v0.26.0

Highlights

• Improved research workflow automation: Updated skill triggers to fire more broadly when research work begins or training scripts are written, reducing manual intervention needs.
• Enhanced capture result visibility: Added begin_bytes_captured field to capture results for better insight into data collection.
• Production reliability improvements: Hardened MCP server for zero-downtime deployments with thread offloading, load shedding, lease management, and probe/drain robustness.

What's Changed

Other changes

• fix(mcp): zero-downtime rollouts — thread offload, load shed, lease guard, probe/drain hardening by @richardwei6 in #110
• feat(sandbox-state): surface begin_bytes_captured on HarborCaptureResult by @mahitoburrito in #109

Full Changelog: v0.25.0...v0.26.0

v0.25.0

Highlights

• Explicit project specification now required: Client initialization now requires projects to be explicitly specified, removing implicit defaults and making project selection more deliberate.
• CLI and plugin versions bumped: CLI updated to v0.25.0 and plugin to v0.13.0 to support the project requirement changes.
• Minimum dependency versions raised: npm package now requires CLI floor of v0.24.0 and launcher v0.10.4 to ensure compatibility with the new project handling behavior.

What's Changed

Other changes

• chore(npm): raise MIN_CLI floor to 0.24.0 (tag verbs), launcher 0.10.4 by @richardwei6 in #108
• v0.25.0 feat: require explicit projects in clients by @richardwei6 in #105

Full Changelog: v0.24.0...v0.25.0

v0.24.0

Highlights

• Tags system overhaul: Added per-entity tag verbs, tag filters, and skew guards with improved skills prose for better tag organization and validation.
• Async CLI writes: Introduced non-blocking outbox operations via a durable journal, improving CLI responsiveness during write-heavy workflows.
• Harbor sandbox improvements: Added labeled state artifacts, begin-bytes (shared per-task before-content), and an SDK-owned capture facade with attach()/finalize() methods for Harbor bridges.
• Documentation refresh: Expanded SDK write-surface documentation to cover both CLI and programmatic agent usage patterns.

What's Changed

Other changes

• feat(harbor): SDK-owned capture facade — attach()/finalize() for Harbor bridges by @mahitoburrito in #99
• docs: teach agents the SDK write surface, not just the CLI by @richardwei6 in #100
• feat(harbor): label sandbox state artifacts by @mahitoburrito in #101
• docs: tighten the SDK write-surface prose by @richardwei6 in #102
• fix(harbor): preserve sample reward identities by @mahitoburrito in #103
• feat(sandbox-state): begin-bytes — shared per-task before-content by @mahitoburrito in #104
• feat: async outbox — non-blocking CLI writes via a durable operation journal by @richardwei6 in #107
• feat(tags): per-entity tag verbs, tags filters, skew guards, skills prose by @richardwei6 in #106

Full Changelog: v0.22.0...v0.24.0

v0.22.0

Highlights

• Ambient run context and span scopes: SDK now supports ambient run execution with span scopes, enabling better context tracking across operations.
• Per-sample rollout and coordinate fixes: Added per-sample rollout rail for miles and fixed trajectory-span coordinate handling.
• Coordinate read surface: New SDK methods, CLI verbs, and MCP tools for reading coordinates across the system.
• Durable queue and coordinate tracking: Miles durable queue records now carry coordinates for better observability.
• Project-direct runs: SDK and CLI support for optional experiments without requiring explicit experiment setup.

What's Changed

Other changes

• fix: archived slugs, positional hypothesis, and split exit codes by @richardwei6 in #89
• v0.16.0 feat: project-direct runs — SDK, CLI, and MCP for optional experiments by @richardwei6 in #92
• v0.17.0 feat: below-run coordinate layer — run.unit ambient context, coords/labels/span_id, capture_trial coordinates by @mahitoburrito in #90
• v0.18.0 feat(harbor-runner): library-mode instrumented trials — hooks + sandbox capture out of the box by @mahitoburrito in #91
• v0.19.0 reconcile(miles): one backend — fold step-key strip + budget plan into the shipped integration by @mahitoburrito in #94
• v0.20.0 feat(miles): durable queue records carry coordinates by @mahitoburrito in #95
• v0.21.0 feat: the coordinate read surface — SDK methods, CLI verbs, MCP tools by @mahitoburrito in #96
• v0.22.0 feat: per-sample rollout rail for miles + trajectory-span coords fix by @mahitoburrito in #97
• feat(sdk): ambient run, span scopes, compare, any-type log, auto-step by @richardwei6 in #98

Full Changelog: v0.15.0...v0.22.0

v0.15.0

Highlights

• Explicit project and experiment creation: Project and experiment creation now requires explicit action rather than implicit creation, clarifying the workflow for organizing research.
• Client capability registration: The setup process now registers installed client capabilities, enabling better tracking of available features.
• Documentation updates: Skills documentation has been refined for clarity.

What's Changed

Other changes

• docs(skills): restore "or create" and cut a restated paragraph by @richardwei6 in #86
• feat!: creation of projects and experiments is explicit by @richardwei6 in #87
• feat(setup): register installed client capabilities by @richardwei6 in #88

Full Changelog: v0.14.5...v0.15.0

v0.14.5

Highlights

• SDK now attributes runs to coding sessions — runs are now linked to the session that produced them, improving traceability and session-level analytics.
• Skills refactored into start/track moments — skills have been reorganized by execution timing (start vs. track phases) with two skills removed as part of the restructuring.
• Documentation clarified SDK/CLI choice — added guidance on selecting between SDK and CLI to make per-step performance curves accessible.
• Probe-research plugin updated to 0.10.0 — underlying probe-research dependency bumped.

What's Changed

Other changes

• refactor(skills): split by moment into start/track, drop two skills by @richardwei6 in #83
• docs(skills): teach the SDK/CLI choice so per-step curves are reachable by @richardwei6 in #85
• feat(sdk): attribute runs to the coding session that produced them by @richardwei6 in #84

Full Changelog: v0.14.4...v0.14.5

v0.14.4

Highlights

• Improved version-check efficiency: Manifest cache TTL reduced from 24 hours to 15 minutes, enabling faster detection of available updates.
• Clarified auto-update messaging: Simplified wizard copy by removing redundant "off means" phrasing.
• Updated probe-research dependency: Bumped to version 0.8.1.

What's Changed

Other changes

• fix(version-check): 15-minute manifest TTL, not 24 hours by @richardwei6 in #81
• copy(wizard): drop the "off means" half of the auto-update line by @richardwei6 in #82

Full Changelog: v0.14.3...v0.14.4

v0.14.3

Highlights

• Auto-update stability: Fixed plugin recording in the stable channel and cleaned up the autoupdate mechanism.
• MCP prompt rebalancing (breaking change): Reorganized prompt surface to prioritize the write path and removed deprecated command aliases.

What's Changed

Other changes

• fix(mcp)!: rebalance the prompt surface toward the write path, remove deprecated aliases by @richardwei6 in #79
• fix(autoupdate): delete the stable channel, record the plugin half by @richardwei6 in #80

Full Changelog: v0.14.2...v0.14.3

v0.14.2

Highlights

• Wizard UI improvements: Renamed tracking category to "CLI + MCP" and centered all wizard pages for better visual consistency and clarity about tool scope

What's Changed

Other changes

• fix(wizard): rename tracking to "CLI + MCP", centre every page by @richardwei6 in #78

Full Changelog: v0.14.1...v0.14.2

v0.14.1

Highlights

• Improved wizard UX: The capture picker now displays both the capture name and location in a single line, making it clearer and more concise to select captures.

What's Changed

Other changes

• fix(wizard): the capture picker says what and where, in one line by @richardwei6 in #77

Full Changelog: v0.14.0...v0.14.1