Skip to content
Chapter 7

The brain steps out of the files

March 18, 2026

Chapter 7 — The brain steps out of the files

Verified sources: git log 2026-03-18 (130 commits) Key commits: 942371d (Cosmos in the nav) · 6e876cb (brain-guardian) · 5e643d7 (BSI v4 intentions) · cadafc6 (pre-flight gate) · ffabc10 (swarm foundation) · 9a8a955 (Cosmos nebula heatmap)

March 18, 2026. The brain has been building itself for six days. The agents are forged. The kernel is locked. The API is online. The MCP talks. But the brain remains a prisoner of text files and the terminal. That day, it steps out.

Cosmos lands

The night begins where the previous one stopped. Cosmos — the 3D map of the embeddings, born as a vision in chapter 4 — must become real. Not a demo in a notebook. A component in brain-ui, reachable from the browser, with the brain’s real memory fragments.

At 01:20, the deploy begins. At 01:42:

942371d — feat: Cosmos 🌌 dans la nav brain-ui — separator + lazy load (“feat: Cosmos 🌌 in the brain-ui nav — separator + lazy load”)

Cosmos is in the navigation. But it does not work yet — 401 on the fetch, UMAP coordinates out of frame, a canvas that does not fill the page. The kind of bugs you only find in real integration.

At 04:56, authentication fix. At 05:07, UMAP coordinates normalized to [-2, 2]. At 05:09, full-height canvas. At 05:12, shader for round points (clamp 1.5-5px). At 05:14, fullscreen button. At 05:18, position absolute chain.

Seven commits in 22 minutes. Each fix surgical, each commit isolated. The brain’s pattern — one problem, one commit, one proof.

At 05:20:

d8d1cd9 — bsi: cosmos-deploy intention — done ✅

Cosmos is live. The brain’s memory fragments float in 3D in the browser. Each point is an embedding chunk. Clusters emerge — the agents group together, the projects form constellations.

Then at 08:00, the aesthetic gesture:

9a8a955 — feat: cosmos heatmap — mode nébuleuse toggle ⬡ (AdditiveBlending, soft gaussian) (“feat: cosmos heatmap — nebula mode toggle ⬡ (AdditiveBlending, soft gaussian)”)

Nebula mode. The points become glowing halos. Additive blending — dense zones shine brighter. Automatic rotation. The brain seen as a galaxy. It is no longer a dashboard — it is a living wallpaper that shows the state of knowledge.

The brain learns to distrust itself

At 03:44, between two Cosmos fixes, a discreet but fundamental agent:

6e876cb — feat: brain-guardian — auto-méfiance structurelle quand brain opère sur lui-même (“feat: brain-guardian — structural self-distrust when brain operates on itself”)

brain-guardian. Not secrets-guardian — that one protects secrets. Brain-guardian protects the brain from itself. When the brain operates on its own kernel files, when it modifies its own rules, a safeguard activates. Structural self-distrust.

The question is not “est-ce que l’agent est compétent ?” (“is the agent competent?”). The question is “est-ce qu’un système qui se modifie lui-même a besoin d’un contrôle supplémentaire ?” (“does a system that modifies itself need an additional check?”). The answer is yes. Always.

At 03:51:

99e4f2b — feat: session-kernel + session-edit-brain — séparation workflow brain/kernel effective (“feat: session-kernel + session-edit-brain — brain/kernel workflow separation in effect”)

Two separate session modes. session-edit-brain: writing allowed on the brain. session-kernel: read-only on the kernel, any write attempt refused. The distinction is architectural — not a convention you can forget, a mode that applies at boot.

BSI v4 — intentions as the unit

At 04:35:

5e643d7 — feat: BSI v4 — intentions/ comme unité de session multi-sessions (“feat: BSI v4 — intentions/ as a multi-session session unit”)

Intentions. Not todos — directions. Not tickets — commitments. Each intention has a life cycle that crosses sessions. An intention opened today can be worked on in three different sessions on two machines. BSI v4 makes it a steering unit.

At 04:36, the TTL check at boot: if an intention has not been touched for longer than its TTL, stale alert. The brain knows what it promised to do — and it knows when it forgets.

The pre-flight gate

At 04:02:

cadafc6 — feat: pre-flight — gate boot step 4.5 BHP, enforce tier_required + write_lock + kerneluser

The pre-flight. A check that runs at boot, after the context is loaded but before the session starts working. It verifies three things: the user’s tier is sufficient for what is about to be loaded. The write locks are respected. The kerneluser is identified.

If the pre-flight fails, the session does not start. No silent degradation — an explicit refusal with the reason. The brain would rather not start than start in an inconsistent state.

The wiki writes itself

Between 06:43 and 07:34, the brain’s wiki goes from 20 pages to 26. Page after page, each concept of the brain receives its documentation.

90f29ef (06:43) — cold-start fix, MCP onboarding, context-loading, brain-access, veracity audit. 1e1a49a (07:15) — Cosmos.md + 21 pages. e5ceb8b (07:19) — brain-engine + tier enforcement + 23 pages. 2977fdf (07:23) — BSI reference + 24 pages. 08566a8 (07:26) — ADR index + 25 pages. e5da2ac (07:34) — integration-schema + 26 pages.

One hour. 26 pages. The brain is no longer content to know — it documents itself. And every page goes through a veracity audit: what is written must be true in the repo, verifiable by commit.

The afternoon of audits

The brain wakes up at 17:05 after a break — the first real break in days. The afternoon is given over to maintenance.

At 18:18:

6d19bec — audit: CATALOG.yml sync — 6 agents manquants ajoutés (“audit: CATALOG.yml sync — 6 missing agents added”)

The CATALOG — the official list of all agents — is synchronized. 73 agents out of 74. Six were missing: coach-boot, time-anchor, audit, pattern-scribe, brain-guardian, pre-flight. All forged recently, none indexed. The kind of silent drift that accumulates when you build fast.

Swarm foundation

At 18:49, a commit that looks far ahead:

ffabc10 — feat: workflows/_template enrichi — execution_mode (ADR-032) + contrats I/O + swarm_ready flag (“feat: workflows/_template enriched — execution_mode (ADR-032) + I/O contracts + swarm_ready flag”)

The swarm. The idea that the brain’s workflows should not be executed one by one, sequentially, with a human copy-pasting between windows. They should be able to coordinate — autonomous agents that launch, exchange results, and converge.

ADR-032 defines three execution modes: manual (the human steers), assisted (the coach steers, the human validates the gates), agentic (the agents coordinate, the human steps in only by exception). The swarm_ready flag in the workflow template says: “ce workflow est prêt pour le mode agentic.” (“this workflow is ready for agentic mode.”)

The foundation is laid. The swarm will not be built today — but when it is, the architecture will be ready.

The guided review

The afternoon ends with a heavy audit. A review plan extended to all 74 agents, batch by batch; about twenty receive a patch. The IPCs corrected — sends_to, receives_from made explicit. The contexts/ paths corrected. The coach receives ADR-032 in its knowledge. The brain-ui-scribe is updated — Sprint 2 delivered, 13 components, 8 hooks.

Eight patch commits in 40 minutes. Nothing new — maintenance. The brain taking care of itself.

The wiki ends the day with 36 indexed ADRs, 26 pages, version 0.9.4.

130 commits. The brain has stepped out of text files and the terminal. Cosmos floats in a browser. brain-guardian protects the kernel from itself. Intentions are multi-session steering units. The pre-flight refuses to start in an inconsistent state. The wiki documents everything. And the swarm waits for its hour.

The title of this chapter has not changed — “Le cerveau sort des fichiers.” (“The brain steps out of the files.”) But now, it steps out for real.

What was built

  • Cosmos live — full deploy, 6 fixes in 22 min, nebula heatmap (942371d→9a8a955, 01:42→08:00)
  • brain-guardian — structural self-distrust when brain operates on itself (6e876cb, 03:44)
  • session-kernel + session-edit-brain — separation of brain/kernel modes (99e4f2b, 03:51)
  • BSI v4 — intentions/ as a multi-session unit + boot TTL check (5e643d7, 04:35)
  • pre-flight gate — tier + write_lock + kerneluser enforcement at boot (cadafc6, 04:02)
  • 26-page wiki — complete documentation in one hour (06:43→07:34)
  • CATALOG sync 73/74 — 6 missing agents added (6d19bec, 18:18)
  • Swarm foundation — ADR-032, execution_mode, swarm_ready flag (ffabc10, 18:49)
  • Guided review — 21 agent files patched: IPC + paths + composition (19:09→20:20)

Key decisions

  • brain-guardian: self-distrust. A system that modifies itself needs a dedicated safeguard. Not because it is incompetent — because modifying its own rules is structurally dangerous.
  • pre-flight: refusal > degradation. The brain would rather not start than start in an inconsistent state. The pre-flight checks the prerequisites before the session begins.
  • BSI v4: intentions cross sessions. A todo ends within one session. An intention can live across ten sessions and two machines. The TTL detects abandonment.
  • ADR-032: three execution modes. Manual → assisted → agentic. The architecture prepares the swarm before the swarm exists.
  • Maintenance is work. About twenty agents patched, CATALOG synced, wiki enriched. Housekeeping days are not lost days — they are the days that keep drift from becoming debt.