Latest dated assessment
October 8, 2026 · Gimbal b6e2c07e · primary intent partially met. The review workflow found the planted bug and made no tracked changes. The console showed only “Completed” at a glance and displayed a stale “Running” label beside the finished run. Six of 46 catalog claims had a passing named check. The newer workflow design pages have separate implementation evidence below.
What was assessed
This assessment was made on October 8, 2026 against Gimbal b6e2c07e. It is a dated observation, not a live health indicator. The complete reconciliation covers all 46 registered claims, source links, checks, screenshots, and the limits of each result.
Twelve exact named checks passed at that revision, covering six claims. The other 40 claims have no passing record from this catalog evaluator. Source presence does not validate behavior, and the absence of a failing check record does not erase an observed product failure.
Since that assessment
Gimbal PR #436 merged on October 9 as 37f8e3a9. The latest public tag checked for this update was v0.12.1, which predates that merge. The October 8 evaluator has not been rerun at the newer revision.
| Capability | Implementation at the named revision | Evidence and limit |
|---|---|---|
| Workflow design pages | Main 37f8e3a9 includes clickable source-derived diagrams, prompt and context inspection, URL deep links, a usage guide, and generated run forms. Development source edits refresh the view. | The PR reports 15 workflow/source-update tests, 115 frontend tests, six production browser scenarios, and a live Go comment edit that retained selection, expansion, and camera. Gimbal View has not rerun those checks or a browser evaluation at this revision. The feature is merged on main, not in the latest tagged release. |
| Project conversations | The October 8 binary includes text conversations with Codex, Claude, or Antigravity, each associated with a Git branch and worktree. The conversation agent can launch a built-in workflow and the UI links its run. | Source and earlier prototype work support implementation. The 46-claim catalog has no dedicated project-conversation claim or current named View evaluation for this whole experience. Native-session resume after restart is implemented for Codex in source; equivalent behavior is not established for the other providers. |
| Backend integration | A separate Temporal consumer uses the public runtime and compiler surface with its own scheduling and deployment choices. | This bounded example supports the consumer-compilers partial stage. It does not establish arbitrary backends, project-local hosted workflows, or general remote execution. |
These capabilities are candidates for clearer catalog promises; the direction states the intended experience. The existing 46-claim inventory remains unchanged while wording and acceptance are settled. A PR’s recorded browser result is attributed to that PR rather than counted as a new Gimbal View evaluator pass.
What the reviewed binary included
At b6e2c07e, the reviewed binary registered seven workflows: sprint-plan, review, implement, research-document, research-eval, pyramid-summary, and validate-product. It also included five harness adapters: Codex, Claude Code, Antigravity, OpenCode, and native Pi. The local runtime had source-derived graphs, persistent run records, controls, project conversations, and a console. A bounded external Temporal consumer demonstrated part of the public compiler contract; it did not establish every infrastructure target or deployment mode.
These are implementation observations. Each workflow’s useful result, model quality, and human legibility still need evidence appropriate to that claim.
What the console showed
A fresh, read-only review run against a small synthetic repository found the planted age-boundary bug and made no tracked changes. The Result inspector rendered the finding clearly. The run list showed only “Completed,” without exposing the finding; Focus displayed the structured response as raw JSON. On the completed run, a User activity card still said “Running,” and a recorded turn offered “Back to live.” Steering was correctly closed for the ended turn.
This one-node completed run does not prove live-stream freshness, complex-map readability, multi-agent legibility, human steering, or cancellation. The September 26 console evaluation records earlier observations; only the issues named above were freshly reproduced in this assessment.
Known gaps and limits
- At the assessed revision, the primary intent was partial: a person had to drill into result details to learn the outcome of this review, and contradictory recorded state weakened first-glance trust. No comparison established the “best-of-class” ambition.
- Gimbal #424 records an
implementrun whose independent QA accepted task-level work while required outcome work remained. Passing scripted checks do not establish reliable complete-outcome judgment in a real run. - Gimbal #389 records a research-document factuality failure on an earlier revision. New auditing implementation exists; this assessment did not rerun the contrary-source case to show the failure is resolved.
- The bounded role-catalog proposal is not delivered by today’s role/model binding alone. Several shipped default role names fall outside
roles.go. - Local CLI and GUI execution are implemented, while the stronger infrastructure-blocked acceptance, multitier isolation, and away-from-developer-machine evaluation remain unproven here.
The feature inventory keeps implementation stage and revision-bound validation separate. The full report names additional capabilities missing from the inventory; proposed promise wording is being reviewed before any new claim is registered.