Dex
AGENTS.md509 starsChanged 2 months agoMITView on GitHub
Instructions for AI agents (and humans) developing this repository. Not the product persona — the root CLAUDE.md is seed prose shipped into user vaults ("You are Dex…"); it is a product surface, not contributor guidance. Edit it like UI copy, not like docs. The founder is non-technical. Lead with the outcome and explain what is happening in plain language; briefly explain an unavoidable technical term the first time it appears. - Use an ADHD-friendly structure: keep steps short and numbered,…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,700 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository153 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- .claude/plugins/compound-engineering/CLAUDE.mdCLAUDE.md
- .cursor/rules/search-routing.mdcCursor rule
- anthropic-algorithmic-artSkill
- anthropic-brand-guidelinesSkill
- anthropic-doc-coauthoringSkill
- anthropic-docxSkill
- anthropic-frontend-designSkill
- anthropic-internal-commsSkill
- anthropic-mcp-builderSkill
- anthropic-pptxSkill
- anthropic-slack-gif-creatorSkill
- anthropic-theme-factorySkill
- anthropic-webapp-testingSkill
- anthropic-web-artifacts-builderSkill
- anthropic-xlsxSkill
- atlassian-setupSkill
- backup-nowSkill
- backup-restoreSkill
- backup-setupSkill
- calendar-setupSkill
- commitmentsSkill
- create-mcpSkill
- daily-planSkill
- daily-reviewSkill
- decision-logSkill
- delegate-checkSkill
- dex-backlogSkill
- dex-doctorSkill
- dex-improveSkill
- dex-obsidian-setupSkill
- dex-rollbackSkill
- dex-updateSkill
- dex-whats-newSkill
- diff-profileSkill
- enable-semantic-searchSkill
- getting-startedSkill
- goal-backlogSkill
- google-workspace-setupSkill
- granola-setupSkill
- identity-snapshotSkill
- industry-truthsSkill
- initiative-kickoffSkill
- integrate-mcpSkill
- journalSkill
- manage-capabilitiesSkill
- meeting-closeoutSkill
- meeting-prepSkill
- ms-teams-setupSkill
- pipedrive-setupSkill
- pipeline-syncSkill
- process-meetingsSkill
- product-briefSkill
- project-healthSkill
- prompt-improverSkill
- relationship-radarSkill
- reviewSkill
- save-insightSkill
- scrapeSkill
- skill-scoreSkill
- The first 60 shown.
The file
Show its rules and text
We split this AGENTS.md into the rules it gives an agent
13 rules. About 1 is common, 7 are distinctive, and 5 are very specific to this project.
≈1
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 14
Use an ADHD-friendly structure: keep steps short and numbered, make the one required action unmistakable, and prefer copy-paste prompts or …
- Distinctive: line 16
When the founder must act, begin with a What I need from you block stating the exact action, where to do it, why the founder is needed, and …
- Distinctive: line 19
When blocked, state the exact blocker, what is and is not affected, the lowest-lift recovery action, and how to prevent the issue recurring.
- Common: line 21
Never ask the founder to diagnose Git, mounts, permissions, credentials, or similar mechanics, and never ask for secrets to be pasted into …
- Distinctive: line 23
Never claim something is fixed, committed, pushed, merged, published, or live without verifying the result the user will actually …
- Distinctive: line 40
If a check fails, classify the failure precisely:
- Distinctive: line 47
Do not collapse these into “Devbox cannot access GitHub.” Do not ask the founder to re-authenticate or paste a token when his normal Devbox …
- Very specific: line 52
If the remedy requires runner bootstrap, host service, DNS, credential-mount, or deployment-key changes, do not edit /etc/resolv.conf as a …
- Very specific: line 60
Run /dex-orient (or python3 scripts/dex_state.py --digest) — released version, merged-but-unreleased work, where the maps live.
- Very specific: line 66
Never hand-edit it; regenerate with python3 scripts/generate-architecture-inventory.py.
- Very specific: line 80
One safe door: Every vault mutation goes through core/lifecycle/service.py → transaction core → portable ownership contract.
- Distinctive: line 82
Never add a code path that writes vault files directly.
- Very specific: line 112
Update/transaction design: docs/transaction-core-design.md, docs/portable-vault-contract-design.md
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Working on dex-core (agent instructions)
Instructions for AI agents (and humans) developing this repository. Not the
product persona — the root CLAUDE.md is seed prose shipped into user vaults
("You are Dex…"); it is a product surface, not contributor guidance. Edit it like
UI copy, not like docs.
Communicating with the founder
The founder is non-technical. Lead with the outcome and explain what is happening in plain language; briefly explain an unavoidable technical term the first time it appears.
- Use an ADHD-friendly structure: keep steps short and numbered, make the one required action unmistakable, and prefer copy-paste prompts or commands.
- When the founder must act, begin with a What I need from you block stating the exact action, where to do it, why the founder is needed, and what will happen next.
- When blocked, state the exact blocker, what is and is not affected, the lowest-lift recovery action, and how to prevent the issue recurring.
- Never ask the founder to diagnose Git, mounts, permissions, credentials, or similar mechanics, and never ask for secrets to be pasted into chat.
- Never claim something is fixed, committed, pushed, merged, published, or live without verifying the result the user will actually experience.
GitHub access in managed Codex runners
The normal Devbox terminal and a managed Codex runner are separate execution contexts. A successful GitHub check in one does not prove that the other has working DNS or credentials. Before any GitHub fetch, push, or remote inspection from a managed runner, run these read-only checks in that same runner:
getent hosts github.com
gh auth status --hostname github.com
gh api user --hostname github.com --jq '"GITHUB_OK: @" + .login'
git ls-remote origin HEADIf a check fails, classify the failure precisely:
- Managed-runner DNS:
getent hosts github.comcannot resolve the name. - Managed-runner HTTPS credentials: DNS works but
gh apior HTTPS Git authentication fails. - SSH key or route: the repository uses SSH and the SSH remote fails.
Do not collapse these into “Devbox cannot access GitHub.” Do not ask the founder to re-authenticate or paste a token when his normal Devbox check already passes. Continue safe local work where possible, and use the already-authenticated Devbox terminal as the fallback for a remote GitHub operation.
If the remedy requires runner bootstrap, host service, DNS, credential-mount,
or deployment-key changes, do not edit /etc/resolv.conf as a temporary fix,
copy secrets, or claim the issue is fixed. Report the exact layer, the exact
configuration change required, and the verification command; ask for approval
only when a host or credential change genuinely needs it.
Orient before you touch anything
- Run
/dex-orient(orpython3 scripts/dex_state.py --digest) — released version, merged-but-unreleased work, where the maps live. docs/architecture/DEX-CORE-MAP.md— the narrative "how it hangs together" map, with SHIPPED / LOCAL / PROTOTYPE / PLANNED status per subsystem.docs/architecture/INVENTORY.md— generated, CI-drift-gated: exact MCP servers/tools, skill catalog + trigger analysis, ownership-class path tables. Never hand-edit it; regenerate withpython3 scripts/generate-architecture-inventory.py.CHANGELOG.md— the single source of release truth, written in house voice.
Merged is not shipped, and shipped is not always live
- Release truth is the version tag + CHANGELOG entry, not
main. - Some merged work is deliberately HELD: customization-migration Lane G
(activation + rewind) and the
/connectdoorway (draft PR #231, security no-go not yet lifted).core/ritual_intelligence/is code-complete but wired to nothing (PARKED). Never describe held/parked work as available, in code comments, docs, or user-facing copy.
Hard rules
- One safe door. Every vault mutation goes through
core/lifecycle/service.py→ transaction core → portable ownership contract. Never add a code path that writes vault files directly. - Generated files are never hand-edited:
docs/architecture/INVENTORY.md,core/paths.json,packages/dex-contracts/dist/*,System/.release-catalog.json. Change the source, re-run the generator (scripts/generate-*.{py,mjs}). - New top-level paths must be classified in
core/portable_contract.py(brain/seed/generated/vault/runtime) or CI fails — this is deliberate. - No PII, no founder-machine content — CI gates (
scripts/check-pii.sh,scripts/check-founder-content.sh) block it; don't allowlist around them casually. - Worktree isolation — never write in a checkout another session is using; work on a branch in its own worktree.
- GitHub Actions are pinned to commit SHAs; keep new workflow edits pinned.
- This repo's issue tracker is public and user-facing only — bug reports and
feature requests from real users, nothing else. Internal roadmap planning,
epics, and architecture-exploration tickets (the "Wayfinder: ..." pattern and
its sub-issues) go in the private
dex-cardsrepo instead. Never open that class of issue here, even in draft form — 26 such issues were found publicly exposed and moved todex-cardson 2026-08-10 after the founder caught it.
Tests
- Python:
pytest core/tests/ core/mcp/tests/ core/migrations/tests/ -m "not fuzz"(flat naming convention:test_<subsystem>_<behavior>.py). - Node:
npm run test:hooks,npm run test:scripts,npm run test:integrations. - Contracts:
npm run check:connections-contract. - CI runs the full gate list in
.github/workflows/ci.yml(macOS runner); nightly quality + fuzz + perf budgets in.github/workflows/nightly-quality.yml. Governance:docs/testing-governance.md.
Deeper reading (on demand)
- Update/transaction design:
docs/transaction-core-design.md,docs/portable-vault-contract-design.md - Customization migration:
docs/customization-migration-threat-model.md,docs/plans/2026-07-24-customization-migration-mcp.md - Doctor:
docs/dex-doctor-spec.md· Merge gates:docs/merge-gates.md - Distribution:
DISTRIBUTION_READY.md,docs/Dex_System/Distribution_Checklist.md - Past root causes and runbooks:
docs/solutions/ - System docs canon:
docs/Dex_System/is canonical; the copies under06-Resources/Dex_System/are compatibility bridge files — keep them byte-identical until the physical move lands (seedocs/Dex_System/README.md).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

