Repository navigation
Conversation
Generated into this briefing from localDNS canonical by tools/sync-briefings.py. Build output — do not hand-edit. The scope-decision instructions pointed at ADR-008 as the next free number. Branching (Yggdrasil / Well of Mimir) has now taken it, so the pointer moves to ADR-009 — the instruction says to check the log first, and this keeps the default it hands a session from being wrong. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DQFfkkSUPXDTgKizj6RfpF
Rendered from localDNS/04-user-services/ai-orchestration/proxy-block.md by tools/sync-briefings.py — build output, never hand-edited. The agent git proxy and the egress proxy stand in front of this repo too, so the briefing now says what actually refuses (enforced), what only asks (declared), and what merely sits in the path (ambient). Carrying the text alone would have shipped "never force-push Yggdrasil or main" with nothing enforcing it — a declared boundary in the language of an enforced one, which is the exact failure the doctrine names. So .claude/hooks/push-guard.sh is installed and wired as PreToolUse(Bash): the agent git proxy blocks --delete but permits --force, and both orphan commits. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bu1wcD4pBe8vK9S53wyQd7
Founder's standing instruction (2026-08-08): every session may list, inspect and spawn sibling sessions without asking. Two halves, because one alone does nothing: - .claude/settings.json permissions.allow — the SITE. A briefing cannot pre-approve a tool call; the permission prompt never reads CLAUDE.md. This file is the grant. Both server spellings are listed (Claude_Code_Remote and claude-code-remote): an entry naming a tool absent from a session is inert, while a missing spelling costs a prompt, so listing both is the cheap direction. - The generated session-visibility block in CLAUDE.md — the DOCTRINE, rendered from localDNS canonical by tools/sync-briefings.py. Build output; do not hand-edit. Work runs in parallel across this portfolio — localDNS has had three sessions on Yggdrasil at once, and two gate.sh conflicts inside a single turn. A session that cannot see its siblings re-derives what they know and edits the file they are editing. Taxing visibility with a prompt taxes the one behaviour that keeps concurrent sessions out of each other's way. interrupt/archive/unarchive are deliberately NOT granted: they reach into another session's running state and can destroy work in flight. Seeing a sibling is not the same as reaching into one. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DQFfkkSUPXDTgKizj6RfpF
# Conflicts: # CLAUDE.md
Founder extended the standing grant (2026-08-08) beyond session visibility to scheduling and repo attach. Added to .claude/settings.json permissions.allow in all ten repos: list_triggers, create_trigger, update_trigger, send_later add_repo, list_repos, register_repo_root set_session_title, set_session_tags Still withheld, and now enforced as such: delete_trigger, fire_trigger, interrupt_session, archive_session, unarchive_session. Each destroys something or fires an effect NOW. Creating a routine is additive and visible; deleting the founders routine, or firing one early, is neither. Triggers are the one grant that acts when nobody is watching — everything else happens in-turn, in view. So the block states that a trigger inherits the one-way door rather than escaping it: it may prepare, report and check, but it may not be the thing that performs an irreversible outward-facing action. Publish, deploy, send, delete, merge to main wait at the * gate whatever the cron says. And the grant may not be used to schedule later what was refused now — that launders the decision as surely as routing it through a sibling. check_session_grant now runs BOTH directions: it fails if a promised tool is missing, and it fails if a WITHHELD tool quietly appears in an allow-list. A grant that shows up unannounced is the same briefing/settings disagreement as one that vanishes, and only one of those was previously catchable. Exercised both ways before landing (drop add_repo -> exit 1 naming it; add fire_trigger -> exit 1 naming it; restored -> exit 0). The section heading widened from "Session visibility" to "Session tooling" across all ten briefings; the marker keeps its original name so existing renders stay addressable. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DQFfkkSUPXDTgKizj6RfpF
Every running session is a Norn (founder's standing instruction, 2026-08-08). The promotion protocol is injected by .claude/hooks/refeed.sh at SessionStart so it enters every session's given-set at token zero: get onto Yggdrasil (a fresh clone lands on main, where the rule saying to use Yggdrasil is invisible), look before you write, claim a lane before the work, and front-load the repos and approvals you will need instead of stalling halfway for one you could have requested up front. The limit is stated hard: promotion grants NO new permissions. Being a Norn is a duty, not a capability. A blocked action never gets routed through a peer session — that launders the founder's permission decision — it goes back to the founder. The hook here is a shared body plus a two-line per-repo config (MANIFEST, QUEUE), so the block was injected into the body rather than overwriting the file, defined once and referenced from both the queue and no-queue branches. Exercised: valid JSON, block present, both branches. Also carries the session-visibility block (rendered) and the grant it describes: list_sessions, get_session, create_session and list_environments in permissions.allow. The block's own check refuses to pass while a repo claims the grant without holding it, so the doctrine and the settings file cannot drift apart. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bu1wcD4pBe8vK9S53wyQd7
Both sessions rendered generated blocks into CLAUDE.md at once. Those blocks are build output, so the conflict was resolved by taking the remote file and re-running the generator rather than merging text by hand. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bu1wcD4pBe8vK9S53wyQd7
The session-tooling block says the grant lives in each repo's settings.json, and its own check refuses while a repo claims a grant it does not hold. localDNS had the full list; the siblings had none of it, then a partial one. Mirrored from localDNS as the reference so doctrine and settings cannot disagree. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bu1wcD4pBe8vK9S53wyQd7
"Every running session is a Norn" was sited only on startup|clear. A re-activated session sends source=resume, and the hook exited early: no promotion, no sync, nothing. Resume is the more dangerous half. A session that went idle carries an IN-CONTEXT briefing from before it stopped — one that may still say "push to main, no branches" — and a stale briefing cannot tell that it is stale. Several idle sessions are parked on retired claude/* branches right now; re-activating one would have pushed to a branch the doom drawer exists to retire, while believing it was following current doctrine. The startup|clear gate is still right: re-pulling and re-injecting mid-work is exactly what it prevents. So resume gets a cheap staleness test instead of the full ritual — read-only (fetch, never pull), and it speaks ONLY when the session is not on Yggdrasil or is behind it. Resumed on Yggdrasil and current, it stays silent as before. When it does speak it leads with the thing the session cannot work out for itself: your briefing may be stale, re-read CLAUDE.md from disk before trusting the branching advice you are already holding. Then the branch move, the weave check, the lane claim, and the permission limit. Exercised in all ten: startup emits, resume-and-current stays silent, and a resume simulated on a retired claude/* branch (153 commits behind) emits with the stale-briefing warning, the retired-branch warning, and the permission limit intact. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bu1wcD4pBe8vK9S53wyQd7
…went stale Two repairs, both from the same cause: doctrine moved overnight and the instruments did not notice. 1. THE LADDER REACHED ONE REPO. The cream-rises promotion ladder (2026-08-09) landed in localDNS canonical, but all nine sibling briefings still taught the superseded "one standing working branch" rule. sync-briefings caught it across sessions — exactly the parallel-session drift it exists for — and the nine are now rendered from canonical. 2. HLIDSKJALF WAS CONFIDENTLY WRONG, on its first full day. Its top-ranked decision told the founder to approve ten Yggdrasil -> main PRs. Those are whole-branch merges into a rail, and tools/check-promotion.py now REFUSES that shape — verified directly: a rail head exits 1. The seat was recommending the one action the lock rejects. That is precisely the failure the seat was built to catch, turned on itself: an instrument reading a world that changed under it, with no way to know. So decide() now detects the lock and emits the ladder instead — cut promote/* from main, cherry-pick the chosen commits, PR that branch — and names the open whole-branch PRs as dead shape to CLOSE, never approve. Re-sitting the seat immediately proved the point twice more: it re-measured claude/* refs at 2, down from 339 (the founder ran the retirement — 337 refs gone, every drawer intact), and it raised its own PR snapshot as 71h stale. An instrument that reports its own staleness is the only kind worth trusting. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01DQFfkkSUPXDTgKizj6RfpF
Owner
Author
|
Closing as superseded shape, not rejected work — the content is wanted; this route to The branch policy moved on 2026-08-09 to the cream rises — promote by cherry-pick, one rung at a time. Nothing here is lost. Every commit stays on The replacement route: git checkout -b promote/<topic> origin/main
git cherry-pick <the chosen commits>
git push -u origin promote/<topic> # then open the PR from that branchlocalDNS #31 is the worked example of the new shape. Generated by Claude Code |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Draws this repo's Yggdrasil work into the Well.
A fresh session clones the default branch, and on
maintoday this repo'sCLAUDE.mdmentionsYggdrasilzero times — so until this merges, the policy does not exist for any new session here.What's in it
tools/sync-briefings.py. Build output — do not hand-edit.This repo remains a stub — no infrastructure added, per its own standing instruction.
🤖 Generated with Claude Code
https://claude.ai/code/session_01DQFfkkSUPXDTgKizj6RfpF
Generated by Claude Code