WritersLogic agent-infra

Signed acceptance receipt: which deterministic tests a delivered output passed — Runs your deterministic acceptance tests (equals, exists, absent, type, contains, length and number ranges, one-of, sha256, a JSON Schema subset) against a sub-agent's or vendor's JSON output at JSON Pointer paths, and returns per-test results plus a hub-signed receipt over the output hash, the tests hash and the results. Use it before paying for delegated work or releasing the next step. It certi

Does a change to the approved thing invalidate the approval under a materiality policy? — Compares what a human or system approved with what is about to be executed (a quote, order, plan, diff summary: any JSON) under a caller-supplied materiality policy: per-path rules exact / ignore / numeric tolerance (absolute or percent) / no-increase / no-decrease, with the longest matching path winning and a default for unlisted paths. Also covers quote revisions (approved revision number must e

Which derived artifacts must be rechecked after these inputs changed? (bounded dependency graph) — Given a dependency graph of artifacts (each lists the artifact ids it was derived from) and the ids that changed, returns every artifact that transitively depends on a change, in an order safe to recompute (inputs before dependents), and the artifacts that are unaffected. A cycle or an unknown input id is a defect in the graph and returns fail with the cycle path. Changed ids that are unknown fail

Find assumptions that disagree across agents (currency, timezone, version, customer id, ...) — Takes assumptions that several agents or steps hold about the same fact (key, value, who) and reports every key where the values differ, naming the conflicting pairs and whether they differ only by case or whitespace (a formatting difference) or in substance (USD vs EUR, a different id, a different version). Optionally requires each listed agent to state each required key; a missing statement is u

Check that what a message says is attached matches the attachments actually supplied — Compares a message's wording with its attachment list: 'see attached' with nothing attached, a named file (report.pdf) that is not in the list, a kind of file mentioned (the spreadsheet, a screenshot, the PDF) that no attachment matches by extension or MIME type, and a stated count ('two files') that differs from the list; attachments the text never mentions are listed. Mentions inside a negation

Check that a result-cache key contains every input that changes the result, and nothing volatile — Given the inputs of a cached computation (each declared as influencing the output or not, optionally volatile) and the list of inputs the cache key is built from, finds influential inputs missing from the key (stale or wrong hits), volatile inputs inside it (the cache never hits), unknown-influence inputs left out, and, for private results, a missing tenant dimension (cross-tenant leak). When valu

Chat completion, deep tier (12k in, 2k out) — OpenAI-style chat completion served by the hub's provider router (strongest available model, up to 12000 input tokens and 2048 output tokens per call), with a signed receipt that binds the provider, model, token usage and the hashes of your messages and the reply. Supports optional strict json_schema output returned both parsed and as a string. Does not stream, call tools or functions, accept imag

Chat completion, fast tier (8k in, 1k out) — OpenAI-style chat completion served by the hub's provider router (small, fast model, up to 8000 input tokens and 1024 output tokens per call), with a signed receipt that binds the provider, model, token usage and the hashes of your messages and the reply. Supports optional strict json_schema output returned both parsed and as a string. Does not stream, call tools or functions, accept images, or us

Chat completion, standard tier (12k in, 2k out) — OpenAI-style chat completion served by the hub's provider router (mid-size general model, up to 12000 input tokens and 2048 output tokens per call), with a signed receipt that binds the provider, model, token usage and the hashes of your messages and the reply. Supports optional strict json_schema output returned both parsed and as a string. Does not stream, call tools or functions, accept images,

Split text into chunks under a hard token cap without cutting a table, sentence or code fence — Plans chunk boundaries for text an agent must feed to a model in pieces: every chunk is at most maxTokens in the named tokenizer (o200k_base or cl100k_base, verified by counting each finished chunk), boundaries fall only between sentences, blocks, whole Markdown or HTML tables and whole code fences, headings stay with the text that follows, and the chunks tile the input exactly (concatenating the

Locate quoted text in a source with exact offsets, or classify it as normalized, elided, paraphrased or absent — Finds each quote an agent attributes to a source text and returns where it sits: exact offsets in UTF-16, code points, UTF-8 bytes and graphemes, so the span slices back to the matched text. Tries an exact match, then a match tolerant of whitespace, curly vs straight quotes, dash types, ellipses, invisible characters and Unicode forms (NFC/NFD, ligatures), naming which normalizations were needed,

Check that every write step in a multi-step plan has an undo step, and give the rollback order — For an agent plan made of steps with declared effects (read or write), checks every write in the plan, run or not, has a compensation step that exists, is itself a write, is not its own compensation, and that the dependencies have no cycle; flags writes declared irreversible; and, given which steps have completed, returns the compensations to run in reverse completion order. Does not judge whether

Classify a confirmation message as submitted, pending, reserved or completed from caller or default phrases — Reads a confirmation page, email or API message and decides which state it states: submitted, pending, reserved or completed, using phrase lists the caller supplies (preferred, because wording is provider-specific) plus an optional default English list. Matches are whole-word, case- and typography-insensitive, returned with offsets in UTF-16, code points, UTF-8 bytes and graphemes, and a match ins

Check that earlier constraints (numbers, dates, codes, quoted terms, negations) survive into a later plan or summary — Takes the constraints an agent was given (budget caps, deadlines, identifiers, prohibitions) and a later artifact (plan, summary, handoff, tool arguments) and reports for each constraint whether its concrete atoms still appear: numbers with their unit or currency, ISO dates, codes, quoted terms, or caller-listed anchor strings, with offsets. For a negated constraint it also reports whether the clo

Pick context blocks that fit a token budget, keeping mandatory blocks and their dependencies — Chooses which context items an agent should put into a prompt without exceeding a token budget: every mandatory item and everything it depends on is always included, then optional items by descending priority (each with the dependencies it pulls in) while they fit. Token counts come from the caller or from o200k_base / cl100k_base; any other tokenizer name returns unsupported (there is no Claude t

Choose which old messages to drop to hit a token target without touching the cached prompt prefix — Given a conversation, the index of the last message inside the provider's cached prefix, and a token target, returns which messages to drop: whole tool call plus tool result pairs (so no result is left without its call) and messages the caller marked droppable, oldest first or largest first, never inside the cached prefix, never the most recent keepRecent messages, never pinned ones. Everything up

Find dangling and duplicated cross-references (Section 3.2, Table 4, Figure 2, [7]) in a document — Scans a document for the targets it defines (numbered headings, Table/Figure/Listing/Equation captions at line start, Appendix/Annex headings, [n] bibliography entries, [^id] footnote definitions) and the references it makes to them ('see Section 3.2', 'Tables 2 and 4', 'Figures 3 to 5', '[7]', '[^note]'), and reports references with no target, targets defined twice, and targets never referenced,

Can this plan of steps finish inside the remaining time, in the expected case and the worst case? — Takes steps with expected and worst-case durations, dependencies (steps with no dependency between them run in parallel), retries with backoff, and the time remaining, and computes the critical path for both cases. Verdict guaranteed (worst case fits: pass), likely (only the expected case fits: unknown) or infeasible (even the expected case does not fit: fail), with the slack and the critical-path

Group webhook deliveries by identity within a time window and decide: process, dedupe, retry, reconcile or abstain — Given previously seen deliveries (identity, receivedAt, payload or digest, handling outcome) and a new incoming one, finds earlier deliveries with the same identity inside the window and returns process (none), dedupe (same payload already processed), retry (earlier ones all failed), reconcile (an earlier one is in flight or unknown, or the same identity arrives with a different payload) or abstai

Find unresolved references (that one, the former, tomorrow, next week) and return clarification questions — Scans a user or tool message for phrases that cannot be understood without context: demonstratives ('that one', 'the latter'), back-references ('as before', 'the usual'), and relative dates or times ('tomorrow', 'next Friday', 'end of month', 'soon'), each with offsets and a clarification question. Given a caller anchor (asOf instant plus IANA time zone) it resolves only unambiguous day offsets (t

Check that a sub-agent delegation states objective, acceptance criteria, budget, evidence and stop conditions — Checks the brief an agent is about to hand to another agent: the required parts are present and non-empty (objective, acceptance criteria, a budget with at least one positive numeric bound, required evidence, stop conditions; output format if required), nothing is duplicated, and acceptance criteria are not made only of vague wording (a small labelled word list: good, appropriate, if possible ...)

DOI check: exists, retracted, or fabricated (batch of 20) — Checks up to 20 DOIs in one call against Crossref (including its Retraction Watch notices), OpenAlex and the doi.org Handle registry. Per DOI: FOUND with title and year, RETRACTED, EXPRESSION_OF_CONCERN, CORRECTED, REGISTERED_NOT_INDEXED, NOT_REGISTERED (no such DOI anywhere: likely fabricated), INVALID_DOI or UNVERIFIED. For research agents and RAG pipelines that cite papers: run every DOI throug

Check whether a saved interaction recipe still applies to the current version of a page — Compares a recipe (ordered steps naming selectors from an earlier dom-skeleton call), the saved skeleton elements those selectors came from, and the current page HTML. Per step it reports matched (same selector, same tag, type, role and name), changed (selector still exists but the element differs), moved (the same element now has a different selector, given as newSelector), ambiguous (several cur

List the interactive elements of an HTML page with stable CSS selectors — Reduces an HTML document to the elements an agent can act on: links with href, buttons, inputs (not type=hidden), selects, textareas, [role], [onclick], contenteditable, each with tag, type, role, best-effort accessible name (including label associations) and a CSS selector that matches exactly that element, preferring a stable id, then a data-testid-style attribute, name, aria-label, then a struc

Check that an edit's old-string occurs exactly once in a file, with match contexts — Counts how many times an old-string occurs in a supplied file text (as a replace-in-file tool would see it) and returns each match's offset, line, column and surrounding lines, for an agent about to issue an edit. pass: exactly one literal match, even counting overlapping ones. unknown: several matches, or one match that overlaps another occurrence of the same text (ambiguous; contexts show how to

Check that every claim in an evidence bundle has enough independent, fresh, quote-verified support — Checks a bundle an agent assembled before answering: each claim must be backed by at least N evidence items from distinct source groups, each item's quote must actually appear in the supplied text (with offsets), publication dates must be within the caller's maximum age relative to a caller-supplied asOf, and duplicates are collapsed. Does not fetch URLs or judge whether a source is reliable or wh

Check that evidence covers the scope of the claim it is used for (geography, population, period, version, basis, unit) — Compares the scope descriptors of a claim and of the evidence offered for it, per dimension: geography and population (every claimed item must be equal to, or contained in a broader evidence item according to the caller's taxonomy), period (the evidence period must contain the claimed period; partial or no overlap is a mismatch), and version, basis and unit (must be equal after normalization). Des

Fleet memory: your agents' shared observations of endpoints, rate limits, IVR paths and ambiguous submissions — Lets every agent in your fleet record what it observed and read what the others saw, with freshness: endpoint brownouts and latency, rate-limit behaviour, IVR menu paths that worked, whether an ambiguous submission was later accepted or rejected, fixes for an error signature, stale feeds, page recipes, merchant workflow failures, document reputation. record() appends timestamped observations for a

Attach footnote texts to their in-text markers from page layout coordinates — Given footnote markers (label, page, x, y) and footnote texts (label, page, x, y) extracted from a laid-out document, pairs each marker with its note by label and page, pairing repeated labels on one page in reading order, optionally letting a note continue on the next page, and reports unattached markers, orphan notes, ambiguous groups (different counts for the same label) and notes placed above

Signed hash receipt (Merkle batch, RFC 9162) — Timestamped, signed receipt for up to 256 SHA-256 hashes in one call: an RFC 9162 Merkle root, a per-hash inclusion proof, and the hub's W3C VC + EIP-191 signatures over both. For agents that need third-party evidence of what they did and when (action logs, generated files, prompts, tool outputs) without sending the data itself. Pure code, no model; verify free at /api/v1/verify-provenance.

Extract hedging and certainty cues (may, probably, clearly, approximately) from text with offsets — Finds hedge and booster cues in an agent's or source's text using a fixed English lexicon (modals, epistemic verbs, adverbs and adjectives, approximators, shield phrases, certainty boosters) plus caller-supplied phrases, and returns each cue with its category, the sentence it sits in, and offsets in UTF-16, code points, UTF-8 bytes and graphemes. Rule-based and English only: a cue is a word match,

Classify an HTTP response as success, retry, reconcile or stop, with the wait it asks for — Classifies one HTTP response (method, status, headers) as success / retry / reconcile / stop for an agent that just called an API, and reads the wait from Retry-After (RFC 9110: delay-seconds or any of the three HTTP-date formats) or from a vendor delay header whose dialect the caller declares (seconds, milliseconds, epoch seconds, epoch milliseconds, HTTP-date). Caller-supplied vendor rules (stat

Monotonic id blocks: the next invoice or ticket numbers for a fleet of agents — Hands out strictly increasing integer blocks from a named sequence (1-1000 ids per call) so parallel agents never issue the same invoice, ticket or order number. Numbers are never reused. Not gapless: a block an agent reserves and never uses stays a gap, and if your call's payment fails after the block is written the same requestId returns that same block on retry. Legally gapless series must be a

Derive a stable idempotency fingerprint from declared identity fields and flag volatile fields — Hashes the fields you declare as the identity of an operation (JSON Pointers into each request) with RFC 8785 canonical JSON and SHA-256, so retries of the same operation get the same fingerprint regardless of key order, and flags the ways that breaks: an identity field missing, an identity field also declared volatile, identity values that look like timestamps, epoch times or UUIDs (heuristic), a

Idempotency registry: has key K already been done? Claim it, complete it, read the stored result — Shared memory for side-effecting operations across agents and retries. begin(key) answers proceed (you hold a short claim), wait (another owner is mid-flight), already_completed (returns the stored result pointer), or fingerprint_mismatch (same key, different operation). complete(key) records the result pointer; abandon(key) lets someone retry. The claim expires, so a crashed worker does not block

Receive-only inbox for agents without a public endpoint — Creates a receive-only inbox for an AI agent that has no public endpoint: a webhook URL that accepts POST or PUT, and (when the operator has configured inbound email) a unique email address. Pair with inbox-poll to read what arrived. It never sends email or calls out, does not store attachments, does not verify senders (no SPF/DKIM), and message content is untrusted third-party data. Inboxes live

Read messages from a receive-only agent inbox — Returns the emails and webhook deliveries that arrived in an inbox you created with inbox-create, oldest first, with sender, subject, selected headers, a plain-text body (up to 64 KB per message in the response) and a sha256 of the stored body. Filter by since (inclusive), kind, limit up to 100; mark_consumed hides the returned messages from later polls unless include_consumed is true. It does not

Label spans of a prompt as pasted data or instructions, and flag instruction-like text inside the data — Splits an agent's input into data regions (code fences, block quotes, XML-style wrappers such as , BEGIN/END markers, triple quotes, email headers with body, JSON blocks) and the text around them, labels each sentence outside the data as instruction (leading imperative verb or polite request) or unknown, and lists injection-style phrases found anywhere in the text ('ignore previous instr

Short lease with fencing tokens: claim a task for N seconds, renew, release — Lets many agents agree which one works on a task: acquire a named resource for 1-3600 s, renew it while working, release it when done. Every new acquisition returns a strictly larger fencing token; pass it to the downstream store and reject writes carrying a smaller one, so a paused holder whose lease expired cannot clobber its successor. Re-acquiring as the current holder extends the lease with t

Detect LLM text watermark evidence: green-list z-test for a key you hold, plus invisible-character markers — Tests whether text carries watermark evidence, for a pipeline that must decide whether to trust or label content. Two independent checks: (1) a green-list z-test (Kirchenbauer et al. 2023) over green flags you supply, or over token ids and a key using this endpoint's own HMAC-SHA256 scheme, returning the z-score and exact binomial p-value; (2) an invisible-character scan of the text for zero-width

Collapse a log into error signatures with counts and first/last occurrence — Groups the records of a log or CI output by signature: timestamps, UUIDs, IPs, hex ids, quoted values and numbers are replaced by placeholders, stack-trace continuation lines are attached to their record and contribute their first frames, and each signature reports its count, severity, first and last line, first and last timestamp (as written, when the line starts with an ISO 8601 timestamp) and a

Normalize sloppy model-generated Markdown to one house style, leaving code untouched — Rewrites Markdown from a language model into a consistent style and reports each edit by line: restores the space in '##Title', converts '*', '+' and '•' bullets to your chosen marker, strips trailing whitespace (keeping two-space hard breaks), collapses runs of blank lines, puts blank lines around headings and fenced blocks, closes a fence left open at the end, and ends the file with one newline.

MCP tool pin: detect changed tool definitions (rug pulls) and hidden Unicode — Send an MCP server's tools/list result (up to 200 tools) and, after the first call, the hashes you approved. Returns a reproducible manifest hash, a per-tool hash, which tools were added, removed or changed since approval (silent description and schema edits are how MCP rug pulls and tool poisoning land), and Unicode smuggling findings in every name, description and schema string. Call it at sessi

Run up to 20 micro checks under one payment ($0.02) — Bundles up to 20 calls to any pure micro check (catalog entries that are neither stateful nor network-reading, e.g. checksum, patch-anchor, allocate, business-deadline) into one x402 settlement, so a full batch costs $0.001 per check. Every call is validated before payment; one invalid call rejects the batch with its index. The request is capped at 128kb and a CPU weight of 40 (an ordinary check w

Diff a before/after JSON write for deletes, null-vs-omit and emptied arrays against protected paths — Compares the JSON document an agent read (before) with the one it is about to write (after) and lists every change by JSON Pointer: removed members, null turned into omitted (and the reverse), arrays emptied or shrunk, objects emptied, type changes. Fails when a change touches a protected path or is destructive and not listed in allowDestructive; use it before a PUT, PATCH or bulk update. Arrays a

Find negation cues and the words they cover; report whether given terms fall inside a negation scope — Applies NegEx-style rules (the English termset of negspacy, MIT, derived from NegEx) to text: finds negation cues, pseudo-negations such as 'not only', and scope terminators such as 'but', and returns each cue's scope window with offsets in UTF-16, code points, UTF-8 bytes and graphemes. Given target terms, reports per occurrence whether it is negated, not negated by the rules, or unknown (near a

Plot a prose intensity curve per sentence and flag where a tense passage loses velocity — Scores each sentence of narrative prose for pace from sentence length, share of one-syllable words, exclamation and question marks and, if you supply a word list, action-word share; smooths the scores into an intensity curve; and reports stalls, the spans where intensity falls well below an earlier peak. For a writing tool or editor agent checking an action or tension scene. A fixed-weight heurist

Detect repeated cursors, duplicate-only pages and no progress in a pagination scan — Takes the pages an agent has fetched so far (cursor used, next cursor returned, item ids) and finds the failures that make a scan loop or lie: a next cursor already used, a page whose items were all seen before, a page that returned nothing new yet offers another cursor, a scan that has not ended. Optionally takes declared ordering and cursor assumptions and lists the omission risks they imply (of

Apply a unified-diff hunk or search/replace edit to a file by content, refusing ambiguity — Applies unified-diff hunks (wrong line numbers are fine) or search/replace blocks to a supplied file text and returns the patched text, for an agent whose edit tool failed or whose diff has drifted line numbers. Each hunk is located by its content against the original file, trying exact, trailing-whitespace, outer-whitespace then collapsed-whitespace matching; the rung used is reported per hunk. I

Effective permissions of a delegated agent: requested ∩ user ∩ parent, minus explicit denials — Computes what a sub-agent may actually do: the permissions it requested, intersected with what the end user holds and what its parent agent holds, minus any explicit deny list, and says for each denied permission which party lacks it. Permission names are compared exactly (no wildcards); caller-declared implications (for example 'admin' grants 'read') can be given and are expanded transitively wit

Policy gate with a signed receipt: your JsonLogic rules vs a proposed tool call or payment — Evaluates your own policy (JsonLogic rules, each allow or deny, deny-overrides or first-match, default deny) against a proposed tool call, payment or action, and returns allow or deny, the rule id that decided, and a receipt signed by the hub over the hashes of the policy, the request and the decision. The receipt is the product: another agent or an auditor can verify offline which rule allowed th

Would this request fit in the remaining quota, given a snapshot and pending reservations? (snapshot-bound) — Computes available quota from a caller-supplied snapshot (limit, used, reserved) minus the unexpired pending reservations and says whether a requested amount fits. This is a read of the numbers you supply as of snapshotAt: it reserves nothing, so concurrent consumers can still use the quota before you do, and every response says so. A stale snapshot gives unknown, not pass; a window that ended at

Content-addressed result cache shared by your agents (short TTL) — Caches a result under the SHA-256 of the canonical (JCS) request, so every agent in your fleet that makes the same request reuses one answer for up to an hour. get returns hit with the value and age, or a miss (status unknown). Scoped to your wallet or tenantId, never shared across customers, so nobody else can poison your cache.

Decide whether a failed HTTP call may be retried, must be reconciled, or must not be repeated — Given a request's method, how it ended (response status, timeout, reset, refused), idempotency-key evidence and optional caller policy, returns retry / do-not-retry / reconcile / unknown for an agent deciding what to do after a failed tool or API call. Encodes RFC 9110 method idempotency, the proxy and retry-of-a-retry rules and Retry-After parsing; a timeout or reset is never read as success or a

Spend ledger with reservations: one budget for a tree of agents — A budget shared by a task tree, in integer minor units. reserve() atomically holds an amount if the remaining budget covers it (expired holds are freed automatically); commit() records what was actually spent, up to the hold; release() frees a hold. These are real reservations, not a snapshot: two workers cannot both reserve the last dollar. It records spending; it does not move money.

Detect near-duplicate passages (syndicated or copied text) so repeated sources are not counted as independent — Compares up to 30 passages an agent collected and reports which pairs share word-shingle content (Jaccard and containment), grouping them into clusters of probable copies so the number of independent sources can be stated honestly. Does not establish who copied whom, does not detect paraphrase or translation, and does not fetch anything; a high similarity is a hint that sources are not independent

Interpret a test run: pass, no-tests, skipped, flaky, infra-failure from runner, exit code and counts — Reads a finished test run (runner, exit code, optional version, flags, caller-parsed counts, optional go test output, optional per-attempt exit codes) and says whether tests actually passed, ran nothing, were all skipped, passed only on retry, or never ran properly (timeout, signal, internal error). Encodes only exit-code semantics the runners document: pytest 0-5, node:test, mocha (failure count

Convert text offsets between UTF-16, code points, UTF-8 bytes and graphemes, and locate hidden Unicode — Reports the length of a string in UTF-16 code units, Unicode code points, UTF-8 bytes and extended grapheme clusters (Intl.Segmenter), converts up to 200 ranges given in any one of those units into all four, and flags ranges that fall inside a surrogate pair, a multi-byte UTF-8 character or a grapheme cluster. It also runs the unicode-check rules (bidi controls, zero-width and invisible characters

Shared token bucket: many agents, one upstream rate limit — One rate limiter shared by every agent in your fleet. take(cost) atomically spends tokens if available and returns allowed plus remaining, or refuses with retryAfterMs until enough tokens refill. capacity and refillPerSecond are fixed when the bucket is first used; a call with different settings is refused so two fleets cannot silently disagree. Use a requestId so a retried take is not charged twi

Check tool-call / tool-result pairing and ordering in an LLM transcript (Anthropic, OpenAI Chat, OpenAI Responses) — Validates a message list before it is sent to Anthropic Messages, OpenAI Chat Completions or OpenAI Responses: every tool call paired with its result, results in the right place and role, ids present and unique. Built for an agent that edits, truncates or replays its own history and would otherwise get a 400. Only rules stated in the providers' documentation fail the check; anything else suspiciou

Check that numbers, units, part numbers, tolerance signs, dates, placeholders and URLs survive translation — For aligned source and target segments, extracts the tokens a translation must not change (numbers with attached units or %, ISO dates, part numbers and versions, tolerance and comparison operators such as ± ≤ ≥, placeholders like {0} or %s, URLs and emails) and reports any that are missing, added or altered, with offsets in both texts in UTF-16, code points, UTF-8 bytes and graphemes. Numbers are

Unicode smuggling check for strings (prompts, tool args, URLs) — Checks up to 100 strings in one call against the 19 deterministic rules of unicode-safety-check: bidi controls (Trojan Source), invisible and zero-width characters, tag-character and variation-selector payloads, homoglyph and mixed-script hosts or names. Built for an agent's hot path: run it on a prompt, a tool argument, a URL or a model output before acting on it. Returns BLOCK / REVIEW / NO_FIND

What do omitted, null, [] and "" mean in this PATCH body under merge-patch, JSON Patch or a declared dialect? — For an agent about to send a PATCH or update body: states, per field, what omitted / null / [] / "" mean under the declared dialect (RFC 7396 merge patch, RFC 6902 JSON Patch, or caller-declared semantics) and checks the body against stated intents such as 'clear this field', 'leave this alone' or 'set to empty array'. With a target document it applies the patch and compares the result; without on

WARM verifiable agent memory — Persistent memory for AI agents as a pay-per-call API, no database to run. op=remember stores subject-predicate-object facts (or extracts them from free text; only facts whose subject and object appear verbatim are kept); a newer value for the same subject and predicate supersedes the old one, and the history stays queryable. op=recall searches by BM25 keyword ranking or exact triple pattern, now

Is this state transition allowed by the caller's workflow definition, and what is allowed instead? — Given a caller-supplied state machine (transitions with optional event names and required facts), the current state and a requested event or target state, says whether the transition is allowed, which required facts are false or unknown, and which transitions are available from the current state. For an agent about to move an order, ticket, deployment or approval to its next state. Terminal states