Skip to content

agents.md

Latest commit

 

History

History
109 lines (74 loc) · 9.3 KB

agents.md

File metadata and controls

109 lines (74 loc) · 9.3 KB
Raw

Account-level .github agent instructions

These instructions apply to this repository. Repository-local instructions may add stricter requirements, but they must not weaken this baseline.

Canonical organization links

Instruction discovery

Lowercase agents.md is canonical. Read every applicable lowercase agents.md from the repository root toward the current working directory before editing. Uppercase AGENTS.md and provider-specific instruction files are compatibility mirrors and must remain aligned with the applicable lowercase policy.

Inspect before editing

Inspect the current branch, complete working tree, remotes, default branch, open pull requests, linked GitHub issues, linked Linear work, repository documentation, tests, schemas, generated artifacts, deployment definitions, and relevant related repositories. Preserve every unfamiliar or uncommitted change.

Use read-only inspection and non-pruning synchronization such as git status --short --branch, git remote -v, git fetch --all, git diff, git log, git show, and git blame. Never treat a dirty worktree or inconvenient branch as permission to discard state.

Mandatory semantic conflict resolution

resolve any and all git conflicts semantically, will full context, even looking back 3-10 commits in git log history for more context - never hastily pick sides in a conflict but merge things conceptually, using max context and complete conceptual awareness for a given github organization's repos and external org repos too

For every conflict:

  1. Read the merge base, both complete sides, surrounding implementation, tests, schemas, generated artifacts, documentation, deployment configuration, and public contracts—not only conflict markers.
  2. Inspect the affected path history and normally review 3–10 relevant commits on each side with git log, git show, and git blame where useful.
  3. Review linked pull requests, issues, Linear work, related repositories in zed-pkg-test, and relevant external-organization repositories whenever behavior or contracts cross boundaries.
  4. Preserve compatible intent and invariants from both sides. Synthesize a conceptual merge; never resolve by selecting ours, theirs, current, or incoming wholesale.
  5. Scan the complete tree for unresolved markers and run the applicable formatter, linter, unit, integration, contract, build, security, and end-to-end checks.
  6. Document incompatible requirements, intentional choices, and any discarded intent in the commit and pull-request description.

Hard denylist for automated agents

Automated agents must never execute or recommend destructive, state-concealing, history-rewriting, purge, revocation, or policy-bypass operations. This is a hard denylist: authorization may support a reviewed human-run procedure, but it does not authorize an automated agent to perform the destructive step.

The blacklist includes, without limitation:

  • every form of git stash, every mode of git reset, every mode of git clean, git filter-repo, git filter-branch, BFG, git rebase, interactive history rewriting, git commit --amend, commit replacement, destructive git checkout -- <path>, destructive git restore, git branch -D, ref or tag deletion, git reflog expire, git gc --prune, git push --force, and git push --force-with-lease;
  • recursive or bulk deletion and destructive filesystem mutation, including rm -rf, find -delete, truncation, shredding, destructive overwrite, formatting, and access-removing ownership or permission changes;
  • destructive data operations, including DROP, TRUNCATE, unbounded DELETE, destructive rollback, irreversible migration, bucket/object purge, queue/topic deletion, and bulk mutation without a bounded reversible plan;
  • destructive infrastructure or identity operations, including kubectl delete, helm uninstall, terraform destroy, pulumi destroy, cloud delete/purge calls, cluster or namespace teardown, and autonomous secret, key, certificate, credential, factor, or session revocation or rotation;
  • deleting repositories, worktrees, submodules, branches, tags, releases, packages, artifacts, registries, environments, evidence, audit logs, customer data, or production state;
  • bypassing hooks, reviews, branch protection, rulesets, required checks, security/compliance gates, approvals, or audit logging, including --no-verify and equivalent bypasses.

Do not use destructive commands merely to make tests pass, clear a conflict, simplify a migration, or hide an inconvenient state.

Required safe alternatives

Use additive branches, separate clean worktrees or clones, explicit path staging, ordinary commits, non-force pushes, patch-based edits, read-only queries, dry runs, backups, additive migrations, and reversible roll-forward changes. Leave unrelated work untouched. When safe progress is impossible, preserve all state and report the exact blocker.

Source ownership and cross-repository context

Edit authoritative sources rather than generated mirrors, vendored copies, caches, or downstream consumers. Identify generators and regenerate derived artifacts from reviewed sources. Never detach, absorb, relocate, remove, or rewrite a submodule or worktree. Cross-repository behavior must be understood across the owning organization and relevant external organizations before contracts are changed.

Secrets and sensitive data

Never print, log, commit, paste into issues, include in fixtures, or expose tokens, passwords, private keys, session material, database URLs, customer data, legal records, private health data, production data, or unpublished security details. Use approved secret stores, placeholders, and redacted diagnostics.

Pull requests, validation, and evidence

Use focused branches and pull requests. Link the relevant Linear issue or project. Explain behavior, risks, migration and roll-forward considerations, security impact, tests run, conflicts and their semantic resolution, and cross-repository dependencies. Never report a branch, commit, pull request, merge, deployment, test run, or external update as complete without authoritative remote evidence.

Primary branch and concurrent-agent policy

This organization policy overrides generic feature-branch and worktree defaults for agent tooling.

  • Highly prefer an existing primary branch, in this order: main, dev, then master.
  • Work directly on the selected primary branch even when other agents are active. Use another branch only when a human or a repository-specific release process explicitly requires it.
  • Never create or use a Git worktree unless a human explicitly instructs you to do so for the current task. Concurrency alone is not permission to use a worktree.
  • Concurrent agents must coordinate repository and file ownership through the available agent communication channel, keep edits scoped, inspect live state before each write, and hand off cleanly. Coordinate instead of isolating routine work in worktrees.
  • Preserve unrelated in-progress changes and never overwrite another agent's work. If safe ownership of overlapping files cannot be established, pause that overlapping edit and coordinate before continuing.

Git and history policy

Prefer merges over rewrites. The rule is: avoid git rebase in favor of git merge. A merge records what actually happened, and the automation across this fleet reads history to decide what has already landed — rewriting that history makes the judgement wrong, and it makes two checkouts of the same work look unrelated.

On any conflict, resolve it semantically. Read at least 3–10 relevant commits of surrounding history on both sides before deciding, then merge the two intents. Picking a side is not a resolution; it silently discards whichever half was dropped, and the loss is invisible afterwards because the conflict marker is gone.

The commands below destroy work that no remote has ever seen, so an agent does not run them without explicit human permission:

  • git stash — stashes live in no remote and appear in neither git status nor ahead/behind counts, so a repository holding thousands of stashed lines reports a clean tree to every tool that scans for unlanded work. Use a wip/<what-it-is> branch instead. If you find someone else's stash, make it reachable with git branch rescue/<id> refs/stash — never pop it.
  • git reset — moves the branch out from under committed work.
  • git clean — deletes untracked files that have never been pushed anywhere.
  • git filter-repo — rewrites every commit id in the repository, which breaks every pin, submodule pointer and open pull request that referenced the old ones.

Stage explicit paths. Never git add -A: most checkouts here carry someone else's work in progress, and -A is how that — plus secrets — gets committed by accident.

Never report work as landed while it is only on local disk. A change is done when it is committed, pushed, and open as a pull request; compiling is not landing.