Skip to content

v0.7.4

Latest
Latest

Choose a tag to compare

View all tags
@cursor cursor released this 26 Jul 20:29
v0.7.4
c4f27e9
This commit was signed with the committer’s verified signature.
cursoragent Cursor Agent
SSH Key Fingerprint: /ZGdxqU+YWGLYEplYKh2c36jVXjB1wQzjGRo7qwO46I
Verified
Learn about vigilant mode.

0.7.4 — Duffel order enrichment + activity/transfer agents

Workspace-wide patch bump 0.7.3 → 0.7.4 so npm picks up #123.

@otaip/adapter-duffel

  • Enrich BookResponse from the live Duffel order already returned by book(): optional ticketNumbers, segments (fare basis), baseAmount / taxAmount, recordLocator, passengerNames, refundable / changeable, timestamps.
  • Add getOrder(orderId) → GET /air/orders/{id} mapped through the same helper.
  • Existing five fields unchanged; new fields optional. Unit tests cover the order mapper.

@otaip/agents-lodging

  • Agent 20.8 ActivitySearchAgent — typed Hotelbeds Activities search wrapper.
  • Agent 20.9 TransferSearchAgent — typed Hotelbeds Transfers search wrapper.
  • agents.manifest.json regenerated (77 agents).

CI / hygiene

  • pnpm overrides for high pnpm audit findings (postcss, brace-expansion, fast-uri, find-my-way, @fastify/static).
  • Public-repo comment guard: no internal product-name refs in lodging agent docs.

0.7.2 — Duffel Cars + auto-publish workflow + post-#93 release fix

Patch bump. Three pieces of substantive work in this window:

@otaip/adapter-duffel — Cars API (#99)

Duffel launched Cars as a full API vertical alongside Flights and Stays. DuffelAdapter now spans both surfaces. Five new direct methods on the same class:

  • Cars (/cars/) — searchCars, quoteCar, bookCar, getCarBooking, cancelCarBooking. Three-step flow (search → quote → book) unlike the two-step flight flow. Geo-coordinate based — no IATA codes — per the brief; IATA-to-coordinate conversion is explicitly deferred (DQ-C8).
  • The shared private request() helper now accepts an optional AbortSignal (pre-flight check; in-flight cancel needs an upstream change to fetchWithRetry, same caveat the Hotelbeds adapter carries).

New canonical types (CarRate, CarQuote, CarBookResponse, etc.), wire types, decimal.js-backed mapper, and a full mock three-step flow on MockDuffelAdapter (two synthetic fixtures: Toyota Corolla compact + VW Tiguan SUV). 18 new test cases covering body shapes, headers, mapper output, abort handling, 429 retry-then-error, and round-trip mock flow. Domain knowledge captured verbatim in docs/knowledge-base/cars.md with 8 numbered DOMAIN_QUESTION markers (DQ-C1..C8) for the gaps the brief leaves open. Per the constitution's no-invent rule, those are surfaced rather than guessed.

Auto-publish workflow wiring (#97, #98)

Post-mortem on why @otaip/core etc. sat at v0.6.4 on npm despite v0.7.0 and v0.7.1 GitHub releases existing:

  • #97 — scripts/count-agents.ts had been importing discoverAgents through the CLI re-export added in #93. The Release workflow's Count agents step ran before any package was built, and the indirection forced tsx to resolve @otaip/core through its package.json exports map → dist/index.js (which doesn't exist on the runner) → ERR_PACKAGE_PATH_NOT_EXPORTED. Result: every push-to-main since #93 was failing the Release workflow silently, blocking tag creation. Fixed by importing directly from the source path.
  • #98 — The Release workflow's gh release create was authenticated with GITHUB_TOKEN, but GitHub Actions intentionally does not trigger downstream workflows for events fired by GITHUB_TOKEN. That's why the Publish workflow never auto-fired off release: published for v0.7.0 or v0.7.1 — packages only made it to npm when someone clicked Run workflow manually. Switched the create-release step to a Personal Access Token (RELEASE_PAT secret) so the event is "user-fired" and propagates. Falls back to GITHUB_TOKEN when the secret is unset, so the workflow still functions in environments without the PAT.

After v0.7.2 merges and tags, this is the first release that will auto-publish to npm without a manual workflow_dispatch.

Verification

  • pnpm exec vitest run packages/adapters/duffel — 58 passed / 3 skipped (sandbox-gated).
  • pnpm -r typecheck — clean across the workspace.
  • pnpm exec eslint packages/adapters/duffel/src — clean.

Versioning note

Patch bump off v0.7.1 per VERSIONING.md. The Cars vendor brief asked for a "next minor" (0.8.0) bump; we honoured the policy and stuck with 0.7.2 instead. The next release is v0.7.3.

0.7.1 — Platform UI + agent-discovery promoted to @otaip/core

Patch bump per the policy reset in v0.7.0. The headline of this window is the Platform UI (#93) — a new React/Vite app at examples/platform-ui/ that gives developers a visual control plane for an OTAIP instance: agent registry, adapter status, health sidebar, and an interactive Playground for searches/agents/adapters. The Platform UI is private (examples/* is not published to npm); the npm-visible delta is one additive export on @otaip/core.

Published-package changes

  • @otaip/core — new discoverAgents() / DiscoveredAgent exports under src/discovery/. Filesystem-only walk of the workspace; no agent code executes. Repo root is resolved by walking up to pnpm-workspace.yaml rather than by counting parent directories, so the helper works the same way from src/ (tsx/vitest) and from dist/ (built consumer).
  • @otaip/cli — agent-discovery.ts now re-exports from @otaip/core. Behavior unchanged; every existing import path keeps working. The pnpm cli agents output is byte-identical.
  • All other published packages bumped 0.7.0 → 0.7.1 with no source changes — workspace-wide version sync.

Reference OTA additions (#93)

examples/ota (private) gains the read-only telemetry routes the dashboard consumes: /api/platform/{agents,adapters,health,stats} (rate limit raised to 5 000/min so dashboard polling cannot trip the global cap) and /api/playground/{catalog,search,agent,adapter}. Catalog surfaces every discovered agent plus an executable_ids whitelist; the playground starts at one wired agent (0.1 AirportCodeResolver — the canonical reference pattern from CLAUDE.md). Non-whitelisted IDs return a clear 501 rather than pretending to run. 12 new tests across platform.test.ts + playground.test.ts follow the existing Fastify inject() + MockOtaAdapter shape.

Dependency hygiene (#95)

Dependabot bumped postcss 8.5.8 → 8.5.10 (dev-only, transitive through tailwindcss).

Verification

  • pnpm -r typecheck — clean across the workspace.
  • pnpm exec vitest run examples/ota — 100 passed across 9 files.
  • pnpm --filter @otaip/platform-ui typecheck — clean.

Versioning note

Patch bump off v0.7.0 per the rule reaffirmed in the v0.7.0 release. The next release is v0.7.2.

0.7.0 — Reference OTA hardening + Hotelbeds Activities/Transfers

Re-sync release. @otaip/adapter-hotelbeds@0.7.0 shipped as a single-package minor bump in #90 ahead of the repo-wide release; this PR aligns root + every other workspace package with that version so the GitHub front page, npm, and package.json files all agree. Per-package, the only API surface change since v0.6.4 lives in @otaip/adapter-hotelbeds. The rest of the work in this window targeted the reference OTA — see below.

@otaip/adapter-hotelbeds — Activities + Transfers (#90)

HotelbedsAdapter now spans the full APItude product family. New direct methods on the same class:

  • Activities (/activity-api/3.0) — searchActivities, bookActivity, cancelActivity. Cancellation policy narrows to 'NOR' | 'NRF'; unknown values default to 'NRF' per DQ-A5.
  • Transfers (/transfer-api/1.0) — searchTransfers, bookTransfer, cancelTransfer. Unknown transferType values pass through verbatim rather than being coerced into the documented PRIVATE | SHARED | LUXURY set.
  • The shared request() helper now accepts an optional AbortSignal (pre-flight check; in-flight cancel needs an upstream change to fetchWithRetry and is left as future work).

New canonical types (ActivityOffer, ActivityModality, TransferOffer, etc.), wire types, mappers (decimal.js for all amounts), capability manifests (hotelbedsActivitiesCapabilities, hotelbedsTransfersCapabilities), MockHotelbedsAdapter fixtures for both surfaces, and unit + sandbox-gated integration tests. Domain knowledge captured verbatim from the vendor brief in docs/knowledge-base/activities.md and docs/knowledge-base/transfers.md with 13 numbered DOMAIN_QUESTION markers (DQ-A1..A5, DQ-T1..T8) for the gaps the brief leaves open. Per the constitution's no-invent rule, those questions are surfaced rather than guessed.

Reference OTA — Path B hardening (#88)

Three pieces:

  • Durable persistence — MockOtaAdapter now optionally stores bookings in SQLite via node:sqlite (Node 24+). Auto-loads when DATABASE_PATH is set; falls back to in-memory otherwise. Survives process restarts.
  • Real Stripe payments — PaymentService runs against Stripe when STRIPE_SECRET_KEY is set. PaymentIntents created at booking time, confirmed at pay time. Mock-mode preserved when no key.
  • Security headers + input schemas — helmet, @fastify/cors (env-driven), @fastify/rate-limit (100/min global, 20/min on /api/book, 10/min on /api/pay), AJV body schemas on every state-mutating route, custom error formatter that preserves the existing { error, details } envelope.

Reference OTA — wire real DuffelAdapter (#89)

When DUFFEL_API_KEY is set, examples/ota's search/price/book flow now runs against the live Duffel sandbox via a new DuffelOtaAdapter wrapper. Without the key it falls back to MockOtaAdapter exactly as before. New BookingLifecycle interface lets Payment / Ticketing / Manage services type against OtaAdapter & BookingLifecycle instead of the concrete mock — drops the as MockOtaAdapter casts in server.ts. Standardizes the env var to DUFFEL_API_KEY (matches the demo); DUFFEL_API_TOKEN still works with a deprecation warning. ADAPTERS=duffel is now a valid value for live multi-source search.

Docs & CI

  • README — "Build on OTAIP" promoted to the hero slot; the domain-flex framing pushed to the bottom (#87).
  • Publish CI — verify step now auto-discovers packages from the workspace instead of being hand-maintained, and repository.url is canonicalized across every package.json (#86).

Verification

  • Workspace typecheck clean.
  • @otaip/adapter-hotelbeds: 100 unit tests pass, 14 sandbox-gated tests skip without credentials.
  • examples/ota: 88 tests pass.
  • All PRs in the v0.7.0 window were merged green individually before this release PR was cut.

Versioning note

The patch-bump-only rule from VERSIONING.md was broken when @otaip/adapter-hotelbeds shipped as 0.7.0 in the per-package PR. This release ratifies that bump for the rest of the workspace and updates VERSIONING.md to acknowledge the deviation. Going forward, all releases are patch bumps off v0.7.x until v1.0 — the next release is v0.7.1.

0.6.4 — Hotelbeds Distribution Adapter

Patch bump per the pre-v1.0 policy in VERSIONING.md. Adds a new adapter package — @otaip/adapter-hotelbeds, the first lodging-distribution adapter on OTAIP. Verified end-to-end against the real Hotelbeds APItude sandbox: full booking lifecycle (search → checkrate → book → retrieve → cancel) returns a real Hotelbeds reference and cleans up after itself.

New

  • @otaip/adapter-hotelbeds — Hotelbeds APItude adapter for the Hotels API. Implements availability, checkRate, book, getBooking, listBookings, and cancelBooking (SIMULATION and CANCELLATION). Search bridges into the lodging search aggregator via the existing HotelSourceAdapter shape. Mock adapter ships in the same package for upstream tests that don't have credentials.
  • SHA256 request signing — signRequest / buildAuthHeaders regenerate the X-Signature per request (Hotelbeds rejects signatures more than ~5 minutes off server time). Backed by node:crypto, no extra dependencies.
  • Field mapper — Hotelbeds wire types → canonical OTAIP lodging types (RawHotelResult, RawRate, CancellationPolicy, BookingSummary). All money stays in Decimal / string land; no number for prices.
  • Demo script at packages/adapters/hotelbeds/demo/index.ts — runnable with HOTELBEDS_API_KEY=… HOTELBEDS_SECRET=… npx tsx demo/index.ts. Walks the full lifecycle end-to-end in a single command.

Domain decisions applied

  • DQ11 (mandatory fees) — when Hotelbeds returns taxes.allIncluded === false, same-currency mandatory fees are folded into the rate's totalRate / nightlyRate so the top-level price is what the traveler actually owes. Fees stay on mandatoryFees for transparency. Cross-currency fees are listed but not folded (the adapter has no FX rate).
  • DQ13 (cancellation markup) — Hotelbeds cancellation amount is net (cost-to-us). Adapter applies a documented HOTELBEDS_CANCEL_FEE_MARKUP = 1.25 and stores both the gross (penaltyValue) and the net (netPenaltyValue) so settlement and reporting can reconcile margin. New optional netPenaltyValue?: number field added to CancellationDeadline in @otaip/agents-lodging — additive, all existing callers still typecheck.
  • DQ14 (booking status) — explicit mapping for CONFIRMED / CANCELLED / ON_REQUEST / PENDING / MODIFIED; unknown values fall back to pending and emit a console.warn so unhandled states don't get silently swallowed.

Verification

  • Unit tests: 61 against mocked fetch (auth, field mapper, full adapter surface, mock adapter).
  • Integration test: 8/8 passing against the real Hotelbeds sandbox at https://api.test.hotelbeds.com. Auto-skipped when credentials aren't set, so CI never hits the live API. Self-cleans the booking it creates.
  • Repo-wide: 3153 tests passing, 0 failed, 11 skipped (the 8 integration + 3 pre-existing).

Deferred

  • Hotelbeds Transfers API and Activities API — tracked as a follow-up so this adapter stays focused on the bedbank content path that unblocks the lodging pipeline.

0.6.3 — Codex review closeout, first npm publish, per-transaction routing

Eleven PRs (#73–#83) addressing all 12 findings from a full-repo Codex review, the first-ever publish of the @otaip/* scope to npm, and a set of CI/publish hardening fixes. Several engines widen their output types and a couple of agents add required input fields — see Potentially-breaking below if you were depending on the previous behaviour.

Published to npm

All 15 @otaip/* packages are now live on npm at 0.6.2 — previously the source existed but nothing had shipped. npm install @otaip/core (and friends) works.

  • @otaip/core, @otaip/connect, @otaip/cli, @otaip/adapter-duffel
  • @otaip/agents-reference, @otaip/agents-search, @otaip/agents-pricing, @otaip/agents-booking, @otaip/agents-ticketing, @otaip/agents-exchange, @otaip/agents-settlement, @otaip/agents-reconciliation, @otaip/agents-lodging
  • @otaip/agents-tmc, @otaip/agents-platform

Removed invented domain logic (CLAUDE.md compliance)

Codex review flagged five engines that were computing fare/penalty/compensation amounts from invented "common industry pattern" data rather than authoritative sources. This release replaces all of them with either caller-supplied authoritative inputs or published-law constants.

  • @otaip/core: EU 261/2004 + US DOT 14 CFR §250 modules (new). applyEU261() encodes the published distance bands (€250/€400/€600), 3h arrival-delay trigger, Article 7(2) rerouting 50% reduction, 14-day cancellation safe harbour, and extraordinary-circumstances exemption. applyUsDotIdb() encodes the current 14 CFR §250.5 denied-boarding tables ($1,075 / $2,150 caps, effective 2025-01-22). Plus greatCircleDistanceKm() haversine helper.
  • @otaip/core: new DomainInputRequired type + helpers (domainInputRequired, isDomainInputRequired) — shared sentinel for engines that refuse to synthesise numbers when authoritative inputs are missing.
  • Fare Construction (2.2): removed the ROE 1.0 fallback (was silently multiplying every non-USD currency by 1), the per-mile-rate HIP heuristic, and the city-revisited BHC heuristic. Engine now returns DomainInputRequired when ROE is missing and populates missing_inputs on HIP/BHC checks that need intermediate-point fare lookups. The EMS mileage-surcharge formula (5%/5%, max 25%) stays — confirmed IATA standard. Output type widened to FareConstructionResult = FareConstructionOutput | DomainInputRequired.
  • Change Management (5.1) + Refund Processing (6.1): removed the "$200 default" fallback, the waiver-=-zero special case, and the residual = original − penalty formula. Engines now read cat31_rules / cat33_rules from input and apply them as filed. When rules are absent, they fall back to the ATPCO default — permitted at no charge for voluntary changes, fee waived for involuntary. Invented rule data moved out of src/data/ into __tests__/fixtures/ with a "TEST FIXTURE — do not use in production" banner.
  • Involuntary Rebook (5.3): removed the hardcoded 60-minute IRROP threshold. Caller now supplies thresholds.time_change_minutes per carrier. Real EU261 compensation is calculated via the core module when eu261_inputs (distance, delay, extraordinary circumstances, notice days, rerouting) are provided; regulatory_flags[].missing_inputs lists what's needed otherwise. Clarified that US DOT IDB applies to denied boarding only — not delays/cancellations.
  • Feedback & Complaint (6.5): replaced ~250 lines of inline EU261 / US DOT compensation math with calls to the core regulation modules. DOT IDB caps updated from pre-amendment $775 / $1,550 to current $1,075 / $2,150. Article 10(2) downgrade reimbursement (30/50/75%) kept — published law not covered by the Article 7 helper.

GDS/NDC router per-transaction routing (CLAUDE.md compliance)

The router previously treated carrier → channel_priority as unconditional. Replaced with per-transaction routing: different transaction types (shopping, booking, ticketing, servicing, group, corporate) route differently for the same carrier. Built-in defaults cover shopping and booking; every other type requires caller-supplied capability_overrides or the engine returns domain_input_required: true. Unknown carriers no longer default silently to GDS/AMADEUS.

HTTP hardening

  • New @otaip/core fetchWithRetry(input, init?, options?) — wraps fetch with per-attempt AbortController timeout (default 30s) + retry on 5xx / 429 / network errors via the existing withRetry. Response stays a Response; callers still inspect response.ok.
  • TripPro defaults now HTTPS. Search and calendar-search URLs were http://mas.trippro.com/... — switched to https://. Reprice/book were already HTTPS.
  • Sabre auth, Navitaire create+refresh, Duffel adapter, TripPro SOAP client all route through fetchWithRetry instead of raw fetch.

Correctness fixes

  • booking/api-abstraction rate limiter — the counter incremented once per execute() call, before the retry loop, so retries against the upstream provider went uncounted. Moved the increment and the rate-limit guard inside the loop so each actual outbound attempt is charged against the quota.
  • Stub agents now throw UnimplementedDomainInputError (new @otaip/core export) instead of raw Error. Four agents migrated: DisruptionResponseAgent (5.4), DynamicPricingAgent (2.6), RevenueManagementAgent (2.7), InterlineSettlementAgent (7.4).

CLI

  • otaip agents registry is now auto-discovered from source metadata (packages/cli/src/agent-discovery.ts) rather than a hand-maintained array that had drifted (claimed 71, listed 69, several names didn't match the exported agent classes). Walks packages/agents/*/src/*/index.ts, packages/agents-platform/src/*/index.ts, packages/agents-tmc/src/*/index.ts, and packages/core/src/agents/shopping/*/index.ts — greps readonly id, readonly name, readonly version. Today: 75 agents across 12 stages.
  • CLI now included in lint (--ignore-pattern 'packages/cli/**' removed). packages/cli/tsconfig.json added to the ESLint parser project list. CLI-scoped override allows console.* (CLI stdout is the contract).
  • 7 new tests for the discovery walk — count floor, unique IDs, file-resolves-to-real-path, stage matches ID prefix, sort order, etc.

Bootstrap + counts

  • New scripts/count-agents.ts (pnpm run count:agents) — single source of truth for agent counts. Replaces a find in release.yml that undercounted by skipping agents-platform and agents-tmc. All agent/stage counts in README, docs, and release notes now derive from one script.
  • Removed root postinstall. With ignore-scripts=true in .npmrc (shipped in 0.5.x for supply-chain safety), the postinstall never ran anyway. Docs updated to instruct an explicit pnpm run data:download.
  • Publish prep — every workspace package now has "type": "module" so tsup's ESM output (index.js/index.d.ts) matches the main/types fields; exports["."].types moved from ./src/index.ts (not in the tarball) to ./dist/index.d.ts; all 15 workspace packages aligned to the root version so the first publish was coherent.

CI / release

  • release.yml no longer swallows test failures. The previous step used pnpm test 2>&1 || true, masking failed tests so a broken release could publish with a misleading test count. Split into two steps that fail fast.
  • publish.yml now verifies packages are actually live. After pnpm -r publish, polls the registry for each of the 15 @otaip/* packages and expects dist-tags.latest to equal the released version. Fails loudly if pnpm reports success but the registry doesn't have the package. (Caught a "published successfully but 404 on registry for 5 minutes" class of bug on the first live publish.)
  • ci.yml builds before typecheck. With exports.types → ./dist/index.d.ts, cross-package @otaip/* imports need dist/ to exist for TS to resolve types. Doubles as clean-tree build validation.

Docs

  • README header numbers synced — 75 agents across 12 stages, 3,092 tests, 16 packages. Test badge updated. Install instructions now say pnpm install --frozen-lockfile && pnpm run data:download.
  • Accurate tsconfig strictness claim — README and CLAUDE.md previously said "all strict flags ON" but exactOptionalPropertyTypes is intentionally off. Replaced with the explicit list of enabled flags.
  • docs/getting-started.md — added the explicit data-download step and documented the supply-chain trade-off.
  • docs/agents.md — header updated to "12 stages" with a note pointing future editors at pnpm run count:agents so the total cannot drift again.

Potentially-breaking

Pre-1.0 policy allows breaking changes in patch bumps; flagged here so downstream consumers can update.

  • FareConstruction#execute() return type widened to AgentOutput<FareConstructionOutput | DomainInputRequired>. Consumers must narrow on result.data.status.
  • GdsNdcRouter input now requires transaction_type (new enum). Previously-working inputs without it will be rejected by the validator.
  • ChangeManagement and RefundProcessing now require cat31_rules / cat33_rules in input to apply filed penalties. Absent rules → ATPCO default (no penalty for voluntary, waived for involuntary) — was a $200 default in 0.6.2.
  • InvoluntaryRebook now requires thresholds.time_change_minutes to mark a time-change as involuntary. Absent → non-involuntary with a warning.
  • @otaip/core exports fetchWithRetry, UnimplementedDomainInputError, DomainInputRequired, applyEU261, applyUsDotIdb, greatCircleDistanceKm and related types.
  • TripPro searchUrl / calendarSearchUrl defaults switched from http:// to https://. Override explicitly for local dev.
  • US DOT IDB compensation caps updated to 14 CFR §250.5 (effective 2025-01-22): $1,075 / $2,150 (was the pre-amendment $775 / $1,550).
  • 11 packages that previously lacked "type": "module" now have it. CommonJS consumers will need to import via import(...) or upgrade their resolution.

Tests

  • 3,092 total passing (was 3,034). 58 new tests:
    • 18 for the new EU261 + US DOT IDB regulation modules
    • 8 for fetchWithRetry
    • 5 new EU261 compensation tests in involuntary-rebook
    • 7 for CLI agent discovery
    • 7 for the ATPCO-default branches in change-management + refund-processing
    • 5 for per-transaction GDS/NDC routing + new validators
    • 3 for the rate-limit retry counting + stub UnimplementedDomainInputError

0.6.2 — Reference OTA Multi-Adapter Fixes

Three bugs in the Sprint H multi-adapter integration caught by Codex review, plus the follow-up to make booking adapter-aware. No behavior change to the single-adapter path; no breaking changes to public interfaces.

Fixed

  • buildApp() now wires MultiSearchService into the search route — ?multi=true was previously unreachable in production. When the ADAPTERS env var is set, a MultiSearchService is constructed automatically; test callers can inject one via buildApp({ multiSearch }).
  • Multi-adapter search now caches its offers — GET /api/offers/:id and POST /api/book no longer 404 on offers returned from the multi path. New SearchService.cacheOffers() is called from the multi branch.
  • returnDate preserved on the multi path — round-trip requests now reach adapters with both segments. Previously the return leg was silently dropped.
  • Adapter-aware booking routing — bookings now route back to the adapter that produced the offer. An offer from a search-only adapter (no book() method) is rejected with HTTP 409 and the adapter name, not silently routed to the default adapter. New AdapterNotBookableError + SearchService.getOfferAdapterSource() + BookingService(defaultAdapter, searchService, bookingAdapters?) registry param.
  • Stale adapterSource cleared on re-cache — a single-adapter search after a multi-adapter search no longer leaves a stale source entry behind that would misroute a subsequent booking.

Documented

  • offer_id collision semantics within a MultiSearchService.search() call are explicitly last-write-wins. Production deployments that need stronger guarantees should namespace IDs with adapterSource at the aggregation boundary.

Tests

  • 12 new tests in examples/ota/src/__tests__/search.test.ts pinning each fix. 3034 total passing, 0 failing.

0.6.1 — Stub Replacements: HotelCarSearch, AITravelAdvisor, SelfServiceRebooking, WaitlistManagement

Four stub agents replaced with real implementations. No new agents; no breaking changes to public interfaces. Existing imports continue to work.

Changed

  • Agent 1.7 HotelCarSearchAgent — Aggregator over injectable hotel/car adapters. Promise.allSettled fan-out with per-adapter timeouts, filters, and sort. Partial-failure tolerant (returns available results with per-source status). New Zod schemas + AgentContract.
  • Agent 1.8 AITravelAdvisorAgent — Rule-based recommender (NOT LLM). Orchestrates AvailabilitySearch + FareShopping, scores offers on price / schedule / airline / connections with business/leisure/default weight profiles. LLMProvider / MockLLMProvider types removed.
  • Agent 5.5 SelfServiceRebookingAgent — Rebooking orchestrator over AvailabilitySearch + ChangeManagement. Ranks priced alternatives by changeFee + fareDifference + taxDifference. Involuntary reasons (schedule change / missed connection / cancellation) waive the change fee.
  • Agent 5.6 WaitlistManagementAgent — Stateful in-memory queue (reference implementation, not durable). Four operations: addEntry, clear, queryStatus, expire. Priority = status tier + fare class type + recency bonus; ties broken by earliest requestedAt. Clearance probability estimated as rate^position with per-booking-class overrides.

Tests

  • 77 new tests across the four agents (19 + 20 + 17 + 21). 3022 total passing, 0 failing.

0.6.0 — Sprint H: Multi-Adapter, OOSD-Native, Full Distribution

The mid-term build plan is complete. Navitaire gains native ONE Order operations, Duffel bridges its order model to OTAIP's AIDM-aligned types, and the Reference OTA searches multiple adapters in parallel with source attribution.

Added

  • Navitaire OrderOperations — NavitaireOrderOperations class implements the full AIDM 24.1 OrderOperations interface: orderCreate, orderRetrieve, orderChange, orderCancel, orderViewHistory. Mock in-memory implementation with NAV-ORD-* IDs, status lifecycle, and OrderEvent tracking. Navitaire is ONE Order certified; this lets OTAIP speak to them natively.
  • Duffel Order Bridge — DuffelOrderBridge class bridges Duffel's native order model to OTAIP's AIDM-aligned Order types. DFL-ORD-* IDs, double-cancel prevention, passenger/payment/offer-item mapping.
  • ChannelCapability Order fields — supportsOrders?: boolean and orderOperations?: ('create' | 'retrieve' | 'change' | 'cancel')[] on ChannelCapability. GdsNdcRouter can use these to decide PNR vs Order path per channel.
  • Multi-adapter search in Reference OTA — MultiSearchService fans out search requests to multiple DistributionAdapter instances via Promise.allSettled, merges results with adapterSource attribution, includes per-source status with timing and error reporting. Activated via ADAPTERS env var (comma-separated) or ?multi=true query param.
  • Updated capability manifests — Navitaire and Duffel now declare supportsOrders: true and the full set of order operations.
  • docs/adapters/oosd-navitaire.md — Navitaire ONE Order adapter documentation.
  • docs/offers-and-orders.md — updated with Sprint H completion section.

Tests

  • 33 new tests (15 Navitaire + 10 Duffel + 8 multi-search). 2985 total passing, 0 failing.

Build plan complete

Sprint Version Delivered
A v0.3.2 Pipeline validator, 9 agent contracts, capability registry
B v0.3.2.1 Tool bridge, catalog generator, EventStore, PnrRetrieval
C v0.3.3 Fallback chain, governance agents, CLI
D+E v0.3.4 Docs overhaul, Reference OTA search flow
F v0.5.0 Reference OTA booking, payment, ticketing
G v0.5.1 Offers & Orders data model (AIDM 24.1)
H v0.6.0 OOSD-native adapters, multi-adapter search

Final stats: 76 agents, 2985 tests, 16 workspace packages, 6 adapters, 14 contracted agents, 2 OOSD-native adapters.

0.5.1 — Sprint G: ONE Order Ready — PNR + Orders Coexist

Native Offers & Orders data model in @otaip/core, aligned with IATA AIDM 24.1 terminology. PNR and Order models coexist through a unified BookingReference bridge — agents accept either and let the adapter decide the underlying model.

Added

  • Order/Offer types — Offer, OfferItem, Order, OrderItem, Service (atomic unit: flight, seat, baggage, meal, lounge, insurance, ancillary), OrderPassenger with TravelDocument + LoyaltyInfo, TicketDocument (ET, EMD-A, EMD-S), OrderPayment, Money (decimal string + ISO 4217)
  • AIDM 24.1 message names — OrderCreate, OrderRetrieve, OrderChange, OrderCancel on the OrderOperations interface. Adapters that support ONE Order implement this directly.
  • OrderEvent — event-driven status changes for Orders (order.created, order.confirmed, order.ticketed, order.changed, order.cancelled, order.payment_received, order.payment_failed, order.refunded). Queue management stays PNR-only.
  • BookingReference bridge — PnrReference | OrderReference union type with constructors (createPnrReference, createOrderReference), type guards (isPnrReference, isOrderReference), accessors (getBookingIdentifier, getBookingOwner), and pnrPassengerToOrderPassenger() converter.
  • Zod schemas for every Order/Offer type — ready for zodToJsonSchema() LLM tool generation.
  • docs/offers-and-orders.md — explains the dual model, AIDM alignment, bridge utilities, and Sprint H roadmap (Navitaire as OOSD adapter target).

Design decisions

  • JSON, not XML. AIDM concepts, not the AIDM XML schema.
  • Queue management stays PNR-only. Orders use OrderEvent.
  • No agent modifications in this release — types and bridge only. Agent integration via BookingReference lands in Sprint H.
  • Navitaire is the target OOSD adapter for Sprint H — they're ONE Order certified and the adapter already exists.

Tests

  • 47 new tests (30 schema validation + 17 bridge utilities). 2952 total passing, 0 failing.

0.5.0 — Sprint F: Reference OTA — Book, Pay, Fly

The reference OTA is now a complete booking application. Users can search flights, select an offer, enter passenger details, pay, and receive a ticket — the full travel e-commerce lifecycle running on OTAIP agents.

Added

  • Booking flow — POST /api/book creates a booking from a search offer with passenger details (title, name, DOB, gender) and contact info. Validates the offer exists in the search cache before booking. Returns a booking reference.
  • Payment flow — POST /api/pay processes a mock payment against a booking reference. Structured for future Stripe integration (PaymentService abstraction) but ships with a mock that always succeeds. No external payment SDK dependency.
  • Ticketing flow (Option B) — POST /api/ticket checks booking status first. If already ticketed, returns existing ticket numbers (idempotent). If not, generates mock 13-digit ticket numbers and updates status. Ticketed bookings cannot be cancelled.
  • Booking management — GET /api/booking/:ref retrieves booking details. POST /api/cancel cancels confirmed (not yet ticketed) bookings.
  • 4 frontend pages — passenger details form (book.html), payment summary + Pay Now (payment.html), full confirmation with tickets + itinerary (confirmation.html), booking lookup + cancel (manage.html). Plain HTML + vanilla JS + Pico CSS.
  • OtaAdapter interface — extends DistributionAdapter with book(), getBooking(), cancelBooking(). MockOtaAdapter extends MockDuffelAdapter with in-memory booking store, reference generation, and status lifecycle (confirmed → ticketed/cancelled).
  • 14 integration tests — booking CRUD, payment, idempotent ticketing, cancellation rules, 2 full end-to-end flows (search → book → pay → ticket, search → book → cancel).

Tests

  • 2905 total passing (14 new + 2891 existing), 0 failing

0.3.4 — Sprint D+E: Docs Overhaul + Reference OTA

Two sprints shipped together: Sprint D rewrites all documentation so the platform's scope is visible at a glance. Sprint E ships a deployable reference OTA that proves OTAIP works end to end — fork it, add your Duffel token, search real flights.

Sprint D — Documentation Overhaul

  • README.md rewrite — 6-adapter comparison table with exact test counts (456 total), 75-agent domain overview across 12 stages, pipeline contract system explanation, CLI usage examples, "Build on OTAIP" section
  • docs/architecture.md — 4 Mermaid diagrams: high-level architecture, pipeline gate sequence, tool bridge flow, EventStore integration
  • docs/agents.md — complete table of all 75 agents with ID, class name, description, contract status (14 contracted)
  • docs/getting-started.md — clone to working demo in 5 minutes
  • docs/adapters/{amadeus,sabre,navitaire,trippro,duffel,haip}.md — 6 adapter docs with capabilities, auth, config, usage, test counts, limitations
  • Agent ID collision fix — PluginManager keeps 9.5, governance agents renumbered to 9.6-9.9

Sprint E — Reference OTA: Search Flow

  • Fastify server (examples/ota/) — POST /api/search, GET /api/offers/:id, GET /health
  • Services — SearchService orchestrates AirportCodeResolver → AvailabilitySearch; OfferService caches offer details
  • Adapter config — DuffelAdapter when DUFFEL_API_TOKEN is set, MockDuffelAdapter when not (works out of the box with zero config)
  • Frontend — plain HTML + vanilla JS + Pico CSS via CDN. Search form with IATA validation, results page with sort-by-price/duration/departure, offer details with price breakdown. No React, no build step.
  • 10 integration tests using Fastify inject + MockDuffelAdapter

Monorepo fix

  • exports.types → src/index.ts — 14 packages had exports["."].types pointing to ./dist/index.d.ts (only exists after build). Changed all to ./src/index.ts matching @otaip/core's pattern. pnpm -r run typecheck now works without a prior build step.

Tests

  • 2891 total passing (10 new OTA tests + 2881 existing), 0 failing

0.3.3 — Sprint C: Governance Agents, Fallback Chain, CLI

The OTAIP build plan is now complete. Sprint C ships the final three steps: the fallback chain engine for automatic channel recovery, four governance agents that monitor the platform's own performance, and a CLI tool for zero-code developer access.

Added

  • Fallback chain engine (executeFallbackChain()) — when the primary routing channel fails, automatically tries each fallback in order. Skips channels whose circuit breaker is open (integrates with ApiAbstraction agent 3.5). Returns a full audit trail of every attempt with durations and error details. 6 tests.
  • Performance Audit Agent (9.5) — reads EventStore agent.executed events, computes success rate, error rate, avg/p95/p99 latency, identifies degraded agents (error rate >15% or p95 >8000ms). AgentContract from day one. 10 tests.
  • Routing Audit Agent (9.6) — reads routing.decided + routing.outcome events, correlates by session, computes per-channel success rates and fallback frequency. AgentContract from day one. 7 tests.
  • Recommendation Agent (9.7) — takes performance + routing audit reports as input, applies deterministic rules, produces typed recommendations (route_adjustment, adapter_health, capacity, config_update). All auto_applicable: false in v1. Confidence based on data volume. AgentContract from day one. 10 tests.
  • Alert Agent (9.8) — configurable threshold monitoring with defaults from the master plan: GDS error rate 5%/15% (warning/critical), NDC 10%/25%, adapter latency p95 >8000ms, 3+ consecutive failures, pipeline rejection rate >20%. AgentContract from day one. 13 tests.
  • CLI tool (@otaip/cli) — new package with 6 commands: otaip search, otaip price, otaip book, otaip adapters, otaip agents (lists all 75 agents with contract status), otaip validate (dry-run pipeline validation). Table format by default, --json for machine output, --verbose for gate details.

Fixed

  • Semver compliance — version 0.3.2.1 (4-part, not valid semver) broke pnpm's workspace:* matching. All packages now use proper 3-part semver 0.3.3.
  • Unused import in chain-engine.ts (FallbackStatus).
  • CLI ESLint config — typescript-eslint can't resolve workspace deps when CWD is a nested package. CLI excluded from root lint glob; covered by pnpm typecheck instead.

Tests

  • 46 net new tests across 5 files (2881 total passing, 0 failing)

Build plan status

All six steps from the master plan are now shipped:

Step Deliverable Sprint
1 Pipeline validator (6 gates) A (v0.3.2)
2 LLM tool layer (bridge + catalog) B (v0.3.2.1)
3 Routing (capability registry + fallback chain) A + C
4 EventStore B (v0.3.2.1)
5 Governance agents (4 agents) C (v0.3.3)
6 CLI tool C (v0.3.3)

0.3.2.1 — Sprint B: LLM Tool Layer + EventStore

The pipeline validator can now talk to LLMs. Sprint B connects the contract infrastructure (shipped in v0.3.2) to the tool-dispatch layer, adds persistent event logging, ships a new agent, and delivers the first demo that runs the full architecture end-to-end.

Added

  • Agent-to-Tool bridge (agentToTool()) — wraps any AgentContract + Agent pair into a ToolDefinition that AgentLoop can register and dispatch. Every tool call runs through the six pipeline gates. Failures throw AgentToolError with structured reason + issues for LLM self-correction. registerAgentTools() batch-converts all contracted agents into a ToolRegistry.
  • Catalog generator — generateMcpTools() (Claude MCP), generateOpenAiFunctions() (OpenAI strict mode, draft-7), generateCatalog() (standalone JSON Schema). All schemas from zodToJsonSchema() — zero hand-written JSON anywhere in the pipeline.
  • EventStore — OtaipEvent discriminated union with 6 event types (agent.executed, routing.decided, routing.outcome, booking.completed, booking.failed, adapter.health). InMemoryEventStore with filter-by-type/session/agent/time-window queries and percentile aggregation (p50/p95/p99). Optional auto-logging from the tool bridge via eventStore option.
  • Agent 3.8 PnrRetrieval — new agent that retrieves an existing PNR/booking by record locator across distribution adapters. AgentContract from day one (actionType: 'query', Zod schemas, semantic validation). Stub retrieval engine — wires to real adapters when ConnectAdapter gains retrieveBooking().
  • GdsNdcRouter registry adapter — buildCarrierCapabilities() converts the existing carrier-channels.json lookup table into ChannelCapability entries for the CapabilityRegistry. 8 equivalence tests prove NDC-preferred, GDS-preferred, and DIRECT-only carriers are correctly encoded. Infrastructure for the full scoring-engine swap.
  • Full pipeline demo (demo/book-flight-full.ts) — first demo that uses the Sprint A/B architecture end-to-end: contract-driven tool definitions, agentToTool() bridge, 6-gate pipeline validator, EventStore logging, pipeline summary. Run with pnpm --filter @otaip/demo book:full.
  • demo/README.md — documents all 5 demo scripts with credential requirements.
  • AGENT_TOOL_NAMES — stable snake_case name map for all 10 contracted agents (e.g. '1.1' → 'availability_search').

Fixed

  • @otaip/connect missing from vitest alias map — the only workspace package without an alias, causing CI to fail when tests imported CapabilityRegistry from @otaip/connect (resolved to dist/ which doesn't exist without a build step).

Tests

  • 39 net new tests across 7 files (2835 total passing, 0 failing)

0.3.2 — Sprint A: Pipeline Contract Foundation

OTAIP moves from a library to a platform: every agent that participates in an LLM-orchestrated or pipeline-composed flow can now declare a machine-verifiable AgentContract, enforced at runtime by six gates (schema, semantic, intent lock, cross-agent consistency, confidence, action classification). Agents without contracts continue to work as direct function calls — the change is purely additive.

Added

  • Pipeline validator (@otaip/core/pipeline-validator) — six-gate runtime with PipelineOrchestrator, session-scoped intent lock, cross-agent consistency checker, confidence gate (floors: query 0.7 / reversible 0.9 / irreversible 0.95 / reference 0.9), action classifier with approval-token enforcement for irreversible mutations, and a 3-retry-per-gate budget
  • Zod → JSON Schema bridge backed by Zod 4's native z.toJSONSchema() — single source of truth for both runtime validation and LLM tool definitions, no new dependencies
  • Shared semantic validators — validateFutureDate, validateIataCode, async resolveAirportStrict, resolveAirlineStrict, resolveFareBasisStrict
  • ReferenceDataProvider interface in @otaip/core with concrete ReferenceAgentDataProvider in @otaip/agents-reference wrapping the Stage 0 agents (no duplicated dataset)
  • Channel capability registry — ChannelCapability types in @otaip/core, CapabilityRegistry class in @otaip/connect, manifests next to each of the 6 adapters (Amadeus, Sabre, Navitaire, TripPro, HAIP, Duffel)
  • 9 agent contracts covering the end-to-end demo flow (search → price → book → ticket):
    • 0.1 AirportCodeResolver, 0.2 AirlineCodeMapper, 0.3 FareBasisDecoder (reference agents)
    • 1.1 AvailabilitySearch, 2.1 FareRuleAgent, 2.4 OfferBuilderAgent (query)
    • 3.1 GdsNdcRouter (query), 3.2 PnrBuilder (mutation_reversible), 4.1 TicketIssuance (mutation_irreversible)
  • Sprint A end-to-end integration test in @otaip/agents-ticketing proving every one of the six gates fires at least once across a full offline run, plus targeted rejection cases (unknown airport → semantic gate; destination drift → intent-lock gate; past date → semantic gate; ticketing without approval → action-class gate)

Fixed

  • OfferEvaluatorAgent time bomb — evaluateOffers() now accepts an optional evaluation_time?: string (ISO 8601). Fixtures dated 2026-04-14 were silently expiring in CI once wall-clock time passed expires_at. Default remains new Date() so existing callers are unaffected.

Scope narrowing

  • GdsNdcRouter (3.1) gets the contract and is now a platform citizen; the internals swap from lookup-table to registry-driven weighted scoring lands in Sprint B (callers unchanged, 467-line test fixture set updates will come with it)

Tests

  • 59 new tests across 9 files (pipeline validator units, capability registry, Sprint A E2E)
  • Full suite: 2796 passing, 3 skipped, 0 failing

0.3.0 — Stage 9: Platform Upgrade

Added

  • Schema-Aware Tool Interface — Zod-validated tool definitions with runtime input/output checking (packages/core/src/tool-interface/)
  • Agent Execution Loop — Deterministic message→tool→response cycle with typed state transitions (packages/core/src/agent-loop/)
  • Lifecycle Hooks — Pre/post hooks on agent actions for logging, metrics, and guardrails (packages/core/src/lifecycle/)
  • Context Budget Manager — Token-aware context management with pluggable compaction strategies (packages/core/src/context/)
  • Retry with Jitter — Shared retry engine with exponential backoff and full jitter, replacing inline retry in BaseAdapter (packages/core/src/retry/)
  • Sub-Agent Spawning — Parent agents can spawn scoped child agents with controlled tool access (packages/core/src/sub-agent/)

Changed

  • BaseAdapter.withRetry() now uses the shared retry engine from @otaip/core (adds jitter, no breaking API changes)
Assets 2
Loading