auto-repo-setup

# Auto Repo Setup Make the repository usable without changing the user's normal way of working. Treat setup as an evidence-driven repository task, not as a special interface for people with a particular job title. ## Entry router Classify the request before changing anything. Do not combine modes merely because they all contain the word "setup". | User outcome | Mode | First action | |---|---|---| | "It won't run", missing dependency, new machine | Environment repair | Read project instructions and detect the actual stack | | "Sync before we work", "pull the latest", or a status question during repository maintenance | Session routine | Determine sync authority from the ongoing task and project instructions, then use Workflow B | | Repeated or unexpected hook output | Hook diagnosis | Count registrations and firing sessions before editing configuration | | A teammate needs a repeatable handoff | Repository handoff | Audit existing onboarding and fill only the missing operational contract | | Commit, push, conflict, leaked secret/history | Git safety | Read [references/git_safety.md](references/git_safety.md) before mutation | | Behavior must happen before the first prompt | Startup automation | Read [references/startup_automation.md](references/startup_automation.md) and pass its hook gate | ## Operating principles 1. **Do not infer competence from role.** Match the user's demonstrated altitude. A collaborator can tell an agent "sync the remote first"; do not invent a launcher, shortcut, or wizard unless the user asked for one. 2. **Use the least mechanism that satisfies the outcome.** Prefer, in order: direct Agent instruction → project instruction → explicit command → lifecycle hook. A hook is not a more professional version of a sentence. 3. **Inspect authority before producing.** Read the current user request, then project instructions, onboarding/runbooks, manifests, lockfiles, and actual command output. Do not make ONBOARDING.md, Python, uv, ffmpeg, or .env mandatory for repositories that do not declare them. 4. **Preserve user work.** Never auto-stash, merge, rebase, force, discard, or overwrite local changes to make setup look successful. Already-authorized delivery of this task's isolated changes may use the repository's declared scoped workflow under [Authorized shared-checkout delivery](references/git_safety.md#authorized-shared-checkout-delivery); this does not authorize integrating unrelated local commits or shared WIP. 5. **Define a falsifiable success state.** "Dependencies installed" is not enough; run the project's real smoke test or startup command and verify the observable result. 6. **Keep diagnosis read-only until the cause is known.** Installation, config edits, hook registration, commits, and external writes require the authority implied by the user's request; destructive or public actions require explicit approval. ## Workflow A — Environment repair ### A1. Read the project map Read only files that exist, in this order: 1. AGENTS.md and AGENTS.override.md 2. CLAUDE.md and scoped project rules 3. ONBOARDING.md or the repository's named setup/runbook 4. README.md 5. Manifests, lockfiles, task runners, CI, and container configuration 6. Existing Claude/Codex settings only when the request concerns agent behavior If no onboarding guide exists, infer the setup path from authoritative manifests and exercised commands. Do not stop to ask whether to create documentation unless a durable handoff is part of the request. ### A2. Run a read-only capability inventory Use the bundled scripts/check_env.py against the repository root when a Python 3.10+ runner is already available. It detects the declared ecosystem and checks only the relevant toolchain; it does not install dependencies or read secret values. If Python is itself the missing prerequisite, perform the same read-only manifest inventory directly instead of installing Python merely to run the audit. ~~~bash uv run python scripts/check_env.py --repo <repo-root> uv run python scripts/check_env.py --repo <repo-root> --json ~~~ Treat the inventory as evidence, not as the project setup specification. A custom runbook or CI workflow can require tools the generic inventory cannot infer. Expected result: - exit 0: all toolchains inferred from manifests are available; - exit 1: one or more inferred prerequisites are missing; - exit 2: the repository or its metadata could not be inspected safely. ### A3. Repair the root cause For each failed requirement: 1. Capture the exact command, exit code, stdout, and stderr. 2. Trace the failure to the declaring source: project guide, manifest, lockfile, configuration, or runtime log. 3. Resolve the affected command's actual runtime from its launch configuration and read-only probes: interpreter and package environment, container image, or service launcher as applicable. Keep an isolated validation environment distinct from that consumer. If the consumer already meets the requirement, retain it and continue to A4 without reinstalling. 4. Before changing that environment, inspect affected in-flight work and existing authority. Setup authorization does not authorize interrupting other work; coordinate or wait when the change would disrupt it, and report the exact unresolved activation step rather than killing or restarting its processes. 5. Apply the smallest project-consistent repair through the project's existing update route, preserving valid non-target dependencies and configuration. 6. Re-run the failed check before moving on. Do not restart, reinstall everything, or switch package managers as a first move. Do not print .env or credential contents; verify presence and behavior without echoing values. ### A4. Verify the product path Run the project's documented smoke test, build, or start command. If none exists, derive one from CI/task-runner configuration and label the derivation. Verify the observable result, not merely a zero exit code. For dependency repairs, use a new process in the resolved consumer environment to read back the installed versions and run that product check. A changed lockfile or successful isolated install does not establish that the consumer was updated. Keep already-running processes and other machines outside the claim unless separately verified; do not restart them just to widen the claim. Read a hosted dependency alert separately before claiming it closed. Report: - what was broken and the evidence; - what changed; - the exact verification and result; - any remaining user action or blocked authority. ## Workflow B — Routine Git sync For "sync before work", prefer a short project instruction shared by the agents that use the repository. The routine is: 1. Determine whether the ongoing task and project instructions already authorize synchronization. Treat a status question during that maintenance as a progress request, not as revocation or completion of the authorized sync. For an explicit read-only audit or when sync authority is absent, inspect and report only. 2. Inspect the target branch, its upstream, and working-tree cleanliness. Fetch and compare the local and freshly fetched remote histories. Treat a behind count as a discovery, not a completed sync. 3. If the task includes already-authorized delivery and the repository declares usable scoped tools that preserve unrelated shared HEAD, index, and WIP, use [Authorized shared-checkout delivery](references/git_safety.md#authorized-shared-checkout-delivery). Dirty or divergent state alone is not a reason to leave that delivery waiting. 4. Otherwise, use repository-declared safe sync tools when applicable; without such a tool, on the target branch with a clean working tree and behind-only histories, run git pull --ff-only. Before reporting synchronized, independently verify the target branch, clean status, and local/remote commit equality. 5. Outside authorized shared-checkout delivery, stop for non-target branches, real WIP, ahead-only or divergent histories, locks, or active-writer conflicts; explain the state. For divergence, show the graph and ask how the local commits should be handled. Do not auto-stash, discard, rebase, create a non-fast-forward merge, force, or absorb others' commits. 6. If the network fails, report the remote state as unknown rather than already current; continue locally only when the user's task can safely proceed on the local version. A collaborator may also simply say: > Sync the latest remote version before starting. That is a normal Agent instruction, not a degraded fallback. Do not register SessionStart merely to automate this routine. Static behavior belongs in AGENTS.md/CLAUDE.md; the Agent can inspect context and handle exceptional Git states instead of hiding them in a shell process. ## Workflow C — Repository handoff Create or revise onboarding only when the user wants a durable handoff and the existing project map does not already provide one. Use [references/onboarding_template.md](references/onboarding_template.md) as a checklist, not as a literal Python/video template: - derive prerequisites from manifests and actual project commands; - include expected output after each command; - separate one-time setup from daily use; - include recovery for real observed failures; - keep project instructions as the operational SSOT and avoid duplicating values. Validate every command on the target operating systems that matter to the user. ## Workflow D — Startup automation and hook diagnosis Read [references/startup_automation.md](references/startup_automation.md) before adding or changing any hook. ### Diagnose repeated output first 1. Enumerate project, local, user, managed, and plugin hook registrations. 2. Count matching entries; do not infer "three registrations" from three outputs. 3. Identify which root sessions, resumes, compactions, or subagents fired them. 4. Verify the event matcher and current runtime payload. 5. State the root cause before proposing a change. ### Install only after the hook gate passes The bundled initializer is Claude Code-specific and installs only a lightweight startup context nudge. It preserves unrelated settings, adds matcher startup, validates the guide path, writes atomically, and is idempotent. Preview first: ~~~bash uv run python scripts/init_session_start_hook.py \ --repo <repo-root> \ --guide ONBOARDING.md \ --dry-run ~~~ After the user confirms that pre-prompt injection is genuinely required, run the same command without --dry-run. Remove only the managed entry with --remove. Never copy this configuration into Codex by analogy. Codex project behavior should normally live in AGENTS.md; if a Codex hook is truly required, verify the installed version's root/subagent behavior and payload first. ## Git safety Read [references/git_safety.md](references/git_safety.md) before commit, push, conflict resolution, or history rewrite. The load-bearing rules are: - stage only intended paths; - never bypass hooks unless the user explicitly typed the bypass in this session; - verify repository visibility from the hosting service before push; - get explicit approval before public push, force push, or history rewrite; - resolve conflicts from project semantics, never by blindly choosing "ours" or "theirs"; - revoke leaked credentials before treating history cleanup as complete. Use [references/pii_guard.md](references/pii_guard.md) for public-distribution content review. A green scanner does not replace semantic review. For automatic artifact validation before ordinary Git commits, use the [commit-time check procedure](references/git_safety.md#commit-time-checks). It covers Git hooks across Agent and manual commits, staged-input verification and existing-hook preservation. Use the [isolated checkout preparation](references/git_safety.md#isolated-checkout-preparation) when task isolation must retain repository-specific LFS and security configuration. For a read-only history scan, the existing scripts/sanitize_history.sh remains available. It never rewrites history; treat its findings as candidates and do not execute rewrite commands without explicit approval. ## Counter-review boundary Use counter-review when the approved work materially changes security policy, shared lifecycle hooks, CI/CD, dependencies, or destructive Git behavior. Do not spawn a review team for an ordinary setup check or a one-line project instruction. Filter every finding by probability, cost, real usage, and direct verification. ## Resources - scripts/check_env.py — stack-aware, read-only repository inventory - scripts/init_session_start_hook.py — guarded Claude Code startup-nudge manager - scripts/sanitize_history.sh — read-only history candidate scan - [references/startup_automation.md](references/startup_automation.md) — project instruction vs command vs hook decision and verification - [references/git_safety.md](references/git_safety.md) — sync, commit, push, conflict, and history rewrite gates - [references/onboarding_template.md](references/onboarding_template.md) — durable handoff checklist - [references/pii_guard.md](references/pii_guard.md) — public-content scan layers