Tool call
Mapped the route boundary
It compared the root experience with explicit workspace routes before changing anything.
$rg "replaceState|workspace/" apps/web packagescode search · 14 matches narrowed to 3 owners
The Long Horizon / 001
We gave an agent a broken product and a moving target.
Its first plan failed. The interesting part was what it noticed next.
A replay-shaped demo of an Armalo mission: an unfamiliar workspace, incomplete clues, and a requirement that changes after the first promising fix.
Mission complete
- mission clock
- 6h 41m
- meaningful moves
- 23
- strategy pivots
- 3
The mission
Three things had to remain true.
- 01
The onboarding route appears healthy in a desktop walkthrough, but new collaborators cannot reach a usable workspace on a phone.
- 02
The agent inherits a codebase, a sparse bug report, and one guardrail: keep existing shared-workspace behavior intact.
- 03
Halfway through, the brief changes: the clean root URL must stay clean until a real first message is accepted.
CHAPTER 01
00:00–01:18
Read the terrain
The agent begins with the obvious question: where does a first visit become a durable workspace?
Discovery
The address bar was part of the product
The URL was not just navigation. Promoting it too early would turn a blank first visit into a shareable-looking workspace before the user had expressed intent.
assumption revised
CHAPTER 02
01:19–03:42
The first good answer was wrong
A clean local fix satisfied the route contract, until the agent tested the smaller viewport that the original report described.
Setback
A green path hid a broken journey
The initial repair made the URL transition correct, but the mobile composer lost its safe-area clearance and blocked the first action.
browser replay contradicted the first model
Tool call
Followed the visual symptom back to its owner
The agent stopped adjusting the route and traced the overlap through the shared workspace shell instead.
$inspect / → mobile viewport → composer boundsrender inspection · one layout owner
Decision
Separated promotion from presentation
Keep the server-accepted turn as the moment of durable intent. Make the visible URL transition a single idempotent presentation update after that acknowledgement.
chosen over eager client-side promotion
CHAPTER 03
03:43–06:41
Find the smaller, truer fix
With the actual constraint visible, the mission becomes smaller: preserve the shared room, repair the first five minutes.
Unexpected lead
The bug was a timing problem wearing a layout costume
The overlap only appeared when a route promotion and a mobile viewport resize arrived in the same short window.
the route was not the whole story
Artifact
First-turn journey patch
A narrow change that preserves the draft URL, promotes only after acknowledgement, and reserves mobile composer space.
diff · route contract · responsive shell
Mission output
The artifact, not just the story
A durable first-turn experience: a clean entry point, a visible acknowledgement, and a workspace that remains usable at the moment a new collaborator needs it most.
- route contract
- mobile shell
- shared workspace
What this replay shows
- The original brief remains visible next to the final scope.
- Each pivot is attached to the event that forced it.
- The output is a replay fixture here—not a claim about a customer deployment.
The verdict
The mission did not get easier. The agent got a better map.
The important moment was not the first correct-looking patch. It was the point where the environment disproved it—and the work changed shape instead of merely continuing.
Outcome
Replay verdict: completed in fixture
Behind this replayWhat is fixture data, what a real mission would add, and why this page avoids pretending otherwise.
This page: static demo fixture content, made to show the intended public replay experience.
A real mission: ordered runtime events, actual artifacts, scoped redaction, and a link back to its originating workspace.
The editorial rule: the story can be exciting without making a demo, simulated run, or partial outcome sound real.
Your turn