[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M78SSDQVRYX64Y16X32TC0NS":3,"skill-same:01M78SSDQVRYX64Y16X32TC0NS":52},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":29,"commitSha":30,"skillDirectory":31,"forksCount":32,"repositoryHtmlUrl":33,"repositoryHomepageUrl":34,"ownerHtmlUrl":35,"ownerBlogUrl":6,"ownerDescription":36,"updatedAt":37,"install":38,"resources":51},"01M78SSDQVRYX64Y16X32TC0NS","steadfaste-fleet-repo-deploy",null,"Use when committing to the steadfaste-fleet repo.","lukemcqueen\u002Fsteadfaste-fleet","lukemcqueen","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F405240?v=4","01M757MHPAGRXTJAT46SKZP98R","Luke McQueen","01M757MHPZ573KP8DT9GB8SBQ7",4,"MIT","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fsteadfaste-fleet-repo-deploy\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: steadfaste-fleet-repo-deploy\nversion: 1.0.0\ncategory: devops\ndescription: \"Use when committing to the steadfaste-fleet repo.\"\nplatforms: [linux, macos]\n---\n\n\u003C!-- Provenance: renamed from skills\u002Fdevops\u002Fhermes-cortex-repo-deploy\u002FSKILL.md\n     on 2026-11-12 (hc→fleet brand sweep, commit 2dc01b0e). Content predates the\n     range; the range swept the brand tokens only. -->\n\n# Steadfaste Fleet Repo — Deploy & Push Gates\n\nWorkflow for landing changes in `~\u002Fsteadfaste-fleet` (shared public repo). The\nrepo's own AGENTS.md governs governance mechanics; this skill carries the\ndeploy-order and manifest pitfalls that cost time when missed.\n\n## Push sequence (gates run in this order)\n\n1. `git pull --rebase origin main` — REQUIRES a clean tree. ANY unstaged\n   change blocks the rebase AND the push's built-in dogfood — your own edits\n   included, not just a peer's. Stash the specific blocking files by name\n   (never `git stash` bare), rebase\u002Fpush, then `git stash pop` immediately —\n   pop your own files back FIRST so your edits are restored before the next\n   step. Never stash, clean, or commit a peer's work, and never stage a\n   file you did not edit.\n\n   **Never let a rebase run over unpushed commits.** With local commits ahead\n   of origin, or on a conflicted rebase, `pull --rebase` moves HEAD to origin\n   and leaves the repo mid-rebase: the local commits drop out of HEAD and\n   survive only in the reflog, while the working tree reverts to origin. The\n   same hazard lives inside tooling — a pre-push gate that pulls for you can\n   orphan your work the moment it fails. Recover with `git rebase --abort`\n   (restores the pre-rebase HEAD exactly); do not re-author the commits. Before\n   pulling on a working branch, check `git log --oneline origin\u002Fmain..HEAD`: if\n   it lists anything, SKIP the pull — the local work is precisely what you are\n   about to verify and push, so it must not be rebased under.\n2. `git push` runs a doctor gate — it fails on ANY doctor ❌. Most common\n   blocker: `Deploy sync` \u002F `Checksum: \u003Cfile>` FAIL because the deployed copy\n   predates the pushed commit. Fix by running `fleet-update.sh`, re-running\n   the doctor (expect 0 fail), then retrying the push. Never bypass.\n3. After the push lands, run `fleet-update.sh` once more so the deployed tree\n   includes the just-pushed commit (the push gate's internal deploy can trail\n   the newest commit).\n\n## The push also runs an always-review RECEIPT gate\n\n`git push` refuses a range that touches an always-review path unless a CLEAN review receipt\nexists for it (`expected receipt: \u003Cstate>\u002F.reviewed-\u003Cslug>-\u003Ctip>.json`). The receipt is keyed\nby the TIP SHA, and the range spans origin's tip..HEAD — which has consequences that are not\nobvious:\n\n- **A rebase invalidates it.** `git pull --rebase` rewrites SHAs, so a receipt written for\n  the pre-rebase tip no longer matches the gate's expected filename. On a shared branch the\n  rebase rewrites a PEER's unpushed commits too, so their receipt coverage dies with yours\n  (their content survives; say so explicitly when you report it). Getting back requires a\n  cycle that ENDS at the new tip and returns CLEAN from a real review — the receipt is\n  written at close.\n- **`fleet-dogfood.sh --force` cannot produce a receipt.** It only passes `--force-all` to\n  the deploy (a full re-deploy); it runs no review. There is no sanctioned way to re-review\n  a CLOSED cycle either — `record-review.py` is the raw sink the reviewer writes to, and the\n  review-queue sweep never touches overrides. Never manufacture a receipt and never use an\n  override: that gate is what keeps an enforcement change from shipping unreviewed.\n- **The range includes a peer's commits**, so when they touch an always-review file the\n  review must cover the whole range, not just your own commits.\n\n**Run the deploy as the LAST step before the push.** Any commit made after a deploy leaves\nthe deployed marker stale and the gate returns `Deploy sync` ❌ however clean your doctor\nran minutes earlier. Do both in ONE command, nothing in between:\n`bash ~\u002F.steadfaste-fleet\u002Fscripts\u002Ffleet-update.sh > \u003Clog> 2>&1; git push origin main`.\n\n## An operational cycle cannot carry evidence of its own success — commit the CHECK\n\nThe push is LAST, so a transcript of it can only be written afterwards; that commit changes\nthe tip, invalidating the per-tip receipt (above) and leaving the deployed tree trailing so\n`Deploy sync` returns ❌. Reviewers are right to demand committed evidence, and the ordering\nrules keep destroying it — committing more evidence repeats the loop.\n\nBreak it by committing a RE-RUNNABLE CHECK before the push, never a transcript:\n\n    python3 ops\u002Fscripts\u002Fmanage\u002Fverify-landed.py --repo . [--base \u003Crev>] [--no-remote]\n\nIt asserts (1) a CLEAN receipt for the tip, (2) every file the range changed that has a\ndeploy-map entry matches its DEPLOYED copy, (3) `origin\u002F\u003Cbranch> == tip`; it exits 0 \u002F 1 (a\nreal mismatch, named) \u002F 2 COULD NOT VERIFY, and prints no path on failure. A later reviewer\nre-runs it against the state that exists AFTER the push, so the evidence cannot go stale and\nneeds no second commit.\n\n**Measure the checker before trusting it.** On its first real range this check reported three\nfalse FAILs: it hashed deployed copies RAW, and the deploy's header (inserted after the\nshebang, closed by a blank line) makes the two sides differ by construction. When a check\ndisagrees with an otherwise-green gate, print the FIRST DIFFERING LINE of the two bodies\ninstead of assuming either side is wrong, and pin the real shape with a test.\n\n## A failed deploy: read the log's FIRST line, and check the lock after\n\n`fleet-update.sh` sources the host env early, so a broken env file kills the run\nbefore any deploy work and the exit code says nothing specific. Diagnose from the\nlog's first line, not the exit code:\n\n- **`rc=127` whose entire output is `...\u002F.env: line N: \u003Ctoken>: command not found`**\n  — the env file is CORRUPT, not the script. A comment line that lost its leading\n  `#` makes the shell execute the remainder (a bare date fragment is the usual\n  first token). Inspect structure WITHOUT printing values: report each line's key\n  (text before the first `=`) with its length, and list the lines that are neither\n  blank, nor start with `#`, nor contain `=`. Watch for one absurdly long line — a\n  botched append glues a whole blob onto line 1 with no newline, so the header\n  comment swallows kilobytes while every real key still sits on later lines.\n  Repair is surgical and lossless: restore ONLY the damaged leading lines from the\n  newest `.env.bak-*`, leave the rest byte-for-byte, then confirm the key count is\n  unchanged and the mode is still `0600`. A corrupt env breaks EVERY deploy on\n  that host, so treat it as urgent.\n- **A quote\u002FEOF error naming a line BEYOND the file's length is not a static error in that\n  file.** `bash -n \u003Cscript>` passes on it — that pass is not a contradiction, it is the\n  tell: bash numbers lines in whatever unit it is reading at that moment, so the reported\n  line can never be the file's own. The failing text is an eval'd or generated fragment, or\n  a file REWRITTEN while it executes (a deploy that syncs the very script it is running).\n  Do not go editing the script to \"fix\" a syntax error that is not there: capture the run's\n  WHOLE output to a file (never a pipe to `tail`, which destroys the one artifact able to\n  identify the unit) and re-run — a transient rewrite failure does not recur.\n- **`✗ Not a git repository: \u003Cpath>`** — a git WORKTREE has a `.git` file, not a\n  directory, and the script's repo check rejects it. Deploy from the primary\n  working tree (or set `REPO_DIR=`); a session worktree can hold the work and\n  push, but it cannot deploy.\n\n**After ANY failed deploy, check the enforcement files before doing anything\nelse** — a run that dies after unlocking leaves them modifiable. `sudo -n\nhermes-plugin-lock lock` relocks every target in one call, and the doctor's\n`Immutable:` lines are the authoritative confirmation.\n\n## Quoting in commit messages\n\nNever put backticks in a shell-quoted commit message: the shell substitutes them\nbefore git sees the text, mangling the message and leaking fragments onto the\ncommand line. Observed consequence: an empty file literally named `=` was created\nat the repo root, tracked, and flagged as scope drift by the reviewer before I\nnoticed it. Write the message to a FILE and use `git commit -F \u003Cfile>` for any message, not only\none containing code formatting: the command is scanned as a whole, so a policy that\npattern-matches command text can refuse the entire commit over words in the MESSAGE\n(a gateway-lifecycle guard rejecting a message that merely mentions a restart), and\nthe refusal names the command rather than the cause. Then run `git status --short`\nand check for surprise paths before pushing.\n\n## Restoring a file you swapped in: `git checkout --` restores the INDEX\n\nTo reproduce a failure against an older revision you may copy that revision over a working\nfile (`git show \u003Crev>:\u003Cpath> > \u003Cpath>`). Restoring it with `git checkout -- \u003Cpath>` gives\nback the **index** copy, not the working tree — so any edit you made but had not staged is\nsilently discarded, and the command reports success. Observed: a comment fix was reverted\nthis way, and the commit that followed carried only the evidence file, so the reviewer's\nfinding it addressed came straight back on the next close.\n\n- Copy the file aside FIRST (`cp \u003Cpath> \u003Cscratch>\u002Fmine.py`) and restore from that copy, or\n  use `git checkout HEAD -- \u003Cpath>` when HEAD's version is the one you want.\n- Then **grep for the edit** before committing. A clean `git status` proves nothing here:\n  a discarded edit leaves the tree cleaner, not dirtier.\n\n## Never ASSUME a deployed path — resolve it; never let a fallback carry a conclusion\n\nThe deployed path of a repo file CANNOT be derived from the repo path: 192 of ~289\n`register()` entries follow no rule (see the header of `ops\u002Fscripts\u002Ffleet_lib\u002Fpaths.py`).\nA probe that mirrors the repo layout under the deploy home therefore checks NOTHING, and\npaired with `2>\u002Fdev\u002Fnull || echo \"\u003Cconclusion>\"` it prints that failure AS a finding — this\nis how the line \"0 (deployed copy predates this commit)\" reached a committed evidence file\nwhile the deployed copy in fact already carried the change.\n\n- **Ask the map, never guess:** `python3 ops\u002Fscripts\u002Fmanage\u002Fdeployed-path.py \u003Crepo-relative-path>`\n  prints the deployed destination (from `deploy-manifest.tsv`, generated from the register()\n  lines) and exits 0 \u002F 2 (no entry) \u002F 3 (entry present, nothing at the destination). On a\n  failure it prints NO path — there is deliberately no fallback value to mistake for an\n  answer.\n- **A repo copy is NOT the deployed copy.** Check the two separately and say which you ran;\n  a green check on the repo side proves nothing about what a host runs.\n- **Never let a fallback string carry a conclusion.** If a check cannot run, print\n  COULD NOT VERIFY and exit non-zero. `|| echo \"\u003Csomething plausible>\"` converts a broken\n  probe into durable \"evidence\", which is worse than having none.\n\n## Edit the REPO source, never the deployed tree\n\n`fleet-update.sh` copies every `register()` source over its deployed destination,\nand the deployed file says so in its own header (`# SOURCE: \u003Crepo path>` \u002F\n`# Do NOT edit this file`). An edit aimed at a deployed path SUCCEEDS silently and\nlooks right — the file is right there, and every test still passes, because the\ntests import the REPO tree. The next deploy reverts it, since repo → deployed is\nthe only direction a deploy copies.\n\n**On an IMMUTABLE target the write does not even succeed** — the enforcement surface\n(the enforcer plugin, the git hooks, the loop-governance MCP server) carries the\nkernel immutable flag, so a direct write fails with `PermissionError: [Errno 1]\nOperation not permitted`, or `EPERM` from a shell. That is the write-protection layer\nworking, not a permissions bug to escalate: you are editing the wrong side. Verify\nwith `lsattr \u003Cdeployed path>` (expect `----i---------e-------`, see\n`enforcement-immutability`), then put the change in the REPO source and let\n`fleet-update.sh` deploy it — the deploy owns the unlock\u002Frelock around the copy, so\nit is the only sanctioned way to change one. A `cp`\u002F`python` write to a deployed path\nis the same wrong move whichever way it fails, loudly or silently.\n\n- **Patch by REPO path only.** Anything under `~\u002F.steadfaste-fleet\u002F` is an artifact;\n  the editable twin is `~\u002Fsteadfaste-fleet\u002Fops\u002Fscripts\u002F…`. A `patch` call resolves the\n  path it is given, so a deployed path applies cleanly and gives no hint it will be\n  thrown away.\n- **The deploy's drift guard is the backstop, and it fails CLOSED.** When a deployed\n  file differs from its source the run stops with `Copy the deployed changes to the\n  repo source first, then commit` (one line per file) and the doctor reports\n  `Deploy sync` FAIL. That is the guard working: a blind overwrite would destroy\n  whichever side would lose. Resolve it by making the REPO carry the change and\n  deploying again — never by copying the deployed file over the source, which\n  discards the edits already in the source.\n- **Work out which side is ahead before choosing a direction.** The deployed header is NOT a\n  fixed block at the top: the deploy inserts it AFTER the shebang (`# SOURCE:` then\n  `# Do NOT edit this file`), and the block is closed by a BLANK line. So `tail -n +4`\n  misaligns the comparison — it drops the shebang and keeps the blank, and the diff shows a\n  difference that is not there. Remove the header LINES and compare what remains; better\n  still, do not hand-roll the comparison at all —\n  `python3 ops\u002Fscripts\u002Fmanage\u002Fverify-landed.py --repo .` compares every changed\n  deploy-map file against its deployed copy and names any real mismatch.\n  Edits only in the deployed copy must be synced INTO\n  the repo; edits only in the repo just need a deploy.\n- **After deploying, confirm the deployed file actually took the repo's version** —\n  the same `diff` above returning nothing — before reporting the change as live.\n  A deploy that printed `rc=0` also prints drift warnings for files it did NOT sync;\n  read the run's summary, not just its exit code.\n\n## New shared module → register() entry in the SAME commit\n\nA new module that an existing deployed script imports (e.g. a helper under\n`fleet_doctor\u002F`) must get its own `register()` line in `fleet-update.sh`,\nplaced next to its importer's. `register()` copies ONLY listed files — a\nmissing entry makes the next deploy break every importer with\n`ModuleNotFoundError: No module named '\u003Cmodule>'` at the deploy's verification\nstep. Symptom to recognize in the deploy log; fix by adding the entry,\ndeploying, re-running the doctor.\n\nThe same rule covers a script the deployed code loads **by path at runtime**\nrather than importing (`importlib.util.spec_from_file_location`, a subprocess\ncall, a plugin lookup): the loader finds nothing on hosts, so the feature is\nsilently absent there while it works perfectly in the repo. Register both the\nscript AND every file it resolves **relative to itself** — a config beside it\n(`Path(__file__).parent \u002F \"\u003Cname>.yaml\"`) is not deployed by registering the\nscript alone. A loader written to try the repo path first then the deployed\npath will pass every local test and fail only on hosts.\n\nNote `importlib.util.spec_from_file_location` returns an Optional: guard\n`spec is None or spec.loader is None` before `module_from_spec`, or a bad path\nbecomes an AttributeError at the first call instead of a clean skip.\n\n**A CLI ENTRYPOINT is the third shape of the same rule.** A script the fleet is\ntold to RUN from the deployed path (an agent-side tool, a reply primitive, a\nmaintenance command) is imported by nothing, so a missing register entry breaks\nno import and no deploy step — the host simply has no such file, and the whole\nsuite stays green because the tests load it by REPO path. The feature is not\n\"silently degraded\" here, it is absent: the documented way to do the job does not\nexist on any host. Add the register line beside its siblings and pin it with a\ntest that asserts the entry AND the exact destination AND the premise (that its\ntest still loads it from the repo — that is what makes a green suite\nuninformative). Then run the DEPLOYED copy before reporting, and say which path\nyou exercised.\n\n**The commit-time audit WARNS about a missing register entry — it does not block.** The\naudit prints the exact line to add (`register \"\u003Csource>\" \"\u003Cdestination>\"`) beside\n`NEW file … is not registered`, and the commit still lands, so a feature can ship with a\ngreen suite, a green commit and committed evidence while the tool simply does not exist on\nany other host (the tests load it by REPO path, which is precisely why the suite says\nnothing). Treat that audit line as a gate rather than as noise: add the entry in the same\ncycle it names the file, and after deploying, exercise the DEPLOYED path — not the repo\ncopy — before reporting the tool as available to the fleet.\n\n**Guard register-vs-directory agreement, and verify on the DEPLOYED tree.** The repo's\ntests import the REPO tree while the host runs the DEPLOYED tree, so a module that is\ncommitted, imported by a deployed script, and never registered keeps a GREEN suite while\nthe host dies at startup (observed: `RuntimeError: fleet_bus queue.py not found (bus core\nmissing)` out of `bot_locks`, whose loader probes `~\u002F.steadfaste-fleet\u002Fqueue.py` — so that\nDESTINATION, not just some deployed copy, is the contract). Assert agreement in BOTH\ndirections (an unregistered module; a register entry whose source no longer exists), assert\nthe guard's PREMISE (that the deployed script really does import what it protects), and add\na control that fails on a wrong destination — a guard that cannot fail is\nindistinguishable from one that is not looking. Then exercise the deployed copy, not the\nrepo one, before reporting the fix.\n\n**A register entry in `fleet_gateway\u002F` must be a `.py` module.** The guards compare\n`registered` against `fleet_gateway\u002F*.py` in BOTH directions, so registering a shell\nhelper beside the modules makes the reverse check fire —\n`register names files that do not exist: ['\u003Cx>.sh']` — even though the file is right\nthere, because the module glob only ever yields `.py`. Shell runners and evidence\nscripts in that directory are therefore repo-only, exactly like the existing ones:\nleave them out of the register and run them from the repo (they read the DEPLOYED\ntree, so running them from the repo still tests what the host runs).\n\n**A new register entry changes numbers the suite asserts.** `docs\u002Fdeploy-registry-pattern.md`\nquotes the register-map total, the derivable count and the no-rule count, and a test\nrecomputes all three from `fleet-update.sh`. Adding one line fails that test until the\ndoc is updated — do it in the SAME change, and derive the numbers from the script rather\nthan adjusting the old ones by hand.\n\n## Editing `fleet-update.sh` itself\n\n- **Insert failure-recording code AFTER `local -a _DEFERRED_FAILURES=()`.** The array is\ndeclared mid-function (after the skills sync, not at the top), and the script runs under\n`set -u`: an `_DEFERRED_FAILURES+=(...)` on an earlier line aborts the whole update with an\nunbound-variable error instead of deferring the failure it was meant to report. Place the new\nper-host steps in that same region and DEFER on their failure — a host whose state is wrong\nmust not print a clean update.\n- **A per-host state change belongs in the update path, and must be safe to re-run.** Make the\nmechanism a registered script with `--check`, have the update call it every run, and design it\nto refuse to guess (exit a distinct code for \"an operator decision is required\") and to\npreserve anything it displaces by rename. A one-off manual fix is not a fix: the next update\nrecreates the old state.\n- **A `NEW file … is not registered` warning is a decision, not noise.** Either add the\n  `register \"\u003Csource>\" \"\u003Cdestination>\"` line the audit prints, or state in the script's own\n  header WHY it is checkout-only (a run-scoped evidence emitter that reads the DEPLOYED tree\n  needs no entry). The warning does not block the commit, so an unaddressed one ships a tool\n  that exists on no other host while every test stays green.\n\n## Register at the RIGHT SCOPE — universal vs orchestrator-only\n\n`register()` deploys to every host; `register_orch()` deploys only to orchestrator\nhosts. A file every agent needs but registered orch-only is simply ABSENT on\nnon-orchestrator hosts, and nothing fails loudly — an installer that expects it warns\nand carries on, so the integration is silently partial (a Pi setup registering 3 of 4\nMCP servers, a bus client with no bus server) and the on-host repair is a hand-copy\nthat the next agent repeats.\n\n- **Decide by who consumes it, and read the consumer's contract.** A server whose\ntools are all CLIENT tools, whose consumers declare client access, and that an\ninstaller requires by path belongs on EVERY host. A dashboard, an orchestration loop,\na launchd plist, or something only the orchestrator's own crons call stays orch-only.\n- **Exactly one class per source.** Registering a file with BOTH `register` and\n  `register_orch` is the natural residue of \"promote this to every host\" — the new\n  line added, the old one forgotten. The two entries look authoritative and disagree;\n  delete the replaced entry in the same edit.\n- **Guard the class with a test**: parse both `register` and `register_orch` out of\n  `fleet-update.sh`, then assert (a) every `mcp-servers\u002F*.py` is registered, (b) the\n  ones every agent needs use the plain form, (c) no source appears in two classes.\n  Add CONTROLS that run the guard against a modified COPY of the script — one that\n  flips a universal registration back to orch-only, one that adds a duplicate — and\n  assert the guard fires. A scope guard that cannot fail is indistinguishable from one\n  that is not looking.\n\n## A registered source can be silently GITIGNORED — confirm it is tracked\n\n`git add -A` skips ignored files with NO error and NO warning, so a `register()`\nentry can ship while its source file does not. The deploy still succeeds on YOUR\nhost — it copies from the working tree — while every other host fails with a\nsource-missing warning, which is exactly why the defect is invisible locally. The\nrepo's gitignore carries deliberate secret guards (`*secret*`, `*credential*`,\n`*cred*`, `*token*`, `*.p12`), so ANY artifact whose NAME matches one is affected —\ntests included: a test named after a secret-ish artifact never reaches CI, and its\ngreen local run becomes evidence nobody else can reproduce.\n\n- **Rename the artifact out of the guard's pattern** (preferred). `git add -f` is\n  never the answer: an ignored file stays invisible to every OTHER agent's\n  `git add -A`, so the manifest breaks again on the next artifact and the special case\n  multiplies. A targeted `!` negation is the repo's EXISTING convention for a script\n  that legitimately carries such a name — `.gitignore` already negates the leak\n  detector and the leak watchdog by full path, each with a comment saying why — so if\n  you negate, negate a FULL PATH beside those and never a wildcard; expect a rename to\n  be the smaller thing to explain later.\n- **Confirm with** `git check-ignore -v \u003Cpath>` (prints the .gitignore line that\n  matched) and `git log --all -- \u003Cpath>` — EMPTY output means it was never committed,\n  no matter how many times you ran it locally.\n- **Guard the whole class with a test**, not just your file: parse `^register \"([^\"]+)\"`\n  out of `fleet-update.sh` and require every source to exist AND satisfy\n  `git ls-files --error-unmatch \u003Cpath>`, naming the matched ignore rule in the failure\n  message. Measure the entry count and the current untracked total BEFORE asserting, so\n  the test is not fighting legitimate exceptions; run it against a still-untracked\n  artifact as a control — if a fresh file passes, the test is not inspecting what you\n  think it is.\n- **After `git add -A`, read `git status --short`.** A path you meant to add but that\n  shows no `A`\u002F`M` was skipped silently; a missing error message is not evidence of a\n  staged file.\n\n## Cron manifest: create_cron vs uninstall arrays\n\n- `parse_expected_crons()` (fleet_doctor\u002Fconfig.py) reads **create_cron**\n  blocks, role-filtered (non-orch hosts exclude orch crons). Uninstall arrays\n  are the CLEANUP list, not the expected list.\n- When a cron moves from universal (`install-crons.sh`) to orchestrator-only\n  (`install-orch-crons.sh`), KEEP its name in the universal uninstall array as\n  a legacy-cleanup entry — non-orch hosts still need to uninstall the stale\n  cron, and the doctor is unaffected. Drop orch-prefixed names that never had\n  a universal create_cron.\n- `fix-cron-duplicates.py --fix` aligns uninstall arrays to create_cron\n  EXACTLY and strips legacy-cleanup entries. Do not run `--fix` blindly after\n  adding one; either accept the warning or whitelist uninstall-only names in\n  the tool first.\n- After any manifest change: `bash -n \u003Cinstall script>`, run\n  `fix-cron-duplicates.py` (no `--fix`) to check sync, deploy, run the doctor.\n\n## Cron scope\u002Fprompt change = three surfaces in one pass\n\nChanging a cron's scope, prompt, or schedule touches three places; a partial\nupdate leaves each layer contradicting the others:\n\n1. **Live prompt** — `~\u002F.hermes\u002Fcron\u002Fjobs.json`, updated via the cronjob tool\n   (`action='update'`); installer edits never rewrite an existing job's\n   prompt.\n2. **Installer `create_cron` block** — the durable source for non-`local-`\n   crons.\n3. **Doc tables** — `docs\u002Ffleet-reference.md` AND `docs\u002Fcron-schedules.md`\n   each carry one row per cron describing its scope. A `local-` cron has NO\n   installer block, so its doc row is the repo's only record of the job's\n   scope — commit the row update in the same push.\n\nVerify the live prompt by reading it back out of jobs.json (the update\nresponse shows only a truncated preview), grep both docs for the old scope\nterm to catch sibling rows, and fire one real run (`cronjob action='run'`)\nso the scheduler's last_status reflects the new prompt.\n\n## Never edit the repo while YOUR OWN ship is running\n\n`fleet-update.sh` → `fleet-dogfood.sh` → `git push` can run for several minutes. Editing\nfiles in that window — even \"just a small fix\" — races the ship: it commits (`git add -A`\nseizes your half-written edit), then its `pull --rebase`\u002Fpush meets a tree that changed\nunder it and fails with `error: Please commit or stash them.` or a non-fast-forward. The\nsymptom looks like a remote race, so the cause is easy to misread.\n\n- **Before editing anything, ask the ship — with a probe that cannot match itself.**\n  A plain `pgrep -f \"fleet-update.sh\"` matches the ASKER: the pattern appears in your own\n  command line and in the shell running your wait loop, so it reports \"STILL RUNNING\" for a\n  ship that already exited, and a bounded wait keyed on it spins its full timeout. Use a\n  pattern that cannot self-match — `pgrep -qf \"[c]ortex-update.sh\"` — or check the pid you\n  started with `ps -o pid,stat,args -p \u003Cpid>`, where \"no such pid\" ends the wait at once.\n  Reading is always safe; writing is not.\n- **After any ship, verify the landing**: `pgrep` empty, then `git log --oneline -1\n  origin\u002Fmain` and `git log --oneline origin\u002Fmain..HEAD | wc -l` == 0. A ship that printed\n  `deploy rc=0` can still have failed to push.\n- **A push blocked with `unpushed: N` and a clean tree is already fixed** — just push again;\n  the failure was the race, not the commit.\n\n## Verify the DELIVERY, not the queue\n\nA message that reached a queue is not a message that reached a human, and `depth: 0` after\na send means *something consumed it* — including your own probe. Two false greens observed\non this path in one night: (1) a queue drained to 0 was read as \"delivered\" while the\nconsumer was the test harness calling `poll_replies()` before `drain_outbound()` saw it;\n(2) a provider `\u002Fmodels` endpoint answering 200 was read as \"credential valid\" — that\nendpoint is public and answers 200 for a revoked key (a real chat\u002Fcompletion call is the\ntest).\n\n- Capture the DOWNSTREAM RECEIPT: spy the transport's `_api` and record the\n  `message_id` the platform returns, or read the value back from the system you wrote to.\n- Measure the thing you claim. \"Drained\" needs a consumed-and-archived fact; \"delivered\"\n  needs the receiver's id; \"queued\" needs the queue to still hold it.\n- **Verify on the endpoint the CONSUMER uses, not the one your write preferred.** A client\n  with a primary URL plus a silent fallback can WRITE to a different instance than the\n  reader READS: a stale credential rejected at the primary sends the write to the fallback,\n  the send still returns a `msg_id`, and a consumer pointed at the primary never sees it —\n  a message that vanishes on a success response. Read the value back from the consumer's\n  endpoint (or log which endpoint served the write) before reporting delivery, and when a\n  write and a read disagree about existence, suspect TWO endpoints before suspecting the\n  queue.\n- When a check can't verify (no receipt available), say COULD NOT VERIFY rather than\n  reporting the nearest green signal.\n\n## Peer in-flight work\n\n`git status` before staging. Modified files that aren't yours belong to a\nconcurrent session — leave them unstaged, commit only your files by explicit\npath. The push-gate dogfood diffs the whole tree, so their in-flight edits\ncan trip doc-friendliness heuristics; coordinate rather than cleaning.\n\n## Perpetually-dirty pipeline file + push races\n\n`ops\u002Finstall\u002Fdeploy\u002Fnginx\u002Fblocked_ips.add` is rewritten by the\n`agent-nginx-threat-pipeline` cron on every tick — it is *always* modified at\ncommit time, and the pipeline both stages it and pushes its own commits. So\non the steadfaste-fleet repo:\n\n- Expect this file to block `git pull --rebase` (\"Please commit or stash\nthem\") on nearly every push. Stash just that path by name, rebase, pop.\n- Expect the push itself to be raced: the pipeline can land a new\n  `auto: block N suspect IPs` commit between your `pull --rebase` and your\n  `push`, so the push is rejected as non-fast-forward. `git pull --rebase`\n  again and retry the push — once it shows `A..B  main -> main` it has\n  landed even if a later command in the same chain printed an earlier\n  failure line. Verify with `git log --oneline -1 origin\u002Fmain`.\n\n## Doctor skill-drift: deployed newer than repo = sync INTO the repo\n\nThe doctor warns `Skill drift: \u003Cskill> — Deployed copy is newer than repo\nsource` when accumulated lessons live only in `~\u002F.hermes\u002Fskills\u002F` (Hermes\nskips Hermes-default skills, but deployed skills sourced from the repo drift\nevery time a session adds a lesson locally without committing). Fix direction\nis deployed → repo, NOT repo → deployed (that would erase the lessons):\n\n1. `diff skills\u002F\u003Ccat>\u002F\u003Cskill>\u002FSKILL.md ~\u002F.hermes\u002Fskills\u002F\u003Ccat>\u002F\u003Cskill>\u002FSKILL.md`\n   and read it — confirm the delta is lesson additions, no PII.\n2. Copy the deployed SKILL.md (and any deployed `references\u002F` files the repo\n   copy lacks) into the repo, stage ONLY those paths, commit through the hook.\n3. `fleet-update.sh`, doctor, push. The drift warning clears when deployed\n   matches repo again.\n\nDo not \"fix\" drift by deleting the deployed copy or re-deploying over it —\nboth discard the lessons the doctor is pointing you at.\n\n## Publishing a new skill to the fleet (end-to-end)\n\nSkills reach every agent only through the repo: identify → vet → copy →\nclobber-check → commit → deploy → doctor → push → FLEET_NOTICE.\n\n1. **Identify the skill.** `find ~\u002F.hermes\u002Fskills -name SKILL.md\n   -newermt '\u003CN days>'` and `cmp` each hit against the repo copy; grep\n   `~\u002F.hermes\u002Flogs\u002Fagent.log` for the session that touched it so you publish\n   the skill the user means, not just the newest-looking file.\n2. **Vet before publishing** — skill-vetting checklist: read the FULL\n   SKILL.md (stub body, scripts, injection, exfil, undeclared services),\n   `bash ops\u002Fscripts\u002Fsecret-leak-detector.sh`. Copy LICENSE and\n   `references\u002F` along with the skill, using `cp` — byte-identical is the\n   verification; a hand-typed rewrite can silently drift.\n3. **Clobber check before deploying.** fleet-update's delta engine pushes\n   the REPO copy over the deployed skill, so any lesson that exists only in\n   `~\u002F.hermes\u002Fskills\u002F` — including ones other sessions added the same day,\n   to skills you are not publishing — is erased by the publish deploy. Diff\n   every recently-changed skill deployed vs repo and sync deployed → repo\n   first. fleet-update's `⚠ SKILL DRIFT: deployed copy is newer than repo\n   source` entries name them; each is content to copy in, never `--force`\n   past.\n4. **Commit → deploy → doctor → push.** Pull with `--autostash`; expect the\n   fleet push race (rebase onto new origin\u002Fmain, retry). After a race-rebase\n   your commit gets a new sha, so Deploy sync FAILs again — run\n   `bash ~\u002F.steadfaste-fleet\u002Fscripts\u002Ffleet-dogfood.sh --force`, push, then\n   verify `git ls-remote origin main` == HEAD.\n5. **Verify the deployed tree** — `diff -r ~\u002Fsteadfaste-fleet\u002Fskills\u002F\u003Cname>\n   ~\u002F.hermes\u002Fskills\u002F\u003Cname>` identical — then broadcast FLEET_NOTICE per the\n   fleet-commands skill (self-test on yourself, one message per agent with\n   `--self-tested`, verify with `fleet inbox` peeks).\n\n## Transient doctor FAILs — verify live before treating as host debt\n\n- **`Bus stuck msgs` FAIL can be a sampling artifact.** The doctor probes\n  the LIVE bus over HTTP, and a message sitting between the handler's read\n  and its early archive fails the check once, then clears. Confirm with the\n  live endpoint before fixing handler debt:\n  `curl -s -u \"$FLEET_BASIC_AUTH\" \"$FLEET_BUS_URL\u002Fapi\u002Fpgmq\u002Fqueue\u002Finbox_\u003Cagent>\"`\n  (both vars in `~\u002Fsteadfaste-fleet\u002F.env`) — `depth: 0,\n  processing: 0` = clean, just push again. The path is singular `queue\u002F`;\n  `\u002Fqueues\u002F\u003Cname>` 404s. Don't conclude from local psql — on the backup\n  orchestrator the local mycortex-postgres is a stale REPORTS MIRROR.\n- **A 401\u002F404 from that endpoint is a PROBE error, not proof the bus is\n  unverifiable.** nginx demands Basic auth and ignores a Bearer token, so a bearer\n  request — or the plural `\u002Fqueues\u002F\u003Cname>` path — answers 401\u002F404 while the queue is\n  perfectly readable with `FLEET_BASIC_AUTH`. Retry the documented form before\n  reporting a dispatch as \"sent, unverified\", and if it still fails, name the exact\n  probe that failed instead of implying the queue cannot be checked. Check the probe\n  before doubting the system.\n- **`fleet` is not on PATH** — call it as `~\u002F.steadfaste-fleet\u002Fscripts\u002Ffleet`.\n\n## Reversing your own wrong fix\n\nIf a just-pushed correction turns out to be the wrong direction (e.g. you\nremoved an uninstall-array entry that was serving legacy cleanup), supersede\nit with a new commit in the same session — a known-wrong commit on main\npropagates to every host on their next pull.\n\n**Before re-pushing into a contested area, check whether upstream still contains your\nchange.** Peers work the same defects from the same symptoms, so a fix can be landed\nAND reverted while you are still mid-cycle; `git log --oneline origin\u002Fmain -5` plus\n`git show --stat \u003Csha>` tell you whether a `Revert` has taken your work out, and\nwhether the current upstream state is coherent (source present, manifest pointing at\nit). Push onto a verified base and raise the revert as a decision — do not silently\nre-land a change someone deliberately reversed, and do not re-argue it by loop.\n","\u003C!-- Provenance: renamed from skills\u002Fdevops\u002Fhermes-cortex-repo-deploy\u002FSKILL.md\n     on 2026-11-12 (hc→fleet brand sweep, commit 2dc01b0e). Content predates the\n     range; the range swept the brand tokens only. -->\n\n# Steadfaste Fleet Repo — Deploy & Push Gates\n\nWorkflow for landing changes in `~\u002Fsteadfaste-fleet` (shared public repo). The\nrepo's own AGENTS.md governs governance mechanics; this skill carries the\ndeploy-order and manifest pitfalls that cost time when missed.\n\n## Push sequence (gates run in this order)\n\n1. `git pull --rebase origin main` — REQUIRES a clean tree. ANY unstaged\n   change blocks the rebase AND the push's built-in dogfood — your own edits\n   included, not just a peer's. Stash the specific blocking files by name\n   (never `git stash` bare), rebase\u002Fpush, then `git stash pop` immediately —\n   pop your own files back FIRST so your edits are restored before the next\n   step. Never stash, clean, or commit a peer's work, and never stage a\n   file you did not edit.\n\n   **Never let a rebase run over unpushed commits.** With local commits ahead\n   of origin, or on a conflicted rebase, `pull --rebase` moves HEAD to origin\n   and leaves the repo mid-rebase: the local commits drop out of HEAD and\n   survive only in the reflog, while the working tree reverts to origin. The\n   same hazard lives inside tooling — a pre-push gate that pulls for you can\n   orphan your work the moment it fails. Recover with `git rebase --abort`\n   (restores the pre-rebase HEAD exactly); do not re-author the commits. Before\n   pulling on a working branch, check `git log --oneline origin\u002Fmain..HEAD`: if\n   it lists anything, SKIP the pull — the local work is precisely what you are\n   about to verify and push, so it must not be rebased under.\n2. `git push` runs a doctor gate — it fails on ANY doctor ❌. Most common\n   blocker: `Deploy sync` \u002F `Checksum: \u003Cfile>` FAIL because the deployed copy\n   predates the pushed commit. Fix by running `fleet-update.sh`, re-running\n   the doctor (expect 0 fail), then retrying the push. Never bypass.\n3. After the push lands, run `fleet-update.sh` once more so the deployed tree\n   includes the just-pushed commit (the push gate's internal deploy can trail\n   the newest commit).\n\n## The push also runs an always-review RECEIPT gate\n\n`git push` refuses a range that touches an always-review path unless a CLEAN review receipt\nexists for it (`expected receipt: \u003Cstate>\u002F.reviewed-\u003Cslug>-\u003Ctip>.json`). The receipt is keyed\nby the TIP SHA, and the range spans origin's tip..HEAD — which has consequences that are not\nobvious:\n\n- **A rebase invalidates it.** `git pull --rebase` rewrites SHAs, so a receipt written for\n  the pre-rebase tip no longer matches the gate's expected filename. On a shared branch the\n  rebase rewrites a PEER's unpushed commits too, so their receipt coverage dies with yours\n  (their content survives; say so explicitly when you report it). Getting back requires a\n  cycle that ENDS at the new tip and returns CLEAN from a real review — the receipt is\n  written at close.\n- **`fleet-dogfood.sh --force` cannot produce a receipt.** It only passes `--force-all` to\n  the deploy (a full re-deploy); it runs no review. There is no sanctioned way to re-review\n  a CLOSED cycle either — `record-review.py` is the raw sink the reviewer writes to, and the\n  review-queue sweep never touches overrides. Never manufacture a receipt and never use an\n  override: that gate is what keeps an enforcement change from shipping unreviewed.\n- **The range includes a peer's commits**, so when they touch an always-review file the\n  review must cover the whole range, not just your own commits.\n\n**Run the deploy as the LAST step before the push.** Any commit made after a deploy leaves\nthe deployed marker stale and the gate returns `Deploy sync` ❌ however clean your doctor\nran minutes earlier. Do both in ONE command, nothing in between:\n`bash ~\u002F.steadfaste-fleet\u002Fscripts\u002Ffleet-update.sh > \u003Clog> 2>&1; git push origin main`.\n\n## An operational cycle cannot carry evidence of its own success — commit the CHECK\n\nThe push is LAST, so a transcript of it can only be written afterwards; that commit changes\nthe tip, invalidating the per-tip receipt (above) and leaving the deployed tree trailing so\n`Deploy sync` returns ❌. Reviewers are right to demand committed evidence, and the ordering\nrules keep destroying it — committing more evidence repeats the loop.\n\nBreak it by committing a RE-RUNNABLE CHECK before the push, never a transcript:\n\n    python3 ops\u002Fscripts\u002Fmanage\u002Fverify-landed.py --repo . [--base \u003Crev>] [--no-remote]\n\nIt asserts (1) a CLEAN receipt for the tip, (2) every file the range changed that has a\ndeploy-map entry matches its DEPLOYED copy, (3) `origin\u002F\u003Cbranch> == tip`; it exits 0 \u002F 1 (a\nreal mismatch, named) \u002F 2 COULD NOT VERIFY, and prints no path on failure. A later reviewer\nre-runs it against the state that exists AFTER the push, so the evidence cannot go stale and\nneeds no second commit.\n\n**Measure the checker before trusting it.** On its first real range this check reported three\nfalse FAILs: it hashed deployed copies RAW, and the deploy's header (inserted after the\nshebang, closed by a blank line) makes the two sides differ by construction. When a check\ndisagrees with an otherwise-green gate, print the FIRST DIFFERING LINE of the two bodies\ninstead of assuming either side is wrong, and pin the real shape with a test.\n\n## A failed deploy: read the log's FIRST line, and check the lock after\n\n`fleet-update.sh` sources the host env early, so a broken env file kills the run\nbefore any deploy work and the exit code says nothing specific. Diagnose from the\nlog's first line, not the exit code:\n\n- **`rc=127` whose entire output is `...\u002F.env: line N: \u003Ctoken>: command not found`**\n  — the env file is CORRUPT, not the script. A comment line that lost its leading\n  `#` makes the shell execute the remainder (a bare date fragment is the usual\n  first token). Inspect structure WITHOUT printing values: report each line's key\n  (text before the first `=`) with its length, and list the lines that are neither\n  blank, nor start with `#`, nor contain `=`. Watch for one absurdly long line — a\n  botched append glues a whole blob onto line 1 with no newline, so the header\n  comment swallows kilobytes while every real key still sits on later lines.\n  Repair is surgical and lossless: restore ONLY the damaged leading lines from the\n  newest `.env.bak-*`, leave the rest byte-for-byte, then confirm the key count is\n  unchanged and the mode is still `0600`. A corrupt env breaks EVERY deploy on\n  that host, so treat it as urgent.\n- **A quote\u002FEOF error naming a line BEYOND the file's length is not a static error in that\n  file.** `bash -n \u003Cscript>` passes on it — that pass is not a contradiction, it is the\n  tell: bash numbers lines in whatever unit it is reading at that moment, so the reported\n  line can never be the file's own. The failing text is an eval'd or generated fragment, or\n  a file REWRITTEN while it executes (a deploy that syncs the very script it is running).\n  Do not go editing the script to \"fix\" a syntax error that is not there: capture the run's\n  WHOLE output to a file (never a pipe to `tail`, which destroys the one artifact able to\n  identify the unit) and re-run — a transient rewrite failure does not recur.\n- **`✗ Not a git repository: \u003Cpath>`** — a git WORKTREE has a `.git` file, not a\n  directory, and the script's repo check rejects it. Deploy from the primary\n  working tree (or set `REPO_DIR=`); a session worktree can hold the work and\n  push, but it cannot deploy.\n\n**After ANY failed deploy, check the enforcement files before doing anything\nelse** — a run that dies after unlocking leaves them modifiable. `sudo -n\nhermes-plugin-lock lock` relocks every target in one call, and the doctor's\n`Immutable:` lines are the authoritative confirmation.\n\n## Quoting in commit messages\n\nNever put backticks in a shell-quoted commit message: the shell substitutes them\nbefore git sees the text, mangling the message and leaking fragments onto the\ncommand line. Observed consequence: an empty file literally named `=` was created\nat the repo root, tracked, and flagged as scope drift by the reviewer before I\nnoticed it. Write the message to a FILE and use `git commit -F \u003Cfile>` for any message, not only\none containing code formatting: the command is scanned as a whole, so a policy that\npattern-matches command text can refuse the entire commit over words in the MESSAGE\n(a gateway-lifecycle guard rejecting a message that merely mentions a restart), and\nthe refusal names the command rather than the cause. Then run `git status --short`\nand check for surprise paths before pushing.\n\n## Restoring a file you swapped in: `git checkout --` restores the INDEX\n\nTo reproduce a failure against an older revision you may copy that revision over a working\nfile (`git show \u003Crev>:\u003Cpath> > \u003Cpath>`). Restoring it with `git checkout -- \u003Cpath>` gives\nback the **index** copy, not the working tree — so any edit you made but had not staged is\nsilently discarded, and the command reports success. Observed: a comment fix was reverted\nthis way, and the commit that followed carried only the evidence file, so the reviewer's\nfinding it addressed came straight back on the next close.\n\n- Copy the file aside FIRST (`cp \u003Cpath> \u003Cscratch>\u002Fmine.py`) and restore from that copy, or\n  use `git checkout HEAD -- \u003Cpath>` when HEAD's version is the one you want.\n- Then **grep for the edit** before committing. A clean `git status` proves nothing here:\n  a discarded edit leaves the tree cleaner, not dirtier.\n\n## Never ASSUME a deployed path — resolve it; never let a fallback carry a conclusion\n\nThe deployed path of a repo file CANNOT be derived from the repo path: 192 of ~289\n`register()` entries follow no rule (see the header of `ops\u002Fscripts\u002Ffleet_lib\u002Fpaths.py`).\nA probe that mirrors the repo layout under the deploy home therefore checks NOTHING, and\npaired with `2>\u002Fdev\u002Fnull || echo \"\u003Cconclusion>\"` it prints that failure AS a finding — this\nis how the line \"0 (deployed copy predates this commit)\" reached a committed evidence file\nwhile the deployed copy in fact already carried the change.\n\n- **Ask the map, never guess:** `python3 ops\u002Fscripts\u002Fmanage\u002Fdeployed-path.py \u003Crepo-relative-path>`\n  prints the deployed destination (from `deploy-manifest.tsv`, generated from the register()\n  lines) and exits 0 \u002F 2 (no entry) \u002F 3 (entry present, nothing at the destination). On a\n  failure it prints NO path — there is deliberately no fallback value to mistake for an\n  answer.\n- **A repo copy is NOT the deployed copy.** Check the two separately and say which you ran;\n  a green check on the repo side proves nothing about what a host runs.\n- **Never let a fallback string carry a conclusion.** If a check cannot run, print\n  COULD NOT VERIFY and exit non-zero. `|| echo \"\u003Csomething plausible>\"` converts a broken\n  probe into durable \"evidence\", which is worse than having none.\n\n## Edit the REPO source, never the deployed tree\n\n`fleet-update.sh` copies every `register()` source over its deployed destination,\nand the deployed file says so in its own header (`# SOURCE: \u003Crepo path>` \u002F\n`# Do NOT edit this file`). An edit aimed at a deployed path SUCCEEDS silently and\nlooks right — the file is right there, and every test still passes, because the\ntests import the REPO tree. The next deploy reverts it, since repo → deployed is\nthe only direction a deploy copies.\n\n**On an IMMUTABLE target the write does not even succeed** — the enforcement surface\n(the enforcer plugin, the git hooks, the loop-governance MCP server) carries the\nkernel immutable flag, so a direct write fails with `PermissionError: [Errno 1]\nOperation not permitted`, or `EPERM` from a shell. That is the write-protection layer\nworking, not a permissions bug to escalate: you are editing the wrong side. Verify\nwith `lsattr \u003Cdeployed path>` (expect `----i---------e-------`, see\n`enforcement-immutability`), then put the change in the REPO source and let\n`fleet-update.sh` deploy it — the deploy owns the unlock\u002Frelock around the copy, so\nit is the only sanctioned way to change one. A `cp`\u002F`python` write to a deployed path\nis the same wrong move whichever way it fails, loudly or silently.\n\n- **Patch by REPO path only.** Anything under `~\u002F.steadfaste-fleet\u002F` is an artifact;\n  the editable twin is `~\u002Fsteadfaste-fleet\u002Fops\u002Fscripts\u002F…`. A `patch` call resolves the\n  path it is given, so a deployed path applies cleanly and gives no hint it will be\n  thrown away.\n- **The deploy's drift guard is the backstop, and it fails CLOSED.** When a deployed\n  file differs from its source the run stops with `Copy the deployed changes to the\n  repo source first, then commit` (one line per file) and the doctor reports\n  `Deploy sync` FAIL. That is the guard working: a blind overwrite would destroy\n  whichever side would lose. Resolve it by making the REPO carry the change and\n  deploying again — never by copying the deployed file over the source, which\n  discards the edits already in the source.\n- **Work out which side is ahead before choosing a direction.** The deployed header is NOT a\n  fixed block at the top: the deploy inserts it AFTER the shebang (`# SOURCE:` then\n  `# Do NOT edit this file`), and the block is closed by a BLANK line. So `tail -n +4`\n  misaligns the comparison — it drops the shebang and keeps the blank, and the diff shows a\n  difference that is not there. Remove the header LINES and compare what remains; better\n  still, do not hand-roll the comparison at all —\n  `python3 ops\u002Fscripts\u002Fmanage\u002Fverify-landed.py --repo .` compares every changed\n  deploy-map file against its deployed copy and names any real mismatch.\n  Edits only in the deployed copy must be synced INTO\n  the repo; edits only in the repo just need a deploy.\n- **After deploying, confirm the deployed file actually took the repo's version** —\n  the same `diff` above returning nothing — before reporting the change as live.\n  A deploy that printed `rc=0` also prints drift warnings for files it did NOT sync;\n  read the run's summary, not just its exit code.\n\n## New shared module → register() entry in the SAME commit\n\nA new module that an existing deployed script imports (e.g. a helper under\n`fleet_doctor\u002F`) must get its own `register()` line in `fleet-update.sh`,\nplaced next to its importer's. `register()` copies ONLY listed files — a\nmissing entry makes the next deploy break every importer with\n`ModuleNotFoundError: No module named '\u003Cmodule>'` at the deploy's verification\nstep. Symptom to recognize in the deploy log; fix by adding the entry,\ndeploying, re-running the doctor.\n\nThe same rule covers a script the deployed code loads **by path at runtime**\nrather than importing (`importlib.util.spec_from_file_location`, a subprocess\ncall, a plugin lookup): the loader finds nothing on hosts, so the feature is\nsilently absent there while it works perfectly in the repo. Register both the\nscript AND every file it resolves **relative to itself** — a config beside it\n(`Path(__file__).parent \u002F \"\u003Cname>.yaml\"`) is not deployed by registering the\nscript alone. A loader written to try the repo path first then the deployed\npath will pass every local test and fail only on hosts.\n\nNote `importlib.util.spec_from_file_location` returns an Optional: guard\n`spec is None or spec.loader is None` before `module_from_spec`, or a bad path\nbecomes an AttributeError at the first call instead of a clean skip.\n\n**A CLI ENTRYPOINT is the third shape of the same rule.** A script the fleet is\ntold to RUN from the deployed path (an agent-side tool, a reply primitive, a\nmaintenance command) is imported by nothing, so a missing register entry breaks\nno import and no deploy step — the host simply has no such file, and the whole\nsuite stays green because the tests load it by REPO path. The feature is not\n\"silently degraded\" here, it is absent: the documented way to do the job does not\nexist on any host. Add the register line beside its siblings and pin it with a\ntest that asserts the entry AND the exact destination AND the premise (that its\ntest still loads it from the repo — that is what makes a green suite\nuninformative). Then run the DEPLOYED copy before reporting, and say which path\nyou exercised.\n\n**The commit-time audit WARNS about a missing register entry — it does not block.** The\naudit prints the exact line to add (`register \"\u003Csource>\" \"\u003Cdestination>\"`) beside\n`NEW file … is not registered`, and the commit still lands, so a feature can ship with a\ngreen suite, a green commit and committed evidence while the tool simply does not exist on\nany other host (the tests load it by REPO path, which is precisely why the suite says\nnothing). Treat that audit line as a gate rather than as noise: add the entry in the same\ncycle it names the file, and after deploying, exercise the DEPLOYED path — not the repo\ncopy — before reporting the tool as available to the fleet.\n\n**Guard register-vs-directory agreement, and verify on the DEPLOYED tree.** The repo's\ntests import the REPO tree while the host runs the DEPLOYED tree, so a module that is\ncommitted, imported by a deployed script, and never registered keeps a GREEN suite while\nthe host dies at startup (observed: `RuntimeError: fleet_bus queue.py not found (bus core\nmissing)` out of `bot_locks`, whose loader probes `~\u002F.steadfaste-fleet\u002Fqueue.py` — so that\nDESTINATION, not just some deployed copy, is the contract). Assert agreement in BOTH\ndirections (an unregistered module; a register entry whose source no longer exists), assert\nthe guard's PREMISE (that the deployed script really does import what it protects), and add\na control that fails on a wrong destination — a guard that cannot fail is\nindistinguishable from one that is not looking. Then exercise the deployed copy, not the\nrepo one, before reporting the fix.\n\n**A register entry in `fleet_gateway\u002F` must be a `.py` module.** The guards compare\n`registered` against `fleet_gateway\u002F*.py` in BOTH directions, so registering a shell\nhelper beside the modules makes the reverse check fire —\n`register names files that do not exist: ['\u003Cx>.sh']` — even though the file is right\nthere, because the module glob only ever yields `.py`. Shell runners and evidence\nscripts in that directory are therefore repo-only, exactly like the existing ones:\nleave them out of the register and run them from the repo (they read the DEPLOYED\ntree, so running them from the repo still tests what the host runs).\n\n**A new register entry changes numbers the suite asserts.** `docs\u002Fdeploy-registry-pattern.md`\nquotes the register-map total, the derivable count and the no-rule count, and a test\nrecomputes all three from `fleet-update.sh`. Adding one line fails that test until the\ndoc is updated — do it in the SAME change, and derive the numbers from the script rather\nthan adjusting the old ones by hand.\n\n## Editing `fleet-update.sh` itself\n\n- **Insert failure-recording code AFTER `local -a _DEFERRED_FAILURES=()`.** The array is\ndeclared mid-function (after the skills sync, not at the top), and the script runs under\n`set -u`: an `_DEFERRED_FAILURES+=(...)` on an earlier line aborts the whole update with an\nunbound-variable error instead of deferring the failure it was meant to report. Place the new\nper-host steps in that same region and DEFER on their failure — a host whose state is wrong\nmust not print a clean update.\n- **A per-host state change belongs in the update path, and must be safe to re-run.** Make the\nmechanism a registered script with `--check`, have the update call it every run, and design it\nto refuse to guess (exit a distinct code for \"an operator decision is required\") and to\npreserve anything it displaces by rename. A one-off manual fix is not a fix: the next update\nrecreates the old state.\n- **A `NEW file … is not registered` warning is a decision, not noise.** Either add the\n  `register \"\u003Csource>\" \"\u003Cdestination>\"` line the audit prints, or state in the script's own\n  header WHY it is checkout-only (a run-scoped evidence emitter that reads the DEPLOYED tree\n  needs no entry). The warning does not block the commit, so an unaddressed one ships a tool\n  that exists on no other host while every test stays green.\n\n## Register at the RIGHT SCOPE — universal vs orchestrator-only\n\n`register()` deploys to every host; `register_orch()` deploys only to orchestrator\nhosts. A file every agent needs but registered orch-only is simply ABSENT on\nnon-orchestrator hosts, and nothing fails loudly — an installer that expects it warns\nand carries on, so the integration is silently partial (a Pi setup registering 3 of 4\nMCP servers, a bus client with no bus server) and the on-host repair is a hand-copy\nthat the next agent repeats.\n\n- **Decide by who consumes it, and read the consumer's contract.** A server whose\ntools are all CLIENT tools, whose consumers declare client access, and that an\ninstaller requires by path belongs on EVERY host. A dashboard, an orchestration loop,\na launchd plist, or something only the orchestrator's own crons call stays orch-only.\n- **Exactly one class per source.** Registering a file with BOTH `register` and\n  `register_orch` is the natural residue of \"promote this to every host\" — the new\n  line added, the old one forgotten. The two entries look authoritative and disagree;\n  delete the replaced entry in the same edit.\n- **Guard the class with a test**: parse both `register` and `register_orch` out of\n  `fleet-update.sh`, then assert (a) every `mcp-servers\u002F*.py` is registered, (b) the\n  ones every agent needs use the plain form, (c) no source appears in two classes.\n  Add CONTROLS that run the guard against a modified COPY of the script — one that\n  flips a universal registration back to orch-only, one that adds a duplicate — and\n  assert the guard fires. A scope guard that cannot fail is indistinguishable from one\n  that is not looking.\n\n## A registered source can be silently GITIGNORED — confirm it is tracked\n\n`git add -A` skips ignored files with NO error and NO warning, so a `register()`\nentry can ship while its source file does not. The deploy still succeeds on YOUR\nhost — it copies from the working tree — while every other host fails with a\nsource-missing warning, which is exactly why the defect is invisible locally. The\nrepo's gitignore carries deliberate secret guards (`*secret*`, `*credential*`,\n`*cred*`, `*token*`, `*.p12`), so ANY artifact whose NAME matches one is affected —\ntests included: a test named after a secret-ish artifact never reaches CI, and its\ngreen local run becomes evidence nobody else can reproduce.\n\n- **Rename the artifact out of the guard's pattern** (preferred). `git add -f` is\n  never the answer: an ignored file stays invisible to every OTHER agent's\n  `git add -A`, so the manifest breaks again on the next artifact and the special case\n  multiplies. A targeted `!` negation is the repo's EXISTING convention for a script\n  that legitimately carries such a name — `.gitignore` already negates the leak\n  detector and the leak watchdog by full path, each with a comment saying why — so if\n  you negate, negate a FULL PATH beside those and never a wildcard; expect a rename to\n  be the smaller thing to explain later.\n- **Confirm with** `git check-ignore -v \u003Cpath>` (prints the .gitignore line that\n  matched) and `git log --all -- \u003Cpath>` — EMPTY output means it was never committed,\n  no matter how many times you ran it locally.\n- **Guard the whole class with a test**, not just your file: parse `^register \"([^\"]+)\"`\n  out of `fleet-update.sh` and require every source to exist AND satisfy\n  `git ls-files --error-unmatch \u003Cpath>`, naming the matched ignore rule in the failure\n  message. Measure the entry count and the current untracked total BEFORE asserting, so\n  the test is not fighting legitimate exceptions; run it against a still-untracked\n  artifact as a control — if a fresh file passes, the test is not inspecting what you\n  think it is.\n- **After `git add -A`, read `git status --short`.** A path you meant to add but that\n  shows no `A`\u002F`M` was skipped silently; a missing error message is not evidence of a\n  staged file.\n\n## Cron manifest: create_cron vs uninstall arrays\n\n- `parse_expected_crons()` (fleet_doctor\u002Fconfig.py) reads **create_cron**\n  blocks, role-filtered (non-orch hosts exclude orch crons). Uninstall arrays\n  are the CLEANUP list, not the expected list.\n- When a cron moves from universal (`install-crons.sh`) to orchestrator-only\n  (`install-orch-crons.sh`), KEEP its name in the universal uninstall array as\n  a legacy-cleanup entry — non-orch hosts still need to uninstall the stale\n  cron, and the doctor is unaffected. Drop orch-prefixed names that never had\n  a universal create_cron.\n- `fix-cron-duplicates.py --fix` aligns uninstall arrays to create_cron\n  EXACTLY and strips legacy-cleanup entries. Do not run `--fix` blindly after\n  adding one; either accept the warning or whitelist uninstall-only names in\n  the tool first.\n- After any manifest change: `bash -n \u003Cinstall script>`, run\n  `fix-cron-duplicates.py` (no `--fix`) to check sync, deploy, run the doctor.\n\n## Cron scope\u002Fprompt change = three surfaces in one pass\n\nChanging a cron's scope, prompt, or schedule touches three places; a partial\nupdate leaves each layer contradicting the others:\n\n1. **Live prompt** — `~\u002F.hermes\u002Fcron\u002Fjobs.json`, updated via the cronjob tool\n   (`action='update'`); installer edits never rewrite an existing job's\n   prompt.\n2. **Installer `create_cron` block** — the durable source for non-`local-`\n   crons.\n3. **Doc tables** — `docs\u002Ffleet-reference.md` AND `docs\u002Fcron-schedules.md`\n   each carry one row per cron describing its scope. A `local-` cron has NO\n   installer block, so its doc row is the repo's only record of the job's\n   scope — commit the row update in the same push.\n\nVerify the live prompt by reading it back out of jobs.json (the update\nresponse shows only a truncated preview), grep both docs for the old scope\nterm to catch sibling rows, and fire one real run (`cronjob action='run'`)\nso the scheduler's last_status reflects the new prompt.\n\n## Never edit the repo while YOUR OWN ship is running\n\n`fleet-update.sh` → `fleet-dogfood.sh` → `git push` can run for several minutes. Editing\nfiles in that window — even \"just a small fix\" — races the ship: it commits (`git add -A`\nseizes your half-written edit), then its `pull --rebase`\u002Fpush meets a tree that changed\nunder it and fails with `error: Please commit or stash them.` or a non-fast-forward. The\nsymptom looks like a remote race, so the cause is easy to misread.\n\n- **Before editing anything, ask the ship — with a probe that cannot match itself.**\n  A plain `pgrep -f \"fleet-update.sh\"` matches the ASKER: the pattern appears in your own\n  command line and in the shell running your wait loop, so it reports \"STILL RUNNING\" for a\n  ship that already exited, and a bounded wait keyed on it spins its full timeout. Use a\n  pattern that cannot self-match — `pgrep -qf \"[c]ortex-update.sh\"` — or check the pid you\n  started with `ps -o pid,stat,args -p \u003Cpid>`, where \"no such pid\" ends the wait at once.\n  Reading is always safe; writing is not.\n- **After any ship, verify the landing**: `pgrep` empty, then `git log --oneline -1\n  origin\u002Fmain` and `git log --oneline origin\u002Fmain..HEAD | wc -l` == 0. A ship that printed\n  `deploy rc=0` can still have failed to push.\n- **A push blocked with `unpushed: N` and a clean tree is already fixed** — just push again;\n  the failure was the race, not the commit.\n\n## Verify the DELIVERY, not the queue\n\nA message that reached a queue is not a message that reached a human, and `depth: 0` after\na send means *something consumed it* — including your own probe. Two false greens observed\non this path in one night: (1) a queue drained to 0 was read as \"delivered\" while the\nconsumer was the test harness calling `poll_replies()` before `drain_outbound()` saw it;\n(2) a provider `\u002Fmodels` endpoint answering 200 was read as \"credential valid\" — that\nendpoint is public and answers 200 for a revoked key (a real chat\u002Fcompletion call is the\ntest).\n\n- Capture the DOWNSTREAM RECEIPT: spy the transport's `_api` and record the\n  `message_id` the platform returns, or read the value back from the system you wrote to.\n- Measure the thing you claim. \"Drained\" needs a consumed-and-archived fact; \"delivered\"\n  needs the receiver's id; \"queued\" needs the queue to still hold it.\n- **Verify on the endpoint the CONSUMER uses, not the one your write preferred.** A client\n  with a primary URL plus a silent fallback can WRITE to a different instance than the\n  reader READS: a stale credential rejected at the primary sends the write to the fallback,\n  the send still returns a `msg_id`, and a consumer pointed at the primary never sees it —\n  a message that vanishes on a success response. Read the value back from the consumer's\n  endpoint (or log which endpoint served the write) before reporting delivery, and when a\n  write and a read disagree about existence, suspect TWO endpoints before suspecting the\n  queue.\n- When a check can't verify (no receipt available), say COULD NOT VERIFY rather than\n  reporting the nearest green signal.\n\n## Peer in-flight work\n\n`git status` before staging. Modified files that aren't yours belong to a\nconcurrent session — leave them unstaged, commit only your files by explicit\npath. The push-gate dogfood diffs the whole tree, so their in-flight edits\ncan trip doc-friendliness heuristics; coordinate rather than cleaning.\n\n## Perpetually-dirty pipeline file + push races\n\n`ops\u002Finstall\u002Fdeploy\u002Fnginx\u002Fblocked_ips.add` is rewritten by the\n`agent-nginx-threat-pipeline` cron on every tick — it is *always* modified at\ncommit time, and the pipeline both stages it and pushes its own commits. So\non the steadfaste-fleet repo:\n\n- Expect this file to block `git pull --rebase` (\"Please commit or stash\nthem\") on nearly every push. Stash just that path by name, rebase, pop.\n- Expect the push itself to be raced: the pipeline can land a new\n  `auto: block N suspect IPs` commit between your `pull --rebase` and your\n  `push`, so the push is rejected as non-fast-forward. `git pull --rebase`\n  again and retry the push — once it shows `A..B  main -> main` it has\n  landed even if a later command in the same chain printed an earlier\n  failure line. Verify with `git log --oneline -1 origin\u002Fmain`.\n\n## Doctor skill-drift: deployed newer than repo = sync INTO the repo\n\nThe doctor warns `Skill drift: \u003Cskill> — Deployed copy is newer than repo\nsource` when accumulated lessons live only in `~\u002F.hermes\u002Fskills\u002F` (Hermes\nskips Hermes-default skills, but deployed skills sourced from the repo drift\nevery time a session adds a lesson locally without committing). Fix direction\nis deployed → repo, NOT repo → deployed (that would erase the lessons):\n\n1. `diff skills\u002F\u003Ccat>\u002F\u003Cskill>\u002FSKILL.md ~\u002F.hermes\u002Fskills\u002F\u003Ccat>\u002F\u003Cskill>\u002FSKILL.md`\n   and read it — confirm the delta is lesson additions, no PII.\n2. Copy the deployed SKILL.md (and any deployed `references\u002F` files the repo\n   copy lacks) into the repo, stage ONLY those paths, commit through the hook.\n3. `fleet-update.sh`, doctor, push. The drift warning clears when deployed\n   matches repo again.\n\nDo not \"fix\" drift by deleting the deployed copy or re-deploying over it —\nboth discard the lessons the doctor is pointing you at.\n\n## Publishing a new skill to the fleet (end-to-end)\n\nSkills reach every agent only through the repo: identify → vet → copy →\nclobber-check → commit → deploy → doctor → push → FLEET_NOTICE.\n\n1. **Identify the skill.** `find ~\u002F.hermes\u002Fskills -name SKILL.md\n   -newermt '\u003CN days>'` and `cmp` each hit against the repo copy; grep\n   `~\u002F.hermes\u002Flogs\u002Fagent.log` for the session that touched it so you publish\n   the skill the user means, not just the newest-looking file.\n2. **Vet before publishing** — skill-vetting checklist: read the FULL\n   SKILL.md (stub body, scripts, injection, exfil, undeclared services),\n   `bash ops\u002Fscripts\u002Fsecret-leak-detector.sh`. Copy LICENSE and\n   `references\u002F` along with the skill, using `cp` — byte-identical is the\n   verification; a hand-typed rewrite can silently drift.\n3. **Clobber check before deploying.** fleet-update's delta engine pushes\n   the REPO copy over the deployed skill, so any lesson that exists only in\n   `~\u002F.hermes\u002Fskills\u002F` — including ones other sessions added the same day,\n   to skills you are not publishing — is erased by the publish deploy. Diff\n   every recently-changed skill deployed vs repo and sync deployed → repo\n   first. fleet-update's `⚠ SKILL DRIFT: deployed copy is newer than repo\n   source` entries name them; each is content to copy in, never `--force`\n   past.\n4. **Commit → deploy → doctor → push.** Pull with `--autostash`; expect the\n   fleet push race (rebase onto new origin\u002Fmain, retry). After a race-rebase\n   your commit gets a new sha, so Deploy sync FAILs again — run\n   `bash ~\u002F.steadfaste-fleet\u002Fscripts\u002Ffleet-dogfood.sh --force`, push, then\n   verify `git ls-remote origin main` == HEAD.\n5. **Verify the deployed tree** — `diff -r ~\u002Fsteadfaste-fleet\u002Fskills\u002F\u003Cname>\n   ~\u002F.hermes\u002Fskills\u002F\u003Cname>` identical — then broadcast FLEET_NOTICE per the\n   fleet-commands skill (self-test on yourself, one message per agent with\n   `--self-tested`, verify with `fleet inbox` peeks).\n\n## Transient doctor FAILs — verify live before treating as host debt\n\n- **`Bus stuck msgs` FAIL can be a sampling artifact.** The doctor probes\n  the LIVE bus over HTTP, and a message sitting between the handler's read\n  and its early archive fails the check once, then clears. Confirm with the\n  live endpoint before fixing handler debt:\n  `curl -s -u \"$FLEET_BASIC_AUTH\" \"$FLEET_BUS_URL\u002Fapi\u002Fpgmq\u002Fqueue\u002Finbox_\u003Cagent>\"`\n  (both vars in `~\u002Fsteadfaste-fleet\u002F.env`) — `depth: 0,\n  processing: 0` = clean, just push again. The path is singular `queue\u002F`;\n  `\u002Fqueues\u002F\u003Cname>` 404s. Don't conclude from local psql — on the backup\n  orchestrator the local mycortex-postgres is a stale REPORTS MIRROR.\n- **A 401\u002F404 from that endpoint is a PROBE error, not proof the bus is\n  unverifiable.** nginx demands Basic auth and ignores a Bearer token, so a bearer\n  request — or the plural `\u002Fqueues\u002F\u003Cname>` path — answers 401\u002F404 while the queue is\n  perfectly readable with `FLEET_BASIC_AUTH`. Retry the documented form before\n  reporting a dispatch as \"sent, unverified\", and if it still fails, name the exact\n  probe that failed instead of implying the queue cannot be checked. Check the probe\n  before doubting the system.\n- **`fleet` is not on PATH** — call it as `~\u002F.steadfaste-fleet\u002Fscripts\u002Ffleet`.\n\n## Reversing your own wrong fix\n\nIf a just-pushed correction turns out to be the wrong direction (e.g. you\nremoved an uninstall-array entry that was serving legacy cleanup), supersede\nit with a new commit in the same session — a known-wrong commit on main\npropagates to every host on their next pull.\n\n**Before re-pushing into a contested area, check whether upstream still contains your\nchange.** Peers work the same defects from the same symptoms, so a fix can be landed\nAND reverted while you are still mid-cycle; `git log --oneline origin\u002Fmain -5` plus\n`git show --stat \u003Csha>` tell you whether a `Revert` has taken your work out, and\nwhether the current upstream state is coherent (source present, manifest pointing at\nit). Push onto a verified base and raise the revert as a decision — do not silently\nre-land a change someone deliberately reversed, and do not re-argue it by loop.\n",{"name":5,"version":24,"category":25,"platforms":26,"description":7},"1.0.0","devops",[27,28],"linux","macos","main","9dc0e92c5fa81d4db5292b205716e8ea0c00d10e","skills\u002Fdevops\u002Fsteadfaste-fleet-repo-deploy",1,"https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet","steadfaste.com","https:\u002F\u002Fgithub.com\u002Flukemcqueen","AI Architect | Building Hermes Cortex: governed multi-agent AI for autonomous software engineering | Enterprise Tech Consultant","2026-11-14 22:33:34",{"source":8,"skillName":5,"commands":39,"prompt":49,"zipAvailable":50},[40,43,46],{"tool":41,"command":42},"npx","npx skills add lukemcqueen\u002Fsteadfaste-fleet --skill steadfaste-fleet-repo-deploy",{"tool":44,"command":45},"bunx","bunx skills add lukemcqueen\u002Fsteadfaste-fleet --skill steadfaste-fleet-repo-deploy",{"tool":47,"command":48},"pnpm","pnpm dlx skills add lukemcqueen\u002Fsteadfaste-fleet --skill steadfaste-fleet-repo-deploy","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fsteadfaste-fleet-repo-deploy\u002FSKILL.md",true,[],[53,60,67,74,81,88],{"skillNo":54,"name":55,"nameZh":6,"description":56,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":19},"01M7DXY4G8NSVB59K8B6S2AQNP","macbook-linux-maintenance","Legacy MacBook Linux fixes: GPU wedge, reboot hang, thermal.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fmacbook-linux-maintenance\u002FSKILL.md",[],[],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M7DXXYJCXFJ3TGB84ZQ8KFE6","governance-regression-triage","Triage a host's governance issues: what changed.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fgovernance-regression-triage\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":19},"01M7DXXSAP1G8FDRJAMXQGBBMK","enforcement-immutability","Use when enforcement files lose their immutable flag.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fenforcement-immutability\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M78SS928F3R2RJAYVW4E30CE","steadfaste-fleet-maintenance","Maintain an installed Steadfaste Fleet instance — update both the upstream Hermes Agent and the fleet repo layer, merge diverged histories, sync skills safely, and troubleshoot common issues. Covers the dual-update workflow (hermes update + fleet-update.sh), encrypted brain data backup, mycortex import, and recovery techniques.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Fsteadfaste-fleet-maintenance\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":19},"01M78SP1CHW0T0BWE5ZT6JBPX8","fleet-preflight","DEPRECATED alias — merged into survey-before-action (2026-08-20). Load survey-before-action instead; its \"Repo-Specific Pre-Flight\" section covers git search, Hermes boundary, and deployment verification.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Ffleet-preflight\u002FSKILL.md",[],[],{"skillNo":89,"name":90,"nameZh":6,"description":91,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":92,"categories":93,"keywords":94,"publishedAt":19},"01M78SNFME3PDYCTFD2W0YJQRF","fleet-deployment-sync","Use when pulling latest or running fleet update.","https:\u002F\u002Fgithub.com\u002Flukemcqueen\u002Fsteadfaste-fleet\u002Fblob\u002Fmain\u002Fskills\u002Fdevops\u002Ffleet-deployment-sync\u002FSKILL.md",[],[]]