Surfbots Dev Platform v0.5.0

Surfbots Dev Platform v0.5.0

This release aligns Cline and Copilot with the scoped task lifecycle and repairs the CLI/API/MCP integration used by that workflow.

• Substantial work uses scoped task begin, investigation/implementation/test/pre-review checkpoints, canonical Git review, verification, and completion.
• Active and bundled Cline/Copilot rules follow one policy. Start a new Cline task after installing updated rules.
• Supervision status uses the correct API endpoint and directly reads paginated workflow state without embedding searches; usage totals and lifecycle retries are corrected. MCP file/range and symbol/reference requests match the API contracts.
• Each profile role can set context_tokens. Defaults are 8,192 for local-lightweight, 32,768 for local-quality, and 128,000 for recognized OpenAI models, within known model limits. Unknown, compatible, and Azure deployment names default to 8,192 until configured.
• Context overrides survive storage, clone, export, and import. Changing a model resets its override. models profile set-context <profile> <role> default restores the model default. Custom inference servers must be configured separately to support the selected window.
• Managed profiles support memory_gib, defaulting to 6 GiB for Qwen and 32 GiB for the optional Devstral Small 2 24B Q4 preset. Clone an active profile, change memory, then activate it. Settings survive clone/export/import; hosted or external servers manage their own RAM. Live 8 GiB activation and restoration to 6 GiB passed.
• Embedding inference now runs on one dedicated worker, preventing overlapping model allocations even when a client cancels. Eight concurrent live requests passed with responsive health checks and no pod restarts.
• Managed local generation uses one parallel slot and disables the extra RAM prompt-state cache (the upstream default was 8 GiB). These bounds prevent sustained reviews from retaining multiple contexts beyond the container limit.
• Local review batches target 16,384 evidence tokens; generation requests allow 600 seconds and the CLI allows 1,800 seconds for a multi-batch review. Timeouts remain blocking and return actionable errors.
• Large reviews use bounded whole-file batches with explicit file scope and report coverage limitations. Reviewer instructions distinguish current diff additions from removed lines and stale indexed snippets. Missing evidence or a failed batch cannot produce an unqualified clean review.
• Runtime packaging now uses committed allowlisted files, includes port-forward units, and validates archive contents, checksums, versions, and source provenance. The established runtime-source distribution format is retained.

Release source: da2814bc76ed325cd5c6764dfb5f4f6e7a1bcf38. The final committed-source archive passed validation of 224 entries, checksum verification, and extracted CLI initialization. Devstral returned no_issues_found across all seven review batches; lifecycle evidence verification passed.

Validation: 425 automated tests passed in the release suite, with 46 focused lifecycle/reviewer tests also passing after the final next-action correction. All seven live services are healthy; MCP code intelligence, memory, supervision, and client integration checks passed. Packaged CLI initialization and bootstrap checksum acceptance/rejection were tested in isolated directories.

The existing external Devstral Small 2 Q4 server is running at 32k context with a 32 GiB systemd limit for release review; Qwen remains the generation model. The optional managed Devstral preset was not activated. Paid-provider requests, Qwen Coder Next performance, and a full installation on a fresh machine were not tested. Batched file coverage does not guarantee joint analysis of cross-batch interactions. Existing dependency deprecation warnings remain.