The canonical .dockerignore and .gitignore both omitted the in-repo agent
scratch directory, which holds one worktree per in-flight agent, so under
`COPY . .` an entire extra checkout of the repo reached the image. The
two entries are deliberately different shapes: anchored in .dockerignore,
where the `**/` form would also delete a legitimately named nested
directory, and unanchored in .gitignore, where a pattern already matches
at every depth. Anchoring leaves a gap where agents run in
subdirectories, stated in the vendored file itself. The second half is
the consequence of excluding .git: `git describe` in a build stage yields
an empty version without erroring, so the version is now computed on the
host and passed in.
Model: opus-5
The canonical .dockerignore was three lines while the canonical
Dockerfile does `COPY . .`, so a local .env, *.pem or *.key shipped into
the build context and could land in an image layer, invisible to every
git-based check. Copying .gitignore's patterns across is not the repair:
.dockerignore anchors an unprefixed pattern at the context root, so that
form protects only the repository root while reading as solved. Every
depth-independent pattern here carries `**/`, and secret names are
character ranges because matching is case-sensitive and an ALL-CAPS twin
still misses `Server.Key`. Public certificates are deliberately left in
as a legitimate build input. Verified by enumerating a probe image.
Model: opus-5
The canonical `if missing <tool>; then install; fi` guard tests PATH
presence and never version, so on any already-provisioned machine a pin
is inert and a version bump is a no-op, while the Dockerfile installs the
pinned version into a clean image and CI then disagrees with local about
what the tool is. Comparing versions alone is not enough either: an
installer writes to its own directory while callers resolve through PATH,
so a shadowing binary lets the install succeed and change nothing anyone
sees. REPO_POLICIES.md now states the whole form — exact whole-token
comparison, mis-parse falling through to a reinstall, re-resolution
through PATH after installing, and a call site that prints the version.
Model: opus-5
script/cibuild was a plain `docker build .` and the Dockerfile does
`COPY . .` followed by `RUN make check`, so on an unchanged tree Docker
served the check layer from cache: the suite never ran and the build
still exited 0. Measured here before the change, a second run on a
byte-identical tree returned in 0.286s with `RUN make check` CACHED.
script/cibuild and script/docker now pass --no-cache. The canonical text
asserting that a bare `docker build .` proves the checks ran was wrong in
REPO_POLICIES.md, both checklists and the Go styleguide, and is corrected
in all of them.
Model: opus-5
`depguard` was in the disable list. It is now enabled with one rule,
`test-support`: in files that are neither test files nor inside a package whose
directory name ends in `test`, the imports named under `deny` are refused.
Canonical denies `net/http/httptest`, which serves tests only and is the same
in every repository.
This replaces `internal/testimportgate` in sneak/homoicon, a package whose only
job was to refuse a non-test Go file importing an in-module package whose last
path segment ends in `test`. sneak ruled on
sneak/homoicon#1029 that the rule moves into linter
configuration here.
depguard cannot express that rule generically. Its package lists are prefix
lists -- its own README says so, and I confirmed it: `*test`, `**test` and
`$gomod/**test` each match nothing, while a full import path matches. "In
module" is not generic either, since the module path differs per repository.
And depguard refuses a rule with no allow or deny list, so this file cannot
ship the rule pre-armed and empty for each repository to fill in.
The closest expressible rule is the one here. The file half is generic and
exact; the package half is a deny list each repository extends with its own
test-support packages, by full import path. REPO_POLICIES.md now says that list
is the one part of a vendored copy a repository may add to.
Tested with golangci-lint v2.12.2 on a scratch module: a production file
importing a denied `*test` package fails, and the same import from a `_test.go`
file and from inside the `*test` package passes. `make check` here is green.
Model: opus-5
Co-authored-by: sneak <sneak@sneak.berlin>
Reviewed-on: #59
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
SPECULATIVE and ahead of your ruling — nothing here is urgent and closing it costs nothing. One sentence of policy prose changed; no config file is touched.
## The contradiction
`REPO_POLICIES.md` line 266 currently reads:
> `.golangci.yml` is standardized and must _NEVER_ be modified by an agent, only manually by the user.
Stated unqualified, that forbids an agent from modifying `.golangci.yml` **anywhere** — including the canonical copy in this repo, which is the only place it can ever be fixed. An agent that wants to remediate a linter problem must either violate the rule or leave the problem standing. A rule that cannot be complied with and satisfied at the same time gets resolved ad hoc, differently by each reader, which is the worst of both outcomes it was trying to produce.
This is not hypothetical. It has already cost real time:
- The `gomodguard` deprecation (#25) has been open since 2026-08-07 and still prints on every lint run in every consuming Go repo.
- One agent read the rule as binding here and **declined to open even a speculative branch**, so the fix was not written at all on that pass.
- #47 exists only because a later request was explicit enough to override the reading, and its lead comment asks for exactly this ruling before the PR itself can be judged on its merits.
- The same warning is refiled downstream as sneak/homoicon#4, where it is correctly marked owner-only and correctly punted upstream.
Each new agent that meets the rule reruns this whole argument.
## The change
Scope the prohibition to the vendored copy, and name the one legitimate path by which the config can change:
```
- `.golangci.yml` is standardized. The vendored copy in a consuming repo must
_NEVER_ be modified by an agent: fetch it from
`https://git.eeqj.de/sneak/prompts/raw/branch/main/.golangci.yml` and keep it
byte-identical, so that no repo can quietly loosen its own linting. Linter
configuration changes are made to the canonical copy in the `prompts` repo and
reach consuming repos by re-vendoring; an agent may open a PR against
canonical, which only the user merges.
```
The version pin sentence that followed is unchanged.
This keeps the property the rule exists for — no repo silently weakens its own linting, and divergence from canonical stays detectable — while removing the reading that freezes canonical itself. Your control is not reduced: an agent may open a PR here, and only you merge it.
## Scope of the wording sweep
I grepped every `.md` in the repo for the absolute phrasing. It appears in **exactly one place**, `prompts/REPO_POLICIES.md` lines 266-267.
`EXISTING_REPO_CHECKLIST.md` (line 39) and `NEW_REPO_CHECKLIST.md` (line 63) both mention `.golangci.yml`, but only as "fetch from `https://git.eeqj.de/sneak/prompts/raw/branch/main/.golangci.yml`" — an instruction to vendor canonical verbatim, which is exactly what the scoped rule says. Neither carries a prohibition, so neither needs changing and neither is left contradicting the other. `REPO_POLICIES.md` line 414 lists `.golangci.yml` as a required file, also unaffected.
Note that the repo-root `REPO_POLICIES.md` is a symlink to `prompts/REPO_POLICIES.md`, so the single edit covers both paths.
## Deliberately NOT included
This PR does **not** change `.golangci.yml`. The `gomodguard` fix stays in #47 so the two can be judged separately — the policy question is worth settling on its own terms regardless of what you decide about that config change, and merging them would collapse two decisions into one.
## Unrelated observation, for the record
While verifying #47 against a scratch `sneak/homoicon` clone, I found that homoicon's vendored `.golangci.yml` is sha256 `391ea68e637432980f1db0776076f51578fa58193bbd17f4b40ad725975e21f8`, while canonical `main` is `021cc83f4e6fc7c31b95b34b846723dfcf20b66b7baeea1dc40406e643346bcb`. The whole difference is a three-line comment recording a one-time agent edit you authorized on 2026-08-07; the config is functionally identical.
That matters only if you ever want a hash-based drift guard against canonical, which #25 floats as an offline alternative to `golangci-lint config verify`: such a guard would already report homoicon as drifted on day one. Worth knowing before building one. No change proposed here.
## Validation
`make check` passes (`prettier --check '**/*.md' --tab-width 4 --prose-wrap always`: all matched files clean). `make fmt` produced no further changes. `last_modified` in the front matter updated to 2026-08-19 per this file's own rule.
Co-authored-by: sneak <sneak@sneak.berlin>
Co-authored-by: Jeffrey Paul <sneak@noreply.example.org>
Reviewed-on: #49
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
Codifies your ruling (2026-08-30, filed as homoicon issue 555): no project logic outside `internal/` or `pkg/`; each `cmd/<name>/` is a single `main.go` whose body is one call into library code.
- `CODE_STYLEGUIDE_GO.md`: strengthens the "keep `main` small" rule to the single-call form and adds the no-logic-outside-`internal/`-or-`pkg/` rule.
- `REPO_POLICIES.md`: annotates `cmd/` in the canonical subdirectory list accordingly. (Root copy is a symlink; one edit covers both.)
Co-authored-by: sneak <sneak@sneak.berlin>
Reviewed-on: #54
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
## The question this answers
Is the 60-second test-time cap the new **org-wide** ceiling, or an approved
**dnswatcher-only** divergence?
- Question: #41
- Origin: sneak/dnswatcher#93
This PR implements **org-wide**.
## The ruling
On sneak/dnswatcher#93 (comment) (2026-08-09),
verbatim:
> make the cap 60s in both and never use mocking, always use live resolvers and
> assume the build and run environments have full unmodified unrestricted
> internet access. it is ok if they fail due to a bad build environment that
> alters dns packets.
And on #41 (comment), disambiguating
the scope:
> org wide. the hard cap is 60 for ci/green, but over 20s should be filed as an
> improvement bug.
That second comment landed after this work was started, and it confirms the
option implemented here. The two-tier shape it describes (60s hard, 20s target,
overage filed as a bug) is encoded in the policy text.
## What changed, and the number chosen
The backstop moves from `30s` to **`90s`**.
The old pairing was incoherent under the new cap: a 60-second ceiling with a
30-second `-timeout` means the timeout kills the suite long before the ceiling
is reached, so the ceiling would never be the thing that fails. The backstop has
to sit above the cap, where it does its actual job of catching a hung test
rather than a merely slow one. `90s` preserves the 1.5x backstop-to-cap ratio
the old `20s`/`30s` pair already had, so the relationship between the two
numbers is unchanged and only the scale moves.
Every place a number changed:
| File | What |
| --- | --- |
| `prompts/REPO_POLICIES.md` | Prose ceiling: `20 seconds` to `60 seconds`, plus the new 20s target / improvement-bug tier |
| `prompts/REPO_POLICIES.md` | Backstop prose: `30-second timeout` to `90-second timeout`, with the rationale for why it exceeds the cap |
| `prompts/REPO_POLICIES.md` | Go example Makefile snippet, first run: `go test -timeout 30s` to `-timeout 90s` |
| `prompts/REPO_POLICIES.md` | Go example Makefile snippet, verbose rerun: `go test -timeout 30s` to `-timeout 90s` |
| `prompts/EXISTING_REPO_CHECKLIST.md` | `make test` has a `30-second` timeout, to `90-second` timeout plus the 60s hard cap and the 20s filing rule |
| `prompts/NEW_REPO_CHECKLIST.md` | `script/test` / `make test` entrypoint line: `30-second timeout` to `90-second timeout, 60-second hard cap on wall time` |
I swept the whole repo for `20 second`, `30-second`, `20s`, `30s`, `timeout 20`,
`timeout 30`, and `under 20`/`under 30`. After this change the only remaining
occurrences of `20` in a test-timing context are the two intentional references
to the new 20-second target. The other `timeout` hits in the repo are unrelated
(`.golangci.yml` lint timeout, HTTP server `ReadTimeout`/`WriteTimeout` examples,
`middleware.Timeout`) and were left alone.
One deliberate non-change: the Python example Makefile snippet in
`prompts/REPO_POLICIES.md` carries no timeout flag today and still carries none.
`pytest` has no built-in timeout, so adding one would mean mandating the
`pytest-timeout` plugin org-wide, which is a new dependency requirement rather
than a renumbering, and outside what was ruled on. It is a pre-existing gap
between the prose and that snippet, not one this PR introduces. Happy to file it
separately or add `--timeout=90` here if you want it in scope.
## The alternative that was not implemented
**Keep canonical at 20s and let `sneak/dnswatcher` carry a documented per-repo
divergence.** That was the recommendation originally written up in
#41, on the reasoning that the 20s
ceiling is doing real work in repos with fast deterministic suites and only
dnswatcher needs the headroom.
Org-wide was chosen instead for two reasons. First, the pressure is not specific
to DNS: any repo whose tests exercise real infrastructure over the network
inherits the same variance, and there is no principled line that admits
dnswatcher and excludes the next such repo. Second, and more decisively,
`REPO_POLICIES.md` is a vendored file. A sanctioned per-repo divergence in a
vendored file is indistinguishable, on inspection, from a stale vendored copy:
the next re-vendoring silently reverts the divergence, and nobody reading a
consuming repo can tell whether the number they are looking at is an intentional
exception or drift. That is the bidirectional-drift problem already tracked in
#31. The two-tier cap gets the same
outcome without the drift, since a fast repo that regresses from 4s to 45s still
generates an improvement bug.
## Status
This PR was opened speculatively, ahead of a decision, on the standing "open it
rather than wait" instruction. The scope question has since been answered
org-wide in the issue, but the specific backstop value of `90s` and the two-tier
wording are still my proposals rather than anything ruled on, so closing this or
sending it back for a different number is a perfectly fine outcome.
`make check` passes; `make fmt` was run and the result is included (it was a
no-op, the edits were already prettier-conformant).
`sneak/dnswatcher` is landing the matching 60s edit to its vendored copy in
parallel, and will match whichever way this is decided.
Co-authored-by: clawbot <clawbot@eeqj.de>
Reviewed-on: #42
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
Requested by sneak.
Sets the canonical `.golangci.yml` to the org-standard v2-schema config already deployed byte-identical across the org's Go repos (vaultik, sfdupes, attrsum, upaas, simplelog, mfer, secret, rgoue, bsfirehose). sha256: `021cc83f4e6fc7c31b95b34b846723dfcf20b66b7baeea1dc40406e643346bcb`.
Why: settings live under `linters.settings`, so the lll/funlen/cyclop/dupl thresholds actually apply under golangci-lint v2. The old canonical file kept them under top-level `linters-settings`, which v2 ignores.
Version note: golangci-lint v2.12.2 tag = commit `c0d3ddc9cf3faa61a4e378e879ece580256d76e5`, recorded for consuming repos. This repo itself has no version pins; the `v2.x.x` / `@sha256:...` strings in `prompts/REPO_POLICIES.md` are intentional placeholders and are unchanged.
Known informational note: this config does not disable the deprecated `gomodguard` linter, so golangci-lint 2.12.x prints a deprecation warning. Harmless, accepted.
Matches dnswatcher PR #96: sneak/dnswatcher#96
Verification: `make check` green (prettier fmt-check on all markdown) after `make fmt`; `.golangci.yml` sha256 verified as `021cc83f...` matching the org-standard file.
Co-authored-by: sneak <sneak@sneak.berlin>
Co-authored-by: clawbot <clawbot@eeqj.de>
Reviewed-on: #24
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
Per sneak's instruction:
- Constructors **must** be `New()`, `From<Something>()`, or `NewThing()` (multi-type packages only)
- Strongly discourage creative names (`Create`, `Make`, `Build`, `Init`)
- Constructors **must** use a `Params` struct (or `ThingParams`) for 2+ arguments — no exceptions
- Single obvious argument (`ctx`, bytes) is the only exception
- `context.Context` does not count against the argument limit (already documented)
Co-authored-by: user <user@Mac.lan guest wan>
Co-authored-by: clawbot <clawbot@noreply.git.eeqj.de>
Reviewed-on: #19
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
## Summary
Adds the conditional verbose test rerun pattern as a policy recommendation in REPO_POLICIES.md.
Per sneak's request from [sneak/chat PR #82](sneak/chat#82): document the pattern where `make test` runs tests without `-v` first, then automatically reruns with `-v` on failure for full diagnostic output.
## Changes
**`prompts/REPO_POLICIES.md`** (root `REPO_POLICIES.md` is a symlink to this):
- Added new policy bullet after the `make test` timeout rule
- Explains the rationale: clean CI/Docker build logs on success, full verbose output on failure
- Includes a generic shell pattern template
- Includes concrete Go and Python examples
- Documents that `exit 1` ensures the target always fails after a rerun (the rerun is solely for diagnostic output)
- Updated `last_modified` from 2026-03-12 to 2026-03-18
## The Pattern
```makefile
test:
@go test -timeout 30s -race -cover ./... || \
{ echo "--- Rerunning with -v for details ---"; \
go test -timeout 30s -race -v ./...; exit 1; }
```
- **On success**: concise package summaries only, no per-test noise
- **On failure**: automatic verbose rerun shows every test case and assertion
- **Always fails**: `exit 1` ensures the build fails regardless of second run's exit code
closes #20
Co-authored-by: clawbot <clawbot@noreply.git.eeqj.de>
Reviewed-on: #21
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
Documents the multistage Docker build pattern we now use across repos (chat, pixa, etc.) where a separate `lint` stage runs `make fmt-check` and `make lint` independently from the build stage.
Key additions to REPO_POLICIES.md:
- Full Dockerfile template showing the lint → build → runtime stage pattern
- Explanation of `COPY --from=lint /src/go.sum /dev/null` as the BuildKit dependency trick
- Handling `//go:embed` placeholders in the lint stage
- CGO/system library notes for the lint stage
- Clarification that tests run in the build stage, not the lint stage
Reference implementations: `sneak/chat`, `sneak/pixa`.
Co-authored-by: user <user@Mac.lan guest wan>
Reviewed-on: #18
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
Closes#16
Adds a comprehensive HTTP/web service security hardening policy to `REPO_POLICIES.md` that must be satisfied before tagging 1.0. The policy covers all items sneak specified (without limitation):
**Security headers** — HSTS (min 1 year, includeSubDomains), CSP (restrictive `default-src 'self'` baseline), X-Frame-Options / frame-ancestors, X-Content-Type-Options: nosniff, Referrer-Policy, Permissions-Policy.
**Request/response limits** — max request body size on all endpoints, max response size for paginated APIs, ReadTimeout + ReadHeaderTimeout (slowloris defense), WriteTimeout, IdleTimeout, per-handler execution time limits.
**Authentication & session security** — rate limiting on password-based auth (API keys exempt as high-entropy), CSRF tokens on state-mutating forms (header-auth APIs exempt), bcrypt/scrypt/argon2 for passwords, session cookies with HttpOnly + Secure + SameSite.
**Reverse proxy awareness** — true client IP detection via X-Forwarded-For/X-Real-IP with trusted proxy allowlist (never trust unconditionally).
**CORS** — explicit origin allowlist for authenticated endpoints; wildcard only for public unauthenticated read-only APIs.
**Error handling** — no leaking stack traces, SQL queries, file paths, or implementation details to clients.
**TLS** — HSTS and secure cookie flags required regardless of whether the service terminates TLS directly or sits behind a reverse proxy.
The policy is explicitly non-exhaustive (defense-in-depth: "when in doubt, harden").
Also adds corresponding checklist sections to `EXISTING_REPO_CHECKLIST.md` and `NEW_REPO_CHECKLIST.md` so that HTTP hardening is verified during repo setup and 1.0 preparation.
Co-authored-by: user <user@Mac.lan guest wan>
Co-authored-by: clawbot <clawbot@eeqj.de>
Reviewed-on: #17
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
Add policy rule: build artifacts and code-derived data must not be committed to repos if they can be generated during the build process.
Notable exception: Go protobuf-generated files (`.pb.go`) may be committed because `go get` downloads source but does not execute build steps.
This addresses feedback from sneak/chat PR [#61](sneak/chat#61).
Co-authored-by: clawbot <clawbot@noreply.git.eeqj.de>
Reviewed-on: #15
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
Updates LLM_PROSE_TELLS.md with three new patterns (two-clause compound sentence, almost-hedge, unnecessary contrast), the lol section with conversation excerpts, fixes for instances of these patterns throughout, and a bracket escaping fix for prettier idempotency. Checklist is now 24 items.
Co-authored-by: user <user@Mac.lan guest wan>
Reviewed-on: #8
Co-authored-by: clawbot <clawbot@noreply.example.org>
Co-committed-by: clawbot <clawbot@noreply.example.org>
Types should live alongside their implementations, not in separate
'types', 'domain', or 'models' packages. Type-only packages cause
alias imports and indicate poor package design.
Prompted by review feedback on upaas PR #126.
Split quickstart into separate repo-policy and code-style passes, clone
prompts repo once instead of per-command, and make each prompt
self-contained so agents don't need memory of prior runs.
- Add test requirement item to both checklists (must not be a no-op)
- Add .dockerignore template URL to Dockerfile items in both checklists
- Add REPO_POLICIES.md symlink in repo root pointing to prompts/
- Add comment to Makefile explaining why prettier flags are repeated
