--- title: "Changelog | Openshard" description: "Everything that's been merged into Openshard, in plain English. We ship every day." url: https://openshard.dev/resources/changelog markdown: https://openshard.dev/resources/changelog.md updated: 2026-10-24 --- # Changelog We ship every day. Everything that's been merged, in plain English. Newest first. ## Sub-agents get their own receipts *Oct 23, 2026* When Claude Code hands work to sub-agents, each one now shows up on the receipt with its own model, tokens, cost and result. **Added** - Each Claude Code sub-agent's model, tokens, cost and outcome are recorded and sent to the dashboard. - Replicas cloud workspaces sync receipts using the workspace's own identity, with no API key to copy. - Receipts record where the agent ran and in which app. - Openshard checks whether a change's new tests would fail without the change, so a test that always passes doesn't count as proof. - Receipts flag pull requests over 400 lines and leftover junk files. **Improved** - The person who approved openshard login is named as the receipt's owner. - Sign-in now verifies your email on Openshard's own page. **Fixed** - A check the command gate refused is recorded as not run, and isn't retried. ## Ready means ready *Oct 22, 2026* Openshard now only says it's set up once a real receipt has reached your dashboard. **Added** - Each organisation's plan is recorded from Stripe checkout, ready for paid plans. - Receipts store a fingerprint of each check's output. **Improved** - Setup says Openshard is ready only when a receipt has reached the dashboard. - openshard last says In progress, not Done, while a turn is still running. **Fixed** - Claude Code work that carries on after an early session end is no longer lost. - A finished cloud turn's receipt is sent while the next turn is still open. ## Merge gate: no receipt, no merge *Oct 21, 2026* A GitHub check that reads each pull request's receipt and only passes verified work. Teams set the rule once in their settings. **Added** - A merge gate GitHub Action that passes or fails a pull request from its receipt. - Owners and admins choose the merge gate rule: agent changes need a receipt, or every change does. - Agent commits are marked with an Openshard-Agent trailer by default, so agent work is provable from git alone. - Receipts record how a run's usage was paid for: an API key or a subscription. **Improved** - Codex and Cursor sessions tested end to end until they reach a Verified receipt. - The task timeline is now a commit graph you can follow turn by turn. - Tested on macOS, with install steps for every platform. ## Rules you can read *Oct 20, 2026* Team policies are now built rule by rule, and Insights shows which agents are worth what they cost. **Added** - A rule builder for team policies, replacing the old form. - A pass rate against cost chart for every agent and model. **Improved** - Each task's timeline shows what happened turn by turn. ## Openshard on GitHub *Oct 19, 2026* The GitHub App can add Openshard to a repository with one pull request and shows each repo's status. **Added** - The Openshard GitHub App, with a one-click Add Openshard pull request. **Improved** - Squash-merged pull requests show a cost worked out from their branch commits. **Fixed** - A missed GitHub webhook no longer leaves a pull request's state out of date. ## Receipts that add up *Oct 18, 2026* Receipts are signed, line counts are recorded for every file, and cost lines now add up to the total. **Added** - Every receipt is signed on your machine, and openshard verify --signature checks it offline. - openshard login links a machine by approving it in the browser. No key to paste. - Lines added and removed are recorded for each file. - Receipts flag test edits, reads of test answer files and hidden failing exits. - Cloud agent commits carry their cost in a usage trailer. **Fixed** - A test that fails then passes counts as passed, and a branch switch no longer inflates the file count. ## Receipts that count only what changed *Oct 17, 2026* 11 changes shipped: 11 improved. **Improved** - Counted only files changed on the receipt's branch after a branch switch - Judged agent-reported checks by their last run - Reported re-runs that couldn't set up as unknown, not failed - Recorded cache-write tokens so cost lines add up - Recorded lines added and removed on each Receipt - Built the Receipt v2 Proof, Control and Cost sections ## Ask questions about your agents *Oct 16, 2026* 4 changes shipped: 1 new, 3 improved. **Added** - An Ask about your agents panel to server Insights **Improved** - Answered Insights questions through allowlisted analytics queries - Showed a delegated task the same way on Home, the list and detail - Matched the task Receipt to the original Receipt design ## Warnings for multi-repo sessions *Oct 15, 2026* 2 changes shipped: 2 improved. **Improved** - Warned when checkout hooks miss multi-repository sessions - Documented supported Claude Code governance controls ## Four new Insights views *Oct 14, 2026* 24 changes shipped: 10 new, 13 improved, 1 fixed. **Added** - A Sessions view to server-backed Insights - A Verification view to server-backed Insights - A verification breakdown of failed checks and incomplete evidence - A Cost & usage view to server-backed Insights - An Agents & models view to server-backed Insights - A server-backed Insights overview over the whole history - Series, usage metrics and sessions to analytics queries - Validated analytics queries over receipt_facts **Improved** - Keep one Receipt per cloud task when hooks send no project directory - Attributed check outcomes by the shell that reported them - Require CI verdicts to agree with their checks before attaching them - Allowed optional Depot builds without breaking Packaging CI - Skipped a redundant full freshness scan in every analytics request - Tested analytics updates when verification and usage evidence arrived later **Fixed** - Depot bake working directory ## Insights and Receipts agree *Oct 13, 2026* 7 changes shipped: 7 improved. **Improved** - Redacted passwords written in prose from Receipt task text - Made Insights and Receipts filters name agents and models the same way - Stopped comparing Insights periods the loaded history does not cover - Tested Claude telemetry ingestion against a real Claude Code export - Counted Insights cost by provenance, lower bounds and session ownership - Tied every Insights finding to the Receipts and number it rests on ## Guidance for cloud coding sessions *Oct 12, 2026* 34 changes shipped: 2 new, 29 improved, 3 fixed. **Added** - Hosted guidance for connected cloud coding sessions - Connected cloud capture identity and hosted guidance **Improved** - Kept capture state out of the worktree tests' commits - Serialised the capture service's state replaces - Synced a session key and the reasons usage is absent - Kept local paths out of tasks and retried refused Receipts - Kept long sessions' events and labelled notifications and idle seals - Counted a session's tokens and cost once across its checkouts **Fixed** - Commit attribution during capture - Windows status-line timing test - Session revocation on sign-out ## Capture stays out of your git *Oct 11, 2026* 8 changes shipped: 8 improved. **Improved** - Kept capture out of the agent's git: no index.lock, no committed .openshard/ - Allowed a Claude workspace under a parent git repository - Kept syncing later Receipts when one record cannot be built - Redacted e-mails, JWTs and private keys in imported and verification text - Stopped append_jsonl joining a new record onto a torn last line - Resolved workspace event paths before matching them to checkouts ## Claude Code cloud sessions, linked *Oct 10, 2026* 11 changes shipped: 9 improved, 2 fixed. **Improved** - Linked Claude cloud commits to captured sessions - Captured multi-repository Claude Code cloud sessions into the checkouts they touched - Sealed Claude cloud turns when SessionEnd is absent - Counted later usage tokens in Insights coverage - Tested Core capture integration and pinned cloud setup - Verified Claude cloud commits against captured sessions **Fixed** - Capture installation before cloud agent launch - Cloud capture prelaunch setup instructions ## An outcome-first Overview *Oct 9, 2026* 19 changes shipped: 3 new, 16 improved. **Added** - Capture-scoped usage evidence delivery endpoints - A fast, sliding Insights focus selector - An outcome-first engineering Overview **Improved** - Recorded explicitly requested unattributed CI/commit Receipts - Reported skipped cloud Receipt deliveries - Collected delayed cloud usage without changing sealed Receipt evidence - Delivered partial cloud Receipts without session-end hooks - Verify Receipt integrity over the stored record on every read path - Attached native Claude usage to immutable cloud Receipt segments ## Permission mode on every receipt *Oct 8, 2026* 20 changes shipped: 20 improved. **Improved** - Let an interrupted OSN worker cancel the run from its own thread - Sent Claude Code's permission mode on the hosted Receipt - Kept the hosted runtime_configuration block within the Platform's contract - Made the TUI's Checks block read the shared verification truth - Recorded Claude Code's permission mode on the Receipt as observed evidence - Kept the repository's .openshard/ runtime state out of git at setup ## Native runs, tested end to end *Oct 7, 2026* 35 changes shipped: 3 new, 31 improved, 1 fixed. **Added** - An OSN acceptance suite that runs representative tasks and summarises the Receipts - `osn diff` and recorded a completed run's outcome on its checkpoint - Unify provider insights and economic cost evidence **Improved** - Give the OSN verifier the repository's project instructions - Outlined plan files too large to show whole for the executor's first turn - Labelled OSN aggregate cost as all roles - Showed the planner's likely files to the executor on its first turn - Named every model a role used instead of crediting the last one with all its calls - Told the OSN executor when its turn budget is running out on inspection alone **Fixed** - The Windows Codex capture validation ## Cloud agents connect to Openshard *Oct 6, 2026* 29 changes shipped: 4 new, 23 improved, 2 fixed. **Added** - Red-team Receipt honesty tests and fix three overclaims - First-class cloud agent connections - Docker images and Railway Terraform for a dogfood stack - WorkOS AuthKit with safe account migration **Improved** - Capture GitHub-connector work from explicit PR metadata - Capture live Codex token usage from its runtime transcript - Name connected captures from the first hook - Accept a proxy-injected connected-capture credential - Attach later Cursor usage to the same Receipt - Measure Receipt evidence when individual capture dimensions are removed **Fixed** - Windows Codex transcript fixture - Agent icon branding: split Grok Bot/Build, add ChatGPT mark ## Connect and set up in one command *Oct 5, 2026* 12 changes shipped: 11 improved, 1 fixed. **Improved** - Connect Openshard and set up capture in one command - Prepare Openshard v0.4.12 - The Next Run Release - Supply hosted verified history to opted-in native prompt hooks - Hand off relevant learning automatically to Claude and Codex prompts - Simplify Insights questions to a compact composer - Give Connect the CLI one safe runnable command **Fixed** - Windows full validation for oversized learning input ## A task timeline you can trust *Oct 4, 2026* 18 changes shipped: 1 new, 17 improved. **Added** - Append-only workflow evidence and hosted task timeline **Improved** - Keep pre-upgrade connected journals eligible for verification delivery - Show connected capture status and validate bounded hosted learning - Keep connected capture sessions isolated and deliver learning evidence - Finish verified releases on existing GitHub release pages - Record workflow delivery evidence and remaining release gates - Refresh cloud capture identity immediately before each Receipt delivery ## Policy files in Settings *Oct 3, 2026* 13 changes shipped: 1 new, 11 improved, 1 fixed. **Added** - Policy file import and JSON export to Settings → Policies **Improved** - Document the openshard config command group - Keep headless Claude Code Receipts complete: transcript model and Stop-time commit/PR - Strip ANSI colour codes before parsing failing test ids - Capture truthful Claude Code Receipts: session tokens, owner, provider, commit and PR - Port the Learning Loop to a bounded, precomputed snapshot - Refresh the model catalog from current provider docs without changing defaults **Fixed** - Full Main Validation after the Learning Loop port ## Insights, graph first *Oct 2, 2026* 1 change shipped: 1 improved. **Improved** - Make Insights visual, compact, and graph-first ## Connected capture *Oct 1, 2026* 89 changes shipped: 17 new, 69 improved, 3 fixed. **Added** - Connected capture transport - Persistent connected capture package - Official-rate token cost estimation - PostHog TTFR analytics with privacy bounds - Production web start command - Hosted MCP OAuth consent screen - Dashboard OAuth consent API client - Hosted read-only MCP endpoint **Improved** - Make GitHub cloud capture provider-neutral - Capture Codex cloud work as progressive hosted Receipts - Restore complete tests with generic privacy sentinels - Remove private company-name guards from public tests - Restore full spool with OAuth secret filtering - Test OAuth secret filtering before capture **Fixed** - Live cloud Receipt workflow YAML - Connected capture lint - Pricing import order ## Remote capture for cloud agents *Sep 30, 2026* 21 changes shipped: 4 new, 17 improved. **Added** - Remote capture: stream evidence out of an ephemeral agent environment - `openshard learn` and measure outcomes with and without learning - Evidence-backed learning signal model - Remote capture: a hosted journal for ephemeral agent runtimes **Improved** - Document remote capture and the cloud provider capability matrix - Document verification evidence sync - Sync later verification evidence to the Platform - Document verification evidence history, CI binding and session close - Let later evidence strengthen a Shard's verification state - Document Learning Loop V1 ## Skills, imports and policy settings *Sep 29, 2026* 28 changes shipped: 6 new, 21 improved, 1 fixed. **Added** - Skill integrity metadata - Truthful MCP settings - Skills and historical Imports settings - Permission policy controls - Policy settings - Organisation policies **Improved** - Sync reconstructed historical Receipts - Enforce organisation permissions - Record per-stage token usage on Receipts - Enforce organisation policies in OSN - Sync recorded run owner - Refresh OpenAI model roster **Fixed** - Dashboard analytics populations and drill-down UX ## Smarter routing for native runs *Sep 28, 2026* 67 changes shipped: 8 new, 59 improved. **Added** - Hosted OSN evidence to changelog - Trajectory-aware Routing V2 for OSN - The OSN dogfood end-to-end test - Friendly organisation display names - Reusable badge primitive - Reusable switch primitive - Real dogfood controls for experimental features - Private controls for experimental features **Improved** - Mark 0.4.9 changelog - Update Supervisor hosted Receipt assertion - Update Routing V2 hosted Receipt assertions - Update dogfood hosted Receipt assertions - Update Agent Budget hosted Receipt assertions - Update adaptive routing hosted Receipt assertion ## Control, Proof and Cost on synced receipts *Sep 26, 2026* 9 changes shipped: 7 improved, 2 fixed. **Improved** - Project recorded control, routing and session evidence into the sync receipt - Show retry details and true run cost - Accept and show recorded Control, Proof and Cost evidence Core already stores - Insights v3: needs attention, what works, patterns, improve - Settings v1: account, organisation, members, API keys, CLI connect - Group Receipt Detail into Control, Proof and Cost **Fixed** - Receipt cost, retries and verification truth - Clipped verification provenance and squeezed card headers ## Policy runs before verification *Sep 25, 2026* 9 changes shipped: 2 new, 7 improved. **Added** - Routing outcome statistics - Bounded OSN execution loop **Improved** - Prepare v0.4.8 release - Enforce policy before Openshard executes verification commands - Classify run outcomes without blaming the wrong model - Run real coding tasks through OSN - Enforce policy before sandbox changes reach the repository - Group agent sessions under the same task ## Import your Claude Code and Codex history *Sep 24, 2026* 7 changes shipped: 2 new, 4 improved, 1 fixed. **Added** - Adaptive routing foundation with deterministic baseline - Dynamic model discovery and controlled routing classes **Improved** - Historical Ingestion v1: import Claude Code + Codex history with git evidence - Improve agent verification and add Openshard verify - Insights v2: decisions from Receipts - Represent Core's evidence faithfully on the dashboard **Fixed** - Avoid Windows race in Antigravity capture test ## Grok Build and Hermes Agent capture *Sep 23, 2026* 11 changes shipped: 3 new, 8 improved. **Added** - First-class Grok Build capture - Grok Bot capture with platform-observed and self-reported evidence - Hermes Agent capture (observation only) **Improved** - Prepare Openshard v0.4.7 - Ensure incomplete capture sessions are closed before sync - Keep balanced parentheses in task titles - Enable structured verification and task title sync - Integrated verification, task titles and Antigravity capture - Keep balanced parentheses in displayed task titles ## The hosted dashboard *Sep 22, 2026* 10 changes shipped: 2 new, 5 improved, 3 fixed. **Added** - Build authenticated dashboard home - CI validation for hosted dashboard **Improved** - Build Repositories experience - Build full receipt detail experience - Polish dashboard Receipt presentation and usage views - Build Receipt history experience - Build dashboard shell and Login experience **Fixed** - Match Receipt Detail to the canonical receipt design - Proxy local API requests through Vite - Pass CORS origin allowlist to Hono directly ## Dashboard foundations *Sep 21, 2026* 1 change shipped: 1 improved. **Improved** - Establish apps/web as the hosted dashboard foundation ## Receipts sync to the dashboard *Sep 17, 2026* 9 changes shipped: 4 new, 5 improved. **Added** - Openshard sync: retry-safe Receipt sync to the Openshard Platform - Task_id: explicit, additive engineering-task identity - Files via upload - Repository identity to the sync contract and a repositories endpoint **Improved** - CI acceleration: tiered PR gates, concurrency cancellation, pip caching, docs fast-path - Consolidated home/git helpers, fixed sanitisation drift, and dropped dead native loop modules - Create .gitkeep - Accept, store, expose and filter Core's explicit task_id - Document the Core sync client and the receipt revisions follow-up ## The hosted receipt foundation *Sep 16, 2026* 2 changes shipped: 1 new, 1 improved. **Added** - Phase 1 hosted receipt foundation **Improved** - Apply pre-merge security and data-model fixes from the architecture review ## Cleaner setup and help *Sep 13, 2026* 1 change shipped: 1 improved. **Improved** - V0.4.3: DX cleanup around setup, help, and receipt presentation ## Clearer telemetry wording *Sep 12, 2026* 2 changes shipped: 2 improved. **Improved** - Make local-only and telemetry wording consistent with the telemetry contract - Default basic telemetry on after agent-driven setup, with a human notice ## Cursor capture and privacy-safe telemetry *Sep 11, 2026* 3 changes shipped: 2 new, 1 fixed. **Added** - Privacy-safe product telemetry (v0.4.2) - Cursor agent capture (v0.4.2) **Fixed** - Capture-service drain livelock that stalled Windows CI ## Openshard v0.4.0 *Sep 5, 2026* 2 changes shipped: 2 improved. **Improved** - V0.4.0 release preparation - Modernized MCP SDK ## Codex and OpenCode capture *Sep 4, 2026* 4 changes shipped: 2 new, 2 improved. **Added** - Codex and OpenCode capture - Evidence-backed recovery **Improved** - Improved relevant context ranking - Make Claude capture near-zero blocking ## Richer Claude Code receipts *Sep 3, 2026* 6 changes shipped: 3 new, 3 improved. **Added** - Richer Claude receipts - Claude auto capture - Claude MCP installation **Improved** - Improved local visibility - Improved first-run onboarding - Improved capture performance ## A local MCP server *Sep 2, 2026* 9 changes shipped: 8 new, 1 fixed. **Added** - Relevant context - Local MCP server - Local history queries - Canonical Events to OSN - Canonical Events to Claude Code - Canonical Events to receipts - Canonical Event model - Persistent Shard attempt grouping **Fixed** - Claude receipt Shard identity ## Routing that learns from failures *Jul 20, 2026* 2 changes shipped: 1 new, 1 improved. **Added** - Failure-memory routing signal active by default **Improved** - Enabled per-role tier dispatch via config ## Safer per-role dispatch *Jul 19, 2026* 2 changes shipped: 2 new. **Added** - Provider-availability guard to per-role dispatch - Zero-result signal to tool-search provenance ## Smoother onboarding *Jun 18, 2026* 2 changes shipped: 1 improved, 1 fixed. **Improved** - Polished onboarding and dashboard UX **Fixed** - First run onboarding trigger ## First-run onboarding *Jun 14, 2026* 2 changes shipped: 2 new. **Added** - Read-only native agent loop foundation - First run onboarding ## Release readiness *Jun 13, 2026* 1 change shipped: 1 fixed. **Fixed** - Release readiness import order ## Model rosters and OpenRouter sync *Jun 12, 2026* 6 changes shipped: 2 new, 2 improved, 2 fixed. **Added** - OpenRouter metadata sync - Custom roster CLI **Improved** - Displayed routing role truth - Recorded per-role dispatch truth **Fixed** - OpenRouter cache import smoke test - Model policy CI lint ## Model policy config *Jun 11, 2026* 3 changes shipped: 2 new, 1 improved. **Added** - Model policy config - Provider-aware routing eligibility **Improved** - Enforced provider-aware model routing ## Advisory routing *Jun 10, 2026* 1 change shipped: 1 improved. **Improved** - Wired advisory routing into dispatch ## Short shard IDs on import *Jun 7, 2026* 3 changes shipped: 3 fixed. **Fixed** - Remove duplicate feedback group and subcommands introduced by rebase - Remove duplicate env_cmd definition introduced by rebase - Show short shard ID in import claude output ## Feedback and memory *Jun 6, 2026* 12 changes shipped: 9 new, 3 fixed. **Added** - Feedback subcommands (accept/reject/retry/note) - Rename --notes flag and add agent path detection - 6 more domain-specific workflow packs - Memory v0 - learn from feedback, openshard memory list/stats - Remove init gate, auto-detect API keys on first run - Openshard note command for attaching notes to runs - 6 new builtin workflow packs - Openshard packs run command **Fixed** - First-run UX - provider auto-detection, clean errors, agent mode - Resolve mypy overload error in _run_verification - Rename --from flag to --notes in openshard import claude ## Shard hash verification *Jun 5, 2026* 16 changes shipped: 10 new, 2 improved, 4 fixed. **Added** - Skills list command - Shard explainer documentation - Shard hash verification - JSONL write locking - Pre-send secret scanning - Model lifecycle tags - Curated model roster update - Demo Shard command **Improved** - Update what-is-a-shard.md - Polished verifier execution **Fixed** - Sandbox tests failing in environments with git commit signing - Stale claude-opus-4.6 reference in config.yml - Model registry drift - Routing truth in proof ## The Shard proof contract *Jun 4, 2026* 10 changes shipped: 5 new, 4 improved, 1 fixed. **Added** - Shard Proof Contract v1 - Changelog - Timeline and policy provenance - Evidence Provenance v0 - Run trust score **Improved** - Hid unavailable CLI options - Cleaned install dependency files - Hardened Shard schema coercion - Changed project license to Apache 2.0 **Fixed** - Sandbox test Git config ## A machine-readable run timeline *Jun 3, 2026* 14 changes shipped: 10 new, 2 improved, 2 fixed. **Added** - Machine-readable run timeline - Repo-aware plan mode - Repo map cache - Failure taxonomy stats - Shard completeness stats - CI policy check mode - GitHub Actions PR receipt outputs - Machine-readable proof outputs **Improved** - Hardened developer interaction trace - Unified runtime gate decision ordering **Fixed** - Repo map path sanitisation on Linux - CI check GitHub output test isolation ## Shard runtime metadata *Jun 2, 2026* 1 change shipped: 1 improved. **Improved** - Populated Shard runtime metadata --- Openshard, Inc. · Delaware corporation · michael@openshard.dev Install: `pipx install openshard && openshard setup` · Dashboard: https://app.openshard.dev · Source: https://github.com/openshard/openshard Agent index: https://openshard.dev/llms.txt · Whole site in one file: https://openshard.dev/llms-full.txt