v0.2.0: Project sync into sandboxes, background processes, and opencode auth login - #2
Open
cooleel wants to merge 12 commits into
Open
v0.2.0: Project sync into sandboxes, background processes, and opencode auth login#2cooleel wants to merge 12 commits into
cooleel wants to merge 12 commits into
Conversation
Fixes the "sandbox is empty" feedback: the local project is now synced into the sandbox at /tmp/workspace/<project-name> on first use. - Git repos: pushWorktree to a TensorLake-hosted repo, clone inside the sandbox, credentials configured so the agent can push changes back. - Plain folders: upload to a cloud volume and mount it into the sandbox (writes persist durably); skips node_modules and other build junk. - Mode auto-detected; TENSORLAKE_SYNC_MODE=git|volume|off overrides. - bash/ls/glob/grep now default to the synced project directory and the system prompt tells the model where the project lives. - Org/project scope resolved via API key introspection when not in env. - Bump tensorlake to 0.5.109 and @opencode-ai/plugin to 1.18.18. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…tatics Sandbox handles are now cached per sandbox id so repeated operations reuse resolved proxy routing. resume() refreshes routing on the same handle; terminate drops the cached handle. Removes the SDK's startup deprecation warning. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
background=true now uses sandbox.startProcess() instead of a fire-and-forget run, returning a pid. New bash_output tool reports process status plus output lines new since the last call; new bash_kill tool stops the process. Processes are started unnamed (non-managed) so the daemon keeps status and output queryable after exit or kill. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- client: route proxy-bound operations through withSandbox(), which drops stale handles (sandbox resumed on another host) and retries idempotent calls once; command execution never retries to avoid double-running. Use typed SDK errors instead of matching message strings. - project-sync: re-syncs fast-forward only instead of git reset --hard, so in-sandbox commits or changes are never clobbered; divergence is logged. - bash_output: cap replay after a plugin restart to the last 200 buffered lines and drop stale offsets when a pid disappears. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…sync - The in-memory-cache branch of _getSandbox now handles 'suspending' (wait for suspend to land, then resume) and waits for any other non-running status, matching the persisted-storage branch. Previously a cached sandbox mid-suspension was returned as if it were live. - Fix a latent deadlock on the cached terminated path: the recursive getSandbox() call returned the session's own still-pending in-flight promise, resolving the promise to itself. Recurse via _getSandbox. - Project sync no longer blocks the first tool call of every process when the sandbox already holds a project copy (e.g. resumed sandbox synced by an earlier process): ensureProjectAvailable() checks for the project dir once and lets the refresh push/clone run in the background; only a sandbox with no copy at all awaits the initial sync. Syncs are deduped per sandbox via an in-flight map. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
tensorlake ^0.5.109 picks up patch fixes but blocks 0.6.x breaking changes; @opencode-ai/plugin is types-only so ^1.18.18 is safe. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Git-mode sync now pushes the repo's actual refs (git push HEAD:main) to the TensorLake-hosted mirror, so the sandbox clone has real commits, authors, and dates — git log/blame/diff work as expected and branches created in the sandbox share ancestry with the original repo. - Uncommitted local changes (modified + untracked) are captured with a temporary index and replayed onto the sandbox working tree uncommitted, but only when the sandbox tree is clean and at the same HEAD — agent work is never overwritten. - Non-fast-forward pushes (local rebase/amend) recreate the disposable mirror and push fresh, with a warning that mirror-only commits are lost. - Credentials go in an HTTP header (redacted from errors), never the URL. - Repos with no commits fall back to the previous pushWorktree snapshot. - Sandbox gets a fallback git identity so agents can commit. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Key the cached org/project scope by API key so a no-restart auth login with a key from another project no longer reuses the stale scope - Distinguish a mirror that is ahead (agent commits) from rewritten local history before recreating the hosted mirror on non-fast-forward pushes - Track uploaded paths in a volume sync manifest so local deletions propagate without touching agent-created files - Use POSIX joins for guest sandbox paths on Windows hosts Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Deletion propagation: track every path still present locally (oversized, symlink, unreadable included) so files skipped from an upload are never deleted from the volume; drop the prototype-chain-prone `in` check. - Move the sync manifest off the volume onto the laptop's storage dir so sandbox agents can't see, commit, or tamper with the file driving remote deletions; validate delete paths and scrub the legacy on-volume manifest. - Mirror ancestry check: treat only merge-base exit 1 as "not an ancestor" (other errors route to the indeterminate path instead of recreating the mirror), and fetch into a throwaway ref with --no-write-fetch-head so the user's FETCH_HEAD is never clobbered. - Mount readiness: always wait for the guest even when the control plane already lists the mount, probe for a real mountpoint instead of `test -d` (the sync-failure fallback mkdirs a decoy at that path), and bound each probe to 5s so the 30s deadline is honest. - Project-dir fast path requires a non-empty directory so the empty decoy left by a failed sync doesn't count as a project copy. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rename TensorLakeClient, TensorLakeSessionManager, and createTensorLakeTools along with all prose, tool descriptions, log messages, and README references. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
New features added:
Moved off the deprecated SandboxClient to the Sandbox.create()/Sandbox.connect() statics.
Dependency pins loosened to caret ranges (tensorlake ^0.5.109, @opencode-ai/plugin ^1.18.18), and the README updated throughout (login flow, testing guide, troubleshooting).