Project timeline

Built the Cohort UI shell

Sidebar, channel thread, composer, context rail. Handled empty, loading, and error states in the first pass without being asked.

Design review — Cohort UI shell

Structurally sound. No state findings this time; remaining notes were about overclaiming, not correctness.

Built the Cohort data layer

Agent roster config, live harness markdown reader, seeded message store, and the GET/POST API routes.

Next.js 16 route params are async

Dynamic route handlers receive params as a Promise. Caught during implementation, logged so the next worker doesn't rediscover it.

Design review — Cohort data layer

Context rail correctly distinguishes 'file missing' from 'file is an unfilled template' — a real empty state, not a generic one.

Reviewed the Cohort data layer

Checked the empty and not-found states in the context rail.

Reviewed the Cohort UI shell

Flagged that seniority badges could be read as an evaluated rank when they are currently static placeholders.

Set the design direction for Cohort

One visual climax per screen (the active thread). Registry seniority as a quiet badge. Context rail reads as reference, not the main event.

A durable log reads differently than a live chat

Recorded in the UX guide: this product is something you review, not something you keep up with. Shapes scroll, density, and notification behavior.

Briefed the Cohort build

Asked for a workspace UI over the real harness rather than a mockup — reading actual repo files where it adds value.

UX.md becomes a harness convention

design.md covers visual style; nothing covered interaction conventions, research, or domain vocabulary. Added UX.md as its counterpart, owned by Hephaestus, with facts staying human-sourced.

Agents need a visible growth record

A seniority badge asserts standing without showing the work behind it. Called for a per-agent timeline that doubles as what someone would bring to a performance review.

Cohort ships as its own repository

Still defining what this is and where it goes. Published standalone to gather feedback, with the harness installed into it so the demo reads a real one.

One timeline for the whole team

Per-agent history answered 'how did this agent grow' but not 'what happened on this project'. The project view treats human and agent work as one stream.

Applied the Parallel Split Gate to Cohort

Split rejected: Julius's components consume the shapes Augustus defines, so the queues fail the producer/consumer check. Sequenced instead of running both.

Recorded current product truth for Cohort

v1 shell over real harness roles with a mock conversation layer. Figma export, edit-sync, and agent wiring named explicitly as out of scope.

Density is not the same as clutter

Recorded after three reviews where simplification would have removed the operational context users needed to act.

Promoted to Mid

Stopped needing states enumerated for him. Two consecutive artifacts shipped without a blocking state finding.

Junior → Mid

Re-review — settings panel

All findings addressed. Noted that states were now designed, not bolted on afterward.

States are part of done, not a follow-up

A screen isn't complete when the happy path renders. Read the state model before building, not after review.

Fixed all three blocking issues

Reworked the panel against the design contract instead of patching around the findings. Asked for the state model up front rather than guessing.

Design review — settings panel

Every required state was missing. The panel looked finished and behaved as if nothing could ever go wrong.

Reviewed the settings panel

Returned Revise. Three blocking state issues; assigned the fixes back to Julius rather than redesigning around them.

Built the settings panel

First assigned UI task. Shipped the layout and the happy path; treated states as a follow-up rather than part of done.

Promoted to Senior

Consistently designs for the failure path without prompting. Now trusted to define data contracts other workers build against.

Mid → Senior

Design review — import pipeline

Partial-success handling was thorough. Users could tell exactly which rows failed and why.

Reviewed the import pipeline

Confirmed partial-success and many-record handling held up at scale. Approved without fixes.

Built the record import pipeline

Batch import with validation and a partial-success report.

Sequenced the import pipeline work

Single worker. Schema and consumers were coupled, so the queue stayed sequential rather than split.