Mirrored verbatim from utopia-models/knowledge wi/decisions/ADR-7799-no-root-claude-md-in-knowledge.md by the allowlisted publish pipeline. Edit it there, not here.

artifact_type: adr
adr_id: ADR-7799
adr_status: accepted
status: active
decision_date: 2026-09-13
supersedes:
superseded_by:
issue: 7799
subprocess: process
references:
  - wi/work/specs/7799-masters-stop-loading-architect-bundle/spec.md
  - wi/work/research/2026-09-13-plan-267-the-remaining-16-workflows-verdicts.md
  - _infra/scripts/kb-lint.py
  - _infra/scripts/task-runner.sh
  - _process/north-star.md

The knowledge repo has no root CLAUDE.md — every agent loads only its own class bundle

Produced by the process subprocess (wi/subprocesses/dispatch/process.md): it changes what a dispatched run loads and what task-runner.sh's brief tells it to read — the execution path itself. Filed from #7788 (PLAN-267 P7), the ADAPT verdict on the OSS-tools / AGENTS.md slice. Parent Feature #7781.

Context

#7316 (PLAN-256 P3, 2026-08-31) made the repo-root CLAUDE.md one line, @_process/_generated/bundle-architect.md, and fixed the root at exactly three files in kb-lint.py R1. Claude Code loads the CLAUDE.md in the working directory and every directory above it, and task-runner.sh runs claude -p inside a knowledge per-run worktree. So every master run in this repo loaded the Architect's bundle on top of its own.

Measured 2026-09-13, before this change:

$ readlink -f ~/.claude/rules/bundle.md          # on each box
win    /home/win/dev/utopia-models/knowledge/_process/_generated/bundle-master.md
nova   /home/nova/dev/utopia-models/knowledge/_process/_generated/bundle-master.md
cloud  /home/cloud/dev/utopia-models/knowledge/_process/_generated/bundle-master.md
dust   /home/dust/dev/utopia-models/knowledge/_process/_generated/bundle-master.md
kxdev  /home/kxdev/dev/utopia-models/knowledge/_process/_generated/bundle-architect.md

$ wc -c _process/_generated/bundle-architect.md _process/_generated/bundle-master.md
 64618 bundle-architect.md
 26550 bundle-master.md

$ cd <knowledge per-run worktree on win> && claude -p 'Print verbatim every line … "**Loaded by:**"' --max-turns 1
**Loaded by:** win · nova · cloud · dust — executes a filed issue.
**Loaded by:** the Architect (`agent`) — plans with Tyler, files, dispatches.

$ ssh kxdev@eth 'cd ~/dev/utopia-models/knowledge && claude -p <same> --max-turns 1'
**Loaded by:** the Architect (`agent`) — plans with Tyler, files, dispatches.

$ cd <detached knowledge worktree at 0854e244a on win> && claude -p /context --max-turns 1   # Memory Files
| User    | /home/win/.claude/CLAUDE.md                                                  | 1.1k  |
| User    | /home/win/dev/utopia-models/knowledge/_process/_generated/bundle-master.md   | 10k   |
| Project | <worktree>/CLAUDE.md                                                         | 22    |
| Project | <worktree>/_process/_generated/bundle-architect.md                          | 24.3k |

🔴 Grade this with claude -p /context, not a "print the lines" prompt. The Architect box's harness refused the verbatim-print prompt 3 times in 4 on 2026-09-13 (reading it as a system-prompt dump) and once answered A=NO to a yes/no variant while the same prompt from ~ answered A=YES. /context's Memory Files table is emitted by the harness, not the model, and does not vary.

Two facts follow. The import added nothing for the Architect — its box already loads bundle-architect.md through ~/.claude/rules/bundle.md, and its read-back prints the line once. On every master it added 64,618 B of audience: architect doctrine to a 26,550 B bundle, ~91 KB against the 30,000 B master ceiling — the widening _process/north-star.md forbids ("Do not widen this audience"). process-build.py measures only ~/.claude/rules/*.md, so its ceiling check could not see it.

Beside it, task-runner.sh resolved $CLONE/<portfolio>/CLAUDE.md into the brief (#6443). That file exists for 0 of 6 portfolios (ls {wi,raw,dmv-research,5am,govcon,sf}/CLAUDE.md), so the block never fired, and both briefs opened with "Read the repo root CLAUDE.md" — a pointer at the import.

Considered Options

• Do nothing. Rejected: ~91 KB on every master run, against a ceiling the build enforces on a different path.
• AGENTS.md at the root. Rejected: "Claude Code reads CLAUDE.md, not AGENTS.md" (code.claude.com/docs/en/memory); one harness means no second reader. 0 of 22 org repos carry one.
• Keep a root CLAUDE.md with repo-wide content instead of the import. Rejected: every agent class that opens this repo (master, architect, cowork, personal, always-on) already has a compiled bundle carrying the shared doctrine (artifact types, core behaviors, the loop). A root file would be a second copy of text the build already delivers — the drift class process-build.py exists to remove.
• Integrate the existing mechanism — delete the root file. ~/.claude/rules/bundle.md is already the one per-class loader, and it is the only one each box needs.

Decision Outcome

Chosen: delete knowledge/CLAUDE.md. The root holds exactly README.md and GOALS.md (kb-lint.py R1). The principle kept from AGENTS.md: a repo-root instruction file carries what ANY agent in that repo needs, never one class's bundle — and in this repo the bundles already carry all of that, so the file has nothing left to hold.

1. task-runner.sh's portfolio-CLAUDE.md block is deleted, and both briefs' first line says the doctrine is already in context rather than sending the run to a root file.
2. kb-lint.py R1 and its negative control assert two root files.
3. A future root CLAUDE.md in this repo is a defect unless it carries something no class bundle does; R1 fails on it, which is the point at which that argument gets made.

Consequences

• Good, because a master run in knowledge loads 26,550 B of bundle instead of ~91 KB, and nothing addressed only to the Architect.
• Good, because the Architect's sessions are unchanged — their bundle never depended on the import.
• Bad, because an agent that opens knowledge WITHOUT a ~/.claude/rules/bundle.md (a box never provisioned by process-build.py's class symlink) now loads no doctrine from the repo at all; before, it at least got the Architect's. Every registered agent box has the symlink (#7362), so this is a provisioning defect to fix on that box, never a reason to restore the import.
• Bad, because dated notes and diagrams that describe the old root file (e.g. wi/work/reference/win/diagrams/dispatched-run-anatomy.gv) now describe a past state. They are evidence with a date on it, not claims about the live tree, and are left as written.

Corrections retired with the code they guarded

• task-runner.sh's "🔴 TWO DEFECTS FIXED 2026-08-28" (SF missing from the portfolio alternation; the grep not case-insensitive, so issue #7149's 26 lowercase sf resolved to nothing). Both guarded the portfolio-CLAUDE.md lookup deleted here. Its measurement — an empty PORTFOLIO_PATH is indistinguishable from a run that needed none — is the reason the block is deleted rather than re-pointed at <portfolio>/README.md: a brief line that silently vanishes on a miss is the same defect with a different file name. The text stays in git history (git show 0854e244a:_infra/scripts/task-runner.sh).
• Preserved, re-worded: README's "🔴 THE ROOT HOLDS EXACTLY THESE 3" → 2; win-docs-index.py's "🔴 RETARGETED #7316" now says the root holds no CLAUDE.md.

What this does NOT change

• The content of any bundle, and process-build.py's ceiling.
• _process/the-loop.md's "Ground first: repo root CLAUDE.md" — it is generic across every repo a master works (utopia, brain repos), where such a file still exists.