Make AGENTS.md the canonical agent guide; CLAUDE.md imports it

AGENTS.md

@@ -0,0 +1,50 @@
+# Core principle: never lose track of content
+
+Squirrel indexes **content** (BLAKE3 hashes), not paths. A hash ever observed
+must stay retrievable. Paths are observations of content; content is the entity.
+
+So `Upsert` never rewrites a row's `blake3` in place: when content at a path
+changes it marks the prior row `superseded` and inserts a new one, keeping at
+most one live (non-`superseded`) row per path. The schema enforces this on
+`files` — the `files_blake3_immutable` trigger and the `uniq_files_live_per_path`
+partial unique index (`store/migrations.go`).
+
+The `runs` table follows the same no-loss spirit by policy, not schema: squirrel
+never auto-prunes runs — they're an audit trail, and any retention is explicit
+and operator-driven.
+
+Any new feature (sync, prune, dedup, GC) must preserve both: no deleting or
+overwriting history without an explicit, opt-in retention policy.
+
+# Code quality
+
+Don't:
+- Export test helpers when tests are in-package
+- Write functions over ~50 lines — decompose by phase
+- Put multiple cobra subcommands in one file
+- Leave unused fields/flags on public types
+- Write to stdout/stderr from library packages — return values instead
+- Concatenate user input into DSNs or URLs
+- Route ambiguous input by syntax alone — check authoritative state first
+- Index low-cardinality columns — prefer partial indexes
+- Forget `go mod tidy` after adding a dependency
+- Keep names/visibility when moving code — re-evaluate
+
+Before pushing: `go vet ./...`, `go test ./...`, `golangci-lint run`.
+
+# Pull requests
+
+- `Closes #N` (one per issue) in the PR body — only when the PR fully closes
+  that issue; otherwise reference it without the keyword.
+- Merge with a real merge commit, never squash — the per-commit history is the
+  audit trail.
+
+# Issue workflow ("implement #N")
+
+Unless told otherwise:
+1. Work on a feature branch; open a PR (see Pull requests).
+2. Self-review the diff against this file: dead code, oversize functions, scope creep.
+3. Watch the PR feed automatically (don't ask) for up to 10 min: fix CI failures,
+   address legitimate review comments, briefly dismiss the rest. If it isn't
+   settled by 10 min, unsubscribe, say so, and wait.
+4. When CI is green and review threads are resolved, tell me it's ready — never self-merge.

CLAUDE.md

@@ -1,38 +1 @@
-# Core principle: never lose track of content
-
-Squirrel indexes **content** (BLAKE3 hashes), not paths. A hash that has ever been observed must remain retrievable from the index. Paths are observations of content; content is the entity.
-
-The v4 schema enforces this: the `files` PK is `(volume_id, path, blake3)`, status is one of `present`/`missing`/`superseded`, and per `(volume_id, path)` at most one row is non-superseded. When content at a path changes, `Upsert` flips the prior row to `superseded` and inserts a new row — `blake3` is never rewritten in place. Any new feature (sync, pruning, dedup, GC) must preserve this rule: don't delete or overwrite historical rows without an explicit, opt-in retention policy.
-
-The same applies to the `runs` table: squirrel never auto-prunes runs — the run history is an audit trail, and any retention is explicit and operator-driven only.
-
-# Code quality reminders
-
-Don't:
-- Export test helpers when tests are in the same package
-- Write functions over ~50 lines — decompose by phase
-- Put multiple cobra subcommands in one file
-- Leave unused fields or flags in public types
-- Write to `os.Stderr`/`os.Stdout` from library packages — surface via return values
-- Concatenate user input into DSNs or URLs
-- Route ambiguous inputs by syntax alone — check authoritative state first
-- Index low-cardinality columns; prefer partial indexes
-- Skip `go mod tidy` after adding a dependency
-- Preserve names or visibility blindly when moving code — re-evaluate
-
-# Pull request conventions
-
-When opening a PR that completes one or more issues, include a closing keyword per issue in the PR body (`Closes #19`, or `Closes #19, Closes #20` for several). GitHub auto-closes the linked issues on merge. Use `Closes` only when the PR completes the issue in full — for partial work, reference the issue without the closing keyword.
-
-Always preserve individual PR commits on merge — never squash. The per-commit history is the audit trail (review fixes, CI fix-ups, refactor steps) and collapsing it loses information that may matter later.
-
-# Default workflow for issue implementations
-
-When asked to implement an issue (e.g. "implement #19"), unless instructed otherwise, do all of the following:
-
-1. Implement on a feature branch, following the principles above.
-2. Open a PR with `Closes #N` in the body.
-3. After pushing, self-review for code quality: dead code, oversize functions, accidental scope creep, anything that violates CLAUDE.md. Push fixes for what you find.
-4. After opening the PR, start watching its activity feed automatically — don't ask first — so CI results, automated reviews (e.g. Copilot), and other events arrive as they happen. Cap the watch at 10 minutes total: if CI and review threads aren't settled by then, unsubscribe, tell the user you stopped watching, and wait for direction. On CI failure within the window, diagnose and push a fix. On review comments, address legitimate findings by pushing fixes; reply briefly to dismiss anything that's incorrect or out of scope.
-5. Once CI is green and any automated review threads are resolved, surface readiness to the user.
-6. Stop and wait for explicit approval before merging — never self-merge.
+@AGENTS.md