Repository navigation
feat(infra): VM deployment without Docker (systemd + install/deploy scripts) - #3
Conversation
…ystemd units) Operators without Docker had no supported production path — the repo shipped Docker Compose only, while the app itself needs nothing containerized (Bun entry points, API serves the SPA via AGRIPPA_WEB_DIST, migrate+seed on boot). infra/vm/ codifies the same topology on one Ubuntu 22.04/24.04 host: - install.sh: idempotent bootstrap (Bun, Postgres 17 via PGDG, optional Redis 7, agrippa system user, /etc/agrippa/agrippa.env with generated secrets, units). Re-runs never rotate AGRIPPA_SECRET_KEY or the DB password — rotating the key would orphan every stored credential. - deploy.sh: git pull --ff-only, frozen install, SPA build, restart api, wait for /healthz (migrations apply on api boot), then restart worker — schema changes always land before code that expects them. The worker unit's ExecStartPre gates on /healthz for the same reason on first boot. - Units mirror the images' security intent: repo root-owned, services run as agrippa. The worker's hardening is deliberately lighter than the api's — bubblewrap needs user/mount/pid/net namespaces, and the SDK sandbox degrades silently without them. Two accepted deviations from Compose, documented in docs/design/08: both services share one user (artifacts not read-only to the api) and one env file (ANTHROPIC_API_KEY visible to the api).
The deployment docs were framed around Docker Compose only; with infra/vm/ in the tree, every page covering deployment needs the third mode so the design doc stays the living source of truth and both manual locales move together (repo convention). - docs/design/08: retitled to Self-Hosted, names the three modes (dev bare-metal / Compose / VM systemd), adds the VM section with the layout mapping to Compose volumes, boot/upgrade ordering, the bubblewrap namespace constraint, and the two accepted deviations from Compose. - manual 01 (en+zh): VM install/update quick start alongside Compose. - manual 06 (en+zh): VM section (units, journalctl, layout, update flow); backup and upgrade sections now cover both modes; troubleshooting rows for the worker gating on /healthz and the Ubuntu 24.04 AppArmor userns restriction that silently degrades the sandbox. Also corrects the backup section's volume name (workdata never existed; the compose volume is artifacts). - README: VM quick-start block; CHANGELOG under [Unreleased].
|
Warning Review limit reached
Next review available in: 18 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (9)
📝 WalkthroughWalkthroughAdds Dockerless Ubuntu VM deployment using an idempotent installer, generated configuration, hardened systemd API/worker services, ordered updates, health checks, optional Redis, Nginx proxy guidance, and corresponding English/Chinese deployment and operations documentation. ChangesVM Deployment
Estimated code review effort: 3 (Moderate) | ~25 minutes Sequence Diagram(s)sequenceDiagram
participant Operator
participant deploy.sh
participant agrippa-api
participant PostgreSQL
participant agrippa-worker
Operator->>deploy.sh: run VM update
deploy.sh->>agrippa-api: restart API
agrippa-api->>PostgreSQL: migrate and seed on boot
deploy.sh->>agrippa-api: poll /healthz
deploy.sh->>agrippa-worker: restart after API health
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
| Filename | Overview |
|---|---|
| infra/vm/agrippa-worker.service | Worker systemd unit: ExecStartPre polls /healthz for 120s but no TimeoutStartSec is set, so systemd's default 90s start timeout would abort the wait prematurely on slow first-boot migrations. |
| infra/vm/deploy.sh | Deploy script: well-structured pull → build → restart ordering; health-check loop lacks --max-time per curl call, so a hung (not refused) healthz could stall indefinitely. |
| infra/vm/install.sh | Idempotent bootstrap script; correctly guards secret regeneration behind an ENV_FILE existence check; Bun is installed via curl |
| infra/vm/agrippa-api.service | API systemd unit: hardening is appropriate (NoNewPrivileges, ProtectSystem=full, PrivateTmp, PrivateDevices, RestrictNamespaces); no issues found. |
| infra/vm/env.example | Env template with placeholder tokens substituted by install.sh; well-commented with backup warning for AGRIPPA_SECRET_KEY. |
| infra/vm/nginx.conf.example | Nginx reverse-proxy example: SSE route correctly sets proxy_buffering off + proxy_http_version 1.1 + Connection ""; proxy_read_timeout 1h is appropriate for long-lived SSE connections. |
| docs/design/08-deployment.md | Design doc updated to cover three deployment modes; documents the two accepted deviations from Compose (shared user, single env file) explicitly rather than hiding them. |
Sequence Diagram
%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
participant Op as Operator (root)
participant IS as install.sh
participant DS as deploy.sh
participant PG as PostgreSQL
participant API as agrippa-api.service
participant W as agrippa-worker.service
Op->>IS: sudo install.sh [--skip-redis]
IS->>IS: apt-get: bun, pg17, redis, bwrap
IS->>PG: create role + database agrippa
IS->>IS: write /etc/agrippa/agrippa.env (secrets)
IS->>IS: install + enable systemd units
IS->>DS: delegate first deploy
DS->>DS: git pull --ff-only
DS->>DS: bun install --frozen-lockfile
DS->>DS: bun run build (SPA)
DS->>DS: copy units → daemon-reload
DS->>API: systemctl restart
API->>PG: migrate + seed on boot
DS->>DS: poll /healthz (≤120 s, 2 s interval)
API-->>DS: 200 OK
DS->>W: systemctl restart
W->>API: ExecStartPre: poll /healthz (≤120 s)
API-->>W: 200 OK
W->>W: start pg-boss consumers
%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
participant Op as Operator (root)
participant IS as install.sh
participant DS as deploy.sh
participant PG as PostgreSQL
participant API as agrippa-api.service
participant W as agrippa-worker.service
Op->>IS: sudo install.sh [--skip-redis]
IS->>IS: apt-get: bun, pg17, redis, bwrap
IS->>PG: create role + database agrippa
IS->>IS: write /etc/agrippa/agrippa.env (secrets)
IS->>IS: install + enable systemd units
IS->>DS: delegate first deploy
DS->>DS: git pull --ff-only
DS->>DS: bun install --frozen-lockfile
DS->>DS: bun run build (SPA)
DS->>DS: copy units → daemon-reload
DS->>API: systemctl restart
API->>PG: migrate + seed on boot
DS->>DS: poll /healthz (≤120 s, 2 s interval)
API-->>DS: 200 OK
DS->>W: systemctl restart
W->>API: ExecStartPre: poll /healthz (≤120 s)
API-->>W: 200 OK
W->>W: start pg-boss consumers
Reviews (1): Last reviewed commit: "docs(manual,design): document the VM (sy..." | Re-trigger Greptile
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| Restart=on-failure | ||
| RestartSec=5 | ||
| TimeoutStopSec=30 | ||
| LimitNOFILE=65536 |
There was a problem hiding this comment.
ExecStartPre script is designed to wait up to 120 seconds (60 × 2 s) for /healthz, but no TimeoutStartSec is set. systemd's default DefaultTimeoutStartSec on Ubuntu 22.04/24.04 is 90 seconds. On a slow first boot where initial migrations take a while, systemd will send SIGKILL to the start hook at 90 s — before the 120-second window closes — causing the worker to enter a failed state even though the API would have become healthy. A value of 150 s gives the full 120-second poll plus a comfortable buffer for the API process itself to initialise.
| Restart=on-failure | |
| RestartSec=5 | |
| TimeoutStopSec=30 | |
| LimitNOFILE=65536 | |
| Restart=on-failure | |
| RestartSec=5 | |
| TimeoutStartSec=150 | |
| TimeoutStopSec=30 | |
| LimitNOFILE=65536 |
| Environment=HOME=/var/lib/agrippa | ||
| # Wait up to 120s for the API to report healthy (schema migrated + seeded). | ||
| # $$ keeps systemd from eating the $ — the shell must see $(seq) and ${PORT}. | ||
| ExecStartPre=/bin/sh -c 'for i in $$(seq 1 60); do curl -sf "http://127.0.0.1:$${PORT:-3000}/healthz" >/dev/null && exit 0; sleep 2; done; echo "API /healthz never became ready" >&2; exit 1' |
There was a problem hiding this comment.
curl calls inside ExecStartPre have no per-attempt timeout. On loopback, a refused connection fails instantly, but if the port is open and the process is stalled (e.g. blocked in migration), the call can hang for the OS's TCP keepalive interval (minutes), making the total wait far exceed 120 seconds. Adding --max-time 2 keeps each attempt bounded to 2 seconds and preserves the intended overall ceiling.
| ExecStartPre=/bin/sh -c 'for i in $$(seq 1 60); do curl -sf "http://127.0.0.1:$${PORT:-3000}/healthz" >/dev/null && exit 0; sleep 2; done; echo "API /healthz never became ready" >&2; exit 1' | |
| ExecStartPre=/bin/sh -c 'for i in $$(seq 1 60); do curl -sf --max-time 2 "http://127.0.0.1:$${PORT:-3000}/healthz" >/dev/null && exit 0; sleep 2; done; echo "API /healthz never became ready" >&2; exit 1' |
| echo "==> wait for /healthz on :${PORT}" | ||
| healthy=0 | ||
| for _ in $(seq 1 60); do | ||
| if curl -sf "http://127.0.0.1:${PORT}/healthz" >/dev/null; then |
There was a problem hiding this comment.
ExecStartPre health loop: curl -sf has no per-attempt timeout. If the API process has opened the port but is blocked mid-migration, a single iteration can stall for minutes and the 120-second budget becomes meaningless. Adding --max-time 2 keeps each attempt bounded.
| if curl -sf "http://127.0.0.1:${PORT}/healthz" >/dev/null; then | |
| if curl -sf --max-time 2 "http://127.0.0.1:${PORT}/healthz" >/dev/null; then |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 3
🧹 Nitpick comments (3)
infra/vm/install.sh (1)
126-126: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low valueFix alignment in the ASCII art banner.
The right border on this line is off by one space, causing a slight visual misalignment in the output banner.
✨ Proposed fix
- │ Losing it orphans every credential the platform has stored. │ + │ Losing it orphans every credential the platform has stored. │🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@infra/vm/install.sh` at line 126, Adjust the spacing on the affected ASCII art banner line so its right border aligns with the surrounding banner lines, without changing the displayed text.infra/vm/deploy.sh (1)
19-20: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low valueStrip carriage returns from the parsed
PORTvariable.If an operator edits
/etc/agrippa/agrippa.envin a Windows environment, the file might contain\r\nline endings. This would causeport_from_envto end with a hidden\r, which breaks thecurlhealth-check URL.Consider stripping carriage returns to make the parsing more robust.
💡 Proposed defensive fix
-port_from_env="$(grep -E '^PORT=' "$ENV_FILE" 2>/dev/null | tail -n1 | cut -d= -f2 || true)" +port_from_env="$(grep -E '^PORT=' "$ENV_FILE" 2>/dev/null | tail -n1 | cut -d= -f2 | tr -d '\r' || true)" PORT="${PORT:-${port_from_env:-3000}}"🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@infra/vm/deploy.sh` around lines 19 - 20, Update the port_from_env parsing in deploy.sh to remove any carriage return characters from the extracted PORT value before applying the fallback chain. Preserve the existing environment-variable precedence and default of 3000.infra/vm/nginx.conf.example (1)
24-30: 🚀 Performance & Scalability | 🔵 Trivial | 💤 Low valueClear the
Connectionheader to enable HTTP/1.1 keep-alive.For standard API routes,
proxy_http_version 1.1;is specified, but theConnectionheader is not explicitly cleared. By default, NGINX setsConnection: closefor proxied connections, which forces a new TCP connection to the backend for every request.To enable true HTTP/1.1 keep-alive performance to the backend, explicitly clear the
Connectionheader as you did for the SSE route.💡 Proposed fix
location / { proxy_pass http://127.0.0.1:3000; proxy_http_version 1.1; + proxy_set_header Connection ""; proxy_set_header Host $host; proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@infra/vm/nginx.conf.example` around lines 24 - 30, Update the standard API proxy location identified by proxy_pass http://127.0.0.1:3000 and proxy_http_version 1.1 to explicitly clear the upstream Connection header, matching the SSE route’s keep-alive configuration. Preserve the existing proxy headers and routing behavior.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@infra/vm/deploy.sh`:
- Around line 44-48: Update the health-check curl invocation in the polling loop
to include a maximum request timeout, such as 5 seconds, so an unresponsive API
causes the iteration to fail and the loop can eventually time out instead of
hanging indefinitely.
- Around line 24-25: Stop the agrippa-worker service before the git pull and
migration steps in the deployment script, ensuring it is paused before any
application files or database schema are modified. Preserve the existing restart
logic at the end of the script.
In `@infra/vm/install.sh`:
- Around line 115-119: Update the environment-file generation in install.sh to
detect the existing skip-Redis option and comment out the REDIS_URL entry when
Redis is skipped, while preserving the template value for normal installations.
Ensure the generated file enables DB polling fallback without changing unrelated
substitutions.
---
Nitpick comments:
In `@infra/vm/deploy.sh`:
- Around line 19-20: Update the port_from_env parsing in deploy.sh to remove any
carriage return characters from the extracted PORT value before applying the
fallback chain. Preserve the existing environment-variable precedence and
default of 3000.
In `@infra/vm/install.sh`:
- Line 126: Adjust the spacing on the affected ASCII art banner line so its
right border aligns with the surrounding banner lines, without changing the
displayed text.
In `@infra/vm/nginx.conf.example`:
- Around line 24-30: Update the standard API proxy location identified by
proxy_pass http://127.0.0.1:3000 and proxy_http_version 1.1 to explicitly clear
the upstream Connection header, matching the SSE route’s keep-alive
configuration. Preserve the existing proxy headers and routing behavior.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 78a97e26-5e41-4269-a613-e1f01d77fa03
📒 Files selected for processing (14)
CHANGELOG.mdREADME.mddocs/design/08-deployment.mddocs/manual/en/01-getting-started.mddocs/manual/en/06-operations.mddocs/manual/zh-CN/01-getting-started.mddocs/manual/zh-CN/06-operations.mdinfra/vm/README.mdinfra/vm/agrippa-api.serviceinfra/vm/agrippa-worker.serviceinfra/vm/deploy.shinfra/vm/env.exampleinfra/vm/install.shinfra/vm/nginx.conf.example
| echo "==> git pull (--ff-only: refuses if the tree has diverged)" | ||
| git pull --ff-only |
There was a problem hiding this comment.
🗄️ Data Integrity & Integration | 🟠 Major | ⚡ Quick win
Stop the worker before updating files and running migrations.
The script currently leaves agrippa-worker running while git pull updates its files and agrippa-api runs database migrations. This creates a dangerous race condition where the worker could execute new codebase logic against an old database schema, read partially built templates, or write to the database while migrations are actively mutating tables.
Stop the worker before modifying the application files to ensure background jobs are safely paused during the update. The script already correctly restarts the worker at the very end.
🛡️ Proposed fix
cd "$APP_DIR"
+echo "==> stop worker (prevent execution during update)"
+systemctl stop agrippa-worker 2>/dev/null || true
+
echo "==> git pull (--ff-only: refuses if the tree has diverged)"
git pull --ff-only📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| echo "==> git pull (--ff-only: refuses if the tree has diverged)" | |
| git pull --ff-only | |
| cd "$APP_DIR" | |
| echo "==> stop worker (prevent execution during update)" | |
| systemctl stop agrippa-worker 2>/dev/null || true | |
| echo "==> git pull (--ff-only: refuses if the tree has diverged)" | |
| git pull --ff-only |
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@infra/vm/deploy.sh` around lines 24 - 25, Stop the agrippa-worker service
before the git pull and migration steps in the deployment script, ensuring it is
paused before any application files or database schema are modified. Preserve
the existing restart logic at the end of the script.
Review findings from PR #3, each verified before applying: - agrippa-worker.service: TimeoutStartSec=150 — the ExecStartPre health wait is designed for 120s, but systemd's 90s default start timeout would SIGKILL it on a slow first boot (initial migrations), leaving the worker flapping through restart cycles instead of waiting. - Both /healthz polls get curl --max-time 2 so a stalled-but-listening API bounds each attempt and the overall ceiling stays honest. - install.sh --skip-redis now comments REDIS_URL out of the generated env file; leaving it set pointed ioredis at a dead server to retry forever instead of taking the intended DB-polling fallback. - nginx example: clear the Connection header on the main location so upstream keep-alive actually works; align the banner box. Rejected (with reasons, for the record): stopping the worker before git pull/migrations — the running worker holds old code in memory and step-granular resume plus advisory-locked boot migration is the platform's rolling-upgrade contract (same as Compose); and CRLF stripping for PORT — the env file is generated and edited on the VM.
… POSTs better-auth's CSRF check trusts only the baseURL origin (http://localhost:3000), but in dev the browser talks to the Vite server on :5173 and the proxy forwarded the Origin header unchanged — every sign-up/sign-in POST got 403 INVALID_ORIGIN, making dev-mode login impossible through the documented three-terminal setup. Fix it in the dev proxy rather than the API: production is same-origin by design (ADR-0001 — the API serves the SPA, no CORS), and the proxy exists to simulate exactly that. Rewriting Origin to the API's own completes the simulation; adding trustedOrigins to the auth config would instead loosen the production code path for a dev-only problem.
The documented dev flow generated AGRIPPA_SECRET_KEY with openssl on every export — but that key encrypts credentials stored in dev (repo tokens, MCP auth), so a fresh key per shell silently orphans them. Bun loads .env.local natively from the cwd and .gitignore already excludes it, so a persistent env file is strictly safer and one less setup step. README keeps the 60-second quick start (now writing .env.local once); CONTRIBUTING carries the detail and the key-stability warning; AGENTS.md gets a one-line pointer so all copies of the boot recipe stay in agreement. Exports still work and take precedence over the file.
There was a problem hiding this comment.
hutusi has reached the 50-credit limit for trial accounts. To continue receiving code reviews, upgrade your plan.
Pushes of PR merge commits fail the push-time commitlint step because "Merge" parses as the commit type (type-case + type-enum), leaving main red and skipping Test/Build — every PR merge (#2, #3, #4) hit this. The PR-range step already lints the branch's real commits; the merge subject is packaging metadata, so commitlint now ignores /^Merge: / (kept narrow so malformed types still fail).
1 participant
Why
The repo shipped Docker Compose as the only production path, but nothing in the app actually requires containers: both services are plain Bun entry points, the API serves the built SPA via
AGRIPPA_WEB_DIST, and migrations + seeding apply on API boot. Operators who run plain VMs (no Docker) had no supported way to deploy. Local dev mode already runs bare and is untouched by this change.What
New
infra/vm/— the same topology as Compose, installed directly on one Ubuntu 22.04/24.04 host:install.sh— idempotent bootstrap: Bun 1.3.14, Postgres 17 (PGDG), optional Redis 7 (--skip-redis; SSE falls back to DB polling), the worker's OS deps (git,ripgrep,bubblewrap), anagrippasystem user, data dirs under/var/lib/agrippa, and/etc/agrippa/agrippa.envwith generated secrets. Re-runs never rotateAGRIPPA_SECRET_KEYor the DB password — rotating the key would orphan every stored credential.deploy.sh— update flow shared with first install:git pull --ff-only→bun install --frozen-lockfile→ SPA build → restart api → wait for/healthz→ restart worker, so schema changes always land before code that expects them.NODE_ENV=productionis set only by the units, never during the build (the SPA build needs devDependencies).agrippa-api.service/agrippa-worker.service— repo stays root-owned at/opt/agrippa(mirrors the images' root-owned/app), services run asagrippa. The worker's hardening is deliberately lighter than the api's: bubblewrap needs user/mount/pid/net namespaces, and the SDK sandbox degrades silently without them, soRestrictNamespaces=/SystemCallFilter=stay off. The worker never migrates, so itsExecStartPregates on the api's/healthz(first boot and upgrades alike).env.example,nginx.conf.example(SSE route withproxy_buffering off), and a pointer README.Two accepted deviations from Compose, documented in
docs/design/08-deployment.mdrather than hidden: both services share one user (artifacts are not read-only to the api) and one env file (ANTHROPIC_API_KEYis visible to the api). Splitting either buys little on a single-org box.Docs move with the change: design doc 08 (retitled "Self-Hosted", now names all three modes), manual 01 + 06 in both locales, README quick-start block, CHANGELOG under
[Unreleased]. The ops manual gains troubleshooting rows for the two VM-specific failure modes (worker waiting on/healthz; Ubuntu 24.04 AppArmor userns blocking bwrap, with a one-line probe).Verification
bun run check✓ ·bun test126 pass / 0 fail, no skipped suites (local Postgres) ·bun run templates:validate✓ ·bun run build✓shellcheck+bash -nclean on both scriptsapparmor_restrict_unprivileged_usernsvs bubblewrap under the unit hardening; a throwawaymultipass launch 24.04smoke run ofinstall.shis the recommended follow-up before trusting a production VM.No code changes — dev mode and the Docker Compose path are byte-identical.
Summary by CodeRabbit