Weekly Product Notes: Pipelines, Invites, and Auditability

This retrospective covers April 21 through April 27, 2026.

The fourth week of April focused on how work moves through Bridge Town.

Pipelines became more explicit, invite and workspace flows became more visible, local CI runs got a cleaner upload path, and the audit backlog closed a large amount of platform debt. It was a week about making work traceable from source data to run output to user-facing notification.

Multi-model pipelines got a stronger runtime contract

Pipeline execution became more concrete.

Bridge Town added a branch-scoped /upstream runtime mount, ordered pipeline execution in the scaffold, upstream materialization, downstream scenario propagation tests, and clearer docs for the multi-model pattern. Output handling also converged around canonical result files so comparisons and downstream steps did not depend on incidental filenames.

Several fixes tightened that contract: exact snapshot prefixes are now preserved, _scaffold_outputs.json is unwrapped into cleaner metric keys, and built-in MCP instructions describe the pipeline and upstream workflow more directly.

The product implication is that model chains can become first-class workflows instead of ad hoc scripts glued together by convention.

Local CI runs became easier to wire into Bridge Town

The CLI and CI path also advanced.

Bridge Town added bt model run --ci headless mode, exit codes, a GitHub Actions guide, an example workflow, and a REST endpoint for uploading local CI run results back to a project. A setup bootstrap endpoint was added for the CLI path as well.

This gives finance engineering teams a better route from source-controlled model work into Bridge Town’s run history, without requiring every run to originate from the web app.

Invites and feed activity became more useful

Workspace and project invites became more visible across the product.

The week added multi-workspace discovery and switching, invite notifications for workspace and project flows, dismissal state, feed rendering, pending invite behavior across workspaces, and docs aligned to the invite notification model. Feed content later moved into the dashboard, and the dashboard getting-started card became dismissible.

These are small interaction details with large cumulative value. Collaboration only works if users can see what they have been invited to do and clear what no longer matters.

Audit and platform debt were cleaned up

A major audit backlog closed this week.

Bridge Town consolidated ORM models, sessions, and Alembic history; centralized the Gitea HTTP gateway with pooled clients, retry policy, and typed errors; adopted a canonical Gitea error path in routers; centralized web fetch behavior with timeouts and consistent error parsing; and hardened model run lifecycle state transitions.

Audit logging became more durable and observable with a bounded sink. API token prefix collisions were handled safely, tenant RLS context was delegated through the canonical helper, and migrations were made more deterministic and idempotent across legacy schema states.

This is the kind of cleanup that reduces strange edge cases months later.

Storage, outputs, and Sheets got practical fixes

Several workflow edges were tightened.

Snapshot listing and deletion paths were scaled, CSV run outputs could be returned directly as response bodies, Google Sheets query_data learned to handle preamble rows and blank-first-cell headers, and plan-copy plus free-plan invite errors were routed toward billing more clearly.

The theme was consistency: data should come back in the shape users expect, and failure states should point to the product surface that can resolve them.

Deploy smoke checks became more dependable

Production smoke work continued.

The deploy runner followed smoke output redirects, stabilized networking, fixed production smoke checks, restored drift checks for MCP tool docs, and recorded EC2 deploy handoff guidance. A handful of Next and web API build issues were also fixed so release confidence depended less on manual interpretation.

That kept the deployment path aligned with the faster-moving product surface.

What this means

By April 27, Bridge Town had better operational continuity:

• Pipelines had clearer runtime inputs and outputs.
• Local CI could publish run results into Bridge Town.
• Invite notifications connected collaboration to the dashboard and feed.
• Audit logging, migrations, gateway behavior, and state transitions were sturdier.
• Production smoke checks were more useful as a release signal.

This week turned a lot of scattered product motion into traceable workflows.