[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7AP5BFDTDN3F1440G79P4YT":3,"skill-same:01M7AP5BFDTDN3F1440G79P4YT":69},{"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":26,"commitSha":27,"skillDirectory":28,"forksCount":29,"repositoryHtmlUrl":30,"repositoryHomepageUrl":6,"ownerHtmlUrl":31,"ownerBlogUrl":32,"ownerDescription":33,"updatedAt":34,"install":35,"resources":48},"01M7AP5BFDTDN3F1440G79P4YT","ephoros",null,"Decide what a step must emit so its behaviour is visible afterwards: the on-call questions first, then structured events, bounded metrics, correlation across the whole path, and alerts on what someone actually feels. Use when adding logging, metrics, tracing or alerting, when shipping anything that runs unattended, or when an incident could not be explained from what was recorded. Do not use it to instrument a failure you are chasing right now, which belongs to elenchus, and do not use it to measure something slow, which belongs to metron.","wildcat-finance\u002Fskills","wildcat-finance","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F113041915?v=4","01M7AP0V1G8X0SWHP6C5V90RM5","The Wildcat Protocol","01M7AP0V24Y50BGVCQ4DV496VD",13,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fephoros\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: ephoros\ndescription: >-\n  Decide what a step must emit so its behaviour is visible afterwards: the\n  on-call questions first, then structured events, bounded metrics, correlation\n  across the whole path, and alerts on what someone actually feels. Use when\n  adding logging, metrics, tracing or alerting, when shipping anything that\n  runs unattended, or when an incident could not be explained from what was\n  recorded. Do not use it to instrument a failure you are chasing right now,\n  which belongs to elenchus, and do not use it to measure something slow, which\n  belongs to metron.\nmetadata:\n  version: \"2.3.0\"\n---\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"..\u002F..\u002Fassets\u002Fcharacters\u002Fephoros.png\" width=\"1200\">\n\u003C\u002Fp>\n\n# Ephoros\n\nFrom *ephoros*, the overseer whose office was to watch and report. Watching is\nthe whole job. It does not make the system faster or safer, and without it the\nfirst incident becomes archaeology.\n\n## Where this sits\n\nEphoros owns the telemetry that stays: what a step emits, in what shape, and\nwhat wakes someone up.\n\nFiat and Mason apply it while implementing anything that runs unattended, and\nWarden reruns its mechanical gate in non-Solidity rounds. Elenchus may add\ntemporary instrumentation while chasing one failure; Ephoros decides what\nremains afterwards. Metron measures performance, and Phylax guards the data and\nsecrets the telemetry boundary crosses.\n\nA Synkrisis finding may suggest that a signal is missing or late. That\nsuggestion cannot decide what telemetry a step keeps.\n\nIts version, held frontier, next job, and maturity state live in\n[EVOLUTION.md](EVOLUTION.md).\n\n**Current state.** Five rules are executable: E001, E002 and E003 read Python and the TypeScript surface, E004 reads the supported block-YAML subset, and E005 reads Python, supported block-YAML label keys and the TypeScript surface through the shared masked lexer. E002, E003 and E005 run clean over this marketplace and the pinned application clone, which holds no metric label container and no mean over a duration; E001 reports 14 interpolated logger messages over that clone.\n\n## Write the questions before the code\n\nTelemetry with no question behind it is volume. Write two to four questions\nsomeone will ask at three in the morning, then emit only what answers them.\n\n```text\nStep: fee submission service\n1. Did the last cycle submit, and did what it submitted land?\n2. When one fails, is it a revert, a timeout, or a stuck nonce?\n3. Is the subgraph behind the chain, and by how many blocks?\n```\n\nCannot name the questions? Then the step is not ready to instrument. You will\nemit everything and learn nothing.\n\n## One signal per question\n\n| Signal | Answers | Cost |\n| --- | --- | --- |\n| Event | What happened in this one case | Grows with volume |\n| Metric | How often, how slow, in aggregate | Fixed per series |\n| Trace | Where the time went across hops | Sampled |\n\nMetrics say something is wrong, traces say where, events say why.\n\n## Events, not sentences\n\nAn event carries a stable name and fields a query can reach. A sentence with\nvalues baked into it carries neither.\n\n```python\n# unreachable by query, and the values are welded into the string\nlog.info(f\"harvested {n} blocks for {market} in {secs}s\")\n\n# stable name, machine-readable fields\nlog.info(\"harvest_interval_done\", extra={\n    \"event\": \"harvest_interval_done\",\n    \"run_id\": run_id,\n    \"venue\": venue,\n    \"blocks\": n,\n    \"duration_s\": round(secs, 3),\n})\n```\n\nLevels carry meaning, so spend them carefully. `error` means an invariant broke\nand somebody may need to act on it now. `warn` means degraded but handled: a\nretry that worked, a fallback taken, a slow dependency ridden out. `info` marks\nan event the business cares about. `debug` stays off outside development.\n\n## Correlation or it is unreadable\n\nInterleaved output without a shared identifier cannot be reassembled. Mint one\nat each boundary and carry it through every line, span and outbound call.\n\nThe identifier differs by surface. A route takes or generates a request id and\nreturns it in the response header. A harvest mints a run id and stamps every\ninterval under it. A submission carries the nonce and, once broadcast, the\ntransaction hash, because that is the identifier the chain itself will use.\n\n## What never appears\n\n`phylax` sets this rule and this skill inherits it. No key, no mnemonic, no RPC\nor database credential, no session token, no signed payload. An exception\nhandler that prints the request it failed on prints the header it failed with.\n\nOne rule belongs here rather than there. Do not index telemetry by wallet\naddress. Recording who asked about which address, or pairing an address with a\nresolved name in a log the team greps, builds exactly the linkage Probitas\nrefuses to produce. Where an address is genuinely needed to diagnose, it goes\nin an event, never in a metric label or a dashboard axis.\n\n## Metrics, and the label that kills them\n\nInstrument rate, errors and duration on every route and every dependency you\ncall. For pools, queues and workers, watch utilisation, saturation and errors.\n\nEvery distinct label combination is a separate series, so labels come from\nsmall fixed sets: the route template, a status class, a venue name, a chain id.\nNever a wallet address, a transaction hash, a run id, a raw URL, or the text of\nan error. Those belong in events, where high cardinality is the point.\n\nRecord durations as histograms and read them at p95 and p99. An average is the\nnumber that hides the people having the worst time.\n\n## The three surfaces\n\n**A route.** Rate, errors and duration per route template, plus the same for\nPrisma and for every call out to the subgraph or an RPC. A route that fails\nbecause a dependency failed should say which.\n\n**A long-running job.** Progress is the signal, not completion. Emit the\ninterval just finished, how far behind head it is, and when it last advanced. A\nharvester that stops is indistinguishable from a slow one unless staleness is\nmeasured.\n\n**A signer.** The highest stakes surface here, because it moves money. Record\nthe intent before broadcast, the hash on broadcast, and the receipt on\ninclusion, each under the same nonce. Count submitted, landed, reverted and\nstill pending, and record the age of the oldest pending item. Under a process\nmanager, logs are files someone will read later, so the rule above about keys\nis not advice.\n\n## Alert on what someone feels\n\nPage on symptoms, put causes on a dashboard. A cause fires when nothing is\nwrong and stays quiet during failures nobody predicted.\n\n- User-facing error rate above a threshold, sustained. Page.\n- Oldest pending submission older than a stated number of blocks. Page.\n- A harvester that has not advanced in a stated interval. Page.\n- Subgraph lag beyond the point where displayed numbers mislead. Page.\n- Host CPU, memory, or one restarted process. Dashboard.\n\nEvery alert earns its place: it is actionable, it links to three lines saying\nwhat to check first, and its threshold comes from an objective or from history\nrather than from taste. Two severities are enough. A third teaches people to\nignore the first.\n\n## Verify the telemetry\n\nInstrumentation is code and can be wrong. Before receipting, exercise it.\n\nForce an error and find it by its correlation id, confirming the fields\narrived as fields. Send traffic and confirm the series appear with the labels\nyou expect and no others. Follow one request end to end with no gap where a\ncontext was dropped. Fire each new alert once, by moving its threshold, and\nconfirm it arrives where it should.\n\nThen read a sample of real output with `phylax`'s rule in hand, looking for a\ncredential or an address that should not be there.\n\n## The mechanical subset\n\nFive of these rules are settled by a parser. Run the lint over the paths a\nstep touched, and require exit 0.\n\n```bash\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fephoros\u002Fscripts\u002Fephoros.py\" src tests\n```\n\nIt reports a log message assembled by formatting, a metric label drawn from an\nunbounded source, a duration summarised as a mean, and E004 for each supported\nblock-YAML list entry starting with `alert:` that lacks its own nested\n`annotations.runbook` Markdown path. Comments, block scalars, top-level keys\nand neighbouring alert entries do not satisfy E004. The YAML pass establishes\npresence only: Hypomnema H003 resolves the path and H007 checks the target's\nanswers.\n\nE001 also reads `.ts`\u002F`.tsx` through the shared masked lexer, under the same\n1 MiB boundary as E005: a log call whose first argument is an interpolated\ntemplate literal or a concatenation involving a string literal reports a\nmessage built by formatting. A template literal carrying no `${}` is a\nconstant string on this surface and stays clean, where Python's E001 fires\non a placeholder-free f-string; that divergence is decided once, in\n`adr\u002Fkeep-a-constant-template-literal-outside-e001`.\n\nBoth halves read the argument's text rather than its grammar, and both fail\ntoward reporting. Interpolation is any `${` in the first argument once it\nopens with a backtick, escaped or not and inside the literal or after it, so\n``logger.debug(`cost is \\${x}`)`` and ``logger.info(`done` \u002F* ${x} *\u002F)``\neach fire. Concatenation is proximity rather than operands: a `+` outside\nany string or comment and a quote character anywhere in the same first\nargument, so `logger.info(labels['k'] + n)`, `logger.info(fmt(a + \"b\"))` and\n`logger.info(a + b \u002F* it's fine *\u002F)` each fire with no string literal as an\noperand. Neither limit is reached in the pinned clone: none of its 21\nlog-call first arguments carries a `+`, and no `.ts` or `.tsx` file the walk\nreads carries `\\${`; the three generated `storybook-static` bundles that do\nare `.js` files the checker never reads.\n\nE002 also reads `.ts`\u002F`.tsx`, from the same label containers E005 reads: the\narray or object literal after a `labels`, `labelNames`, `label_names`, `tags`\nor `attributes` property inside a metric-named call, and the object argument\nof a `.labels(...)` call. Each key is split into words at a case change and\nat each `_`, the way E005 reads `walletAddress` and `wallet_address` alike,\nand matched against the Python rule's fragments as whole words: `hash`,\n`tx`, `nonce`, `url`, `path`, `email`, `user`, `account`, `session`, `error`,\n`message`, `id` and their compounds and plurals. An address-shaped key\nreports E005 alone, so one key carries one code. Two consequences of that\nvocabulary differ from Python, deliberately. Every plural is a word of its\nown, so `hashes` reports E002 and `addresses` reports E005 here, where the\nPython `s?` suffix passes both; that gap stays open on the Python surface.\nAnd the split reaches camel case, so every key carrying `id` as a word\nreports E002 here, `chainId`, `marketId` and `tokenId` as much as\n`requestId`; Python's `_` anchor reaches `chain_id` and `market_id` but not\n`chainId`, while its `request_?id`, `run_?id` and `trace_?id` fragments\nalready read `requestId`, `runId` and `traceId`. A chain id is a bounded\nlabel by the rule above, and the `id` word claims it on both surfaces; that\nreach is the vocabulary's, not this surface's. The pinned clone holds no\nlabel container, so its E002 count is zero and the fixtures are the whole\nguard.\n\nE003 also reads `.ts`\u002F`.tsx`, from the bracket table the other TypeScript\nrules already walk rather than from a scan of its own: a `mean`, `average`,\n`avg` or `fmean` call, or a `reduce(...)` call followed by `\u002F` and a `.length`\nchain, standing as the whole right-hand side of a declaration or assignment,\na bare or dotted name with a plain `=` before it and at most a simple or\ndotted type annotation between. The duration vocabulary is the Python rule's\nfragments as whole words after the same case and `_` split: `duration`,\n`latency`, `elapsed`, `seconds`, `secs`, `millis`, `ms`, `runtime`,\n`responsetime`, `took`, `wait`, `time` and their plurals, read from the\ntarget name and from every identifier and string literal inside the\nexpression, so `avgWait`, `stats.mean(requestDurationsMs)` and\n`waits.length` each carry one. `latencies` fires here where the Python `s?`\nsuffix passes it; `timeout` and `timestamp` carry no duration word on either\nsurface. The rule is lexical and follows no dataflow, on either surface: a\nduration that reaches a mean through an innocently named variable passes,\n`const summary = mean(samples)` reports nothing however the samples were\nmeasured, and the rule does not widen to close that gap, because the\nvariable's name is the only evidence the parser has. Four shapes stay\noutside it, as they do in Python: a mean used as an operand, `mean(latencies)\n* 1000` or `Math.round(mean(latencies))`; an object property, `{ meanLatency:\nmean(latencies) }`; a `reduce` reached through another call, `ds.map(toMs)\n.reduce(...) \u002F ds.length`; and a `sum` helper divided by a length. A comment\ninside the expression carries no word, because the word table is built from\nthe code and string spans alone, so `mean(xs \u002F* latency *\u002F)` stays clean. The\npinned clone holds no mean call and no `reduce` divided by a length, so its\nE003 count is zero and the fixtures are the whole guard.\n\nE005 reports telemetry keyed by wallet address: an address-shaped name or a\n40-hex literal used as a metric label, a dashboard key or a log index. A\nsubscript on a logger-named receiver is a log index only where the code writes\nit: assigned, deleted, incremented, or the receiver of an element mutator such\nas `append` or `push`. A read passes, so `log[\"address\"]` taken from a decoded\non-chain log is not telemetry. E005 reads Python, address-named keys directly under a supported block-YAML\n`labels:` mapping, and `.ts`\u002F`.tsx` source through the shared masked lexer\nphylax already uses, with the same input boundary: at most 1 MiB per\nTypeScript file, and a file the lexer cannot read or terminate reports E000\nand fails the run. Address-shaped metric labels that E002 used to claim now\nreport E005, so one concern carries one code; every other unbounded fragment\nkeeps E002. Where the line between this lint and phylax runs over the same\nTypeScript files is decided once, in\n[ADR-010](..\u002F..\u002F..\u002F..\u002Fdocs\u002Fdecisions\u002FADR-010-split-address-telemetry-from-boundary-control.md).\n\nFour limits are part of the rule rather than defects in it. The finding\nmessage says wallet address for any address-fragment key, so `ip_address`\ndraws the same words. Recognition under a YAML `labels:` mapping is\ndirect-children-only, so a key nested one mapping deeper passes silently. On\nthe Python and block-YAML surfaces the `s?` suffix family E005 shares\nwith E002 misses `-es` plurals, so `addresses` passes where `address` fires;\nthe TypeScript word set lists `addresses` and fires on it. And a write to a\nfield of a decoded record named like a logger, `log[\"address\"] = x`, still\nreports E005, because the name cannot tell that record from a log store keyed\nby address; it takes a reasoned pragma.\n\nTwo things it deliberately leaves alone. A `print` in Python and `console.*`\nin TypeScript are command-line output rather than telemetry, and this\nmarketplace writes a great deal of the first. A mean of something that is not\na duration is arithmetic, so sentence lengths, layout positions and prices\npass untouched.\n\nDeliberate exceptions state a reason: `# ephoros: allow \u003Cwhy>` in Python and\nYAML, `\u002F\u002F ephoros: allow \u003Cwhy>` as a genuine line comment in TypeScript, on\nthe finding line or the one above it. A bare pragma suppresses nothing, in\neither form. Everything else here stays judgement, and a clean exit says only\nthat these five found nothing.\n\n## Rationalisations\n\n- \"Logging goes in after it works.\" After means after the first incident,\n  which is the most expensive moment to discover you are blind.\n- \"More logs, more visibility.\" Unstructured volume slows an incident down.\n  Three queryable events beat three hundred lines of prose.\n- \"A print statement is fine for now.\" It cannot be filtered, correlated or\n  alerted on, and it outlives the afternoon that added it.\n- \"We will look at the dashboard when it breaks.\" A dashboard built without\n  questions shows everything except the answer.\n- \"Alert on it all and tune later.\" The tuning does not happen. The ignored\n  page does.\n- \"Address as a label makes debugging easier.\" It also multiplies the series\n  without bound and builds the linkage the marketplace refuses.\n- \"Tracing is overkill for two services.\" Two services is already a question\n  about which one spent the time.\n\n## Red flags\n\n- A step that adds retries, a queue or an external call and emits nothing new.\n- Log lines assembled by formatting instead of fields.\n- No correlation identifier, so every line stands alone.\n- Metric labels carrying addresses, hashes, run ids or error text.\n- Latency reported as an average.\n- Alerts that fire daily and get acknowledged without action.\n- Pages on host metrics while user-facing errors go unwatched.\n- Credentials, signed payloads or wallet addresses sitting in output.\n- Long-running jobs with no staleness signal.\n\n## Before the step is receipted\n\nReport the count, then name every item that failed.\n\n- [ ] The on-call questions are written down, and each signal answers one.\n- [ ] Output is structured, with stable event names and a correlation id.\n- [ ] Rate, errors and duration exist for each new route and dependency.\n- [ ] Label sets are bounded, with no address, hash or free text among them.\n- [ ] Durations are histograms, and p95 and p99 are queryable.\n- [ ] Long-running jobs report progress and staleness, not just completion.\n- [ ] A signer records intent, broadcast and receipt under one nonce.\n- [ ] Each new alert is symptom-based, carries a runbook link, and was fired once.\n- [ ] Sampled real output contains no credential and no address used as a key.\n\n## Hand back\n\nLead with the state: instrumented against the questions you wrote, or open on a\nnamed gap. List the questions and the signal answering each.\n\nSeparate what you verified from what you added. A signal you triggered and then\nfound in the output is established. One that merely exists in the diff is\nasserted, and saying which is which costs a sentence.\n\nEnd with one action: the alert still needing a threshold, the question with no\nsignal behind it, or the sampled output someone should read.\n\n## Promise Machine contract\n\n### ephoros-mechanical-gate\n\n- Promise: A zero-exit Ephoros lint establishes that the bounded parser found none of its specified formatted-log, unbounded-metric-label or mean-duration patterns in the selected Python and TypeScript paths, no supported block-YAML alert entry without its own nested runbook annotation, and no wallet-address key on a metric label, dashboard key or log index in the selected Python, TypeScript and supported block-YAML label surfaces.\n- Evidence: The exact lint version, arguments, selected paths, structured findings and zero exit status.\n- Evidence classes: checked\n- Boundary: A clean lint covers only the four implemented Python rules, E004 annotation presence and E005 address-key recognition in the supported block-YAML subset, and E001, E002, E003 and E005 on the TypeScript surface in the lexical shapes stated above; it does not prove useful observability, safe output, correct alerting, a resolving or useful runbook, general YAML semantics or any other rule in another language.\n- Authorises: Passing the mechanical Ephoros gate for the exact paths and checker version recorded.\n- Consequence: 1\n- Refuses: Unreadable or oversized input, an unexplained suppression, a non-zero result or any broader observability claim.\n- Recovery: Repair the offending signal or add a narrowly reasoned suppression when the rule is inapplicable, then rerun the same bounded lint.\n- Exceptions: none\n\n### ephoros-observability-review\n\n- Promise: A completed observability review establishes that the step's on-call questions have bounded signals, correlation and user-symptom alerts, and that emitted samples were checked for sensitive or unbounded data.\n- Evidence: The question-to-signal map, event and metric definitions, trace correlation, alert test, sampled output, staleness signal for long jobs and unresolved-gap list.\n- Evidence classes: checked, inferred, recorded\n- Boundary: The review covers the named step and exercised signals; it does not prove future telemetry availability, alert thresholds under every workload or absence of sensitive data outside the samples reviewed.\n- Authorises: Operating the reviewed step with the recorded signals and escalating through the named runbooks.\n- Consequence: 2\n- Refuses: A new unattended path with no signal, unbounded labels, mean-only duration, missing correlation, untested alerts, absent runbooks or sensitive values in emitted output.\n- Recovery: Write the unanswered on-call question, add or bound the missing signal, exercise the alert, inspect real output and repeat the review.\n- Exceptions: none\n","\u003Cp align=\"center\">\n  \u003Cimg src=\"..\u002F..\u002Fassets\u002Fcharacters\u002Fephoros.png\" width=\"1200\">\n\u003C\u002Fp>\n\n# Ephoros\n\nFrom *ephoros*, the overseer whose office was to watch and report. Watching is\nthe whole job. It does not make the system faster or safer, and without it the\nfirst incident becomes archaeology.\n\n## Where this sits\n\nEphoros owns the telemetry that stays: what a step emits, in what shape, and\nwhat wakes someone up.\n\nFiat and Mason apply it while implementing anything that runs unattended, and\nWarden reruns its mechanical gate in non-Solidity rounds. Elenchus may add\ntemporary instrumentation while chasing one failure; Ephoros decides what\nremains afterwards. Metron measures performance, and Phylax guards the data and\nsecrets the telemetry boundary crosses.\n\nA Synkrisis finding may suggest that a signal is missing or late. That\nsuggestion cannot decide what telemetry a step keeps.\n\nIts version, held frontier, next job, and maturity state live in\n[EVOLUTION.md](EVOLUTION.md).\n\n**Current state.** Five rules are executable: E001, E002 and E003 read Python and the TypeScript surface, E004 reads the supported block-YAML subset, and E005 reads Python, supported block-YAML label keys and the TypeScript surface through the shared masked lexer. E002, E003 and E005 run clean over this marketplace and the pinned application clone, which holds no metric label container and no mean over a duration; E001 reports 14 interpolated logger messages over that clone.\n\n## Write the questions before the code\n\nTelemetry with no question behind it is volume. Write two to four questions\nsomeone will ask at three in the morning, then emit only what answers them.\n\n```text\nStep: fee submission service\n1. Did the last cycle submit, and did what it submitted land?\n2. When one fails, is it a revert, a timeout, or a stuck nonce?\n3. Is the subgraph behind the chain, and by how many blocks?\n```\n\nCannot name the questions? Then the step is not ready to instrument. You will\nemit everything and learn nothing.\n\n## One signal per question\n\n| Signal | Answers | Cost |\n| --- | --- | --- |\n| Event | What happened in this one case | Grows with volume |\n| Metric | How often, how slow, in aggregate | Fixed per series |\n| Trace | Where the time went across hops | Sampled |\n\nMetrics say something is wrong, traces say where, events say why.\n\n## Events, not sentences\n\nAn event carries a stable name and fields a query can reach. A sentence with\nvalues baked into it carries neither.\n\n```python\n# unreachable by query, and the values are welded into the string\nlog.info(f\"harvested {n} blocks for {market} in {secs}s\")\n\n# stable name, machine-readable fields\nlog.info(\"harvest_interval_done\", extra={\n    \"event\": \"harvest_interval_done\",\n    \"run_id\": run_id,\n    \"venue\": venue,\n    \"blocks\": n,\n    \"duration_s\": round(secs, 3),\n})\n```\n\nLevels carry meaning, so spend them carefully. `error` means an invariant broke\nand somebody may need to act on it now. `warn` means degraded but handled: a\nretry that worked, a fallback taken, a slow dependency ridden out. `info` marks\nan event the business cares about. `debug` stays off outside development.\n\n## Correlation or it is unreadable\n\nInterleaved output without a shared identifier cannot be reassembled. Mint one\nat each boundary and carry it through every line, span and outbound call.\n\nThe identifier differs by surface. A route takes or generates a request id and\nreturns it in the response header. A harvest mints a run id and stamps every\ninterval under it. A submission carries the nonce and, once broadcast, the\ntransaction hash, because that is the identifier the chain itself will use.\n\n## What never appears\n\n`phylax` sets this rule and this skill inherits it. No key, no mnemonic, no RPC\nor database credential, no session token, no signed payload. An exception\nhandler that prints the request it failed on prints the header it failed with.\n\nOne rule belongs here rather than there. Do not index telemetry by wallet\naddress. Recording who asked about which address, or pairing an address with a\nresolved name in a log the team greps, builds exactly the linkage Probitas\nrefuses to produce. Where an address is genuinely needed to diagnose, it goes\nin an event, never in a metric label or a dashboard axis.\n\n## Metrics, and the label that kills them\n\nInstrument rate, errors and duration on every route and every dependency you\ncall. For pools, queues and workers, watch utilisation, saturation and errors.\n\nEvery distinct label combination is a separate series, so labels come from\nsmall fixed sets: the route template, a status class, a venue name, a chain id.\nNever a wallet address, a transaction hash, a run id, a raw URL, or the text of\nan error. Those belong in events, where high cardinality is the point.\n\nRecord durations as histograms and read them at p95 and p99. An average is the\nnumber that hides the people having the worst time.\n\n## The three surfaces\n\n**A route.** Rate, errors and duration per route template, plus the same for\nPrisma and for every call out to the subgraph or an RPC. A route that fails\nbecause a dependency failed should say which.\n\n**A long-running job.** Progress is the signal, not completion. Emit the\ninterval just finished, how far behind head it is, and when it last advanced. A\nharvester that stops is indistinguishable from a slow one unless staleness is\nmeasured.\n\n**A signer.** The highest stakes surface here, because it moves money. Record\nthe intent before broadcast, the hash on broadcast, and the receipt on\ninclusion, each under the same nonce. Count submitted, landed, reverted and\nstill pending, and record the age of the oldest pending item. Under a process\nmanager, logs are files someone will read later, so the rule above about keys\nis not advice.\n\n## Alert on what someone feels\n\nPage on symptoms, put causes on a dashboard. A cause fires when nothing is\nwrong and stays quiet during failures nobody predicted.\n\n- User-facing error rate above a threshold, sustained. Page.\n- Oldest pending submission older than a stated number of blocks. Page.\n- A harvester that has not advanced in a stated interval. Page.\n- Subgraph lag beyond the point where displayed numbers mislead. Page.\n- Host CPU, memory, or one restarted process. Dashboard.\n\nEvery alert earns its place: it is actionable, it links to three lines saying\nwhat to check first, and its threshold comes from an objective or from history\nrather than from taste. Two severities are enough. A third teaches people to\nignore the first.\n\n## Verify the telemetry\n\nInstrumentation is code and can be wrong. Before receipting, exercise it.\n\nForce an error and find it by its correlation id, confirming the fields\narrived as fields. Send traffic and confirm the series appear with the labels\nyou expect and no others. Follow one request end to end with no gap where a\ncontext was dropped. Fire each new alert once, by moving its threshold, and\nconfirm it arrives where it should.\n\nThen read a sample of real output with `phylax`'s rule in hand, looking for a\ncredential or an address that should not be there.\n\n## The mechanical subset\n\nFive of these rules are settled by a parser. Run the lint over the paths a\nstep touched, and require exit 0.\n\n```bash\npython3 \"$PLUGIN_ROOT\u002Fskills\u002Fephoros\u002Fscripts\u002Fephoros.py\" src tests\n```\n\nIt reports a log message assembled by formatting, a metric label drawn from an\nunbounded source, a duration summarised as a mean, and E004 for each supported\nblock-YAML list entry starting with `alert:` that lacks its own nested\n`annotations.runbook` Markdown path. Comments, block scalars, top-level keys\nand neighbouring alert entries do not satisfy E004. The YAML pass establishes\npresence only: Hypomnema H003 resolves the path and H007 checks the target's\nanswers.\n\nE001 also reads `.ts`\u002F`.tsx` through the shared masked lexer, under the same\n1 MiB boundary as E005: a log call whose first argument is an interpolated\ntemplate literal or a concatenation involving a string literal reports a\nmessage built by formatting. A template literal carrying no `${}` is a\nconstant string on this surface and stays clean, where Python's E001 fires\non a placeholder-free f-string; that divergence is decided once, in\n`adr\u002Fkeep-a-constant-template-literal-outside-e001`.\n\nBoth halves read the argument's text rather than its grammar, and both fail\ntoward reporting. Interpolation is any `${` in the first argument once it\nopens with a backtick, escaped or not and inside the literal or after it, so\n``logger.debug(`cost is \\${x}`)`` and ``logger.info(`done` \u002F* ${x} *\u002F)``\neach fire. Concatenation is proximity rather than operands: a `+` outside\nany string or comment and a quote character anywhere in the same first\nargument, so `logger.info(labels['k'] + n)`, `logger.info(fmt(a + \"b\"))` and\n`logger.info(a + b \u002F* it's fine *\u002F)` each fire with no string literal as an\noperand. Neither limit is reached in the pinned clone: none of its 21\nlog-call first arguments carries a `+`, and no `.ts` or `.tsx` file the walk\nreads carries `\\${`; the three generated `storybook-static` bundles that do\nare `.js` files the checker never reads.\n\nE002 also reads `.ts`\u002F`.tsx`, from the same label containers E005 reads: the\narray or object literal after a `labels`, `labelNames`, `label_names`, `tags`\nor `attributes` property inside a metric-named call, and the object argument\nof a `.labels(...)` call. Each key is split into words at a case change and\nat each `_`, the way E005 reads `walletAddress` and `wallet_address` alike,\nand matched against the Python rule's fragments as whole words: `hash`,\n`tx`, `nonce`, `url`, `path`, `email`, `user`, `account`, `session`, `error`,\n`message`, `id` and their compounds and plurals. An address-shaped key\nreports E005 alone, so one key carries one code. Two consequences of that\nvocabulary differ from Python, deliberately. Every plural is a word of its\nown, so `hashes` reports E002 and `addresses` reports E005 here, where the\nPython `s?` suffix passes both; that gap stays open on the Python surface.\nAnd the split reaches camel case, so every key carrying `id` as a word\nreports E002 here, `chainId`, `marketId` and `tokenId` as much as\n`requestId`; Python's `_` anchor reaches `chain_id` and `market_id` but not\n`chainId`, while its `request_?id`, `run_?id` and `trace_?id` fragments\nalready read `requestId`, `runId` and `traceId`. A chain id is a bounded\nlabel by the rule above, and the `id` word claims it on both surfaces; that\nreach is the vocabulary's, not this surface's. The pinned clone holds no\nlabel container, so its E002 count is zero and the fixtures are the whole\nguard.\n\nE003 also reads `.ts`\u002F`.tsx`, from the bracket table the other TypeScript\nrules already walk rather than from a scan of its own: a `mean`, `average`,\n`avg` or `fmean` call, or a `reduce(...)` call followed by `\u002F` and a `.length`\nchain, standing as the whole right-hand side of a declaration or assignment,\na bare or dotted name with a plain `=` before it and at most a simple or\ndotted type annotation between. The duration vocabulary is the Python rule's\nfragments as whole words after the same case and `_` split: `duration`,\n`latency`, `elapsed`, `seconds`, `secs`, `millis`, `ms`, `runtime`,\n`responsetime`, `took`, `wait`, `time` and their plurals, read from the\ntarget name and from every identifier and string literal inside the\nexpression, so `avgWait`, `stats.mean(requestDurationsMs)` and\n`waits.length` each carry one. `latencies` fires here where the Python `s?`\nsuffix passes it; `timeout` and `timestamp` carry no duration word on either\nsurface. The rule is lexical and follows no dataflow, on either surface: a\nduration that reaches a mean through an innocently named variable passes,\n`const summary = mean(samples)` reports nothing however the samples were\nmeasured, and the rule does not widen to close that gap, because the\nvariable's name is the only evidence the parser has. Four shapes stay\noutside it, as they do in Python: a mean used as an operand, `mean(latencies)\n* 1000` or `Math.round(mean(latencies))`; an object property, `{ meanLatency:\nmean(latencies) }`; a `reduce` reached through another call, `ds.map(toMs)\n.reduce(...) \u002F ds.length`; and a `sum` helper divided by a length. A comment\ninside the expression carries no word, because the word table is built from\nthe code and string spans alone, so `mean(xs \u002F* latency *\u002F)` stays clean. The\npinned clone holds no mean call and no `reduce` divided by a length, so its\nE003 count is zero and the fixtures are the whole guard.\n\nE005 reports telemetry keyed by wallet address: an address-shaped name or a\n40-hex literal used as a metric label, a dashboard key or a log index. A\nsubscript on a logger-named receiver is a log index only where the code writes\nit: assigned, deleted, incremented, or the receiver of an element mutator such\nas `append` or `push`. A read passes, so `log[\"address\"]` taken from a decoded\non-chain log is not telemetry. E005 reads Python, address-named keys directly under a supported block-YAML\n`labels:` mapping, and `.ts`\u002F`.tsx` source through the shared masked lexer\nphylax already uses, with the same input boundary: at most 1 MiB per\nTypeScript file, and a file the lexer cannot read or terminate reports E000\nand fails the run. Address-shaped metric labels that E002 used to claim now\nreport E005, so one concern carries one code; every other unbounded fragment\nkeeps E002. Where the line between this lint and phylax runs over the same\nTypeScript files is decided once, in\n[ADR-010](..\u002F..\u002F..\u002F..\u002Fdocs\u002Fdecisions\u002FADR-010-split-address-telemetry-from-boundary-control.md).\n\nFour limits are part of the rule rather than defects in it. The finding\nmessage says wallet address for any address-fragment key, so `ip_address`\ndraws the same words. Recognition under a YAML `labels:` mapping is\ndirect-children-only, so a key nested one mapping deeper passes silently. On\nthe Python and block-YAML surfaces the `s?` suffix family E005 shares\nwith E002 misses `-es` plurals, so `addresses` passes where `address` fires;\nthe TypeScript word set lists `addresses` and fires on it. And a write to a\nfield of a decoded record named like a logger, `log[\"address\"] = x`, still\nreports E005, because the name cannot tell that record from a log store keyed\nby address; it takes a reasoned pragma.\n\nTwo things it deliberately leaves alone. A `print` in Python and `console.*`\nin TypeScript are command-line output rather than telemetry, and this\nmarketplace writes a great deal of the first. A mean of something that is not\na duration is arithmetic, so sentence lengths, layout positions and prices\npass untouched.\n\nDeliberate exceptions state a reason: `# ephoros: allow \u003Cwhy>` in Python and\nYAML, `\u002F\u002F ephoros: allow \u003Cwhy>` as a genuine line comment in TypeScript, on\nthe finding line or the one above it. A bare pragma suppresses nothing, in\neither form. Everything else here stays judgement, and a clean exit says only\nthat these five found nothing.\n\n## Rationalisations\n\n- \"Logging goes in after it works.\" After means after the first incident,\n  which is the most expensive moment to discover you are blind.\n- \"More logs, more visibility.\" Unstructured volume slows an incident down.\n  Three queryable events beat three hundred lines of prose.\n- \"A print statement is fine for now.\" It cannot be filtered, correlated or\n  alerted on, and it outlives the afternoon that added it.\n- \"We will look at the dashboard when it breaks.\" A dashboard built without\n  questions shows everything except the answer.\n- \"Alert on it all and tune later.\" The tuning does not happen. The ignored\n  page does.\n- \"Address as a label makes debugging easier.\" It also multiplies the series\n  without bound and builds the linkage the marketplace refuses.\n- \"Tracing is overkill for two services.\" Two services is already a question\n  about which one spent the time.\n\n## Red flags\n\n- A step that adds retries, a queue or an external call and emits nothing new.\n- Log lines assembled by formatting instead of fields.\n- No correlation identifier, so every line stands alone.\n- Metric labels carrying addresses, hashes, run ids or error text.\n- Latency reported as an average.\n- Alerts that fire daily and get acknowledged without action.\n- Pages on host metrics while user-facing errors go unwatched.\n- Credentials, signed payloads or wallet addresses sitting in output.\n- Long-running jobs with no staleness signal.\n\n## Before the step is receipted\n\nReport the count, then name every item that failed.\n\n- [ ] The on-call questions are written down, and each signal answers one.\n- [ ] Output is structured, with stable event names and a correlation id.\n- [ ] Rate, errors and duration exist for each new route and dependency.\n- [ ] Label sets are bounded, with no address, hash or free text among them.\n- [ ] Durations are histograms, and p95 and p99 are queryable.\n- [ ] Long-running jobs report progress and staleness, not just completion.\n- [ ] A signer records intent, broadcast and receipt under one nonce.\n- [ ] Each new alert is symptom-based, carries a runbook link, and was fired once.\n- [ ] Sampled real output contains no credential and no address used as a key.\n\n## Hand back\n\nLead with the state: instrumented against the questions you wrote, or open on a\nnamed gap. List the questions and the signal answering each.\n\nSeparate what you verified from what you added. A signal you triggered and then\nfound in the output is established. One that merely exists in the diff is\nasserted, and saying which is which costs a sentence.\n\nEnd with one action: the alert still needing a threshold, the question with no\nsignal behind it, or the sampled output someone should read.\n\n## Promise Machine contract\n\n### ephoros-mechanical-gate\n\n- Promise: A zero-exit Ephoros lint establishes that the bounded parser found none of its specified formatted-log, unbounded-metric-label or mean-duration patterns in the selected Python and TypeScript paths, no supported block-YAML alert entry without its own nested runbook annotation, and no wallet-address key on a metric label, dashboard key or log index in the selected Python, TypeScript and supported block-YAML label surfaces.\n- Evidence: The exact lint version, arguments, selected paths, structured findings and zero exit status.\n- Evidence classes: checked\n- Boundary: A clean lint covers only the four implemented Python rules, E004 annotation presence and E005 address-key recognition in the supported block-YAML subset, and E001, E002, E003 and E005 on the TypeScript surface in the lexical shapes stated above; it does not prove useful observability, safe output, correct alerting, a resolving or useful runbook, general YAML semantics or any other rule in another language.\n- Authorises: Passing the mechanical Ephoros gate for the exact paths and checker version recorded.\n- Consequence: 1\n- Refuses: Unreadable or oversized input, an unexplained suppression, a non-zero result or any broader observability claim.\n- Recovery: Repair the offending signal or add a narrowly reasoned suppression when the rule is inapplicable, then rerun the same bounded lint.\n- Exceptions: none\n\n### ephoros-observability-review\n\n- Promise: A completed observability review establishes that the step's on-call questions have bounded signals, correlation and user-symptom alerts, and that emitted samples were checked for sensitive or unbounded data.\n- Evidence: The question-to-signal map, event and metric definitions, trace correlation, alert test, sampled output, staleness signal for long jobs and unresolved-gap list.\n- Evidence classes: checked, inferred, recorded\n- Boundary: The review covers the named step and exercised signals; it does not prove future telemetry availability, alert thresholds under every workload or absence of sensitive data outside the samples reviewed.\n- Authorises: Operating the reviewed step with the recorded signals and escalating through the named runbooks.\n- Consequence: 2\n- Refuses: A new unattended path with no signal, unbounded labels, mean-only duration, missing correlation, untested alerts, absent runbooks or sensitive values in emitted output.\n- Recovery: Write the unanswered on-call question, add or bound the missing signal, exercise the alert, inspect real output and repeat the review.\n- Exceptions: none\n",{"name":5,"metadata":24,"description":7},{"version":25},"2.3.0","main","6b6406ecd922c4127bc688e0b5122bbe395be0eb","plugins\u002Fhexaemeron\u002Fskills\u002Fephoros",6,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills","https:\u002F\u002Fgithub.com\u002Fwildcat-finance","https:\u002F\u002Fwildcat.finance","Banking, But Worse","2026-11-07 17:14:07",{"source":8,"skillName":5,"commands":36,"prompt":46,"zipAvailable":47},[37,40,43],{"tool":38,"command":39},"npx","npx skills add wildcat-finance\u002Fskills --skill ephoros",{"tool":41,"command":42},"bunx","bunx skills add wildcat-finance\u002Fskills --skill ephoros",{"tool":44,"command":45},"pnpm","pnpm dlx skills add wildcat-finance\u002Fskills --skill ephoros","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fephoros\u002FSKILL.md",true,[49,55,59,63],{"path":50,"resourceType":51,"resourceTypeDesc":52,"sizeBytes":53,"sourceHtmlUrl":54},"agents\u002Fopenai.yaml","REFERENCE","参考文档",221,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fephoros\u002Fagents\u002Fopenai.yaml",{"path":56,"resourceType":51,"resourceTypeDesc":52,"sizeBytes":57,"sourceHtmlUrl":58},"DEMONSTRATION.md",2365,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fephoros\u002FDEMONSTRATION.md",{"path":60,"resourceType":51,"resourceTypeDesc":52,"sizeBytes":61,"sourceHtmlUrl":62},"EVOLUTION.md",8067,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fephoros\u002FEVOLUTION.md",{"path":64,"resourceType":65,"resourceTypeDesc":66,"sizeBytes":67,"sourceHtmlUrl":68},"scripts\u002Fephoros.py","SCRIPT","脚本",60679,"https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Fplugins\u002Fhexaemeron\u002Fskills\u002Fephoros\u002Fscripts\u002Fephoros.py",[70,77,83,90,96,102],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7APAWXHFSKVQEJBAHKASH25","example","A deliberately unsupported evidence class fixture.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funsupported-evidence-class\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[],{"skillNo":78,"name":72,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"publishedAt":19},"01M7APATBK93107VF5NMPA3RJH","A router fixture target.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funresolved-router\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[],{"skillNo":84,"name":85,"nameZh":6,"description":86,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"publishedAt":19},"01M7APARV9C0MQB8WTB0GGAAMG","promise-machine","A deliberately unresolved router.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funresolved-router\u002F.agents\u002Fskills\u002Fpromise-machine\u002FSKILL.md",[],[],{"skillNo":91,"name":72,"nameZh":6,"description":92,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":93,"categories":94,"keywords":95,"publishedAt":19},"01M7APAQMDG4977HXGSKMB29KV","A deliberately unclassified skill fixture.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funclassified-skill\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[],{"skillNo":97,"name":72,"nameZh":6,"description":98,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":99,"categories":100,"keywords":101,"publishedAt":19},"01M7APAPE6DE05A24J4BYPK6WY","A deliberately unattributed exception fixture.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Funattributed-exception\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[],{"skillNo":103,"name":72,"nameZh":6,"description":104,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":19},"01M7APAKYXRTABBSN7Q637CC6M","A deliberately conflated version fixture.","https:\u002F\u002Fgithub.com\u002Fwildcat-finance\u002Fskills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fpromise-machine\u002Fpackage-as-skill-version\u002Fplugins\u002Fexample\u002Fskills\u002Fexample\u002FSKILL.md",[],[]]