jackson2w/claude-code-skills / infisical-secrets-manager
infisical-secrets-manager
FlaggedThis skill should be used when migrating a credential off a plaintext .env file onto self-hosted Infisical (the fleet secrets-management platform — distinct from Agent Vault, its separate MITM-proxy product), when provisioning a new host's Infisical machine identity, when debugging an `infisical-wrapper.sh`/`infisical-get.sh` fetch failure, or when auditing which fleet credentials still live in plaintext. Trigger phrases include "infisical secrets get empty", "secret not found exit 0", "*not found* infisical", "infisical secrets manager", "infisical-wrapper.sh", "infisical-get.sh", "infisical machine identity", "infisical login universal-auth", "homelab-fleet project", "migrate credential to infisical", "retire .env file", "Injecting N Infisical secrets", "infisical secrets get", "lookup('pipe', '/root/bin/infisical-get.sh", "deployed-copy vs git-tracked-source".
Infisical secrets manager — fleet credential migration
Self-hosted Infisical (VM 111 infisical, 192.168.50.29, https://infisical.tail922cee.ts.net),
project homelab-fleet (ID 4655aace-2e75-4c2a-8d29-9bd438868396), environment dev. Path / held everything until 2026-09-05, when a second agent host needed the same secret NAMES with different values and moved to a per-host folder (--path=/hermes) -- read the folder-vs-prefix gotcha below before adding a host. Built and fully rolled out
2026-09-02 across ansible-ctrl, pbs, n8n, immich, dfw, and Claude Code's own interactive
Cloudflare admin tokens (30 secrets total) — see project_infisical_secrets_manager memory for
the full incident-by-incident history. This skill is the durable "how," not the narrative.
Not the same as Agent Vault. Agent Vault (also Infisical, separate product) is a MITM proxy that injects credentials into an agent's own outbound API calls at the network layer. This skill's Infisical is a real secrets store — a scoped, authenticated pull of one named value. Genuinely independent products, no shared backend. Don't conflate deployment steps between them.
Before migrating anything: verify it's real and still live
Two real findings this session, both caught by verifying instead of assuming:
- A
.envfile existing doesn't mean it holds a real secret.pihole-exporter.envlooked like a normal credential file (right permissions, right shape) but held only explanatory comments — Pi-hole's password auth had been disabled months earlier, so the exporter had been running unauthenticated by design the whole time. A first-pass playbook edit to render this "credential" from Infisical was written, then reverted, once the live file was actually checked. Always read the current live file's content (or have the user do it) before writing migration code around it, not just its existence/permissions. - A credential can go dead with zero automated consumer to notice, for weeks. Two Cloudflare
admin tokens (
claude-code-cf-account/claude-code-cf-zone) were purely interactive-use — no cron/service ever read them, only an ad-hoc Claude Code session doing Cloudflare admin work. They went stale sometime after their last real use and nobody knew until this migration's own functional-verification step (curl .../user/tokens/verify) caught it. For any credential being migrated, do a real functional check against the provider before trusting the migration is meaningful — an Infisical secret that round-trips a dead value isn't actually fixing anything. If a credential turns out dead, that's a separate problem from the migration itself: diagnose the real cause (check for a related rotation event, check if a broker like Agent Vault or another host already holds a live replacement) before minting a blind replacement.
Two consumer mechanisms — pick based on how the credential is currently used
Mechanism A — direct-execution host, systemd-invoked. The consuming process runs on the same
host that will fetch the secret, invoked by a systemd unit's ExecStart=. Wrap it with a shared
infisical-wrapper.sh instead of the unit sourcing a local .env (via EnvironmentFile= or the
script's own source):
ExecStart=/root/bin/infisical-wrapper.sh /root/bin/some-script.sh [args...]
The wrapper logs in via the host's machine identity, then exec infisical run ... -- "$@" —
every secret in the project lands as an env var, same shape EnvironmentFile= gave before. In the
consumer script itself, delete the source /root/.config/X.env lines and change any
[[ -f /root/.config/X.env ]] existence-gate into [[ -n "${VAR:-}" ]] (the script may still run
standalone/ad-hoc without the wrapper — keep the graceful-degrade warn path).
For a script invoked directly by something other than systemd (a cron job, a fail2ban
action hook, anything that can't have its ExecStart= wrapped), use the sibling
infisical-get.sh inline instead, fetching one value at a time:
TELEGRAM_BOT_TOKEN="$(/root/bin/infisical-get.sh TELEGRAM_BOT_TOKEN)"
TELEGRAM_CHAT_ID="$(/root/bin/infisical-get.sh TELEGRAM_CHAT_ID)"
export TELEGRAM_BOT_TOKEN TELEGRAM_CHAT_ID
Mechanism B — Ansible-templated push. A controller (e.g. ansible-ctrl) renders a credential
onto a different target host via ansible.builtin.copy/template, historically sourced via
lookup('file', '/root/.config/<name>-creds.env') + regex_findall. Replace the vars: entry
with a lookup('pipe', ...) call to infisical-get.sh on the controller instead — the rendered
output on the target host is byte-identical, only where Ansible sources the value from changes:
r2_access_key_id: "{{ lookup('pipe', '/root/bin/infisical-get.sh R2_ACCESS_KEY_ID') }}"
This needs no changes on the target host at all — it keeps its real rendered .env file and
EnvironmentFile= wiring exactly as before. Don't try to eliminate the target host's file too
unless there's a real reason (a new machine identity there is extra ongoing maintenance for
marginal benefit — this was explicitly declined for pbs's R2/B2 backup jobs).
The two shared helper scripts (copy these verbatim per host/repo)
infisical-wrapper.sh and infisical-get.sh — same content shape everywhere, only the
INFISICAL_PROJECT_ID/INFISICAL_DOMAIN constants are shared (identical across every
consumer), while IDENTITY_FILE is always the local host's own
/root/.config/infisical-machine-identity.env. Deploy via a small dedicated playbook
(infisical-tooling-install.yml) that other credential-consuming playbooks depend on running
first — don't fold the deploy into an existing playbook, since multiple consumers need these
scripts independently.
# infisical-wrapper.sh
#!/usr/bin/env bash
set -euo pipefail
INFISICAL_DOMAIN="https://infisical.tail922cee.ts.net/api"
INFISICAL_PROJECT_ID="4655aace-2e75-4c2a-8d29-9bd438868396"
IDENTITY_FILE="/root/.config/infisical-machine-identity.env"
[[ $# -eq 0 ]] && { echo "usage: infisical-wrapper.sh <command> [args...]" >&2; exit 2; }
set -a; source "$IDENTITY_FILE"; set +a
# Neither the token NOR the client credential goes on the command line -- see the argv
# exposure gotcha below. These two env vars are undocumented in `infisical login --help`.
export INFISICAL_UNIVERSAL_AUTH_CLIENT_ID="$INFISICAL_CLIENT_ID"
export INFISICAL_UNIVERSAL_AUTH_CLIENT_SECRET="$INFISICAL_CLIENT_SECRET"
INFISICAL_TOKEN=$(infisical login --method=universal-auth \
--domain="$INFISICAL_DOMAIN" --silent --plain)
export INFISICAL_TOKEN
exec infisical run --projectId="$INFISICAL_PROJECT_ID" \
--env=dev --domain="$INFISICAL_DOMAIN" --silent -- "$@"
# infisical-get.sh
#!/usr/bin/env bash
set -euo pipefail
INFISICAL_DOMAIN="https://infisical.tail922cee.ts.net/api"
INFISICAL_PROJECT_ID="4655aace-2e75-4c2a-8d29-9bd438868396"
IDENTITY_FILE="/root/.config/infisical-machine-identity.env"
[[ $# -ne 1 ]] && { echo "usage: infisical-get.sh SECRET_NAME" >&2; exit 2; }
set -a; source "$IDENTITY_FILE"; set +a
# Neither the token NOR the client credential goes on the command line -- see the argv
# exposure gotcha below. These two env vars are undocumented in `infisical login --help`.
export INFISICAL_UNIVERSAL_AUTH_CLIENT_ID="$INFISICAL_CLIENT_ID"
export INFISICAL_UNIVERSAL_AUTH_CLIENT_SECRET="$INFISICAL_CLIENT_SECRET"
INFISICAL_TOKEN=$(infisical login --method=universal-auth \
--domain="$INFISICAL_DOMAIN" --silent --plain)
export INFISICAL_TOKEN
infisical secrets get "$1" --projectId="$INFISICAL_PROJECT_ID" \
--env=dev --domain="$INFISICAL_DOMAIN" --plain --silent
The infisical login ... --silent banner line ("Injecting N Infisical secrets into your
application process") goes to stderr, not stdout — a script that pipes infisical-wrapper.sh's
child process output into jq/JSON parsing is safe by default. Only breaks if you explicitly
merge stderr into stdout (2>&1) while testing — that's a test artifact, not a real bug.
Provisioning a new host (one not already on ansible-ctrl's identity)
A host outside homelab-ansible's own reach (e.g. dfw, which runs its own separate
dfw-ansible repo locally, ansible_connection=local) needs its own Universal Auth machine
identity — reusing ansible-ctrl's identity file across hosts is not how Infisical scopes access.
- Will creates it directly in the Infisical UI (cannot be automated/provisioned by an agent):
Organization → Access Control → Identities → Create Identity (name it after the host, org role
no-access, auth method Universal Auth) → open it → Add to Project →homelab-fleet→ same access level asansible-ctrl's identity → generate a Client Secret with TTL=0/max-uses=0. No further per-secret association step is needed — a project-level identity can read every secret in that project immediately, including ones added to the project after the identity was created. - Client ID/Secret saved to
/root/.config/infisical-machine-identity.envon the new host by Will directly (never through chat). - Install the Infisical CLI via the same apt repo used everywhere else:
curl -1sLf https://artifacts-cli.infisical.com/setup.deb.sh | bash && apt-get install -y infisical. - Deploy the two helper scripts (see above), pointed at the same project/domain constants.
Least privilege on the free plan = a separate project (2026-10-05, hermes). Custom project roles
(and so path conditions like /hermes/**) need Infisical Enterprise; the free plan offers only Admin,
Member, Viewer, No Access, and each applies to a whole project. Anything given a role on homelab-fleet
can read every fleet secret. For a host that must see only its own credentials (an agent host), create
its own project, give its identity Viewer there and no homelab-fleet membership, and point its
helper scripts' INFISICAL_PROJECT_ID at it (path /). To load values without them reaching chat or
argv, add ansible-ctrl's identity to the new project as Member and write via the API
(POST /api/v3/secrets/raw/<name>, JSON body built by jq into a 600 file, token in a curl -H @file
header). GET /api/v1/workspace lists the projects an identity can see, IDs included. Done this way for
hermes: project hermes = 7a848701-b64d-412b-b6b9-de7f530940a4.
Once a host's watchers run through the wrapper, every alert depends on Infisical being reachable:
the wrapper exits non-zero (a failed unit, visible in the sweep) rather than running without
credentials. dfw has accepted that since 2026-09. Prove a migration with the old file gone (rename
to .bak), plus a negative control: the same script run without the wrapper must refuse.
Upgrading the self-hosted server (2026-10-06)
homelab-ansible playbooks/infisical-server.yml owns /opt/infisical/docker-compose.prod.yml and the
unit (.env is asserted, never managed). To upgrade: read Infisical's upgrade-impact note for every
release in between (upgrade-impact/data/releases/<version>.yaml in the GitHub repo; gh api repos/Infisical/infisical/contents/upgrade-impact/data/releases lists them), and prefer a release that
already has one: the newest often doesn't on its release day. Bump the tag in
files/infisical-server/docker-compose.prod.yml and run the playbook; it pulls, pg_dumps, restarts, waits
for /api/status and checks the migration line. Schema migrations are one-way: rollback is the previous
compose plus the dump. The whole fleet's fetches fail for the ~30 s restart, so stay clear of the
02:30–03:30 CT backup hour. v0.162.20 → v0.165.16 on 2026-10-06 took 32 s with zero error logs. v0.165.16 → v0.167.10 on 2026-11-16 (playbook 113 s end to end, 0 error logs, 0 "policy not migrated"). From v0.167.8, reading audit logs (API and UI) needs a paid licence on self-hosted; nothing in the fleet reads them, Will accepted losing the UI page. Most v0.166–0.167 breaking changes are Infisical's built-in Agent Vault, PKI and approval policies, which this fleet does not use (our Agent Vault is the separate product). The restart window fails every wrapped timer job that fires during it (11-16: post-reboot-verify, roaming-dns-canary, wan-failover-watch, wp-static-publish). watcher-failed-notify does not page on a single failure; afterwards run systemctl reset-failed <unit>; systemctl start <unit> on each so systemctl --failed is clean, then check a fetch from ansible-ctrl and hermes (infisical-wrapper.sh restic cat config).
Verification discipline — do all of this, in order, for every credential
- Non-revealing length check:
infisical-get.sh NAME | wc -con the new host — confirms the secret exists and is fetchable before touching any consumer. - For Mechanism B: capture a
sha256sumof the target host's currently-rendered file before re-running the playbook, re-run it, confirm Ansible reportsok(notchanged) on the credential-file task, and confirm the checksum is unchanged after. - Real functional check against the actual consumer/provider — an API call that authenticates, a
real triggered service run that produces its real side effect (an email actually delivered, a
backup that actually completes with
no errors were found), not just "exited 0." A clean exit code alone doesn't prove the credential value is correct, only that nothing crashed. - Retire the old plaintext file only after step 3 passes — rename to
.bak, never delete outright (non-destructive-automation-default; keep for at least one full week of clean runs). - Re-run the same real functional checks from step 3 with the old file genuinely gone. This
is the step most likely to be skipped and most likely to catch a real bug — twice this session
it caught a genuine miss: once, editing the git-tracked source script but forgetting to
redeploy it to the host's actual
/root/bin//usr/local/bincopy (see the next section); the fix in each case was mechanical once found, but neither would have been caught by step 3 alone since the old file was still present when step 3 ran.
Gotchas hit building this out
-
infisical run/infisical secrets get --token=...leaks the token via process argv —ps,systemctl status, andjournalctl -u <unit>all print it in full. Confirmed 2026-09-03: a routinesystemctl status weekly-housekeeping.service(run to check on an in-progress sweep, nothing unusual) printed a live 30-day Infisical machine-identity access token straight into a Claude Code session transcript, because the unit'sExecStart=wasinfisical-wrapper.sh, whoseexec infisical run --token="$TOK" ...put the token in the child process's command line.--helpdoesn't document it, butINFISICAL_TOKENas an exported env var works identically for bothinfisical runandinfisical secrets get(confirmed live) and never appears in argv. Fixed in bothinfisical-wrapper.shandinfisical-get.sh, on bothansible-ctrl(homelab-ansible7a82244) anddfw(dfw-ansible027b862) — if either script is ever rewritten from scratch, keep the token in an env var, not a--token=flag. General lesson: any wrapper that execs a subprocess with a secret as a CLI flag has this same exposure tops/systemctl status/journalctl, not just to a carelesscat— check for an env-var alternative before accepting--token=/--password=-shaped flags as the only option. -
The same fix stopped one step short for a year of habit's worth of scripts: the CLIENT CREDENTIAL was still in argv. Found 2026-09-05 while porting these scripts to a third host. The 09-03 fix moved
INFISICAL_TOKENoff the command line but leftinfisical login --client-id=... --client-secret=...— the call that mints the token — fully exposed to the sameps//proc/systemctl status/journalctlread, on a credential that does not expire on its own.INFISICAL_UNIVERSAL_AUTH_CLIENT_IDandINFISICAL_UNIVERSAL_AUTH_CLIENT_SECRETwork and are undocumented ininfisical login --help, exactly asINFISICAL_TOKENis undocumented ininfisical run --help. Fixed fleet-wide:homelab-ansible@7b17b50,dfw-ansible@898afe0, hermes's copy correct from the start. Verify an undocumented env var in BOTH directions before relying on it —infisical loginalso succeeds from a cached local session, so a positive test alone proves nothing. The negative control isenv -u INFISICAL_UNIVERSAL_AUTH_CLIENT_ID -u INFISICAL_UNIVERSAL_AUTH_CLIENT_SECRET infisical login --method=universal-auth ..., which must exit 1 with no output. General lesson: after fixing a secret-in-argv leak, look one call UPSTREAM — the thing that produced the secret you just protected is usually passed the same way. -
The helper's install path is not the same on every host, and the mismatch fails silently.
ansible-ctrlkeeps it at/root/bin/infisical-get.sh;dfwinstalls to/usr/local/bin/viainfisical-tooling-install.yml. A consumer that hardcoded/root/bin/...on dfw therefore never invoked Infisical at all — its[[ -x ... ]]test simply failed and it fell through to a plaintext file, for every run since it was built, while the comment above the branch claimed Infisical was in use. Confirmed 2026-09-05 inagent-mail-dispatch.sh. Nothing errored, which is the point: a fallback chain that silently succeeds by the worse route is indistinguishable from one that took the good route. Two fixes worth copying: probe both locations, and have the consumer log which route actually answered (cred_source=infisical:/usr/local/bin/infisical-get.sh) so the question is answerable from the log instead of by re-reading the chain. -
Use a per-host FOLDER, never a name prefix, when two hosts need the same secret names.
infisical runinjects each secret into the environment under its stored name, so renamingRESTIC_PASSWORDtoHERMES_RESTIC_PASSWORDto dodge a collision silently delivers nothing to a consumer that readsRESTIC_PASSWORD. Pass--path=/<host>in both helper scripts instead and keep the names the consumers require. Trade-off to state up front: a genuinely shared value (e.g. a Postmark server token) then exists in two folders and a rotation must hit both — the tidy end state is/<host>folders for host-specific values with shared ones at the root. -
A host's "onboarding complete" record can drift from live reality without anything erroring.
infisical's own build memory said "fullnew-host-closing.ymlrun, all 8 touchpoints closed" (2026-09-02) with node-exporter "installed, confirmed up via a real API query" — but by 2026-09-03 the real systemd-managedprometheus-node-exporter.servicehad never actually started (a leftover manually-runnode_exporterbinary from the build session was squatting on port 9100, so Prometheus scraping looked healthy the whole time regardless), and the Pi-hole DNS drop-in (/etc/systemd/resolved.conf.d/10-pihole.conf) didn't exist on disk at all despite being a scoped, supposedly-applied task. Both surfaced only because the weekly sweep's ownansible_check-style drift detection flagged them — re-running the two relevant idempotent playbooks (pihole-dns-client.yml,node-exporter.yml, both--limit infisical) fixed both. Lesson: "closed all touchpoints" at build time is a snapshot, not a guarantee — a host's first post-build weekly sweep is worth checking for real, not just trusting the build session's own self-report.Correction, 2026-09-07: the 9/3 fix above was itself incomplete, and the same bug recurred. "Killed the rogue binary" stopped the process but never removed its systemd unit (
/etc/systemd/system/node_exporter.service,WantedBy=multi-user.target, still enabled). Left enabled, it silently racedprometheus-node-exporter.servicefor port 9100 on every subsequent boot — invisible the whole time because a stopped-but-enabled unit isn't "failed," sosystemctl --failed(what both the daily drift-check and weekly sweep actually query) had nothing to report. Four days of clean checks later, the first real reboot (an unrelatedpvehost apt/kernel upgrade, 2026-09-07) started every enabled unit at once and crash-looped the real service tostart-limit-hit. Fixed properly this time by deleting the unit file, the orphaned binary, and its associated system user outright — not just disabling. General lesson for this whole fleet, not just Infisical: on a rushed or manually-touched build, "stop the process" and "remove what would restart it" are different claims, and a fix that only does the first is invisible between reboots and silently recurs at the next one. If a build session ever ran anything manually before the real Ansible role converged the host, checksystemctl list-unit-files --state=enabledfor a hand-created duplicate before calling the host settled, not justsystemctl --failed. This exact collision pattern is now caught automatically going forward — seeweekly-housekeeping-checks.sh'senabled_unit_dup_*check (service category) and thepost-reboot-verify.timeronansible-ctrl, which runs it fleet-wide within ~5 minutes of any detected reboot (project_post_reboot_verifymemory). -
Deployed-copy vs. git-tracked-source split. Several hosts in this fleet keep a git-tracked script (e.g.
homelab-ansible/scripts/weekly-housekeeping-checks.sh) separate from its actually running deployed copy (/root/bin/weekly-housekeeping-checks.shonansible-ctrl), synced only by re-running the deploying playbook. Editing the source and testing against the deployed copy without redeploying first will silently test stale code — confirmed this exact mistake mid-migration: a post-retirement verification run showed realfail/warnresults still referencing the old (already-renamed) filenames, because the deploy playbook hadn't been re-run after the script edit. Always redeploy before the post-retirement re-check, not just before the first functional check. -
Audit a playbook's
hosts:target before running it, especially a multi-member group. A real pre-existing bug (n8n-install.yml: hosts: productivityinstead ofhosts: n8n) silently applied n8n's install tasks to two unrelated hosts (immich,console) sharing that inventory group, caught only because the run was taking unusually long andps auxfon the controller showed the live SSH target was a host with no business being touched. Seefeedback_verify_playbook_hosts_scopememory for the full incident and the fleet-wide audit pattern (grep -A5 '\[groupname\]' inventory.iniagainst every playbook'shosts:line) that confirmed it was an isolated mistake, not systemic. -
An uncommitted local diff on a target repo isn't automatically safe to overwrite or ignore. Check
git status/git diffbefore touching any file a migration needs to edit. If it turns out to be the agent's own earlier, unfinished, but real and already-deployed work (confirmed here by checking whether the live deployed copy matched the uncommitted source, and whether a project memory already described the feature), it's safe to fold in and commit alongside the migration edit — pull the live file down (ssh host "sudo cat <path>", not a fresh clone, which would be missing the diff), edit on top of it, push, then on the target hostgit checkout -- <file>(only after confirming the local diff is a strict content subset of the incoming commit) followed by a cleangit pull. -
The global credential-dump guard hook (
block-credential-dump.sh, aPreToolUseBash hook) matches on path text, not actual content sensitivity. A completely non-secret file whose name happens to contain "secret"/"credential"/"token" (e.g. this very memory file,project_infisical_secrets_manager.md) will trigger the same block as a real.envdump. Work around it withls -la(metadata),wc -l(line count),cut -d= -f1(key names only), or theReadtool instead ofBash cat/grep/tailon such a path — don't fight it by trying to rephrase the samecatcommand.
A secret can leak through the request path, and your own log read can re-leak it (2026-10-31)
pbpaste | ssh root@ansible-ctrl /root/bin/infisical-set.sh HA_CLAUDE_CODE_TOKEN failed with "unable to
process new secret creations ... POST /api/v3/secrets/raw/". The backend log showed the JWT had landed
in the URL as the secret key, so Infisical logged it. The root cause isn't found yet; the same script
worked the day before. Until it is, don't hand the pbpaste | infisical-set.sh path out for one-off
secrets: use the Infisical web UI. When reading any API server log, redact JWT-shaped tokens first
(sed -E 's/eyJ[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+/<JWT>/g') and drop /secrets/raw/<x>
path tails. A grep -v token= filter still prints a token that arrived in the URL path. Full incident:
reference_infisical_set_key_became_value memory.
Full secret inventory as of 2026-09-02 (for reference, not exhaustive going forward)
homelab-fleet holds: Tailscale OAuth, Postmark, shared Telegram bot token, Grafana API token,
Cloudflare billing + DNS + admin (account/zone) tokens, Immich admin + DB password, R2/B2 backup
keys, Vultr, Anthropic (both admin and headless-Claude-Code keys), Proxmox/Terraform, n8n's
encryption key, and dfw's restic backup credentials. Deliberately not migrated: Hermes's
credentials (uses Agent Vault instead, a different broker — see agent-vault-credential-broker
skill); Olu's own openclaw-anthropic.env bundle on dfw (a hard classifier-blocked file,
different ownership boundary); Infisical's own bootstrap .env and Agent Vault's master password
(unavoidable chicken-and-egg — the credential that proves identity to the store can't itself
live in the store); n8n-import-key.env (confirmed zero real consumers, orphaned).
secrets set NAME=@/path/to/file is documented and does not work (2026-09-06)
infisical secrets set --help advertises <secretName=@/path/to/file>. On the CLI version running
here it is not interpreted: it stores the literal string @/root/.config/whatever.key and
exits 0. Nothing about the command's output distinguishes this from a real write.
Caught only by reading the value back and comparing lengths — 31 characters stored against a 50-character source. Had the round-trip check been skipped, a credential in the vault would have been the path to the credential, and the failure would have surfaced later as an auth error at whatever consumed it.
Use --file with an .env-format staging file instead, which keeps the value out of argv the
same way the @file form was supposed to:
STAGE=$(mktemp); chmod 600 "$STAGE"
trap 'shred -u "$STAGE" 2>/dev/null || rm -f "$STAGE"' EXIT
{ printf 'NAME='; tr -d '\r\n' < "$KEYFILE"; printf '\n'; } > "$STAGE"
infisical secrets set --file "$STAGE" --projectId=... --env=dev --domain=... --silent
Always verify the round trip without printing the value: read it back and compare lengths and
equality in a single shell invocation ([ "$v" = "$w" ]), never by echoing either side.
set and delete disagree about what a secret is by default
infisical secrets set defaults to --type shared. infisical secrets delete defaults to
--type personal. Same noun namespace, same project, opposite defaults — so a set followed
later by a delete with otherwise-matching arguments silently addresses two different objects,
and the cleanup fails while looking entirely correct.
The failure message is the expensive part. Deleting something that does not exist returns:
• Unable to complete your delete request
If this issue continues, get support at https://infisical.com/slack
No cause, no object name, no status. Against a scoped machine identity this reads as an
authorization denial — such a plausible story that it gets accepted without a second invocation.
Confirmed 2026-09-06: it was reported to the user and to another agent as a capability limit
("this identity can create and update but not delete") before --type=shared deleted the same
value on the first try.
Always pass --type=shared explicitly to delete. Anything written by set, by --file, or
through the UI is shared unless someone deliberately chose otherwise.
The transferable tell: a real authorization failure from this CLI names authorization. A message that names nothing is evidence of a missing object far more often than a missing permission. Before concluding an identity lacks a permission, try the operation once more with the object's type, path and environment stated explicitly — a scoped identity is such a satisfying explanation for a destructive operation failing that it substitutes for the check.
Corollary for cleanup after any test that writes a value: confirm it is gone by listing, rather than trusting the delete's exit status or its error text. A leftover under a plausible-looking name is worse than no cleanup at all, because the next reader cannot tell it from something live.
A wrapper that injects a whole path gives every job every secret — name new secrets distinctly (2026-09-12)
infisical run --projectId=… --env=… with no --path injects every secret at the root path into
the job's environment, and so does every other host's wrapper reading the same project and env. On
this fleet that meant ansible-ctrl's report scripts and seven wrapper-based units on another host
all received a newly added secret.
So a generic name is a collision risk. Adding LUNAROUTE_API_KEY for one consumer could shadow a
variable of the same name that another consumer, on another host, expects to hold its own key.
Before adding a secret:
- Check which identities read that project, env and path, and which units use the wrapper
(
grep -l infisical-wrapper /etc/systemd/system/*.serviceon each). - Grep those consumers' EnvironmentFiles and scripts for the proposed name (names only).
- Prefer a consumer-prefixed name (
CLAUDE_CODE_LUNAROUTE_API_KEY), even when no collision exists today.
Longer-term, per-consumer folders read with --path are the real isolation. A distinct name is the
cheap guard until then. Every secret at a shared path is readable by every job that wrapper runs, so
say so when adding one.
Scoped and on-demand secrets for frequent jobs (2026-11-17)
The full-project injection (infisical run, ~63 secrets) still suits daily jobs. For watchers that run every few minutes it meant ~2,600 logins a day, every secret in every child's environment, and the watcher dying whenever Infisical did. On ansible-ctrl there are now three shapes:
infisical-wrapper.sh --only A,B -- cmd: one login and oneinfisical export --format=json. Only the named values are exported, so the env drops from ~131 to ~19 vars. A missing name fails closed (rc 1).- No wrapper at all, for jobs that need secrets only to alert.
homelab-report-lib.sh'screds_on_demandfetches Telegram/Postmark throughinfisical-get.shon the first send. It never overwrites a set variable and does nothing on hosts without that tool.claude-send-email.shalready did the same for Postmark. The hermes watchers do it inline in their ownsend_telegram. --fallback CRED(with--only), for a job that must survive an Infisical outage (proton-watch). The unit hasLoadCredentialEncrypted=CRED:/etc/credstore.encrypted/CRED.cred, and the playbook rebuilds that file from Infisical on every run:printf '%s=%q\n'piped intosystemd-creds encrypt --name=CRED -. When the login or fetch fails, the wrapper logs it and sources$CREDENTIALS_DIRECTORY/CRED.
Gotchas:
- In an LXC,
systemd-credswarns "credential.secret is not located on encrypted media" and works anyway. - Test the fallback with a copy of the wrapper whose
INFISICAL_DOMAINpoints at a dead port, run undersystemd-run --wait -p LoadCredentialEncrypted=.... A plain shell has no$CREDENTIALS_DIRECTORY. - A job whose own code checks
TELEGRAM_*up front (wan-failover-watch, agent-peer-watch, the Python watchers) needs those names in--only.
secrets get --plain exits 0 with an empty line for a secret that does not exist
Infisical CLI 0.43.x: infisical secrets get NAME --plain for a missing NAME prints an empty line and
exits 0. The table and JSON forms show the API's sentinel instead (--output json gives
[{"secretKey":"NAME","secretValue":"*not found*"}]). So any lookup('pipe', 'infisical-get.sh NAME')
with a typo templated a blank credential onto a host with no error. Fixed 2026-10-04 in all three
infisical-get.sh copies (homelab-ansible 704dcbf, dfw-ansible 7ff249f, hermes-ansible b9fd7d0):
they read --output json, check the shape (exit 5 if unexpected) and the *not found* sentinel (exit 4,
nothing on stdout, a message on stderr), then print with jq -r. That output is byte-identical to
--plain, verified by sha256 over all 38 secrets on ansible-ctrl and dfw, and the value never goes
through argv or a $( ). An Ansible pipe lookup on a missing name now fails the task ("returned 4").
Anything new that calls the CLI directly needs the same check; don't test existence by exit code alone.
infisical-set.sh flattened multi-line values, and its verification still said "verified" (fixed 2026-11-17)
The old infisical-set.sh ran tr -d '\r\n' on the input value AND on the read-back before comparing. A multi-line value (a WireGuard .conf) was stored as one line, both copies were mangled the same way, and the script reported "set and verified" on a corrupted secret (caught 2026-11-13 with PROTON1_PREV_CONF; worked around as b64:<base64>). Fixed in hermes-ansible f76a6ab (script review S8-08): it strips only trailing newlines and one CR, normalizes both sides identically and losslessly, exits 3 on an embedded line break with a hint to use the b64: prefix or the web UI, and validates the name against ^[A-Z][A-Z0-9_]*$.
General lesson: a verify step that normalizes both sides the same lossy way matches even when the stored value is wrong. When the point of the check is to catch corruption, compare raw bytes (or normalize one side only).
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 467k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 467k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 467k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 193k | Updated | Claude CodeCodex | Verified | Install |