provider precedence and retry safety.

Cover local configuration, stale accounts, cancellation, and startup typing in
primary-screen and fullscreen terminal sessions.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
…ter Use engines

Use the shared Autohand Developer ID certificate and App Store Connect API credentials, block publication without signing, and pin the separately notarized macOS engine revision. Verify managed native signatures before reusing an installed engine and preserve explicit custom driver paths.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
The command result can resolve before the gate's final ownership probe. Synchronize the regression with finishLaunch settlement so it still detects repeated polling and verifies retained recovery reservations.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Add opt-in bounded discovery, canonical package references, complete pinned task instructions and request-backed receipts through the existing code-mode tools. Preserve local authoring, registry installation and qualified same-name packages.

Keep scoped learning drafts, independent verification and held-out preference gates separate from activation observations. Restore paused goal intent and exact resources under current ownership and permissions, and bind optional Weka decisions to eligible active Autohand Cloud execution routes.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Defer the full-screen origin reset to Ink's next substantive paint and write them together. Remove the extra resize listener and preserve drafts and cursor through streaming and burst resizing.

Add unit and real PTY coverage for normal/full-screen idle/streaming, wrapped drafts, editing and clean exit.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Reserve a non-shrinking warning row above the composer so long drafts and wrapped status text cannot hide why editing is blocked. Cover goal disappearance, draft retention, rejected submission and Escape cancellation.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Find the input prompt below its top border before checking modal-return geometry. Retain exact position and liveness assertions when earlier user messages also begin with the prompt glyph.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Defer Ink resize paints until React commits the new viewport and cancel stale queued output. Publish clear and repaint bytes together while retaining stdout backpressure, callbacks, and terminal handoff cleanup.

Keep the composer and completion menus within the resized viewport. Cover intermediate frames, retained drafts, inline mode, and terminals that ignore synchronized output with real terminal regressions.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Route standalone built-in discovery through a cancellable local preview. Collect real factual answers under approval auto-confirm and report unanswered questions when no response channel exists.

Compact native code-mode policy and descriptions while retaining goal contracts, installed-agent metadata, and native parameters. Preserve complete provider failure details once in the Ink transcript. Cover request size, real choice input, discovery cancellation, and give exhaustive catalog checks an explicit execution budget.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
igorcosta and others added 28 commits November 4, 2026 15:23
Retain every matching command and subcommand for keyboard navigation while
rendering five options around the active selection. Cover wrapping, retained
input, and Tab acceptance beyond the initial visible options in terminal tests.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Distinguish empty conversations from those exceeding the existing 500-message
limit before upload. Explain how to continue without truncating the snapshot or
changing the local conversation, and cover the empty flow in a real terminal.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Preserve provider defaults and the Autohand request floor while migrating legacy
configuration with atomic replacement and an intervening-edit check. Keep readable
configurations usable when optional persistence is unavailable. Preserve both
settings test suites during integration.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Propagate a bounded per-turn denial tracker through child agents and preserve authorization failures through serial and parallel delegation.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Clear reused outcomes for every accepted instruction, stop later specialist batches after unattended denials, preserve expired YOLO time limits, and avoid duplicate managed-branch CI runs.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Enable changed-content checks by default, retain optional unread and partial-read enforcement, and reject revision races during hashing.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Include Sakana in provider configuration and keep first-party account authentication outside API-key settings.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Count Claude folder skills without duplicating discovered packages, stage alternate-profile resources before publication, retain existing packages, and preserve configuration links during atomic migration and saves.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Use native rename after complete staging and fail closed on cross-device moves so failed imports cannot expose partial packages.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Reuse the terminal driver deadline loop so a later redraw cannot invalidate an already checked completion frame. Retain the approved and failed completion assertions.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Record provider deadlines, unattended approvals, specialist reuse, stale-write guards, Claude package imports, panel recovery, sync warnings, handoff guidance, and the remote branch audit. Preserve the incomplete aggregate proof and publication boundaries.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Update the validation report after the normal fetch and fast-forward succeeded in the original checkout. Retain the incomplete proof and unpublished CI boundaries.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Preserve the validated permission implementation and equivalent regression test while including the refreshed CTO branch tip. Update the audit after normal Git fetching recovered.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Display the active provider budget, explain provider-raised minimums, and retain the reviewed setup copy across locales.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Retain nested denial propagation while reporting each denied tool and one deduplicated stderr summary. Record the final captured branch tips, effective timeout follow-ups, and validation limits.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Complete resume initialization, start MCP handshakes in the background, and release exit handlers on every path. Cover stalled and synchronous MCP failures plus real inline and fullscreen resumed turns.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
… validation

Serialize shared configuration writes, scope startup authentication updates to the checked credential, and retain newer preferences and sign-ins. Repair native tool dependencies, provider fixtures, installer discovery, and first-run assertions across CI platforms. Document the stable release gates.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Set the package and legacy formula version to 0.9.10, publish curated release notes, and keep the previously released goal defaults and continue alias in the 0.9.9 changelog section.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Allow an explicit two-worker terminal proof while preserving serial defaults and per-case deadlines. Give the opt-in complete proof a 90-minute job ceiling after the expanded suite exceeded one hour, and test invalid worker settings. Clarify scoped startup authentication persistence.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Exercise quoted project paths containing -D and --force, and assert the safe worktree removal and branch deletion argument sequences. Keep real command execution coverage while removing random-path false positives.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
Hold the mocked response until a nonzero working-status frame appears, then verify cursor visibility and draft retention through completion. Recognize elapsed frames with token metrics instead of relying on a fixed sleep, and document the observable-condition pattern for terminal regressions.

Co-authored-by: Autohand Evolve <code-noreply@autohand.ai>
@igorcosta igorcosta added the full-proof Run the complete lint, unit, build, and terminal proof on GitHub label Nov 4, 2026
This pull request was closed.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

full-proof Run the complete lint, unit, build, and terminal proof on GitHub

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant