collaboraitCASE STUDY · DRAFT ← site
case study 02 · dogfood, 2026-08-14

Three repos. Seven agents. 4.5 hours.

7 sessions2×claude5×codex3 repos~4.5 hours1 independent validation
evidence classes   ● receipt-verified — retrieved from session receipts/summaries by the authoring lane  ·  ● operator-attested — stated by the operator; machine re-verification pending

Read the numbers again: seven AI sessions across two providers restructured three interlocking repositories in one afternoon — and a separate lane signed off on the result. That is roughly a week of a well-coordinated team’s work, compressed into the time between lunch and dinner, with an audit trail.

The mission

A structural reorganization across three foundational repositories of the Sociail platform — the kind of cross-cutting change that usually blocks a team for a week, because every repo touches the others and nothing can move alone.

The fleet

Seven agent sessions — two Claude, five Codex — coordinated on the cait substrate: work split into bounded lanes per repository and concern, handoffs delivered fail-closed between them, and a separate CEO lane holding independent validation — the implementing lanes never graded their own work. Wall-clock from start to validated finish: about 4.5 hours.

CLAIMDETAILEVIDENCE
Fleet mix2 × Claude, 5 × Codex sessions, coordinated as one program of work● attested
ScopeStructural reorg across three foundational repositories (private)● attested
Wall-clock~4.5 hours, start to validated finish● attested
ValidationIndependent acceptance by a separate lane (CEO@macstudio), not the implementers● verified
Substrateicodex v0.3.0-rc.43-era coordination: handoffs, inbox, cross-mac summaries, receipts● verified

The shape of the work

7 sessions · 2× claude + 5× codex ›_ ›_ ›_ ›_ ›_ ›_ ›_ clone · repo A1 writer · bounded lanes clone · repo B1 writer · bounded lanes clone · repo C1 writer · bounded lanes cross-cutting sequence · repos change together, in order bounded handoffs · idle-only · one writer per repo GitHub · origin/dev — LocalDev + 9 repositories validated clean ›_ CEO codex · independent validation FOUNDATION_GREEN_ACCELERATED_DEVELOPMENT_RELEASED · ~4.5 hours wall-clock · receipt-verified receipts converge · validation is a separate lane
human act observation · read-only receipt claude session codex session ╌╌ clone fence / origin correlation

Why this one matters

The validation receipt

Retrieved from the CEO lane’s session summary over the workstation mesh, 2026-08-14 23:08 UTC:

Confirmed: the foundation-release goal is achieved.
project state: FOUNDATION_GREEN_ACCELERATED_DEVELOPMENT_RELEASED
LocalDev + nine repositories clean on dev@origin/dev · one checkout each, zero stashes
setup dry-run · first execute · byte-stable repeat · doctor: zero failures
binding mode 0600 · revision 1 · sha-256 1b51072c…

The DOCS lane, observed the same evening, had already moved to post-release guardrails — “stop agents recreating legacy topology… prove the clean foundation under ordinary development pressure” — the reorg not just done, but holding.

Honest limits

The fleet mix, wall-clock, and three-repo scope are operator-confirmed (2026-08-14; the estate-wide validation covered LocalDev + nine repositories); the outcome and validation are now receipt-verified — the validation record above was retrieved directly from the studio and its marker flipped, exactly as this page promised it would be. The scheduling was human: an operator authored the lane split and the handoffs. Making that faster and safer — measured acceptance seams, isolated workcells, bounded coordination — is precisely the roadmap.

Maturity note. Version at time of record: private predecessor icodex v0.3.0-rc.43 (pre-cait rename), before the v0.4 improvements. A 4.5-hour, seven-agent, three-repo structural reorg on the old version is the baseline, not the ceiling.
collaborait case studies are receipts-first: every claim carries its evidence class. corrections and re-verifications land as updates, never silent edits.