Skip to content

Ship: redactor and deploy-app fixes, plus the day's bug findings - #337

Merged
vsilent merged 13 commits into
mainfrom
dev
Nov 12, 2026
Merged

vsilent merged 13 commits into
mainfrom
dev

Conversation

@vsilent

@vsilent vsilent commented Nov 12, 2026

Copy link
Copy Markdown
Collaborator

Two code fixes that unblock the CLI video scripts, the day's BUGS.md
findings, and the release-CI pin. Both fixes ship with regression tests
written first and confirmed failing against the previous code.

The fixes

fix(redact) 0bfd3c0 — never mask a boolean. is_sensitive_env_key
matches by substring, so marketplace.no_secrets_confirmation - a flag whose
name contains "secret" - came back from the public template API as the string
***REDACTED***. That is what stacker install writes to disk, so the
generated stacker.yml stopped parsing and the deploy the CLI itself
suggests refused to run. Booleans are now left alone in all three redaction
paths. Numbers are still redacted on purpose: a numeric token is conceivable,
a boolean secret is not. BUGS.md 24.

fix(agent) 3de52ed — deploy-app kept the container's ports.
stacker agent deploy-app <app> --image <ref> --force recreated the
container with PORTS: - where it had been 5678:5678, and the app stopped
answering. The request carried no compose at all: the local-config helper
returns early when no environment resolves, and a template-based project
resolves none. With only --image, the agent rebuilds from the bare image
and every port, volume and network from the compose is gone. It now sends the
compose the deploy generated into .stacker/. One branch. BUGS.md 31.

Also in this range

  • BUGS.md: 20 promoted to CONFIRMED, 24 and 31 marked fixed, 25 to 30 added
    from a live run. Addresses and local paths are redacted - this repo is
    public.
  • ci(release) 261134f: Linux runners pinned to ubuntu-22.04 and an arm64
    build added. Already on main via Release prep 0.3.6: pin the Linux build runners, bump the version #334; it appears here as a duplicate and
    git will collapse it.
  • Install docs pointed at stacker.try.direct/install.sh (403) and
    get.stacker.dev (does not resolve); both now point at the file that is
    actually served.
  • docs/HOMEBREW_TAP.md: the runbook for the tap, which now exists and
    serves brew install trydirect/tap/stacker on macOS and Linux.

Not mine, flagging it: 661d34d fix(marketplace): derive template deploy_count from deployment records and its merge 0af81d0 came in through
#330 and are in this range. If they are not meant to ship yet, this PR is the
wrong vehicle.

Verification

Full lib suite on the tip: 2314 passed, 0 failed, 6 ignored. clippy reports
nothing on either changed file.

Merging builds and pushes the image; it does not update the running
container. Two checks are worth running once it is live:

curl -s https://try.direct/stacker/api/templates/n8n | grep -c REDACTED

should be 0, and one stacker agent deploy-app against a running stack
should leave the published port answering. Neither is covered by the tests -
they exercise the request the CLI builds, not the deployed service.

🤖 Generated with Claude Code

robotizeit and others added 13 commits November 6, 2026 10:25
stack_template.deploy_count was 0 for every template on /api/templates and
the storefront: the counter is only incremented by the marketplace
deploy-complete hooks, which never fire from the web pipeline (the
stack_template_deployment event table is empty and the counter untouched),
while real deployments are recorded as project rows with source_template_id.

Compute deploy_count as GREATEST of the stored counter, the number of
project records with source_template_id, and the number of verified
stack_template_deployment events. Existing deployments are counted
immediately and future ones keep counting even when a hook is missed. The
'popular' sort now orders by the derived value instead of the stale
counter. All queries are runtime query_as, so no .sqlx data changes.
fix(marketplace): derive template deploy_count from deployment records
The Quick Start advertised stacker.try.direct/install.sh, which returns
HTTP 403 with a zero-byte body, so the documented one-liner could never
install anything. Use the installer that returns 200:
https://raw.githubusercontent.com/trydirect/stacker/main/install.sh

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
install.sh and scripts/install.sh both opened with

    curl -fsSL https://get.stacker.dev/install.sh | bash

get.stacker.dev does not resolve at all - curl returns 000, not even a
refusal. These two lines are the first thing a careful person reads before
piping a script into bash, and our audience is exactly the kind that reads
first. They now point at the URL that actually serves this file:

    https://raw.githubusercontent.com/trydirect/stacker/main/install.sh

verified 200, 5998 bytes, and its sha256 matches this repository's root
install.sh exactly.

Comments only - bash -n passes on both files and the executable bit is
unchanged.

Noted while here, not fixed: scripts/install.sh is a DIVERGED copy of the
root install.sh, 37 lines shorter, missing write_user_config among other
things. The raw URL serves the root one. Two installers that differ is a
trap; one of them should become a symlink or go away, which is a decision
rather than a typo fix.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
brew install trydirect/tap/stacker needs no new build work: release.yml
already publishes three tarballs on every release and each holds one binary
named stacker. What is missing is the tap repository, which does not exist -
the GitHub API returns 404 for trydirect/homebrew-tap.

homebrew-core is not an option yet: it applies a notability bar of roughly
tens of forks and watchers, and this repository has 7 stars, 2 forks and 2
watchers. A tap has no such bar.

The formula in the runbook carries real values - the v0.3.5 asset URLs and
the sha256 of each, computed by downloading them on 2026-11-12.

Two blockers are written down rather than fixed, because both are decisions:

- the v0.3.5 binary reports 0.3.4, so Cargo.toml was bumped after the tag was
  cut; a formula declaring 0.3.5 next to a binary saying 0.3.4 fails the
  sniff test
- there is no aarch64-unknown-linux-gnu build, so brew and the curl installer
  both fail on arm64 Linux, which is now the cheap default at several VPS
  providers and exactly where our self-hosting audience is

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two problems, one cause.

The published v0.3.5 Linux binary requires GLIBC_2.39 - read off the asset
itself. Ubuntu 22.04 ships glibc 2.35 and Debian 12 ships 2.36, so that
binary refuses to start on both. Those are the two distributions a
self-hosting audience is most likely to be running, which makes this worse
than a missing architecture: we have been shipping a Linux binary that a
large part of our audience cannot execute at all. The cause is ubuntu-latest
in the matrix drifting to Ubuntu 26.04 - a release binary inherits the build
image's glibc. Pinned to ubuntu-22.04, which sets the floor at 2.35 and keeps
every newer distribution working.

And there was no arm64 Linux build at all, while arm64 instances are now the
cheap default at several VPS providers. Added as a native ubuntu-22.04-arm
runner rather than a cross-compile: the existing protoc step, the toolchain
action and the cache keys all work unchanged that way.

The installer needs no change - install.sh:73 already builds the archive name
from uname -m and maps aarch64/arm64 to aarch64, so the new asset is picked
up as soon as it exists.

Unverified until the next release runs: that the ubuntu-22.04-arm label is
available to this repository. If it is not, the fallback is ubuntu-24.04-arm,
which raises the arm64 glibc floor to 2.39.

docs/HOMEBREW_TAP.md updated to match: the Linux arm64 block in the formula
stays commented out until an asset exists, and the glibc rule is written down
so ubuntu-latest does not come back.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… deploy

Walked the n8n quickstart end to end on a fresh Hetzner server with
stacker-cli v0.3.6, from a clean clone of awesome-selfhosted-stacker. Three
findings, evidence quoted in each entry. Host addresses and local key paths
are redacted - this repository is public.

#20 promoted to CONFIRMED. The deployment ended paused [port_conflict] on a
server the same run had just created, so nothing could have held the port
first. The stack was serving: the app port answers 200, the HTML carries
n8n's own config marker and the readiness endpoint returns 200. The port is
held by this deploy's own healthy container, which is what the preflight
reports as a conflict. That removes the last innocent reading of the three
earlier occurrences.

#24, new. stacker status, run in that project directory immediately
afterwards, reported a completed deployment from 30 July on a different
server and a different project id. It is the first command in SKILL.md's
verification checklist and the one the deploy itself suggests, so following
that advice after a failed deploy returns a green completed from months ago.

#25, new. The same run printed that the SSH key was authorized and saved a
backup key. Neither authenticates, while the host is up and serving.
AGENTS.md treats that line as the gate for testing SSH, which makes the gate
useless and leaves nobody able to investigate or clean up the machine.

The one inference not verified in code - that status resolves by
project.identity - is marked as a guess.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The CLI-16 and CLI-17 video scripts already cite section 24 as the redactor
bug. My entries from this morning had taken that number, so they move down
and 24 is written up properly.

24, new: the secret redactor masks any key whose name contains the substring
"secret". marketplace.no_secrets_confirmation is a boolean flag and comes
back as the string ***REDACTED***, so the served template no longer parses.
Read straight off the public template API, not inferred. This is the root
cause behind gate 2 of the CLI-16 script - the reason that video cannot be
filmed as a real run - and the same response shows the n8n template fix
merged in awesome-selfhosted-stacker has not reached what the API serves.

25 was 24: stacker status reports a months-old deployment from another
server.

26 was 25: the deploy says the SSH key is authorized and neither key opens.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`is_sensitive_env_key` matches by substring, so any key containing "secret"
is redacted. `marketplace.no_secrets_confirmation` is a boolean flag whose
name contains it, and the redactor replaced `true` with the string
`***REDACTED***`.

Read off the public template API today:

    no_secrets_confirmation  ->  ***REDACTED***

That is served by routes/marketplace/public.rs and written to disk by
`stacker install`, so the generated stacker.yml no longer parses as the
schema expects. A first-time user installs a template, runs the deploy
command the CLI itself printed, and it refuses. Logged as BUGS.md 24 and it
is gate 2 of the CLI-16 video script, the reason that video cannot be filmed
as a real run.

A boolean carries no secret material, and masking it changes the value's
type, which is the actual damage. Three conditions, one per redaction path:
the JSON key/value entry form, the JSON direct-key form, and the YAML
mapping. The YAML `VAR=value` sequence form only ever sees strings and is
untouched. Substring matching itself is left alone - narrowing it is a
bigger change than this bug needs.

Numbers are deliberately still redacted: a numeric token is conceivable, a
boolean secret is not.

Two regression tests, one per format, written first and confirmed failing
against the previous code. Full lib suite: 2313 passed, 0 failed. clippy
reports nothing on this file.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The fix is in 0bfd3c0. Records what was changed, what was deliberately left
alone (substring matching, and redaction of numbers), and that the deployed
API has not been re-checked yet - that needs the change to ship first.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
None of these stop a deploy, so they go here rather than into a fix.

27: `stacker install --domain` is accepted and discarded - `--region` and
`--size` land in the written file, `commonDomain` keeps the template default.
`--key` is ignored at install and honoured at deploy. A flag the CLI accepts
and drops is worse than one it rejects.

28: install printed project #818 and the deploy that followed resolved
id=817, in one unbroken run. Unexplained; distinct from 25, which explains a
stale resolution rather than two ids inside one run.

29: `stacker agent install` refuses when `deploy.server.host` is empty, which
is its normal state right after a cloud deploy created the server. It is the
step the deploy itself suggests next, and continuing needs a hand edit of the
file the tool generates.

30: the marketplace serves a template older than the repository - the n8n
fix merged on `main` earlier the same day is still absent from the API, so
`stacker install` keeps writing the broken SSH key path that fix removed.

Also replaces the guess in 25 with the mechanism observed in the run: the
deploy renames the project to the `stacker.yml` name, colliding with an
older project of that name, and status then resolves to the older one.

Evidence for all of these is the CLI-16 capture set in the blog repo; no
addresses or local paths are reproduced here.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`stacker agent deploy-app <app> --image <ref> --force` recreated the
container without its published ports: `5678:5678` became `PORTS: -` and the
app stopped answering. A one-command update that takes the app down is the
opposite of what the command is for, and it is the blocker on the CLI-17
video script.

The request carried no compose at all.
`local_config_files_for_agent_deploy` returns early when no environment
resolves, and a template-based project resolves none: it declares no
`deploy.environment` and no `environments:` block. Without
`compose_content`, `--image` is all the agent has, so it recreates the
container from the image alone and every port, volume and network from the
compose is gone.

The compose describing the running stack is sitting in `.stacker/` the whole
time - the deploy that created the stack generated it there. When no
environment resolves, send it.

One branch, nine lines. Projects that do resolve an environment are
untouched, and so is every other early return: no stacker.yml, no compose
file configured, configured file missing.

Regression test written first and confirmed failing against the previous
code, with the n8n template's exact shape. Full lib suite: 2314 passed, 0
failed. The one clippy warning in the file is at line 2470, outside both
hunks.

Two earlier attempts at this fix were wrong and are not in the diff: the
missing-`compose_file` branch and the missing-file branch are both
unreachable here, because `resolve_environment_config` always supplies a
conventional `docker/<env>/compose.yml` and the function bails on the
environment first. Instrumented the early returns rather than keep guessing.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The fix is in 3de52ed. Records the cause, the fix, and the two wrong
diagnoses that read as correct from the source - both unreachable, because
resolve_environment_config always supplies a conventional compose path and
the function bails on the environment before it ever looks.

Also records what the regression test does not cover: a live update on a
running stack, which needs the fix shipped and a server.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@vsilent
vsilent merged commit 4cdea7c into main Nov 12, 2026
23 of 24 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants