procurement-ingest/lambdas/po/email_processor/handler.py

149 lines
6.2 KiB
Python
Raw Normal View History

"""
PO email processor Lambda.
Triggered by S3 events when SES delivers a Coupa PO email.
feat: template-first PO parser with fail-closed gate and Bedrock fallback (#105) * Add deterministic template parser for WO emails The workorder-email-processor sends every one of ~22.9k emails/month to an LLM, but ~93.6% are the plain-text "AMAZON UPDATE WO DETAILS" comment template and ~6.4% the HTML "AMAZON assign Work Order" template. Parse those two shapes deterministically, offline, so the AI call is reserved for the long tail. The module is pure (no boto3, no network). try_deterministic_parse classifies by subject, extracts the shared contract fields, and returns a result ONLY when it passes a strict fail-closed validation gate: exact contract-key set, subject/id agreement, the literal "Work Order: <id>" double space, per-type required fields, site-code shape, and a label-bleed guard so a value that over-ran into the next field fails. Any miss, drift, or extractor exception yields None so the caller falls back to the AI extractor -- data is never corrupted, only the fallback rate rises. Refs: #23 * Migrate WO processor to Bedrock and fix comment_id collision Switch the AI path from the Anthropic SDK to bedrock-runtime InvokeModel on the inference profile us.anthropic.claude-haiku-4-5-20251001-v1:0 (BEDROCK_MODEL_ID env), so parsing no longer needs a provider API key or Secrets Manager secret. The EXTRACTION_PROMPT and JSON contract are kept byte-identical, so the AI-fallback output is unchanged. Try the new deterministic template parser first and only call Bedrock on a miss/invalid result. Fix issue #23: the WorkOrderComments range key was work_order_id#<comment_time>, so two emails on one WO with an identical or absent comment time collided and overwrote each other. Derive a 12-hex suffix from the S3 object key alone -- deterministic, so an async retry of the same object is byte-identical (idempotent) while distinct emails get distinct keys -- and keep wall-clock now() out of the key (literal 'nocomment' segment when comment_time is absent). Also emit one CloudWatch EMF line per record (Seahaven/WorkorderIngest ParseOutcome, dimensioned by ParseMethod/TemplateId) for parse-outcome observability, replace the deprecated datetime.utcnow() with datetime.now(timezone.utc), and drop the anthropic dependency. Refs: #23 * Migrate PO processor to Bedrock Switch the PO email processor's AI extraction from the Anthropic SDK to bedrock-runtime InvokeModel on the inference profile us.anthropic.claude-haiku-4-5-20251001-v1:0 (BEDROCK_MODEL_ID env), so it no longer needs a provider API key or Secrets Manager secret. PO parsing stays fully AI -- only the provider changes. The EXTRACTION_PROMPT is kept byte-identical and the Bedrock text output is still decoded with json.loads(..., parse_float=Decimal), which DynamoDB requires (it rejects floats). Replace the deprecated datetime.utcnow() with datetime.now(timezone.utc) and drop the anthropic dependency. * Grant Bedrock IAM, drop Anthropic secrets, add fallback alarm Both stacks moved their processors from the Anthropic API to the Bedrock inference profile us.anthropic.claude-haiku-4-5-20251001-v1:0. Grant each processor role bedrock:InvokeModel + bedrock:InvokeModelWithResponseStream on BOTH the inference-profile ARN AND the per-region foundation-model ARNs for us-east-1/us-east-2/us-west-2 (empty-account) -- the us.* profile routes cross-region, so a profile-only grant AccessDenies at runtime. Remove both anthropic-api-key Secret constructs, their grant_read, and the ANTHROPIC_API_KEY_SECRET_ARN env; add BEDROCK_MODEL_ID. The secrets had RemovalPolicy.RETAIN so they are orphaned, not deleted -- flagged in the README for manual post-deploy deletion and key revocation. Add the workorder-email-processor-template-fallback-rate alarm: a FILL(0) + >=10-sample volume-floor MathExpression over the EMF ParseOutcome metric (15-min periods) that pages when the AI-fallback share exceeds 15% sustained, catching Hexagon template drift. ALARM-only SnsAction to site-alerts, no OK action, NOT_BREACHING, matching the existing stack idiom. * Add offline WO parser test suite Cover the deterministic parser with golden-file tests over 55 real scrubbed .eml fixtures (both comment sub-shapes, username Submitted-By, address present/absent, br+CRLF assign addresses), fail-closed validation-gate rules, adversarial and prompt-injection cases that must route to ai_fallback or parse without corrupting other fields, the issue #23 comment_id idempotency invariants, and the Bedrock-fallback dispatch plus EMF-metric emission with a mocked invoke_model. Extend pytest.ini testpaths to discover the co-located suite, and update tests/conftest.load_handler to put a handler's own directory on sys.path so the WO handler's new `from template_parser import ...` resolves under the existing shared handler tests. Point test_local.py at the new template-first + Bedrock flow. Refs: #23 * Document Bedrock migration and WO parse flow in README Record the provider switch to the Bedrock inference profile (no Anthropic API key or Secrets Manager secret, with the retired secrets flagged for manual deletion), the WO deterministic-template-first + AI-fallback flow, the new ParseOutcome EMF metric and template-fallback-rate alarm, the issue #23 comment_id format change, the +00:00 aware-UTC timestamp shift, and offline test instructions. Refs: #23 * Fix f-string lint and formatting in backfill scripts Drop the f prefix from two f-strings that carry no placeholders (F541) and apply ruff format, so `ruff check` / `ruff format --check` pass in CI. * Emit ParseMethod-only EMF set so fallback alarm can fire The fallback-rate alarm queries the ParseOutcome series keyed on ParseMethod alone, but the emitter published only the joint (ParseMethod, TemplateId) dimension set. CloudWatch materializes exactly the listed dimension sets and does not auto-aggregate, so the alarm's series never received data: it evaluated a constant 0 and could never page on template-drift coverage collapse. Publish both ["ParseMethod"] and ["ParseMethod","TemplateId"] and update the EMF regression test to assert both sets are present. * Commit WO parser .eml fixtures for executable coverage The parser test suite globbed for input .eml fixtures that the repo's `*.eml` ignore rule kept uncommitted, so every parametrized golden and fail-closed test collected zero cases and CI could not exercise the deterministic parser that handles 100% of WO email volume. Add a fixtures-only negation to .gitignore and commit the 55 scrubbed positive samples (50 update-plaintext, 5 assign-html) plus 14 ai-fallback and 3 adversarial fixtures. The ai-fallback set covers each fail-closed reason code (subject_no_match, single_space_work_order, malformed_site_code, label_bleed, creation_time_unparseable, wo_id_mismatch, missing_required_field) and the adversarial set proves the parser is total and confines prompt-injection payloads to comment_text without steering the structured fields. * feature: Add PO template parser scaffold and design doc Mirror WO PR #99's template-first approach for the Coupa PO processor. Two templates identified from a full 3,448-email triage: - coupa_new_po (95.5%): scaffolded; fails closed to the LLM until extract_new_po lands. - coupa_cancellation (2.9%): implemented. Nested contract with recursive validation, Decimal money, and a fail-closed gate. Derived fields (site_code/trade/fiscal_year) are deferred to a shared post-stage. Comments, revisions, multi-line, and non-USD emails fall back to Bedrock. docs/po-template-parser.md records the investigation, decisions, and remaining work. Signed-off-by: Adam Moussa <166072409+amoussa1229@users.noreply.github.com> * Implement PO new_po extraction and value-level gate Replace the extract_new_po scaffold stub with the full section-windowed extractor (duplicate-label anchoring, sentinel ship-to, label-keyed U+2022 bullet split, Decimal money from three anchored contexts only) and add value-level gate rules V1-V13. Both new_po_not_implemented scaffold guards are removed; rules 6-8 (unrecognized_status, multiline_unsupported, non_usd) go live. The gate re-derives every byte proof from the email body so an extractor bug cannot vouch for itself: amount re-serialization with a digit/comma border check (the thousands-separator truncation kill switch), sum(lines)==total against both Total blocks, anchor/supplier identity proofs, USPS address shape on the raw pre-enrichment zip, bullet label discipline, and sentinel/artifact hygiene. Any failure falls closed to the LLM; a validation failure is never a parsed result. Refs: #99 * Wire template-first parse into PO handler with EMF metric Run try_deterministic_parse ahead of the Bedrock extractor and fall back only on a miss/invalid (fail-closed) result. The shared enrich_parsed post-stage and the save_cancellation/save_revision/ save_new_po routing are untouched, so both paths write identical DynamoDB shapes and the po-ingest-site-extractor stream contract is preserved. Each record emits one ParseMethod EMF line (Seahaven/PoIngest/ ParseOutcome, dimension sets [ParseMethod] and [ParseMethod,TemplateId], ReasonCode/po_number ride-alongs) mirroring the WO idiom. The metric fires before the Bedrock call so a Bedrock-side error still records the ai_fallback outcome. Refs: #99 * Add PO fallback-rate alarm retuned for ~57 emails/day The WO alarm's 15-min period and >=10-sample floor assume ~760/day and would be structurally dead at PO volume (a 15-min period holds ~0.6 emails, so the floor is never met). Retune: 6-hour periods (~14.25 expected emails), IF((fb+tmpl)>=8,...) volume floor so a single email can never breach a datapoint (1/8 = 12.5% < 20%), threshold >20% against a ~1% expected baseline, eval 4 / datapoints 2 (24h span) so noise self-clears while total template drift pages within ~12h. No element-wise MAX in the math expression (post-#102 rule); ALARM-only SnsAction to site-alerts, NOT_BREACHING. Gated with 'npx cdk synth po-ingest'. Also add template_parser.py to the bundling cp list -- without it every deployed invocation would ImportError (unit tests cannot catch an asset-bundling omission). Refs: #99, #102 * Add offline PO parser suite with scrubbed fixture corpus 132 tests: golden-file comparison for all 25 positive fixtures (17 single-line new-PO + 8 cancellations, Decimal-exact via parse_float=Decimal), every fail-closed gate reason code covered (body-level triggers via 17 synthetic adversarial .eml mutations, candidate-level via direct validate() unit tests), real multi-line and comment/non-Coupa fallback fixtures, dual line-ending parse identity, two-path enrich/save parity (site-extractor stream guard), V10 URL-id corpus sweep, fixture hygiene (ses_auth pass + scrub-marker leak sweep), and Bedrock dispatch/EMF assertions. The suite loads handler/template_parser via importlib under unique module names and binds the handler's bare sibling imports around exec (tests/conftest.py load_handler gets the same treatment) -- the WO suite caches bare 'handler'/'template_parser' names in sys.modules, and bare imports here would silently bind to the wrong pipeline. moto is imported before the handler so its botocore stubber hook precedes boto3 session creation (the PO conftest chain now loads at pytest session start). Fixtures are scrubbed real S3 samples: transport/auth header values replaced with same-shape placeholders (structure kept so ses_auth still passes), per-file digit ciphers, amounts remapped with sum==total re-established. The .gitignore exception is scoped to the PO fixtures path only. Refs: #99 * Document PO template-first parser and retuned alarm README: PO flow is now template-first with Bedrock fallback; parser/gate section mirroring the WO writeup; Seahaven/PoIngest ParseOutcome namespace and the fallback-rate alarm numbers with their volume justification (deliberately not WO's settings); test-suite and repo-layout updates. Design doc: mark PR #1 complete in progress/checklist sections; document the six value-level gate reason codes and the scaffold guard removal; correct the stale data-access note (default CLI session is 328440206208) and note the ~90-day S3 lifecycle aging of the corpus; record the 2.3 layout addendum (leading Supplier bullet segment, EA evidence lines, summary unit-price tokens, decode-path line endings), the fixture-build pins (address join convention, quantity/unit/price source), the V10 sweep outcome, and resolutions for open questions Q3/Q6. Cross-family review and the Confluence architecture-map update are flagged outstanding for merge. Refs: #99 * Record cross-family review outcome for handler wiring GPT-4.1 cross_review.py run against the real handler diff returned no BLOCK and no security findings; both FIX items verified as no-change-needed (fallback logging already correct; non-dict AI output is the pre-existing issue #101 pattern this PR deliberately does not touch). Refs: #99 * Pin line-item currency to USD in the PO gate The non_usd rule only checked the Total-block top-level currency, so a new_po whose line item read 'for 55,206.00 CAD' under a USD Total block still template-parsed as ok -- a fail-open hole in the fail-closed gate. Every line item's captured currency and its re-derived body token must now byte-equal the proven-USD top-level currency; covered by a line-level CAD adversarial fixture (the existing adv-non-usd only exercised the Total-block variant) and a candidate-mutation unit test. * Scrub residual transport tokens from PO fixtures The first-pass harvest scrub sanitized only the primary SES/DKIM header blocks, leaving the real SES Feedback-ID sender-identity hash in 49 committed fixtures and, on the two non-Coupa fixtures, an embedded second SES block's X-Ses-Receipt, the Exchange cross-tenant UPN ciphertext, and Gmail ARC fh= / X-Gm-* tokens -- exactly the token classes the PR #99 fixture lesson requires placeholdered. Replace each with a same-shape ScrubbedFixture value (byte-safe, CRLF and folding preserved) so header structure and ses_auth behavior are unchanged. * Converge quantity/price to Decimal on both paths EXTRACTION_PROMPT declares quantity and price as JSON strings, so a prompt-obedient Bedrock response stores DynamoDB Strings where the template parser stores Numbers -- divergent attribute types for the same email on the purchase-orders stream. Coerce numeric strings to Decimal in the shared enrich_parsed post-stage (thousands-separator safe; non-numeric strings kept verbatim) so both paths converge; prompt rewording itself remains PR #2 scope. The two-path parity test was circular -- it replayed the parser- derived golden as 'the LLM output', so it could never see the type divergence. It now feeds a prompt-shaped payload (string quantity/ price, LLM-filled site_code) through enrich_parsed and save_new_po, and the fixture-hygiene test now asserts the scrubbed transport-token header classes so fixture regressions are caught. * Coerce bare-int quantity/price to Decimal in enrich_parsed GPT-4.1 cross-family review of the final PR diff (no BLOCK) flagged residual type drift: parse_float=Decimal rules out floats on the LLM path, but a bare JSON int survived as Python int. Coerce it so both parse paths emit one canonical Decimal type. * Scrub fixture-body PII and harden cancellation gate (sec review) /sh-security-review of PR #105 (5 fresh-context detectors + proof-or-kill verifier) confirmed two diff-introduced findings; both fixed here. F3 (medium, real PII in new fixtures): the harvest scrub replaced header tokens but left real third-party PII in message BODIES -- an Amazon contact's name/phone/personal email in non-coupa-02.eml and an internal t.corp.amazon.com ticket URL in comment-02.eml, plus real submitter/attn names recurring across the new_po corpus. Replaced every personal name, phone, personal email, and internal URL with synthetic placeholders (QP-soft-wrap aware) across both .eml bodies and expected goldens. Extended test_fixture_hygiene to scan BODIES (phone shapes, corp URLs, the leaked tokens), closing the header-only gap that let this through. F1 (medium, cancellation gate): _CANCELLATION_SUBJECT was unanchored and matched with .search(), unlike the anchored new_po pattern -- a subject merely ending with the cancellation phrase could be routed to the sticky- Cancelled write. Fully anchored it and switched to .match, and added a body-corroboration gate (the real Coupa body independently restates 'Purchase Order #<po> ... has been cancelled'); a near-miss/misrouted subject whose body does not corroborate now fails closed to the LLM (new reason code cancellation_body_unconfirmed). Pre-existing (advisory, not this PR): the LLM-fallback else->save_new_po dispatch and undelimited extraction prompt (issue #101 family) are byte-identical to main and unchanged here. 401 tests pass; ruff/format clean; cdk synth po-ingest clean. --------- Signed-off-by: Adam Moussa <166072409+amoussa1229@users.noreply.github.com>
2026-07-16 17:50:59 -04:00
Parses the raw email, tries the deterministic template parser first, falls
back to Claude on Bedrock for structured extraction on a miss/invalid result,
then writes the result to the purchase-orders DynamoDB table.
feat: decompose email-processor handlers into flat siblings + lazy boto3 clients (refactor phase 5) (#113) Both email-processor God-handlers split along the seams that already work in the flat-sibling pattern established by lambdas/shared/, so bare-name imports keep working under the existing bundling glob. PO (5-way split): handler.py keeps only the event loop, fail-closed auth, and email_type routing. extraction.py holds extract_with_claude and _EMAIL_TAG_RE, importing EXTRACTION_PROMPT from prompts.py and parse_raw_email from shared/email_parsing.py rather than recreating a PO-local copy. enrichment.py is a pure code move of enrich_parsed and pad_zip (PO-only; WO has no enrichment stage) with zero behavior change. telemetry.py holds the EMF ParseMethod emit wrappers. persistence.py holds _write_fields/_merge_update/save_*, collapsing the byte-identical save_new_po/save_revision bodies into one _save_merge helper that both now call through, preserving the sticky Cancelled ConditionExpression guard for both callers; save_cancellation stays separate. WO (5 concerns, no enrichment stage): the handler loop keeps validate_ai_fallback and the re.fullmatch(r"[0-9]+", work_order_id) key guard ahead of both save_work_order and save_event, since the guard protects the DynamoDB partition key and the '#'-delimited comment_id range-key segment. _header_date_iso and comment_id determinism stay colocated with persistence.py's save_event for the retry-idempotent event_id key. EXTRACTION_PROMPT (PO) moves to prompts.py with cross-reference headers to derived_fields.py's authoritative trade/site/fiscal rule tables; handler.py re-exports it (from prompts import EXTRACTION_PROMPT) since four tests dereference handler.EXTRACTION_ PROMPT directly. WO's prompt moves the same way. I/O modules (extraction.py's bedrock client, persistence.py's dynamodb resource, handler.py's s3 client) get lazy cached boto3 accessors; pure modules (enrichment.py, prompts.py, telemetry.py) import no boto3. Test monkeypatch surfaces move to the module that now owns the client (e.g. persistence.dynamodb) everywhere tests patch it, and the moto-before-handler-import ordering in _po_parser_support.py is preserved so the moto-backed suites don't hit real AWS. Behavior-preservation pins, verified with tests: PO still emits ParseMethod=ai_fallback before the Bedrock call, with ai_fallback_rejected as the additive second datapoint on rejection. WO still emits after its gate with mutually-exclusive ai_fallback / ai_fallback_rejected. Shadow DerivedFieldAgreement telemetry stays ai_fallback-only. derived_fields.py is untouched (diff against feature/phase-3-shared-extraction is empty). handler(event, context) signatures and the save_* public contract are unchanged on both pipelines; goldens unchanged. PO_EXPECTED_TOP_LEVEL_MODULES and its WO equivalent in tests/test_bundle_consistency.py are updated for the new sibling modules so the AST bundle-consistency test still fails on an unshipped or uncommented-out sibling.
2026-07-20 15:34:53 -04:00
The concerns are split across flat sibling modules (all bundled into the same
Lambda asset, so bare-name imports resolve):
* extraction.py -- extract_with_claude + the Bedrock client
* enrichment.py -- enrich_parsed + pad_zip + derived-field classification
* telemetry.py -- the EMF ParseMethod / derived-agreement emit wrappers
* persistence.py -- the purchase-orders merge-writes (save_*)
* prompts.py -- EXTRACTION_PROMPT (re-exported below for tests)
This module keeps the S3 event loop, fail-closed sender auth, and email_type
routing.
"""
import logging
import boto3
import sentry_init # noqa: F401
feat: extract lambdas/shared/ — single-source ses_auth, web_ui auth, email parsing, EMF emitter (refactor phase 3) (#111) Four modules move into the handbook-mandated lambdas/shared/ location, collapsing duplicated logic that had to be kept in sync by hand across the PO and WO pipelines: - ses_auth.py: the PO and WO copies were verified sha256-identical against the feature/phase-7-ops-recovery baseline before the move (no drift since the last audit). shared/ses_auth.py is the exact bytes of that one copy; both originals are git rm'd (the PO copy via rename, the WO copy as a straight delete). Bundling lands the module flat in /asset-output for both email processors, so the handlers keep `from ses_auth import authenticate_inbound_email` unchanged — zero handler diff for this move, which is what keeps fail-closed auth byte-identical through the change. - web_ui_auth.py: extracts the byte-identical _get_auth_token / _header / is_authenticated block plus the four token-cache globals out of both web_ui handlers. The per-stack INFRA-74 comments stay in each handler as-is (deliberately drifted wording, stack-specific) rather than being unified into the shared module. Fail-closed semantics (unset ARN or Secrets Manager exception -> deny) are unchanged. - email_parsing.py: parse_raw_email ships as the superset version that returns cc unconditionally. WO's output is bit-identical to before; PO simply ignores the cc field rather than being "cleaned up" to consume it. No second variant is kept. - emf.py: a generic emitter parameterized by namespace, dimension sets, and properties. Every call site's emitted EMF envelope is unchanged, including the load-bearing [["ParseMethod"],["ParseMethod","TemplateId"]] dimension-set shape the alarms and metric filters depend on. Emission ordering is untouched: PO still emits ai_fallback before the Bedrock call, WO still emits its mutually-exclusive ai_fallback/ai_fallback_rejected after its gate. The deliberate-double-count comments survive. _emit_derived_agreement_metric was found living inside derived_fields.py, so per the DERIVED-FIELDS exception it is left as a third, unconverted copy (derived_fields.py and the shadow DerivedFieldAgreement telemetry stay untouchable while that bake runs) — a comment there points at shared/emf.py for the eventual follow-up. Bundling: both email-processor cdk bundling commands gain a trailing `cp shared/*.py /asset-output/` (they were already cp-only post-Phase 7, so no pip step or manylinux pin is reintroduced). Both web_ui functions gain the same widened-root staging so web_ui_auth.py ships beside their handler; site_extractor's from_asset is untouched. Tests: PO_EXPECTED_TOP_LEVEL_MODULES gains the shared modules that now ship, the AST sibling-import check resolves imports whose source now lives under shared/, and the new shared cp line has its own revert/mutation detection. _SIBLING_MODULES resolution and _po_parser_support.py now load ses_auth/email_parsing/emf from shared/; the two-copy ses_auth byte-identity fixture-hygiene test is retired as obsolete now that there is one copy, and the ses_auth fixture parameterization over two identical copies is dropped. The sys.modules save/restore dance for template_parser (still duplicated per-pipeline) is left in place.
2026-07-20 13:38:23 -04:00
from email_parsing import parse_raw_email
feat: decompose email-processor handlers into flat siblings + lazy boto3 clients (refactor phase 5) (#113) Both email-processor God-handlers split along the seams that already work in the flat-sibling pattern established by lambdas/shared/, so bare-name imports keep working under the existing bundling glob. PO (5-way split): handler.py keeps only the event loop, fail-closed auth, and email_type routing. extraction.py holds extract_with_claude and _EMAIL_TAG_RE, importing EXTRACTION_PROMPT from prompts.py and parse_raw_email from shared/email_parsing.py rather than recreating a PO-local copy. enrichment.py is a pure code move of enrich_parsed and pad_zip (PO-only; WO has no enrichment stage) with zero behavior change. telemetry.py holds the EMF ParseMethod emit wrappers. persistence.py holds _write_fields/_merge_update/save_*, collapsing the byte-identical save_new_po/save_revision bodies into one _save_merge helper that both now call through, preserving the sticky Cancelled ConditionExpression guard for both callers; save_cancellation stays separate. WO (5 concerns, no enrichment stage): the handler loop keeps validate_ai_fallback and the re.fullmatch(r"[0-9]+", work_order_id) key guard ahead of both save_work_order and save_event, since the guard protects the DynamoDB partition key and the '#'-delimited comment_id range-key segment. _header_date_iso and comment_id determinism stay colocated with persistence.py's save_event for the retry-idempotent event_id key. EXTRACTION_PROMPT (PO) moves to prompts.py with cross-reference headers to derived_fields.py's authoritative trade/site/fiscal rule tables; handler.py re-exports it (from prompts import EXTRACTION_PROMPT) since four tests dereference handler.EXTRACTION_ PROMPT directly. WO's prompt moves the same way. I/O modules (extraction.py's bedrock client, persistence.py's dynamodb resource, handler.py's s3 client) get lazy cached boto3 accessors; pure modules (enrichment.py, prompts.py, telemetry.py) import no boto3. Test monkeypatch surfaces move to the module that now owns the client (e.g. persistence.dynamodb) everywhere tests patch it, and the moto-before-handler-import ordering in _po_parser_support.py is preserved so the moto-backed suites don't hit real AWS. Behavior-preservation pins, verified with tests: PO still emits ParseMethod=ai_fallback before the Bedrock call, with ai_fallback_rejected as the additive second datapoint on rejection. WO still emits after its gate with mutually-exclusive ai_fallback / ai_fallback_rejected. Shadow DerivedFieldAgreement telemetry stays ai_fallback-only. derived_fields.py is untouched (diff against feature/phase-3-shared-extraction is empty). handler(event, context) signatures and the save_* public contract are unchanged on both pipelines; goldens unchanged. PO_EXPECTED_TOP_LEVEL_MODULES and its WO equivalent in tests/test_bundle_consistency.py are updated for the new sibling modules so the AST bundle-consistency test still fails on an unshipped or uncommented-out sibling.
2026-07-20 15:34:53 -04:00
from enrichment import enrich_parsed
from extraction import extract_with_claude
from persistence import save_cancellation, save_new_po, save_revision
# EXTRACTION_PROMPT is re-exported so handler.EXTRACTION_PROMPT still resolves
# for the tests that dereference it as a module attribute.
from prompts import EXTRACTION_PROMPT # noqa: F401
Add fail-closed SES sender authentication (INFRA-107) (#98) * Add fail-closed SES sender authentication The From header and any raw-MIME Authentication-Results copies are attacker-forgeable, so a forged email to apm@int.seahaven.com or amazon_po@int.seahaven.com could create or mutate a WO/PO (INFRA-107, CRITICAL). Both S3-triggered email processors now authenticate the sender against the Authentication-Results header SES itself prepends at delivery: only the topmost header is consulted, its authserv-id must be amazonses.com, and it must carry dkim=pass for a domain in the per-pipeline ALLOWED_DKIM_DOMAINS env var (comma-separated, set in CDK so ops can adjust without code changes). Allowlists come from live traffic observed 2026-07-15 on both ingest buckets: WO mail arrives via the apm@ Google Groups forward, which re-signs as seahaven.com (the hxgnsmartcloud.com signature does not survive the forward); PO mail passes for amazon.coupahost.com. amazonses.com also passes on PO mail but is deliberately excluded -- every SES customer's outbound mail passes for it. Every failure path (env var unset, header missing or unparseable, verdict fail, unaligned domain) rejects the email: a structured warning with the reason and S3 key is logged and the record skipped without erroring the invocation, so rejected mail causes no Lambda retries or DLQ messages. Handler signatures and event sources are unchanged. Refs: INFRA-107 * Harden AR parser per cross-family review Cross-family (GPT-4.1) review findings: terminate the dkim result token at end-of-clause, whitespace, or a comment so a value like "dkim=pass-fake" can never be read as a pass; normalize trailing dots off allowlist entries so "seahaven.com." matches; make the compat32 parser policy explicit. Adds tests for result-token boundaries, comments after the result, quoted domain values, and folding inside a dkim clause. Refs: INFRA-107 * Harden AR parsing and alarm on sender-auth rejects The SES-stamped Authentication-Results value echoes attacker-controlled SMTP-session tokens (envelope-from, helo, header.from) as their own semicolon-delimited property clauses. A naive split(";") tore an RFC 5321 quoted-local-part MAIL FROM apart and manufactured a forged dkim=pass clause, so a fully spoofed email was accepted on the genuinely SES-stamped topmost header. Tokenise comment- and quoted-string-aware (RFC 8601 / RFC 5322): strip CFWS comments, split clauses only on semicolons outside a quoted-string, and fail closed on unbalanced quotes/comments so a ';' inside a quoted pvalue can never start a clause. Rejected mail returns normally (no error, no retry, no DLQ message), so a signing-domain drift or a wrong allowlist would silently discard 100% of legitimate mail while every alarm stayed green. Add a CloudWatch Logs metric filter + alarm on the sender_auth_rejected warning to both stacks so a false-reject storm pages instead of vanishing. This is also the safety net for the WO seahaven.com allowlist assumption, which must be validated against a live SES-stamped header (a plain Gmail auto-forward re-signs under the sending Workspace domain, not seahaven.com). Refs: INFRA-107 * chore: retrigger CI (no run recorded for 7c74ac1) * Fix quoted-AUID DKIM domain spoof in sender auth Resolve three confirmed /sh-security-review findings on the fail-closed SES sender-authentication control. HIGH: header.i/header.d domain extraction was not quoted-string aware. An attacker with a valid DKIM key for their own domain could set an RFC 6376-legal AUID such as i="@seahaven.com"@attacker.com; the naive extractor stopped at the closing quote and returned seahaven.com, accepting forged mail. Extraction now tokenises the clause with the same quoted-string discipline already used for clause splitting: header.d (the plain signing domain) is authoritative when present, otherwise the header.i domain is the part after the AUID's LAST top-level "@", so a "@" inside a quoted local-part is treated as signer-controlled label text and yields the true signer (attacker.com), not seahaven.com. LOW: the topmost-header parse ran outside evaluate_sender_authentication's try/except, so an unexpected parser exception on crafted input could propagate into the handler and Lambda async retries/DLQ. The parse now fails CLOSED with an authentication_results_unparseable reason. MEDIUM: the sender_auth_rejected alarm used Sum>=3 over 15 min, blind to a low-volume total-reject outage (a trickle that never sums to 3). Both stacks now alarm on >=1 reject per 5-min period with evaluation_periods=3 / datapoints_to_alarm=2, so a sustained reject condition pages even at one reject per period while a lone stray probe self-clears. Refs: INFRA-107 * Load Lambda function dir on sys.path in tests Rebasing INFRA-107 onto main folded #95's pytest suite into this branch's tests. The unified conftest loads the PO/WO handlers by file path, and handler.py now does `from ses_auth import authenticate_inbound_email` -- a bare sibling import that resolves in the Lambda only because the runtime puts each function's own directory on sys.path. The shared load_handler now adds that directory so the handler tests import correctly alongside the sender-auth tests. Refs: INFRA-107 * Note #97 test files in README directory tree The rebase onto main brought in #97's tests/requirements.txt and tests/test_po_merge.py. List both in the directory tree so it matches the tree on disk. Refs: INFRA-107 * Document INFRA-107 forwarder-binding risk acceptance Record the accepted risk that WO sender auth binds to the apm@ forward's re-signing domain (seahaven.com) rather than the Hexagon originator; the apm@ Google Group's restricted posting policy is the load-bearing control (escalates to HIGH if the group is opened to external posting). Also correct the sender-auth-rejected alarm docs to match the shipped config (>=1 per 5-min, 2-of-3 datapoints, not the superseded >=3/15min) and note the SES-AR-01/02 parser hardening follow-ups. Refs: INFRA-107
2026-07-15 20:58:47 -04:00
from ses_auth import authenticate_inbound_email
feat: decompose email-processor handlers into flat siblings + lazy boto3 clients (refactor phase 5) (#113) Both email-processor God-handlers split along the seams that already work in the flat-sibling pattern established by lambdas/shared/, so bare-name imports keep working under the existing bundling glob. PO (5-way split): handler.py keeps only the event loop, fail-closed auth, and email_type routing. extraction.py holds extract_with_claude and _EMAIL_TAG_RE, importing EXTRACTION_PROMPT from prompts.py and parse_raw_email from shared/email_parsing.py rather than recreating a PO-local copy. enrichment.py is a pure code move of enrich_parsed and pad_zip (PO-only; WO has no enrichment stage) with zero behavior change. telemetry.py holds the EMF ParseMethod emit wrappers. persistence.py holds _write_fields/_merge_update/save_*, collapsing the byte-identical save_new_po/save_revision bodies into one _save_merge helper that both now call through, preserving the sticky Cancelled ConditionExpression guard for both callers; save_cancellation stays separate. WO (5 concerns, no enrichment stage): the handler loop keeps validate_ai_fallback and the re.fullmatch(r"[0-9]+", work_order_id) key guard ahead of both save_work_order and save_event, since the guard protects the DynamoDB partition key and the '#'-delimited comment_id range-key segment. _header_date_iso and comment_id determinism stay colocated with persistence.py's save_event for the retry-idempotent event_id key. EXTRACTION_PROMPT (PO) moves to prompts.py with cross-reference headers to derived_fields.py's authoritative trade/site/fiscal rule tables; handler.py re-exports it (from prompts import EXTRACTION_PROMPT) since four tests dereference handler.EXTRACTION_ PROMPT directly. WO's prompt moves the same way. I/O modules (extraction.py's bedrock client, persistence.py's dynamodb resource, handler.py's s3 client) get lazy cached boto3 accessors; pure modules (enrichment.py, prompts.py, telemetry.py) import no boto3. Test monkeypatch surfaces move to the module that now owns the client (e.g. persistence.dynamodb) everywhere tests patch it, and the moto-before-handler-import ordering in _po_parser_support.py is preserved so the moto-backed suites don't hit real AWS. Behavior-preservation pins, verified with tests: PO still emits ParseMethod=ai_fallback before the Bedrock call, with ai_fallback_rejected as the additive second datapoint on rejection. WO still emits after its gate with mutually-exclusive ai_fallback / ai_fallback_rejected. Shadow DerivedFieldAgreement telemetry stays ai_fallback-only. derived_fields.py is untouched (diff against feature/phase-3-shared-extraction is empty). handler(event, context) signatures and the save_* public contract are unchanged on both pipelines; goldens unchanged. PO_EXPECTED_TOP_LEVEL_MODULES and its WO equivalent in tests/test_bundle_consistency.py are updated for the new sibling modules so the AST bundle-consistency test still fails on an unshipped or uncommented-out sibling.
2026-07-20 15:34:53 -04:00
from telemetry import _emit_parse_method_metric
PO ai-fallback fail-closed gate + prompt hardening (refactor phase 1) (#108) * feat: PO ai-fallback fail-closed gate + prompt hardening, parity with #104 (refactor phase 1) Ports WO's #104 AI-fallback security hardening to the PO email processor, adapted for PO's nested contract instead of copying the WO gate verbatim. validate_ai_fallback() (template_parser.py) fail-closes raw Bedrock output before it reaches enrich_parsed or any dispatch/save: recursive key-set check with missing-key normalization (nested contract: supplier{}, ship_to{}, line_items[]); po_number checked against the same hardened prefix+hyphen+digits regex family that guards the DynamoDB partition key the handler builds from it (rejects fullwidth-digit and trailing-artifact injection); email_type enforced against the {new_po, revision, cancellation} allow-list before dispatch so a miss can never fall into the else -> save_new_po branch; money fields accept Decimal/int/None only, matching PO's parse_float=Decimal decode (a float-typed check would be wrong here). A gate failure emits ParseMethod=ai_fallback_rejected and `continue`s to the next record -- it never raises, so attacker-controlled input can't churn the retry/DLQ path. extract_with_claude() wraps the untrusted email in an <email> data block and neutralizes forged <email>-tag lookalikes in the body with the same linear-time regex approach as WO's _EMAIL_TAG_RE, and sets temperature=0 on the Bedrock call. Deliberate double-count: PO emits ParseMethod=ai_fallback before the Bedrock call (so a Bedrock-side error still records the outcome), so a rejected email always produces both an ai_fallback datapoint (pre-call) and an ai_fallback_rejected datapoint (post-gate). This is intentional, not a bug -- documented in handler.py, template_parser.py, and the README. cdk/po_stack.py: in-place property update to the existing po-email-processor-template-fallback-rate alarm (same logical ID, no rename/replacement) -- the fb/(fb+tmpl) expression is left byte-identical to its pre-Phase-1 form and ai_fallback_rejected is deliberately excluded from the numerator/denominator/volume floor, since folding it in as WO does would double-count every rejection (PO's pre-call emit already counts it once via fb). A net-new EmailProcessorAiFallbackRejectedAlarm watches the rejected series on its own, retuned for ~57 emails/day with the 6h/IF-floor/eval-4/ datapoints-2 idiom (not WO's 5-minute sparse idiom, which is structurally dead at PO volume). Both alarms remain ALARM-only to site-alerts, NOT_BREACHING, with no element-wise MAX in the math (post-#102 rule). * Block "Cancelled" po_status off the AI cancellation route The AI-fallback gate type-checked po_status but let any string through, unlike the template path which never emits "Cancelled" on a new_po. Dispatch routes on email_type, so an AI-path new_po or revision carrying po_status="Cancelled" would reach save_new_po/save_revision and cancel a live PO via _merge_update's sticky-cancel write without ever hitting save_cancellation. Reject the exact sticky marker on any non-cancellation email_type so the AI path matches the template path's guard; arbitrary non-marker status strings still pass. email_type is already validated to the enum before this check, and a cancellation reaches save_cancellation (which hardcodes the status), so po_status stays irrelevant on that route.
2026-07-17 14:50:47 -04:00
from template_parser import try_deterministic_parse, validate_ai_fallback
logger = logging.getLogger()
logger.setLevel(logging.INFO)
feat: decompose email-processor handlers into flat siblings + lazy boto3 clients (refactor phase 5) (#113) Both email-processor God-handlers split along the seams that already work in the flat-sibling pattern established by lambdas/shared/, so bare-name imports keep working under the existing bundling glob. PO (5-way split): handler.py keeps only the event loop, fail-closed auth, and email_type routing. extraction.py holds extract_with_claude and _EMAIL_TAG_RE, importing EXTRACTION_PROMPT from prompts.py and parse_raw_email from shared/email_parsing.py rather than recreating a PO-local copy. enrichment.py is a pure code move of enrich_parsed and pad_zip (PO-only; WO has no enrichment stage) with zero behavior change. telemetry.py holds the EMF ParseMethod emit wrappers. persistence.py holds _write_fields/_merge_update/save_*, collapsing the byte-identical save_new_po/save_revision bodies into one _save_merge helper that both now call through, preserving the sticky Cancelled ConditionExpression guard for both callers; save_cancellation stays separate. WO (5 concerns, no enrichment stage): the handler loop keeps validate_ai_fallback and the re.fullmatch(r"[0-9]+", work_order_id) key guard ahead of both save_work_order and save_event, since the guard protects the DynamoDB partition key and the '#'-delimited comment_id range-key segment. _header_date_iso and comment_id determinism stay colocated with persistence.py's save_event for the retry-idempotent event_id key. EXTRACTION_PROMPT (PO) moves to prompts.py with cross-reference headers to derived_fields.py's authoritative trade/site/fiscal rule tables; handler.py re-exports it (from prompts import EXTRACTION_PROMPT) since four tests dereference handler.EXTRACTION_ PROMPT directly. WO's prompt moves the same way. I/O modules (extraction.py's bedrock client, persistence.py's dynamodb resource, handler.py's s3 client) get lazy cached boto3 accessors; pure modules (enrichment.py, prompts.py, telemetry.py) import no boto3. Test monkeypatch surfaces move to the module that now owns the client (e.g. persistence.dynamodb) everywhere tests patch it, and the moto-before-handler-import ordering in _po_parser_support.py is preserved so the moto-backed suites don't hit real AWS. Behavior-preservation pins, verified with tests: PO still emits ParseMethod=ai_fallback before the Bedrock call, with ai_fallback_rejected as the additive second datapoint on rejection. WO still emits after its gate with mutually-exclusive ai_fallback / ai_fallback_rejected. Shadow DerivedFieldAgreement telemetry stays ai_fallback-only. derived_fields.py is untouched (diff against feature/phase-3-shared-extraction is empty). handler(event, context) signatures and the save_* public contract are unchanged on both pipelines; goldens unchanged. PO_EXPECTED_TOP_LEVEL_MODULES and its WO equivalent in tests/test_bundle_consistency.py are updated for the new sibling modules so the AST bundle-consistency test still fails on an unshipped or uncommented-out sibling.
2026-07-20 15:34:53 -04:00
# Lazily-built, cached S3 client. Public name ``s3`` is preserved so the
# monkeypatch attribute is unchanged; building at first CALL (not import) keeps
# the moto-before-handler invariant and honors any patched fake.
s3 = None
Land safe fixes from 2026-06-17 security sweep (#97) * Remove gratuitous KMS grant on shared DynamoDB CMK wo-email-processor held grant_encrypt_decrypt on the shared seahaven-dynamodb CMK, but the WorkOrders/WorkOrderComments tables are not encrypted with that CMK. The grant was dead weight that extended the WO processor's decrypt reach to the CMK protecting the purchase-orders table (cross-stack decrypt). Drop it to restore least privilege; re-add as part of the table CMK migration (INFRA-6). Refs: INFRA-6 * Require Secrets Manager key for Anthropic client Remove the silent fallback to a plaintext ANTHROPIC_API_KEY env var in both email processors; require ANTHROPIC_API_KEY_SECRET_ARN and raise if absent so a misconfigured deploy fails loudly instead of using an unmanaged key. Adapted from f175323 on security/sweep-2026-06-17. The From-header sender-domain allowlist from that commit is intentionally dropped: the From header is spoofable (INFRA-107, confirmed critical) and sender authentication is being reworked in a separate PR. Refs: INFRA-107 * Merge PO revisions and handle out-of-order events save_revision did a full put_item overwrite, so a revision omitting line_items/supplier permanently deleted them. save_new_po used a conditional put that silently dropped the PO when an out-of-order cancellation had already created a skeleton row. Switch both to field-level merge update_items: a revision now SETs only the fields it carries, and a new_po backfills data into a pre-existing Cancelled skeleton while preserving the Cancelled status. No email can now delete data established by an earlier one. * Gate web UIs behind auth and escape currency XSS The po-web-ui and workorder-web-ui handlers had no auth: any invocation path returned the full PO/WO DB. Add a fail-closed shared-secret gate (X-Auth-Token / Bearer, constant-time compared to WEB_UI_AUTH_TOKEN) so a future re-attached Function URL cannot re-expose the data (URLs removed under INFRA-74). Wire the token from the SSM String param /procurement-ingest/web-ui-auth-token. Also fix stored XSS in po-web-ui fmt_currency: the non-numeric fallback returned str(val) unescaped, so a prompt-injected email could make Claude emit total_amount as <script>. Escape it. Refs: INFRA-74 * Document sweep security fixes and merge semantics Update the README for the 2026-06-17 security sweep: required Secrets Manager key (no plaintext env fallback), web UI auth gate + SSM token setup step, output-escaping note, and the new PO revision/cancellation merge behavior. Adapted from d91f45e on security/sweep-2026-06-17; the sender allowlist documentation is dropped along with the allowlist itself (deferred to the INFRA-107 sender-authentication rework). Refs: INFRA-107 * fix: resolve web UI auth token from Secrets Manager at runtime Replace the plaintext SSM String parameter with a Secrets Manager secret referenced by ARN only. The token is fetched and cached at module level on first invocation, keeping shared secrets out of CloudFormation templates and Lambda environment variables. Refs: PR-97 * Add TTL to web UI auth token cache for rotation The web-ui handlers cached the Secrets Manager auth token at module level with no expiry, so a rotated secret was only picked up when the warm container recycled — an emergency rotation could take hours to take effect. Cache the fetched value for a 5-minute TTL instead, so a rotated token propagates within the TTL while still avoiding a Secrets Manager call on every request. Still fails closed when the secret is unset or unreadable. Refs: INFRA-74 * Log Secrets Manager failures in web UI auth token fetch The web UI auth gate correctly fails closed when the shared token cannot be read, but _get_auth_token() swallowed every exception silently. A Secrets Manager permission or config error then made every request 401 with no operational signal, leaving an outage indistinguishable from ordinary unauthenticated traffic. Add a module-level logger to both web_ui handlers and log the fetch failure with logger.exception() in the except block before returning None. Behavior is unchanged (still fails closed); the failure is now visible in CloudWatch. The secret value is never logged. The two handlers stay byte-consistent in the mirrored _get_auth_token() region. The companion finding on the CDK import of the shared procurement-ingest/web-ui-auth-token secret was evaluated and left as-is: the token is a single secret shared by both the PO and WO stacks, so from_secret_name_v2 (which scopes grant_read via the standard 6-char suffix wildcard) is correct; making it a CDK-managed Secret in both stacks would collide the two stacks on the same explicit secret name at deploy time. Refs: INFRA-74 * Make Cancelled PO status sticky via atomic write The PO merge path read status with a get_item (_is_cancelled) and then wrote with an unconditional update_item. Two defects followed from this: - Race (Issue A): a cancellation landing between the read and the write was silently un-cancelled by a revision carrying a non-cancelled po_status — a TOCTOU on a table with concurrent email processing. - Over-broad strip (Issue B): save_revision dropped po_status whenever the PO was Cancelled, so legitimate status updates on non-cancelled POs and status-less revisions were affected rather than only the true un-cancel transition. Enforce the invariant server-side instead. "Cancelled" is a sticky, authoritative status: once set, later new_po/revision emails may enrich other fields but must never move it to a non-cancelled status. When the payload carries a non-cancelled po_status, _merge_update issues the update_item guarded by ConditionExpression "attribute_not_exists(po_status) OR po_status <> :marker", evaluated atomically at write time, so a cancellation that lands first always wins. On ConditionalCheckFailedException the same fields are re-written without po_status/cancelled_at, enriching the record while Cancelled sticks. Payloads with no status change, or an already -Cancelled status, take a plain merge — the status is only ever suppressed on a real un-cancel. This removes the non-atomic get_item from the write path; _is_cancelled is deleted. Key schema and attribute names are unchanged, so the cross-stack purchase-orders contract (read-only by seahaven-slack-bot) holds. Add moto-backed tests covering un-cancel suppression with field enrichment, status-less merge onto a Cancelled PO, legitimate status updates on non-cancelled POs, new_po backfill of a Cancelled skeleton, fresh create/merge, and authoritative save_cancellation. Refs: #97
2026-07-15 20:17:46 -04:00
feat: decompose email-processor handlers into flat siblings + lazy boto3 clients (refactor phase 5) (#113) Both email-processor God-handlers split along the seams that already work in the flat-sibling pattern established by lambdas/shared/, so bare-name imports keep working under the existing bundling glob. PO (5-way split): handler.py keeps only the event loop, fail-closed auth, and email_type routing. extraction.py holds extract_with_claude and _EMAIL_TAG_RE, importing EXTRACTION_PROMPT from prompts.py and parse_raw_email from shared/email_parsing.py rather than recreating a PO-local copy. enrichment.py is a pure code move of enrich_parsed and pad_zip (PO-only; WO has no enrichment stage) with zero behavior change. telemetry.py holds the EMF ParseMethod emit wrappers. persistence.py holds _write_fields/_merge_update/save_*, collapsing the byte-identical save_new_po/save_revision bodies into one _save_merge helper that both now call through, preserving the sticky Cancelled ConditionExpression guard for both callers; save_cancellation stays separate. WO (5 concerns, no enrichment stage): the handler loop keeps validate_ai_fallback and the re.fullmatch(r"[0-9]+", work_order_id) key guard ahead of both save_work_order and save_event, since the guard protects the DynamoDB partition key and the '#'-delimited comment_id range-key segment. _header_date_iso and comment_id determinism stay colocated with persistence.py's save_event for the retry-idempotent event_id key. EXTRACTION_PROMPT (PO) moves to prompts.py with cross-reference headers to derived_fields.py's authoritative trade/site/fiscal rule tables; handler.py re-exports it (from prompts import EXTRACTION_PROMPT) since four tests dereference handler.EXTRACTION_ PROMPT directly. WO's prompt moves the same way. I/O modules (extraction.py's bedrock client, persistence.py's dynamodb resource, handler.py's s3 client) get lazy cached boto3 accessors; pure modules (enrichment.py, prompts.py, telemetry.py) import no boto3. Test monkeypatch surfaces move to the module that now owns the client (e.g. persistence.dynamodb) everywhere tests patch it, and the moto-before-handler-import ordering in _po_parser_support.py is preserved so the moto-backed suites don't hit real AWS. Behavior-preservation pins, verified with tests: PO still emits ParseMethod=ai_fallback before the Bedrock call, with ai_fallback_rejected as the additive second datapoint on rejection. WO still emits after its gate with mutually-exclusive ai_fallback / ai_fallback_rejected. Shadow DerivedFieldAgreement telemetry stays ai_fallback-only. derived_fields.py is untouched (diff against feature/phase-3-shared-extraction is empty). handler(event, context) signatures and the save_* public contract are unchanged on both pipelines; goldens unchanged. PO_EXPECTED_TOP_LEVEL_MODULES and its WO equivalent in tests/test_bundle_consistency.py are updated for the new sibling modules so the AST bundle-consistency test still fails on an unshipped or uncommented-out sibling.
2026-07-20 15:34:53 -04:00
def _get_s3():
global s3
if s3 is None:
s3 = boto3.client("s3")
return s3
def handler(event, context):
"""Lambda entry point. Triggered by S3 ObjectCreated events."""
feat: deploy-pipeline guards — healthcheck, smoke gate, bundle glob + AST test (refactor phase 0) (#107) * feat: deploy-pipeline guards — healthcheck, smoke gate, bundle glob + AST test (refactor phase 0) Deploys of po-email-processor and workorder-email-processor had no verification step, so an init-time ImportError in the bundled zip could ship silently and only surface on the next real S3 event. This adds a synchronous post-deploy smoke gate wired into the deploy workflow: both Lambdas are invoked with {"healthcheck": true} and the FunctionError field is checked, since an Unhandled init error still returns HTTP 200 on RequestResponse invokes and would false-pass a plain exit-code check. The healthcheck branch is the first statement in each handler, before any boto3/S3 use or ses_auth, and only fires on a top-level direct invoke ("healthcheck" is not a key AWS ever sets on a real S3 ObjectCreated event, so mail content can't reach this path). It emits no EMF metrics and no log text that could match the sender-auth-rejected metric filter, so two deploys in one window won't trip the alarm. Separately, the PO stack's asset bundling copied a hand-maintained four-file allowlist into the zip, so every new sibling module handler.py imports had to be added by hand or the deploy shipped a Lambda that ImportErrors at cold start (bit us for template_parser in PR #105 and nearly for derived_fields in PR #2). Replaced it with a non-recursive ./*.py glob so top-level source files ship automatically while tests/ and the stale package/ dir still cannot, and added an AST-based bundle-consistency test that parses each handler's first-party imports and fails CI if the bundling command would omit any of them (a revert to an incomplete allowlist, or code moved into a subdirectory the glob doesn't cover). Includes the refactor-evaluation report that scoped this phase. * fix: review nits — unambiguous bundling-command extraction, smoke payload-parse message, dead asserts - tests/test_bundle_consistency.py: _extract_bundling_command now collects all command=[...] matches and demands exactly one per stack file, instead of silently returning whichever ast.walk visits first if a second bundled function is ever added. - scripts/post-deploy-smoke.sh: distinguish an unparseable response payload from a payload mismatch so the failure message says what actually happened (the previous "could not parse" branch was unreachable — the inline python always exited 0). - test_po_healthcheck.py: drop the substring assertions on stdout that were dead behind the stricter `captured.out == ""` assertion; keep the stderr filter-pattern check. Review follow-up on PR #107; no behavior change to any shipped code path.
2026-07-17 13:18:45 -04:00
# Direct-invoke healthcheck (post-deploy smoke). This MUST be the very first
# thing handler() does -- before any S3 fetch, before ses_auth, before the
# Records loop -- so it (a) creates no accept path for mail (real mail is an
# S3 ObjectCreated event whose top-level keys AWS controls; email content
# can never set a top-level "healthcheck" key), and (b) emits no EMF metric
# and no log line that could match the sender_auth_rejected metric-filter
# pattern, so two deploys in ~30 min never page that alarm.
if isinstance(event, dict) and event.get("healthcheck") is True:
return {"healthcheck": "ok"}
for record in event.get("Records", []):
bucket = record["s3"]["bucket"]["name"]
key = record["s3"]["object"]["key"]
s3_key = f"s3://{bucket}/{key}"
logger.info(f"Processing email: {s3_key}")
feat: decompose email-processor handlers into flat siblings + lazy boto3 clients (refactor phase 5) (#113) Both email-processor God-handlers split along the seams that already work in the flat-sibling pattern established by lambdas/shared/, so bare-name imports keep working under the existing bundling glob. PO (5-way split): handler.py keeps only the event loop, fail-closed auth, and email_type routing. extraction.py holds extract_with_claude and _EMAIL_TAG_RE, importing EXTRACTION_PROMPT from prompts.py and parse_raw_email from shared/email_parsing.py rather than recreating a PO-local copy. enrichment.py is a pure code move of enrich_parsed and pad_zip (PO-only; WO has no enrichment stage) with zero behavior change. telemetry.py holds the EMF ParseMethod emit wrappers. persistence.py holds _write_fields/_merge_update/save_*, collapsing the byte-identical save_new_po/save_revision bodies into one _save_merge helper that both now call through, preserving the sticky Cancelled ConditionExpression guard for both callers; save_cancellation stays separate. WO (5 concerns, no enrichment stage): the handler loop keeps validate_ai_fallback and the re.fullmatch(r"[0-9]+", work_order_id) key guard ahead of both save_work_order and save_event, since the guard protects the DynamoDB partition key and the '#'-delimited comment_id range-key segment. _header_date_iso and comment_id determinism stay colocated with persistence.py's save_event for the retry-idempotent event_id key. EXTRACTION_PROMPT (PO) moves to prompts.py with cross-reference headers to derived_fields.py's authoritative trade/site/fiscal rule tables; handler.py re-exports it (from prompts import EXTRACTION_PROMPT) since four tests dereference handler.EXTRACTION_ PROMPT directly. WO's prompt moves the same way. I/O modules (extraction.py's bedrock client, persistence.py's dynamodb resource, handler.py's s3 client) get lazy cached boto3 accessors; pure modules (enrichment.py, prompts.py, telemetry.py) import no boto3. Test monkeypatch surfaces move to the module that now owns the client (e.g. persistence.dynamodb) everywhere tests patch it, and the moto-before-handler-import ordering in _po_parser_support.py is preserved so the moto-backed suites don't hit real AWS. Behavior-preservation pins, verified with tests: PO still emits ParseMethod=ai_fallback before the Bedrock call, with ai_fallback_rejected as the additive second datapoint on rejection. WO still emits after its gate with mutually-exclusive ai_fallback / ai_fallback_rejected. Shadow DerivedFieldAgreement telemetry stays ai_fallback-only. derived_fields.py is untouched (diff against feature/phase-3-shared-extraction is empty). handler(event, context) signatures and the save_* public contract are unchanged on both pipelines; goldens unchanged. PO_EXPECTED_TOP_LEVEL_MODULES and its WO equivalent in tests/test_bundle_consistency.py are updated for the new sibling modules so the AST bundle-consistency test still fails on an unshipped or uncommented-out sibling.
2026-07-20 15:34:53 -04:00
response = _get_s3().get_object(Bucket=bucket, Key=key)
raw_email = response["Body"].read()
Add fail-closed SES sender authentication (INFRA-107) (#98) * Add fail-closed SES sender authentication The From header and any raw-MIME Authentication-Results copies are attacker-forgeable, so a forged email to apm@int.seahaven.com or amazon_po@int.seahaven.com could create or mutate a WO/PO (INFRA-107, CRITICAL). Both S3-triggered email processors now authenticate the sender against the Authentication-Results header SES itself prepends at delivery: only the topmost header is consulted, its authserv-id must be amazonses.com, and it must carry dkim=pass for a domain in the per-pipeline ALLOWED_DKIM_DOMAINS env var (comma-separated, set in CDK so ops can adjust without code changes). Allowlists come from live traffic observed 2026-07-15 on both ingest buckets: WO mail arrives via the apm@ Google Groups forward, which re-signs as seahaven.com (the hxgnsmartcloud.com signature does not survive the forward); PO mail passes for amazon.coupahost.com. amazonses.com also passes on PO mail but is deliberately excluded -- every SES customer's outbound mail passes for it. Every failure path (env var unset, header missing or unparseable, verdict fail, unaligned domain) rejects the email: a structured warning with the reason and S3 key is logged and the record skipped without erroring the invocation, so rejected mail causes no Lambda retries or DLQ messages. Handler signatures and event sources are unchanged. Refs: INFRA-107 * Harden AR parser per cross-family review Cross-family (GPT-4.1) review findings: terminate the dkim result token at end-of-clause, whitespace, or a comment so a value like "dkim=pass-fake" can never be read as a pass; normalize trailing dots off allowlist entries so "seahaven.com." matches; make the compat32 parser policy explicit. Adds tests for result-token boundaries, comments after the result, quoted domain values, and folding inside a dkim clause. Refs: INFRA-107 * Harden AR parsing and alarm on sender-auth rejects The SES-stamped Authentication-Results value echoes attacker-controlled SMTP-session tokens (envelope-from, helo, header.from) as their own semicolon-delimited property clauses. A naive split(";") tore an RFC 5321 quoted-local-part MAIL FROM apart and manufactured a forged dkim=pass clause, so a fully spoofed email was accepted on the genuinely SES-stamped topmost header. Tokenise comment- and quoted-string-aware (RFC 8601 / RFC 5322): strip CFWS comments, split clauses only on semicolons outside a quoted-string, and fail closed on unbalanced quotes/comments so a ';' inside a quoted pvalue can never start a clause. Rejected mail returns normally (no error, no retry, no DLQ message), so a signing-domain drift or a wrong allowlist would silently discard 100% of legitimate mail while every alarm stayed green. Add a CloudWatch Logs metric filter + alarm on the sender_auth_rejected warning to both stacks so a false-reject storm pages instead of vanishing. This is also the safety net for the WO seahaven.com allowlist assumption, which must be validated against a live SES-stamped header (a plain Gmail auto-forward re-signs under the sending Workspace domain, not seahaven.com). Refs: INFRA-107 * chore: retrigger CI (no run recorded for 7c74ac1) * Fix quoted-AUID DKIM domain spoof in sender auth Resolve three confirmed /sh-security-review findings on the fail-closed SES sender-authentication control. HIGH: header.i/header.d domain extraction was not quoted-string aware. An attacker with a valid DKIM key for their own domain could set an RFC 6376-legal AUID such as i="@seahaven.com"@attacker.com; the naive extractor stopped at the closing quote and returned seahaven.com, accepting forged mail. Extraction now tokenises the clause with the same quoted-string discipline already used for clause splitting: header.d (the plain signing domain) is authoritative when present, otherwise the header.i domain is the part after the AUID's LAST top-level "@", so a "@" inside a quoted local-part is treated as signer-controlled label text and yields the true signer (attacker.com), not seahaven.com. LOW: the topmost-header parse ran outside evaluate_sender_authentication's try/except, so an unexpected parser exception on crafted input could propagate into the handler and Lambda async retries/DLQ. The parse now fails CLOSED with an authentication_results_unparseable reason. MEDIUM: the sender_auth_rejected alarm used Sum>=3 over 15 min, blind to a low-volume total-reject outage (a trickle that never sums to 3). Both stacks now alarm on >=1 reject per 5-min period with evaluation_periods=3 / datapoints_to_alarm=2, so a sustained reject condition pages even at one reject per period while a lone stray probe self-clears. Refs: INFRA-107 * Load Lambda function dir on sys.path in tests Rebasing INFRA-107 onto main folded #95's pytest suite into this branch's tests. The unified conftest loads the PO/WO handlers by file path, and handler.py now does `from ses_auth import authenticate_inbound_email` -- a bare sibling import that resolves in the Lambda only because the runtime puts each function's own directory on sys.path. The shared load_handler now adds that directory so the handler tests import correctly alongside the sender-auth tests. Refs: INFRA-107 * Note #97 test files in README directory tree The rebase onto main brought in #97's tests/requirements.txt and tests/test_po_merge.py. List both in the directory tree so it matches the tree on disk. Refs: INFRA-107 * Document INFRA-107 forwarder-binding risk acceptance Record the accepted risk that WO sender auth binds to the apm@ forward's re-signing domain (seahaven.com) rather than the Hexagon originator; the apm@ Google Group's restricted posting policy is the load-bearing control (escalates to HIGH if the group is opened to external posting). Also correct the sender-auth-rejected alarm docs to match the shipped config (>=1 per 5-min, 2-of-3 datapoints, not the superseded >=3/15min) and note the SES-AR-01/02 parser hardening follow-ups. Refs: INFRA-107
2026-07-15 20:58:47 -04:00
# Fail-closed sender authentication (INFRA-107): only mail with an
# SES-stamped dkim=pass verdict for an allowlisted domain may create
# or update purchase orders. Rejected mail is logged and skipped
# without erroring the invocation (no retries / DLQ spam).
if not authenticate_inbound_email(raw_email, s3_key):
continue
email_data = parse_raw_email(raw_email)
logger.info(f"Subject: {email_data['subject']}")
feat: template-first PO parser with fail-closed gate and Bedrock fallback (#105) * Add deterministic template parser for WO emails The workorder-email-processor sends every one of ~22.9k emails/month to an LLM, but ~93.6% are the plain-text "AMAZON UPDATE WO DETAILS" comment template and ~6.4% the HTML "AMAZON assign Work Order" template. Parse those two shapes deterministically, offline, so the AI call is reserved for the long tail. The module is pure (no boto3, no network). try_deterministic_parse classifies by subject, extracts the shared contract fields, and returns a result ONLY when it passes a strict fail-closed validation gate: exact contract-key set, subject/id agreement, the literal "Work Order: <id>" double space, per-type required fields, site-code shape, and a label-bleed guard so a value that over-ran into the next field fails. Any miss, drift, or extractor exception yields None so the caller falls back to the AI extractor -- data is never corrupted, only the fallback rate rises. Refs: #23 * Migrate WO processor to Bedrock and fix comment_id collision Switch the AI path from the Anthropic SDK to bedrock-runtime InvokeModel on the inference profile us.anthropic.claude-haiku-4-5-20251001-v1:0 (BEDROCK_MODEL_ID env), so parsing no longer needs a provider API key or Secrets Manager secret. The EXTRACTION_PROMPT and JSON contract are kept byte-identical, so the AI-fallback output is unchanged. Try the new deterministic template parser first and only call Bedrock on a miss/invalid result. Fix issue #23: the WorkOrderComments range key was work_order_id#<comment_time>, so two emails on one WO with an identical or absent comment time collided and overwrote each other. Derive a 12-hex suffix from the S3 object key alone -- deterministic, so an async retry of the same object is byte-identical (idempotent) while distinct emails get distinct keys -- and keep wall-clock now() out of the key (literal 'nocomment' segment when comment_time is absent). Also emit one CloudWatch EMF line per record (Seahaven/WorkorderIngest ParseOutcome, dimensioned by ParseMethod/TemplateId) for parse-outcome observability, replace the deprecated datetime.utcnow() with datetime.now(timezone.utc), and drop the anthropic dependency. Refs: #23 * Migrate PO processor to Bedrock Switch the PO email processor's AI extraction from the Anthropic SDK to bedrock-runtime InvokeModel on the inference profile us.anthropic.claude-haiku-4-5-20251001-v1:0 (BEDROCK_MODEL_ID env), so it no longer needs a provider API key or Secrets Manager secret. PO parsing stays fully AI -- only the provider changes. The EXTRACTION_PROMPT is kept byte-identical and the Bedrock text output is still decoded with json.loads(..., parse_float=Decimal), which DynamoDB requires (it rejects floats). Replace the deprecated datetime.utcnow() with datetime.now(timezone.utc) and drop the anthropic dependency. * Grant Bedrock IAM, drop Anthropic secrets, add fallback alarm Both stacks moved their processors from the Anthropic API to the Bedrock inference profile us.anthropic.claude-haiku-4-5-20251001-v1:0. Grant each processor role bedrock:InvokeModel + bedrock:InvokeModelWithResponseStream on BOTH the inference-profile ARN AND the per-region foundation-model ARNs for us-east-1/us-east-2/us-west-2 (empty-account) -- the us.* profile routes cross-region, so a profile-only grant AccessDenies at runtime. Remove both anthropic-api-key Secret constructs, their grant_read, and the ANTHROPIC_API_KEY_SECRET_ARN env; add BEDROCK_MODEL_ID. The secrets had RemovalPolicy.RETAIN so they are orphaned, not deleted -- flagged in the README for manual post-deploy deletion and key revocation. Add the workorder-email-processor-template-fallback-rate alarm: a FILL(0) + >=10-sample volume-floor MathExpression over the EMF ParseOutcome metric (15-min periods) that pages when the AI-fallback share exceeds 15% sustained, catching Hexagon template drift. ALARM-only SnsAction to site-alerts, no OK action, NOT_BREACHING, matching the existing stack idiom. * Add offline WO parser test suite Cover the deterministic parser with golden-file tests over 55 real scrubbed .eml fixtures (both comment sub-shapes, username Submitted-By, address present/absent, br+CRLF assign addresses), fail-closed validation-gate rules, adversarial and prompt-injection cases that must route to ai_fallback or parse without corrupting other fields, the issue #23 comment_id idempotency invariants, and the Bedrock-fallback dispatch plus EMF-metric emission with a mocked invoke_model. Extend pytest.ini testpaths to discover the co-located suite, and update tests/conftest.load_handler to put a handler's own directory on sys.path so the WO handler's new `from template_parser import ...` resolves under the existing shared handler tests. Point test_local.py at the new template-first + Bedrock flow. Refs: #23 * Document Bedrock migration and WO parse flow in README Record the provider switch to the Bedrock inference profile (no Anthropic API key or Secrets Manager secret, with the retired secrets flagged for manual deletion), the WO deterministic-template-first + AI-fallback flow, the new ParseOutcome EMF metric and template-fallback-rate alarm, the issue #23 comment_id format change, the +00:00 aware-UTC timestamp shift, and offline test instructions. Refs: #23 * Fix f-string lint and formatting in backfill scripts Drop the f prefix from two f-strings that carry no placeholders (F541) and apply ruff format, so `ruff check` / `ruff format --check` pass in CI. * Emit ParseMethod-only EMF set so fallback alarm can fire The fallback-rate alarm queries the ParseOutcome series keyed on ParseMethod alone, but the emitter published only the joint (ParseMethod, TemplateId) dimension set. CloudWatch materializes exactly the listed dimension sets and does not auto-aggregate, so the alarm's series never received data: it evaluated a constant 0 and could never page on template-drift coverage collapse. Publish both ["ParseMethod"] and ["ParseMethod","TemplateId"] and update the EMF regression test to assert both sets are present. * Commit WO parser .eml fixtures for executable coverage The parser test suite globbed for input .eml fixtures that the repo's `*.eml` ignore rule kept uncommitted, so every parametrized golden and fail-closed test collected zero cases and CI could not exercise the deterministic parser that handles 100% of WO email volume. Add a fixtures-only negation to .gitignore and commit the 55 scrubbed positive samples (50 update-plaintext, 5 assign-html) plus 14 ai-fallback and 3 adversarial fixtures. The ai-fallback set covers each fail-closed reason code (subject_no_match, single_space_work_order, malformed_site_code, label_bleed, creation_time_unparseable, wo_id_mismatch, missing_required_field) and the adversarial set proves the parser is total and confines prompt-injection payloads to comment_text without steering the structured fields. * feature: Add PO template parser scaffold and design doc Mirror WO PR #99's template-first approach for the Coupa PO processor. Two templates identified from a full 3,448-email triage: - coupa_new_po (95.5%): scaffolded; fails closed to the LLM until extract_new_po lands. - coupa_cancellation (2.9%): implemented. Nested contract with recursive validation, Decimal money, and a fail-closed gate. Derived fields (site_code/trade/fiscal_year) are deferred to a shared post-stage. Comments, revisions, multi-line, and non-USD emails fall back to Bedrock. docs/po-template-parser.md records the investigation, decisions, and remaining work. Signed-off-by: Adam Moussa <166072409+amoussa1229@users.noreply.github.com> * Implement PO new_po extraction and value-level gate Replace the extract_new_po scaffold stub with the full section-windowed extractor (duplicate-label anchoring, sentinel ship-to, label-keyed U+2022 bullet split, Decimal money from three anchored contexts only) and add value-level gate rules V1-V13. Both new_po_not_implemented scaffold guards are removed; rules 6-8 (unrecognized_status, multiline_unsupported, non_usd) go live. The gate re-derives every byte proof from the email body so an extractor bug cannot vouch for itself: amount re-serialization with a digit/comma border check (the thousands-separator truncation kill switch), sum(lines)==total against both Total blocks, anchor/supplier identity proofs, USPS address shape on the raw pre-enrichment zip, bullet label discipline, and sentinel/artifact hygiene. Any failure falls closed to the LLM; a validation failure is never a parsed result. Refs: #99 * Wire template-first parse into PO handler with EMF metric Run try_deterministic_parse ahead of the Bedrock extractor and fall back only on a miss/invalid (fail-closed) result. The shared enrich_parsed post-stage and the save_cancellation/save_revision/ save_new_po routing are untouched, so both paths write identical DynamoDB shapes and the po-ingest-site-extractor stream contract is preserved. Each record emits one ParseMethod EMF line (Seahaven/PoIngest/ ParseOutcome, dimension sets [ParseMethod] and [ParseMethod,TemplateId], ReasonCode/po_number ride-alongs) mirroring the WO idiom. The metric fires before the Bedrock call so a Bedrock-side error still records the ai_fallback outcome. Refs: #99 * Add PO fallback-rate alarm retuned for ~57 emails/day The WO alarm's 15-min period and >=10-sample floor assume ~760/day and would be structurally dead at PO volume (a 15-min period holds ~0.6 emails, so the floor is never met). Retune: 6-hour periods (~14.25 expected emails), IF((fb+tmpl)>=8,...) volume floor so a single email can never breach a datapoint (1/8 = 12.5% < 20%), threshold >20% against a ~1% expected baseline, eval 4 / datapoints 2 (24h span) so noise self-clears while total template drift pages within ~12h. No element-wise MAX in the math expression (post-#102 rule); ALARM-only SnsAction to site-alerts, NOT_BREACHING. Gated with 'npx cdk synth po-ingest'. Also add template_parser.py to the bundling cp list -- without it every deployed invocation would ImportError (unit tests cannot catch an asset-bundling omission). Refs: #99, #102 * Add offline PO parser suite with scrubbed fixture corpus 132 tests: golden-file comparison for all 25 positive fixtures (17 single-line new-PO + 8 cancellations, Decimal-exact via parse_float=Decimal), every fail-closed gate reason code covered (body-level triggers via 17 synthetic adversarial .eml mutations, candidate-level via direct validate() unit tests), real multi-line and comment/non-Coupa fallback fixtures, dual line-ending parse identity, two-path enrich/save parity (site-extractor stream guard), V10 URL-id corpus sweep, fixture hygiene (ses_auth pass + scrub-marker leak sweep), and Bedrock dispatch/EMF assertions. The suite loads handler/template_parser via importlib under unique module names and binds the handler's bare sibling imports around exec (tests/conftest.py load_handler gets the same treatment) -- the WO suite caches bare 'handler'/'template_parser' names in sys.modules, and bare imports here would silently bind to the wrong pipeline. moto is imported before the handler so its botocore stubber hook precedes boto3 session creation (the PO conftest chain now loads at pytest session start). Fixtures are scrubbed real S3 samples: transport/auth header values replaced with same-shape placeholders (structure kept so ses_auth still passes), per-file digit ciphers, amounts remapped with sum==total re-established. The .gitignore exception is scoped to the PO fixtures path only. Refs: #99 * Document PO template-first parser and retuned alarm README: PO flow is now template-first with Bedrock fallback; parser/gate section mirroring the WO writeup; Seahaven/PoIngest ParseOutcome namespace and the fallback-rate alarm numbers with their volume justification (deliberately not WO's settings); test-suite and repo-layout updates. Design doc: mark PR #1 complete in progress/checklist sections; document the six value-level gate reason codes and the scaffold guard removal; correct the stale data-access note (default CLI session is 328440206208) and note the ~90-day S3 lifecycle aging of the corpus; record the 2.3 layout addendum (leading Supplier bullet segment, EA evidence lines, summary unit-price tokens, decode-path line endings), the fixture-build pins (address join convention, quantity/unit/price source), the V10 sweep outcome, and resolutions for open questions Q3/Q6. Cross-family review and the Confluence architecture-map update are flagged outstanding for merge. Refs: #99 * Record cross-family review outcome for handler wiring GPT-4.1 cross_review.py run against the real handler diff returned no BLOCK and no security findings; both FIX items verified as no-change-needed (fallback logging already correct; non-dict AI output is the pre-existing issue #101 pattern this PR deliberately does not touch). Refs: #99 * Pin line-item currency to USD in the PO gate The non_usd rule only checked the Total-block top-level currency, so a new_po whose line item read 'for 55,206.00 CAD' under a USD Total block still template-parsed as ok -- a fail-open hole in the fail-closed gate. Every line item's captured currency and its re-derived body token must now byte-equal the proven-USD top-level currency; covered by a line-level CAD adversarial fixture (the existing adv-non-usd only exercised the Total-block variant) and a candidate-mutation unit test. * Scrub residual transport tokens from PO fixtures The first-pass harvest scrub sanitized only the primary SES/DKIM header blocks, leaving the real SES Feedback-ID sender-identity hash in 49 committed fixtures and, on the two non-Coupa fixtures, an embedded second SES block's X-Ses-Receipt, the Exchange cross-tenant UPN ciphertext, and Gmail ARC fh= / X-Gm-* tokens -- exactly the token classes the PR #99 fixture lesson requires placeholdered. Replace each with a same-shape ScrubbedFixture value (byte-safe, CRLF and folding preserved) so header structure and ses_auth behavior are unchanged. * Converge quantity/price to Decimal on both paths EXTRACTION_PROMPT declares quantity and price as JSON strings, so a prompt-obedient Bedrock response stores DynamoDB Strings where the template parser stores Numbers -- divergent attribute types for the same email on the purchase-orders stream. Coerce numeric strings to Decimal in the shared enrich_parsed post-stage (thousands-separator safe; non-numeric strings kept verbatim) so both paths converge; prompt rewording itself remains PR #2 scope. The two-path parity test was circular -- it replayed the parser- derived golden as 'the LLM output', so it could never see the type divergence. It now feeds a prompt-shaped payload (string quantity/ price, LLM-filled site_code) through enrich_parsed and save_new_po, and the fixture-hygiene test now asserts the scrubbed transport-token header classes so fixture regressions are caught. * Coerce bare-int quantity/price to Decimal in enrich_parsed GPT-4.1 cross-family review of the final PR diff (no BLOCK) flagged residual type drift: parse_float=Decimal rules out floats on the LLM path, but a bare JSON int survived as Python int. Coerce it so both parse paths emit one canonical Decimal type. * Scrub fixture-body PII and harden cancellation gate (sec review) /sh-security-review of PR #105 (5 fresh-context detectors + proof-or-kill verifier) confirmed two diff-introduced findings; both fixed here. F3 (medium, real PII in new fixtures): the harvest scrub replaced header tokens but left real third-party PII in message BODIES -- an Amazon contact's name/phone/personal email in non-coupa-02.eml and an internal t.corp.amazon.com ticket URL in comment-02.eml, plus real submitter/attn names recurring across the new_po corpus. Replaced every personal name, phone, personal email, and internal URL with synthetic placeholders (QP-soft-wrap aware) across both .eml bodies and expected goldens. Extended test_fixture_hygiene to scan BODIES (phone shapes, corp URLs, the leaked tokens), closing the header-only gap that let this through. F1 (medium, cancellation gate): _CANCELLATION_SUBJECT was unanchored and matched with .search(), unlike the anchored new_po pattern -- a subject merely ending with the cancellation phrase could be routed to the sticky- Cancelled write. Fully anchored it and switched to .match, and added a body-corroboration gate (the real Coupa body independently restates 'Purchase Order #<po> ... has been cancelled'); a near-miss/misrouted subject whose body does not corroborate now fails closed to the LLM (new reason code cancellation_body_unconfirmed). Pre-existing (advisory, not this PR): the LLM-fallback else->save_new_po dispatch and undelimited extraction prompt (issue #101 family) are byte-identical to main and unchanged here. 401 tests pass; ruff/format clean; cdk synth po-ingest clean. --------- Signed-off-by: Adam Moussa <166072409+amoussa1229@users.noreply.github.com>
2026-07-16 17:50:59 -04:00
# Deterministic template parse first; fall back to the Bedrock AI
# extractor only on a miss or an invalid (fail-closed) result. The
# metric is emitted before the Bedrock call, so a Bedrock-side error
# still records the ai_fallback outcome (the invocation then errors
# into the errors alarm / DLQ as before).
parsed, parse_method, template_id, parse_reason = try_deterministic_parse(
email_data
)
_emit_parse_method_metric(
parse_method,
template_id,
parse_reason,
parsed.get("po_number") if parsed else None,
)
if parsed is None:
parsed = extract_with_claude(email_data)
PO ai-fallback fail-closed gate + prompt hardening (refactor phase 1) (#108) * feat: PO ai-fallback fail-closed gate + prompt hardening, parity with #104 (refactor phase 1) Ports WO's #104 AI-fallback security hardening to the PO email processor, adapted for PO's nested contract instead of copying the WO gate verbatim. validate_ai_fallback() (template_parser.py) fail-closes raw Bedrock output before it reaches enrich_parsed or any dispatch/save: recursive key-set check with missing-key normalization (nested contract: supplier{}, ship_to{}, line_items[]); po_number checked against the same hardened prefix+hyphen+digits regex family that guards the DynamoDB partition key the handler builds from it (rejects fullwidth-digit and trailing-artifact injection); email_type enforced against the {new_po, revision, cancellation} allow-list before dispatch so a miss can never fall into the else -> save_new_po branch; money fields accept Decimal/int/None only, matching PO's parse_float=Decimal decode (a float-typed check would be wrong here). A gate failure emits ParseMethod=ai_fallback_rejected and `continue`s to the next record -- it never raises, so attacker-controlled input can't churn the retry/DLQ path. extract_with_claude() wraps the untrusted email in an <email> data block and neutralizes forged <email>-tag lookalikes in the body with the same linear-time regex approach as WO's _EMAIL_TAG_RE, and sets temperature=0 on the Bedrock call. Deliberate double-count: PO emits ParseMethod=ai_fallback before the Bedrock call (so a Bedrock-side error still records the outcome), so a rejected email always produces both an ai_fallback datapoint (pre-call) and an ai_fallback_rejected datapoint (post-gate). This is intentional, not a bug -- documented in handler.py, template_parser.py, and the README. cdk/po_stack.py: in-place property update to the existing po-email-processor-template-fallback-rate alarm (same logical ID, no rename/replacement) -- the fb/(fb+tmpl) expression is left byte-identical to its pre-Phase-1 form and ai_fallback_rejected is deliberately excluded from the numerator/denominator/volume floor, since folding it in as WO does would double-count every rejection (PO's pre-call emit already counts it once via fb). A net-new EmailProcessorAiFallbackRejectedAlarm watches the rejected series on its own, retuned for ~57 emails/day with the 6h/IF-floor/eval-4/ datapoints-2 idiom (not WO's 5-minute sparse idiom, which is structurally dead at PO volume). Both alarms remain ALARM-only to site-alerts, NOT_BREACHING, with no element-wise MAX in the math (post-#102 rule). * Block "Cancelled" po_status off the AI cancellation route The AI-fallback gate type-checked po_status but let any string through, unlike the template path which never emits "Cancelled" on a new_po. Dispatch routes on email_type, so an AI-path new_po or revision carrying po_status="Cancelled" would reach save_new_po/save_revision and cancel a live PO via _merge_update's sticky-cancel write without ever hitting save_cancellation. Reject the exact sticky marker on any non-cancellation email_type so the AI path matches the template path's guard; arbitrary non-marker status strings still pass. email_type is already validated to the enum before this check, and a cancellation reaches save_cancellation (which hardcodes the status), so po_status stays irrelevant on that route.
2026-07-17 14:50:47 -04:00
# Fail-closed gate on the raw model output (#104 parity): runs
# BEFORE enrich_parsed and BEFORE any dispatch/save. INTENTIONAL
# DOUBLE-COUNT: ParseMethod=ai_fallback was already emitted above,
# BEFORE the Bedrock call (deliberate -- a Bedrock-side error must
# still record the outcome), so a rejected email produces BOTH an
# ai_fallback and an ai_fallback_rejected datapoint. The po_stack
# fallback-rate alarm therefore EXCLUDES the rejected series from
# its rate math (fb already counts these emails once); see
# cdk/po_stack.py and README.
ok, val_reason, normalized = validate_ai_fallback(parsed)
if not ok:
logger.warning(
f"AI-fallback output rejected by validation gate "
f"({val_reason}); skipping: {key}"
)
_emit_parse_method_metric(
"ai_fallback_rejected",
template_id,
val_reason,
str(parsed.get("po_number"))[:64]
if isinstance(parsed, dict) and parsed.get("po_number")
else None,
)
# Skip, never raise: attacker-controlled input must not churn
# the retry/DLQ path.
continue
parsed = normalized
logger.info(
feat: template-first PO parser with fail-closed gate and Bedrock fallback (#105) * Add deterministic template parser for WO emails The workorder-email-processor sends every one of ~22.9k emails/month to an LLM, but ~93.6% are the plain-text "AMAZON UPDATE WO DETAILS" comment template and ~6.4% the HTML "AMAZON assign Work Order" template. Parse those two shapes deterministically, offline, so the AI call is reserved for the long tail. The module is pure (no boto3, no network). try_deterministic_parse classifies by subject, extracts the shared contract fields, and returns a result ONLY when it passes a strict fail-closed validation gate: exact contract-key set, subject/id agreement, the literal "Work Order: <id>" double space, per-type required fields, site-code shape, and a label-bleed guard so a value that over-ran into the next field fails. Any miss, drift, or extractor exception yields None so the caller falls back to the AI extractor -- data is never corrupted, only the fallback rate rises. Refs: #23 * Migrate WO processor to Bedrock and fix comment_id collision Switch the AI path from the Anthropic SDK to bedrock-runtime InvokeModel on the inference profile us.anthropic.claude-haiku-4-5-20251001-v1:0 (BEDROCK_MODEL_ID env), so parsing no longer needs a provider API key or Secrets Manager secret. The EXTRACTION_PROMPT and JSON contract are kept byte-identical, so the AI-fallback output is unchanged. Try the new deterministic template parser first and only call Bedrock on a miss/invalid result. Fix issue #23: the WorkOrderComments range key was work_order_id#<comment_time>, so two emails on one WO with an identical or absent comment time collided and overwrote each other. Derive a 12-hex suffix from the S3 object key alone -- deterministic, so an async retry of the same object is byte-identical (idempotent) while distinct emails get distinct keys -- and keep wall-clock now() out of the key (literal 'nocomment' segment when comment_time is absent). Also emit one CloudWatch EMF line per record (Seahaven/WorkorderIngest ParseOutcome, dimensioned by ParseMethod/TemplateId) for parse-outcome observability, replace the deprecated datetime.utcnow() with datetime.now(timezone.utc), and drop the anthropic dependency. Refs: #23 * Migrate PO processor to Bedrock Switch the PO email processor's AI extraction from the Anthropic SDK to bedrock-runtime InvokeModel on the inference profile us.anthropic.claude-haiku-4-5-20251001-v1:0 (BEDROCK_MODEL_ID env), so it no longer needs a provider API key or Secrets Manager secret. PO parsing stays fully AI -- only the provider changes. The EXTRACTION_PROMPT is kept byte-identical and the Bedrock text output is still decoded with json.loads(..., parse_float=Decimal), which DynamoDB requires (it rejects floats). Replace the deprecated datetime.utcnow() with datetime.now(timezone.utc) and drop the anthropic dependency. * Grant Bedrock IAM, drop Anthropic secrets, add fallback alarm Both stacks moved their processors from the Anthropic API to the Bedrock inference profile us.anthropic.claude-haiku-4-5-20251001-v1:0. Grant each processor role bedrock:InvokeModel + bedrock:InvokeModelWithResponseStream on BOTH the inference-profile ARN AND the per-region foundation-model ARNs for us-east-1/us-east-2/us-west-2 (empty-account) -- the us.* profile routes cross-region, so a profile-only grant AccessDenies at runtime. Remove both anthropic-api-key Secret constructs, their grant_read, and the ANTHROPIC_API_KEY_SECRET_ARN env; add BEDROCK_MODEL_ID. The secrets had RemovalPolicy.RETAIN so they are orphaned, not deleted -- flagged in the README for manual post-deploy deletion and key revocation. Add the workorder-email-processor-template-fallback-rate alarm: a FILL(0) + >=10-sample volume-floor MathExpression over the EMF ParseOutcome metric (15-min periods) that pages when the AI-fallback share exceeds 15% sustained, catching Hexagon template drift. ALARM-only SnsAction to site-alerts, no OK action, NOT_BREACHING, matching the existing stack idiom. * Add offline WO parser test suite Cover the deterministic parser with golden-file tests over 55 real scrubbed .eml fixtures (both comment sub-shapes, username Submitted-By, address present/absent, br+CRLF assign addresses), fail-closed validation-gate rules, adversarial and prompt-injection cases that must route to ai_fallback or parse without corrupting other fields, the issue #23 comment_id idempotency invariants, and the Bedrock-fallback dispatch plus EMF-metric emission with a mocked invoke_model. Extend pytest.ini testpaths to discover the co-located suite, and update tests/conftest.load_handler to put a handler's own directory on sys.path so the WO handler's new `from template_parser import ...` resolves under the existing shared handler tests. Point test_local.py at the new template-first + Bedrock flow. Refs: #23 * Document Bedrock migration and WO parse flow in README Record the provider switch to the Bedrock inference profile (no Anthropic API key or Secrets Manager secret, with the retired secrets flagged for manual deletion), the WO deterministic-template-first + AI-fallback flow, the new ParseOutcome EMF metric and template-fallback-rate alarm, the issue #23 comment_id format change, the +00:00 aware-UTC timestamp shift, and offline test instructions. Refs: #23 * Fix f-string lint and formatting in backfill scripts Drop the f prefix from two f-strings that carry no placeholders (F541) and apply ruff format, so `ruff check` / `ruff format --check` pass in CI. * Emit ParseMethod-only EMF set so fallback alarm can fire The fallback-rate alarm queries the ParseOutcome series keyed on ParseMethod alone, but the emitter published only the joint (ParseMethod, TemplateId) dimension set. CloudWatch materializes exactly the listed dimension sets and does not auto-aggregate, so the alarm's series never received data: it evaluated a constant 0 and could never page on template-drift coverage collapse. Publish both ["ParseMethod"] and ["ParseMethod","TemplateId"] and update the EMF regression test to assert both sets are present. * Commit WO parser .eml fixtures for executable coverage The parser test suite globbed for input .eml fixtures that the repo's `*.eml` ignore rule kept uncommitted, so every parametrized golden and fail-closed test collected zero cases and CI could not exercise the deterministic parser that handles 100% of WO email volume. Add a fixtures-only negation to .gitignore and commit the 55 scrubbed positive samples (50 update-plaintext, 5 assign-html) plus 14 ai-fallback and 3 adversarial fixtures. The ai-fallback set covers each fail-closed reason code (subject_no_match, single_space_work_order, malformed_site_code, label_bleed, creation_time_unparseable, wo_id_mismatch, missing_required_field) and the adversarial set proves the parser is total and confines prompt-injection payloads to comment_text without steering the structured fields. * feature: Add PO template parser scaffold and design doc Mirror WO PR #99's template-first approach for the Coupa PO processor. Two templates identified from a full 3,448-email triage: - coupa_new_po (95.5%): scaffolded; fails closed to the LLM until extract_new_po lands. - coupa_cancellation (2.9%): implemented. Nested contract with recursive validation, Decimal money, and a fail-closed gate. Derived fields (site_code/trade/fiscal_year) are deferred to a shared post-stage. Comments, revisions, multi-line, and non-USD emails fall back to Bedrock. docs/po-template-parser.md records the investigation, decisions, and remaining work. Signed-off-by: Adam Moussa <166072409+amoussa1229@users.noreply.github.com> * Implement PO new_po extraction and value-level gate Replace the extract_new_po scaffold stub with the full section-windowed extractor (duplicate-label anchoring, sentinel ship-to, label-keyed U+2022 bullet split, Decimal money from three anchored contexts only) and add value-level gate rules V1-V13. Both new_po_not_implemented scaffold guards are removed; rules 6-8 (unrecognized_status, multiline_unsupported, non_usd) go live. The gate re-derives every byte proof from the email body so an extractor bug cannot vouch for itself: amount re-serialization with a digit/comma border check (the thousands-separator truncation kill switch), sum(lines)==total against both Total blocks, anchor/supplier identity proofs, USPS address shape on the raw pre-enrichment zip, bullet label discipline, and sentinel/artifact hygiene. Any failure falls closed to the LLM; a validation failure is never a parsed result. Refs: #99 * Wire template-first parse into PO handler with EMF metric Run try_deterministic_parse ahead of the Bedrock extractor and fall back only on a miss/invalid (fail-closed) result. The shared enrich_parsed post-stage and the save_cancellation/save_revision/ save_new_po routing are untouched, so both paths write identical DynamoDB shapes and the po-ingest-site-extractor stream contract is preserved. Each record emits one ParseMethod EMF line (Seahaven/PoIngest/ ParseOutcome, dimension sets [ParseMethod] and [ParseMethod,TemplateId], ReasonCode/po_number ride-alongs) mirroring the WO idiom. The metric fires before the Bedrock call so a Bedrock-side error still records the ai_fallback outcome. Refs: #99 * Add PO fallback-rate alarm retuned for ~57 emails/day The WO alarm's 15-min period and >=10-sample floor assume ~760/day and would be structurally dead at PO volume (a 15-min period holds ~0.6 emails, so the floor is never met). Retune: 6-hour periods (~14.25 expected emails), IF((fb+tmpl)>=8,...) volume floor so a single email can never breach a datapoint (1/8 = 12.5% < 20%), threshold >20% against a ~1% expected baseline, eval 4 / datapoints 2 (24h span) so noise self-clears while total template drift pages within ~12h. No element-wise MAX in the math expression (post-#102 rule); ALARM-only SnsAction to site-alerts, NOT_BREACHING. Gated with 'npx cdk synth po-ingest'. Also add template_parser.py to the bundling cp list -- without it every deployed invocation would ImportError (unit tests cannot catch an asset-bundling omission). Refs: #99, #102 * Add offline PO parser suite with scrubbed fixture corpus 132 tests: golden-file comparison for all 25 positive fixtures (17 single-line new-PO + 8 cancellations, Decimal-exact via parse_float=Decimal), every fail-closed gate reason code covered (body-level triggers via 17 synthetic adversarial .eml mutations, candidate-level via direct validate() unit tests), real multi-line and comment/non-Coupa fallback fixtures, dual line-ending parse identity, two-path enrich/save parity (site-extractor stream guard), V10 URL-id corpus sweep, fixture hygiene (ses_auth pass + scrub-marker leak sweep), and Bedrock dispatch/EMF assertions. The suite loads handler/template_parser via importlib under unique module names and binds the handler's bare sibling imports around exec (tests/conftest.py load_handler gets the same treatment) -- the WO suite caches bare 'handler'/'template_parser' names in sys.modules, and bare imports here would silently bind to the wrong pipeline. moto is imported before the handler so its botocore stubber hook precedes boto3 session creation (the PO conftest chain now loads at pytest session start). Fixtures are scrubbed real S3 samples: transport/auth header values replaced with same-shape placeholders (structure kept so ses_auth still passes), per-file digit ciphers, amounts remapped with sum==total re-established. The .gitignore exception is scoped to the PO fixtures path only. Refs: #99 * Document PO template-first parser and retuned alarm README: PO flow is now template-first with Bedrock fallback; parser/gate section mirroring the WO writeup; Seahaven/PoIngest ParseOutcome namespace and the fallback-rate alarm numbers with their volume justification (deliberately not WO's settings); test-suite and repo-layout updates. Design doc: mark PR #1 complete in progress/checklist sections; document the six value-level gate reason codes and the scaffold guard removal; correct the stale data-access note (default CLI session is 328440206208) and note the ~90-day S3 lifecycle aging of the corpus; record the 2.3 layout addendum (leading Supplier bullet segment, EA evidence lines, summary unit-price tokens, decode-path line endings), the fixture-build pins (address join convention, quantity/unit/price source), the V10 sweep outcome, and resolutions for open questions Q3/Q6. Cross-family review and the Confluence architecture-map update are flagged outstanding for merge. Refs: #99 * Record cross-family review outcome for handler wiring GPT-4.1 cross_review.py run against the real handler diff returned no BLOCK and no security findings; both FIX items verified as no-change-needed (fallback logging already correct; non-dict AI output is the pre-existing issue #101 pattern this PR deliberately does not touch). Refs: #99 * Pin line-item currency to USD in the PO gate The non_usd rule only checked the Total-block top-level currency, so a new_po whose line item read 'for 55,206.00 CAD' under a USD Total block still template-parsed as ok -- a fail-open hole in the fail-closed gate. Every line item's captured currency and its re-derived body token must now byte-equal the proven-USD top-level currency; covered by a line-level CAD adversarial fixture (the existing adv-non-usd only exercised the Total-block variant) and a candidate-mutation unit test. * Scrub residual transport tokens from PO fixtures The first-pass harvest scrub sanitized only the primary SES/DKIM header blocks, leaving the real SES Feedback-ID sender-identity hash in 49 committed fixtures and, on the two non-Coupa fixtures, an embedded second SES block's X-Ses-Receipt, the Exchange cross-tenant UPN ciphertext, and Gmail ARC fh= / X-Gm-* tokens -- exactly the token classes the PR #99 fixture lesson requires placeholdered. Replace each with a same-shape ScrubbedFixture value (byte-safe, CRLF and folding preserved) so header structure and ses_auth behavior are unchanged. * Converge quantity/price to Decimal on both paths EXTRACTION_PROMPT declares quantity and price as JSON strings, so a prompt-obedient Bedrock response stores DynamoDB Strings where the template parser stores Numbers -- divergent attribute types for the same email on the purchase-orders stream. Coerce numeric strings to Decimal in the shared enrich_parsed post-stage (thousands-separator safe; non-numeric strings kept verbatim) so both paths converge; prompt rewording itself remains PR #2 scope. The two-path parity test was circular -- it replayed the parser- derived golden as 'the LLM output', so it could never see the type divergence. It now feeds a prompt-shaped payload (string quantity/ price, LLM-filled site_code) through enrich_parsed and save_new_po, and the fixture-hygiene test now asserts the scrubbed transport-token header classes so fixture regressions are caught. * Coerce bare-int quantity/price to Decimal in enrich_parsed GPT-4.1 cross-family review of the final PR diff (no BLOCK) flagged residual type drift: parse_float=Decimal rules out floats on the LLM path, but a bare JSON int survived as Python int. Coerce it so both parse paths emit one canonical Decimal type. * Scrub fixture-body PII and harden cancellation gate (sec review) /sh-security-review of PR #105 (5 fresh-context detectors + proof-or-kill verifier) confirmed two diff-introduced findings; both fixed here. F3 (medium, real PII in new fixtures): the harvest scrub replaced header tokens but left real third-party PII in message BODIES -- an Amazon contact's name/phone/personal email in non-coupa-02.eml and an internal t.corp.amazon.com ticket URL in comment-02.eml, plus real submitter/attn names recurring across the new_po corpus. Replaced every personal name, phone, personal email, and internal URL with synthetic placeholders (QP-soft-wrap aware) across both .eml bodies and expected goldens. Extended test_fixture_hygiene to scan BODIES (phone shapes, corp URLs, the leaked tokens), closing the header-only gap that let this through. F1 (medium, cancellation gate): _CANCELLATION_SUBJECT was unanchored and matched with .search(), unlike the anchored new_po pattern -- a subject merely ending with the cancellation phrase could be routed to the sticky- Cancelled write. Fully anchored it and switched to .match, and added a body-corroboration gate (the real Coupa body independently restates 'Purchase Order #<po> ... has been cancelled'); a near-miss/misrouted subject whose body does not corroborate now fails closed to the LLM (new reason code cancellation_body_unconfirmed). Pre-existing (advisory, not this PR): the LLM-fallback else->save_new_po dispatch and undelimited extraction prompt (issue #101 family) are byte-identical to main and unchanged here. 401 tests pass; ruff/format clean; cdk synth po-ingest clean. --------- Signed-off-by: Adam Moussa <166072409+amoussa1229@users.noreply.github.com>
2026-07-16 17:50:59 -04:00
f"Parsed ({parse_method}/{template_id}/{parse_reason}): "
f"type={parsed.get('email_type')}, po={parsed.get('po_number')}"
)
if not parsed.get("po_number"):
logger.warning(f"No PO number found in email, skipping: {key}")
continue
feat: Python derived-field classifier with shadow telemetry (PO PR 2) (#106) * feat: Python derived-field classifier with shadow telemetry for PO ingest Port the site_code/trade/fiscal_year rules from EXTRACTION_PROMPT into a pure, total derived_fields module applied in the shared enrich_parsed() post-stage. Python fills gaps on both parse paths (the template path has no LLM values, closing the derived-field gap opened by the PR #105 two-PR split) and never overwrites a non-null LLM value; on ai_fallback a DerivedFieldAgreement EMF record per field shadows Python against the LLM during the bake. Rules hardened against a full-corpus backtest (3,422 real emails vs the LLM-written baseline): site_code 99.4% with zero Python-wrong cases, fiscal_year 100%, trade 96.8% ex-deliberate. Also: quantity/price now declared numeric in the prompt, and derived_fields.py added to the po_stack bundling copy (deploy-time ImportError otherwise). * fix: security-review hardening — EMF value length clamp, aggregate trade CPU budget sh-security-review (4 detectors + proof-or-kill verifier): PASS, 0 confirmed critical/high. Fixes the one confirmed low (unbounded LLM-value str() into the DerivedFieldAgreement EMF log line, clamped to 64 chars) and adds the verifier-recommended defense-in-depth aggregate character budget across line items in derive_trade (per-item caps alone allowed ~10s full-core on a pathological direct-call input; unreachable through the deployed handler but cheap to bound). Bundling cp list now carries a warning comment (GPT-4.1 cross-review FIX).
2026-07-17 11:47:33 -04:00
parsed = enrich_parsed(
parsed, s3_key, email_data["subject"], parse_method=parse_method
)
email_type = parsed.get("email_type")
if email_type == "cancellation":
save_cancellation(parsed)
elif email_type == "revision":
save_revision(parsed)
else:
save_new_po(parsed)
return {"statusCode": 200, "body": "OK"}