Changelog

What's new in Watari

Every meaningful change to the support-ticket-to-draft-PR pipeline, newest first.

Subscribe via RSS.

September 2026

Before Watari pushes a draft PR, it now runs the build your repository already uses, so Java, Kotlin, Rust, Ruby and every other language get the same check TypeScript, Python, Go and C# had. A draft nothing could check says so on the PR.

  • Watari reads your build from your CI definition (GitHub Actions or Azure Pipelines) and runs it in an isolated sandbox, once on your branch and once with the fix. When a fix breaks your build, Watari sees the errors and tries again before your team sees the PR.
  • Prefer to say it yourself? Add a build command, and an optional setup command, in the repository's settings. It is used instead of Watari's built-in checks.
  • Every draft PR now has a Validation section that says what was checked. When nothing could be checked, it says the draft is not validated and how to change that.
  • Your CI still comes first: where your pipeline already builds Watari's pull requests, its result is the check.

When your CI fails on a draft PR Watari opened, Watari reads the error and pushes a fix commit to that branch, on GitHub and Azure DevOps alike. On by default.

  • A draft PR your team opens should already build. When a compiler error, lint error or failing test shows up on a Watari draft, Watari pushes one follow-up commit that fixes it and says so in a comment on the PR.
  • It only touches its own work: Watari fixes a failure on a commit it pushed itself and never builds on top of anyone else's. If someone pushes while Watari is working, nothing is written.
  • It only acts on code: a package feed, network or build agent problem gets no commit, and a failure it already tried to fix is not tried again.
  • The same behaviour on GitHub and Azure DevOps. Turn it off or allow up to five fix commits per PR in Settings, AI and pipeline, under Fix failing CI on draft PRs.

The billed unit is renamed Mapped Issue, because a code-actionable feature request now counts the same as a bug. Tickets and Requests are one Issues list with a type switch.

  • The unit you are billed for is called a Mapped Issue: a bug, or a code-actionable feature request (a specific change to code you already have), paired with the file and function Watari ranks as the most likely cause. Nothing about your bill changes with the name.
  • An open-ended feature request, and any question, is never mapped and never billed. It is still filed to your issue tracker.
  • Tracker sync is never billed. Filing to Jira, Linear, GitHub Issues or Azure Boards stays bundled, whatever those tools call the record they create.
  • The sidebar used to carry Tickets and Requests as two entries pointing at the same list. They are now one Issues entry with a Bugs / Requests / All switch on the page itself, the way Linear, Jira and GitHub all shape it.
  • The list moved from /tickets to /issues so the address bar matches what the product calls it. An old /tickets link no longer resolves. The meter doc is now /docs/billing/mapped-issue-meter, and the old doc URL redirects.

Before writing a fix, Watari now searches your repository to confirm which file actually carries the defect, so a fix can land in a file the mapping stage never picked.

  • Searches the repository before drafting, instead of trusting the ranked guess
  • A fix can now edit a file that was not in the mapped locations
  • Skipped on follow-up commits, which already know the file they are changing

Four changes to how Watari searches your codebase, measured on the published benchmark: the top ranked file is one the merged fix actually changed in 36 of 56 cases, 64.3%, with a 95% interval of 51.2% to 75.5%. The first published run scored 8 of 17, 47.1%, on a smaller sample.

  • A whole class of function was invisible to the indexer: anything defined by wrapping a call, which is how most handlers, hooks and factories are written. Those bodies are now indexed.
  • A ticket describes a symptom and code implements a mechanism, so the search now also looks for the mechanism the symptom implies rather than matching the customer's words against source.
  • Each repository gets enough search depth for the right file to reach the ranking step, instead of being cut off by a shared budget.
  • Files a fix routinely edits are indexed rather than skipped: Vue, Svelte and Astro components, stylesheets, tree-sitter queries and XML build files. Golden-file test snapshots are excluded, because they were crowding out real code.
  • The number and every case behind it are at [/proof](/proof), including what got worse: repository routing went from 56 of 56 to 55 of 56. The change was chosen by looking at these same cases, so read it as a ceiling rather than as evidence it generalizes.

August 2026

A pre-registered measurement of how often Watari reads a bug report and points at the file that actually had to change, with the protocol, the frozen sample and the unedited results in a public repository.

  • Ground truth is the maintainer's own merged fix on public issues from six open-source projects in six languages, so a third party produced the answers before we ran anything.
  • The protocol was committed before the sample was frozen, and the sample before the results. That ordering is checkable with git log in the public repository, not something you have to take on trust.
  • Every case is published, hits and misses alike, with the file Watari chose next to the file the maintainer changed. Every rate carries its denominator and its confidence interval.
  • It measures one step of six. It says nothing about whether the draft pull request is any good or whether the customer-facing incident report reads well, and we do not put a number on those.
  • Read it at [/proof](/proof).

A second Jira connect path using an Atlassian API token, plus alerts when a tracker stops accepting what Watari files.

  • Jira sites that require admin approval before an app can be installed can now be connected with an Atlassian account email and API token instead, which sits under a different admin policy. Credentials only ever go to your Atlassian Cloud site.
  • A Jira project without a Bug issue type (team-managed Scrum projects have none) no longer rejects every filing: the issue type is chosen from the project's own schema.
  • If a destination stops accepting issues, whether the site was deleted, the project is gone or the credentials were revoked, the integration is marked disconnected and you get told once, rather than silently failing on every bug.

July 2026

Watari now reads whatever your customer sent: PDFs, screen recordings, voice notes, logs, HAR captures, CSVs and Office documents, not only images.

  • Screenshots and screen recordings are read visually, and a recording's narration is transcribed too. The clearest description of a bug is often something the customer said out loud while reproducing it.
  • PDFs are read page by page, including scans with no text layer, so an invoice or an exported report attached to a ticket is evidence rather than a download link.
  • Logs, HAR captures, stack traces, CSVs and Office documents are read as text and quoted verbatim into the bug report. A stack trace's real error is at the end, so long files are trimmed from the middle and the omission is marked.
  • Zendesk tickets used to lose any file attached to the opening description when an agent's reply fired the webhook, and Freshdesk tickets lost every reply past the tenth. Both now reconcile against the full conversation.
  • Archives are still stored for download and deliberately not opened.

Azure DevOps (Azure Repos) joins GitHub as a supported code host, the full ticket-to-draft-PR pipeline now runs on Azure DevOps at parity with GitHub.

  • Connect Azure Repos in Settings → Integrations through Microsoft Entra ID: tenant-backed OAuth with short-lived, admin-revocable tokens, not a personal access token. Pick your organization, project, and repository.
  • Full pipeline parity with GitHub: index the repo, map each bug to the responsible file and function, open a draft pull request, watch build validation, and hold the customer-facing RCA until an Azure Pipelines production stage confirms the deploy.
  • Your branch policies run untouched: Watari publishes the draft so build validation and required reviewers apply exactly as they would on a human-opened PR. Watari never bypasses a policy.
  • Optionally add Watari's service principal to your organization to have pull requests authored by "Watari" instead of the connecting user, the way Renovate and Copilot operate on Azure. Skip it and PRs are simply human-authored; nothing depends on it.
  • Mix GitHub and Azure DevOps repositories in one workspace, a bug is routed to whichever repositories its code actually lives in, regardless of provider.

A production-readiness pass that makes billing, recovery, and dashboard numbers trustworthy under real-world failure.

  • Mapped-Bug metering and Razorpay webhooks are now replay- and retry-safe, so a redelivered event can never double-bill you.
  • Watchdogs recover repositories stuck mid-index or mid-mapping instead of leaving them wedged.
  • Every outbound provider call is wrapped in a timeout and circuit breaker, so one flaky third party can't cascade.
  • Dashboard metrics show real deltas: placeholder and estimated numbers were removed.

Freshdesk joins Zendesk and Intercom as a supported support tool, escalated tickets flow into the same ticket-to-draft-PR pipeline.

  • Connect in Settings → Integrations by pasting your Freshdesk domain and API key, validated live against your account.
  • Inbound tickets arrive over a per-tenant verified webhook you wire into one Freshdesk Automation Rule, so only escalated tickets reach Watari.
  • The API key is AES-256-GCM encrypted at rest and tenant-isolated by row-level security, like every other integration.
  • The customer-facing RCA is published back as a public reply on the originating Freshdesk ticket once the fix ships.

June 2026

File mapped bugs into GitHub Issues alongside Linear and Jira, an explicit, opt-in destination you choose per workspace.

  • Pick the destination repository explicitly, like Linear and Jira; it's off until you turn it on.
  • Only bugs that clear the Mapped-Bug confidence bar are filed, and each can route to the repo it actually maps to.
  • Fixes the old behavior of filing into the oldest-connected repo for multi-repo workspaces.

Connect more repositories as you grow, and let a single bug fan out into coordinated draft PRs across every repo it touches.

  • Repo count scales with your plan; Growth and above unlock cross-repo PR fan-out.
  • One bug that spans several repos opens a coordinated draft PR in each, routed to the right reviewers.
  • Downgrades deactivate excess repos instead of deleting them, with an in-app chooser for which to keep; upgrading reactivates them.

Bugs now route to the repositories whose code actually matches, replacing a summary-based guess that could miss the right repo.

  • Routing is driven by a cross-repo search over your indexed code, not a prose summary of each repo.
  • PR fan-out targets are derived from the bug's ranked code locations.
  • Each mapped location shows a source-repo badge, and re-mapping takes your hints into account.

A bounded-memory streaming indexer lets big repositories index without running out of memory, with live progress.

  • The indexer streams and flushes in bounded batches, so memory stays flat regardless of repo size.
  • Repos up to tens of thousands of files index cleanly; larger ones are flagged gracefully instead of failing silently.
  • Watch indexing progress live, with automatic recovery if a run stalls.

A calmer, faster workspace: a borderless sidebar on a paper canvas, one floating content card, and a real Inbox.

  • A dedicated Inbox route replaces the old notification bell.
  • A sidebar search row opens the ⌘K command palette from anywhere.
  • Settings moved into the profile menu, and a Help launcher sits in the footer.

The pipeline keeps only vector embeddings and structural metadata, your actual source is fetched on demand and discarded.

  • Source file contents are no longer persisted anywhere in Watari.
  • When code is needed to map a bug or draft a fix, it's fetched from GitHub at the indexed commit and thrown away.
  • Tokens and secrets stay AES-256-GCM encrypted, and tenant data stays isolated by row-level security.

May 2026

Watari extracts bugs from tickets in any language and publishes the RCA back in the customer's original language.

  • Customer verbatims are preserved untouched, including mixed languages like Hinglish or Spanglish.
  • Screenshots and short screen-share clips are read by a vision-aware model.

Every Watari-drafted PR shows the same green/red CI signal GitHub does, mirrored into the dashboard and Slack.

  • Reads the check and status events your branch protection already trusts, no new CI integration.
  • The PR tab shows passing, running, or failing with a per-check breakdown.
  • The Slack notification carries a live CI line that updates in place and pages reviewers on the first failure.

Leave a review comment on a Watari-drafted PR and it proposes the follow-up commit on the same branch within minutes.

  • A classifier reads each comment, and only actionable feedback triggers a regeneration.
  • Follow-ups are committed to the same branch: no force push, no history rewrite.
  • A per-workspace iteration cap, with a Slack heads-up if the fix isn't converging.

Pay only for Mapped Bugs (a bug paired with the exact file and function responsible) with PR drafts, RCAs, and tracker sync bundled in.

  • A bug is billable only when both extraction and code-mapping clear a confidence bar; the meter fires once, at persist time.
  • PR drafting, customer RCAs, and Slack, Linear, Jira, and GitHub sync are bundled, never metered.
  • Billed in INR for India (with GSTIN) or USD elsewhere, with GST-compliant invoices and a seven-day mismapped-credit window.

April 2026

Watari now turns real support tickets into structured bugs, Zendesk tickets and Intercom conversations flow in by webhook.

  • Scope what reaches Watari with a Zendesk trigger tag or an Intercom team assignment.
  • Each ticket is normalized and structured against a strict schema: severity, repro steps, impact, and the customer's own words.
  • Non-bug tickets like billing and account questions are classified and skipped.

Watari writes the fix and the test, checks them, and opens a draft PR, the core support-ticket-to-PR loop.

  • An extended-thinking model drafts the patch plus a test, validated before the PR opens.
  • The draft PR lands in your repository with live status in the PR tab.
  • Slack notifies you when the PR is created and when it merges.

March 2026

The first release: a multi-tenant SaaS foundation with the dashboard, notification delivery, billing, and team management the rest of the product builds on.

  • Multi-tenant workspaces with role-based access and team management.
  • Delivery to Slack, Linear, and Jira.
  • Razorpay billing with a free trial.
  • The dashboard shell, authentication, and onboarding.