feat: add Iteration 3e for memory security and integrity (OWASP ASI06) by theagenticguy · Pull Request #5 · aws-samples/sample-autonomous-cloud-coding-agents

theagenticguy · 2026-04-02T19:50:39Z

Summary

Adds a new roadmap iteration (3e) and updates design documents to address memory poisoning — classified as OWASP ASI06 (Memory & Context Poisoning) in the 2026 Top 10 for Agentic Applications. Based on a deep research report covering 33 sources (academic papers, CVE disclosures, OWASP standards, and industry analyses).

Changes across 3 files (docs only, no code changes):

ROADMAP.md — New Iteration 3e with 4 implementation phases (input hardening → trust-aware retrieval → detection & response → advanced protections), addressing 9 identified memory-layer security gaps
MEMORY.md — New "Memory Security Analysis" section: full threat taxonomy (intentional + emergent corruption), 9-gap table with severity ratings, 6-layer defense architecture diagram, existing mitigations catalog, academic references
SECURITY.md — OWASP ASI06 classification, expanded attack vectors (MINJA, GitHub issue injection, experience grafting, RAG poisoning, emergent self-corruption), 6-layer defense requirements, updated Known Limitations

Key findings from research

MINJA achieves 95%+ injection success against undefended agent memory (query-only, no direct access needed)
Emergent self-corruption (hallucination crystallization, feedback loops) is equally dangerous — no attacker signature
Defense requires 6 layers working together; no single control is sufficient
The platform has strong network security but 9 memory-layer gaps — most critically, loadMemoryContext() injects retrieved memory without any validation

9 identified gaps

#	Gap	Severity
1	No memory content validation in `loadMemoryContext()`	Critical
2	No source provenance tracking on memory entries	Critical
3	GitHub issue content injected without trust differentiation	Critical
4	No trust scoring at retrieval	High
5	No memory integrity checking (hashing/signatures)	High
6	No anomaly detection on memory patterns	High
7	No memory rollback capability	High
8	No write-ahead validation (guardian pattern)	Medium
9	No circuit breaker for behavioral anomalies	Medium

4-phase implementation roadmap

Phase 1 — Input hardening: content sanitization, GitHub issue filtering, provenance metadata, SHA-256 integrity hashing
Phase 2 — Trust-aware retrieval: trust scoring, temporal decay TTLs, memory validation Lambda
Phase 3 — Detection & response: CloudWatch anomaly detection, circuit breaker, quarantine API, snapshot rollback
Phase 4 — Advanced protections: guardian model, cross-task drift detection, Merkle tree provenance, MINJA/AgentPoison red teaming

Test plan

Verify ROADMAP.md renders correctly with new Iteration 3e section and updated summary
Verify MEMORY.md threat taxonomy tables, gap analysis table, and defense architecture diagram render correctly
Verify SECURITY.md expanded attack vectors and updated Known Limitations are consistent with MEMORY.md
Verify cross-references between ROADMAP.md ↔ MEMORY.md ↔ SECURITY.md resolve correctly
Confirm no breaking changes to existing content (docs-only PR)

Address OWASP ASI06 (Memory & Context Poisoning) with a 4-phase implementation plan based on deep research (33 sources). ROADMAP.md: - Add Iteration 3e between 3d and 4 with 4 phases: Phase 1 (input hardening), Phase 2 (trust-aware retrieval), Phase 3 (detection and response), Phase 4 (advanced protections) - Update summary section with 3e entry MEMORY.md: - Add Memory Security Analysis section with full threat taxonomy (intentional and emergent corruption vectors) - Document 9 identified gaps with severity ratings - Define 6-layer defense architecture - Catalog existing partial mitigations - Add academic and industry references SECURITY.md: - Add OWASP ASI06 classification and context - Expand attack vectors beyond PR review comments (MINJA, GitHub issue injection, experience grafting, RAG poisoning, emergent self-corruption) - Document 6-layer defense architecture requirements - Update Known Limitations with memory security gaps

theagenticguy · 2026-04-02T22:51:28Z

CI Note: The build job is failing due to a transient GitHub API rate limit (403) when mise tries to install trivy@0.69.2. This is unrelated to the docs-only changes in this PR — main branch builds pass fine. Re-running to clear the flake.

…rokoko review aws-samples#1, aws-samples#5, aws-samples#9, aws-samples#12) Four related fixes on the fanout + github-comment surface, from the code review on PR aws-samples#52. Grouped because they share the narrative "defense-in-depth on the fanout dispatcher" — any one landing without the others leaves a hole. ## Findings addressed **aws-samples#1 — Fanout handler returns void despite reportBatchItemFailures: true** The ``FanOutConsumer`` construct (``cdk/src/constructs/fanout-consumer.ts:146``) has ``reportBatchItemFailures: true`` on its DDB Stream event-source mapping. The handler returned ``void``, so Lambda retried the entire batch on any unhandled throw instead of isolating the poisonous record. Combined with aws-samples#5 this could cascade into retry storms and violated the per-task ordering guarantee we rely on (§6.4, AD-9). Fix: handler return type becomes ``Promise<DynamoDBBatchResponse>``. Per-record processing is wrapped in try/catch; caught throws push ``{ itemIdentifier: record.eventID }`` to ``batchItemFailures`` and emit ``fanout.record.failed`` warn. Final ``fanout.batch.complete`` log grows a ``failed`` count. Note: ``DynamoDBStreamHandler`` constrains return to ``void | Promise<void>``, so the handler is typed as a plain 3-arg async function. Lambda's runtime accepts either shape; existing tests (passing ``event, context, cb``) work unchanged. **aws-samples#5 — Unhandled exception in routeEvent crashes batch** ``routeEvent`` uses ``Promise.allSettled`` internally, but ``resolveTokenSecretArn`` can throw ``AccessDeniedException`` SYNCHRONOUSLY before the ``allSettled`` guard is reached. The new per-record try/catch from aws-samples#1 catches these too. **aws-samples#9 — renderCommentBody not self-defending against uncoerced DDB strings** The ``.toFixed(4)`` call on ``costUsd`` is the same bug class as the ``toFixed is not a function`` crash we fixed at the fanout boundary in commit 9fe704e. Today the sole call site coerces via the shared helper; a future caller that forgets to would crash. Fix: ``renderCommentBody`` coerces ``durationS`` and ``costUsd`` internally via the shared ``coerceNumericOrNull`` helper (second line of defense; caller's coercion remains the first). Widened ``CommentBodyInput`` fields to ``number | string | null`` to honestly model the DDB Document-client boundary. **aws-samples#12 — Markdown injection possible via prUrl in GitHub comment body** ``prUrl`` was interpolated directly into a Markdown link target (``[link](${input.prUrl})``). A crafted URL containing ``)`` / ``|`` / ``\n`` could break the table layout or inject content, and a ``javascript:`` scheme could produce a click-to-execute link on some Markdown renderers. Fix: new exported ``sanitizeMarkdownLinkTarget`` helper in ``shared/github-comment.ts`` rejects URLs containing ``\r\n\t\s)|]"<>`` characters, validates via ``new URL()``, and rejects non-http(s) schemes. Returns ``null`` on rejection so ``renderCommentBody`` omits the Pull-request row entirely rather than emitting a broken or unsafe link. ## Tests +22 regression tests net (fanout 7 for aws-samples#1+aws-samples#5 + 3 for aws-samples#9; github-comment 12 for aws-samples#12): - Fanout partial-batch: poison-pill isolation, mixed-batch (good record NOT in failures), observability warn, empty-failures regression guard, baseline pin that today's ``Promise.allSettled`` containment still works. - renderCommentBody numeric self-defense: string-typed values render correctly; non-finite strings collapse to null with warn; null does NOT warn. - sanitizeMarkdownLinkTarget unit tests: accept clean http/https, reject 9 injection patterns, reject 4 non-http schemes (``javascript:``, ``data:``, ``file:``, ``ftp:``), reject malformed, handle null/undefined. Plus end-to-end assertions on ``renderCommentBody`` proving the PR row is omitted on rejection. CDK suite: 1029 passing (was 1001). Refs: krokoko code review on PR aws-samples#52 (findings 1, 5, 9, 12)

…gs (aws-samples#79 review aws-samples#5) The reaction / delete helpers (``addReaction``, ``removeReaction``, ``deleteMessage``) used to log every catch at warn with a single generic event key, lumping API-level rejections (e.g. ``no_reaction``) together with infrastructure failures (DNS lookup, TLS handshake, fetch timeout, JSON parse error from a hostile gateway). Operators who alarmed on the warn rate saw a flat signal that masked genuine infra problems. Split the boundary: - API-level (``!result.ok`` after a successful HTTP call) stays at warn with channel-specific event keys (``fanout.slack.reaction_add_api_error``, ``fanout.slack.reaction_remove_api_error``, ``fanout.slack.message_delete_api_error``). These are per-message UX problems; operators don't page. - Network errors (the outer ``catch (err)`` after ``fetch``) promote to ``logger.error`` with dedicated event keys (``fanout.slack.reaction_add_network_error``, ``fanout.slack.reaction_remove_network_error``, ``fanout.slack.message_delete_network_error``) and ``error_id``s (``FANOUT_SLACK_REACTION_NETWORK``, ``FANOUT_SLACK_DELETE_NETWORK``) so each has its own alarmable signal. User-visible symptoms when these fire silently: stale emoji reactions (hourglass never swaps to ✅) and orphaned intermediate messages. Behaviour unchanged: errors are still swallowed (per-message reactions and intermediate cleanup are best-effort by design; they must not fail the batch), but operators now get distinct metrics for each failure class.

#79) * feat(fanout): migrate SlackNotifyFn to FanOutConsumer subscriber (#64) Move the Slack outbound delivery off its own DynamoDB Streams consumer onto FanOutConsumer as a per-channel dispatcher. Drops TaskEventsTable from 2 concurrent stream readers to 1, restoring headroom for future channels (Email, Teams, etc.) without exceeding the documented DynamoDB Streams 2-reader-per-shard practical limit. The PR also addresses an adversarial code review on the original migration; the body below walks through each piece in the order it landed. ## (a) Migration - `cdk/src/handlers/slack-notify.ts` — rewritten as exported `dispatchSlackEvent(event, ddb)` plus a tagged `SlackApiError` class. The standalone `handler(event)` stream entrypoint is gone; the FanOutConsumer is now the only TaskEventsTable stream reader. Behaviour preserved bit-for-bit: channel_source==='slack' gate, terminal-event dedup via conditional UpdateItem on `channel_metadata.slack_notified_terminal`, threaded replies under the @mention or task_created message, emoji transitions (eyes -> hourglass -> ✅/❌/🚫/⏲), DM channel_id -> user_id rewrite, intermediate session+created message cleanup on terminal events. - `cdk/src/handlers/fanout-task-events.ts` — replaces the log-only `dispatchToSlack` stub with a wrapper that calls dispatchSlackEvent and routes errors via the new typed contract (see (b) below). Slack defaults gain task_created, session_started, task_timed_out so the router fans out the lifecycle events the old SlackNotifyFn handled; the dispatcher's channel_source gate keeps non-Slack tasks unaffected. - `cdk/src/constructs/fanout-consumer.ts` — adds a scoped `secretsmanager:GetSecretValue` grant on `bgagent/slack/*` so the fanout Lambda can fetch per-workspace bot tokens. Same scope the old SlackNotifyFn role held. - `cdk/src/constructs/slack-integration.ts` — deletes SlackNotifyFn, its DynamoEventSource, its IAM policy, and its NagSuppressions entry. Drops the now-unused StartingPosition / FilterCriteria / FilterRule / lambdaEventSources imports. After this lands, `aws lambda list-event-source-mappings` shows exactly one consumer of the TaskEventsTable stream (FanOutFn); verified on the dev stack with end-to-end @mention + cancel + CLI isolation scenarios. ## (b) Review fix #1 — partial-batch retry semantics (BLOCKER) The first review pass found that the post-migration handler silently dropped Slack-side infra errors (DDB throttle on the GetItem, Secrets Manager 5xx, transient Slack timeout). Pre-migration the SlackNotifyFn handler rethrew non-SlackApiError so Lambda retried the batch; post-migration `Promise.allSettled` swallowed the rejection and routeEvent returned an empty list with no escalation path to `batchItemFailures`. routeEvent's return type changed from `NotificationChannel[]` to `{ dispatched, infraRejections }`. The handler now pushes the record into `batchItemFailures` whenever `infraRejections.length>0`, so Lambda replays the record under the partial-batch contract. The warn line on rejection is tagged `retryable: true` so operators can alert distinctly from the channel-terminal swallow path. GitHub got the symmetric treatment: 4xx (excluding the existing 401 and 404 handling) is now treated as a channel-terminal swallow via `fanout.github.api_error` instead of escalating to retry. ## (c) Review fix #2 — split SlackApiError into terminal + retryable Originally any `!result.ok` Slack response was wrapped in SlackApiError and swallowed. That collapsed retryable codes (`ratelimited`, `service_unavailable`, `internal_error`, `fatal_error`, `request_timeout`) into the same swallow as `channel_not_found` — a tier-1 Slack outage would silently drop every message. Introduced `TERMINAL_SLACK_API_ERRORS` set + `classifySlackError` helper. Terminal codes still throw SlackApiError (router swallows). Retryable codes throw a plain Error so the router classifies them as infra rejections and Lambda replays. ## (d) Review fix #3 — NOTIFIABLE_EVENTS / CHANNEL_DEFAULTS drift The original migration added task_created/session_started/task_timed_out to CHANNEL_DEFAULTS.slack but the dispatcher's NOTIFIABLE_EVENTS gate already excluded several events the router was subscribing Slack to (agent_error, pr_created, task_stranded). Result: Slack was reported as `dispatched` for events it silently dropped — telemetry lied, agent_error never reached operators on Slack-origin tasks, and task_stranded rendered the generic "Event: task_stranded for owner/repo" fallback (UX regression). Added render cases for task_stranded and agent_error in slack-blocks.ts and added them to NOTIFIABLE_EVENTS. Forward-compat approval_required and status_response stay out of NOTIFIABLE_EVENTS until their emitters ship; a new cross-file consistency test in fanout-task-events.test.ts fails if anyone re-introduces the drift. The Slack dispatcher wrapper now passes `effectiveEventType` so an agent_milestone(pr_created) wrapper is unwrapped before NOTIFIABLE_EVENTS matching. Without the rewrite, the dispatcher would short-circuit on the wrapper string `agent_milestone`. ## (e) Review fix #4 — conditional UpdateItem on lifecycle persists Once the BLOCKER fix made batches retry, the original task_created and session_started UpdateItem calls became hazardous: a Slack POST that succeeded but whose follow-up UpdateItem failed transiently would, on retry, post a second root and overwrite slack_thread_ts — orphaning every threaded reply that had threaded under the first ts. Both UpdateItems now carry an `attribute_not_exists` ConditionExpression on the relevant `channel_metadata.slack_*_msg_ts`. On ConditionalCheckFailedException the handler logs at info, deletes the duplicate Slack message via `chat.delete`, and returns. Sibling retry wins the race; the duplicate is cleaned up. ## (f) Dev-stack regression: drop pr_created from Slack defaults Live verification surfaced a UX duplication: pr_created (subscribed in CHANNEL_DEFAULTS.slack as the original §6.2 design called for) and task_completed both rendered messages with View PR buttons, posted seconds apart. The original SlackNotifyFn had silently dropped pr_created (NOTIFIABLE_EVENTS gate), so users hadn't relied on it. Removed pr_created from CHANNEL_DEFAULTS.slack and from NOTIFIABLE_EVENTS, and removed the prCreatedMessage renderer. GitHub keeps pr_created (its edit-in-place comment surface genuinely benefits from the early checkpoint). ## Verification - mise //cdk:compile — clean - mise //cdk:test — 1183 / 1183 pass (8 net-new tests added for the review fixes: NOTIFIABLE_EVENTS drift guard, retryable Slack codes, GitHub 4xx swallow, infra rejection escalation, SlackApiError swallow, task_stranded render) - mise //cdk:eslint — clean - mise //cdk:synth — confirms exactly one Lambda::EventSourceMapping on TaskEventsTable, pointing at FanOutFn - Dev-stack scenarios — @mention happy path, Cancel button, CLI submit (channel_source=api -> zero Slack dispatches, GitHub edit-in-place still fires) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(fanout): retry GitHub 403/429 instead of swallowing as terminal (#79 review #1) PR #79 review found that the new 4xx terminal-swallow path treats HTTP 403 and 429 as channel-terminal — but on GitHub these are transient rate-limit responses (403 with "API rate limit exceeded", 429 "Too Many Requests"). Under a reconciliation wave that touches many tasks, an entire window of GitHub comment updates would be permanently dropped with only a warn log. Carve out 403 and 429 from the swallow guard so they propagate as infra rejections through ``Promise.allSettled``. The record lands in ``batchItemFailures`` and Lambda replays until the rate-limit window clears (or DLQs after ``retryAttempts``). Test coverage: parametrized over 403 + 429 with a GitHubCommentError mock at the helper boundary, asserting the record's eventID surfaces in ``batchItemFailures`` rather than being absorbed. * fix(fanout): guard Slack Secrets Manager grant on a prop (#79 review #2) Every other external-service grant in FanOutConsumer (taskTable, repoTable, githubTokenSecret) is gated by ``if (props.X)``, so a deployment that hasn't onboarded the corresponding service stays free of dangling IAM permissions. The original migration broke the pattern with an unconditional ``bgagent/slack/*`` Secrets Manager grant — dev stacks without Slack onboarding ended up holding read permission on a resource pattern they never use, with a misleading ``cdk-nag AwsSolutions-IAM5`` suppression reason. Adds an optional ``slackSecretArnPattern`` prop on ``FanOutConsumerProps``; the policy statement is only attached when the prop is set. ``cdk/src/stacks/agent.ts`` now computes the ``bgagent/slack/*`` ARN inline and passes it through, mirroring the other guarded props. ``ArnFormat`` and ``Stack`` imports moved out of fanout-consumer.ts since the construct no longer needs them. No changes to live behaviour — agent.ts always passes the prop, so the IAM policy still attaches in production. The dispatcher will log-and-fail-retry on a missing pattern (covered by review #3 fix). Test gap covering the construct itself ships in a follow-up commit (test gap #34). * fix(fanout): throw on missing TASK_TABLE_NAME env var (#79 review #3) Pre-fix: when ``TASK_TABLE_NAME`` was unset on a Slack-subscribed event, ``dispatchSlackEvent`` returned silently after a warn line. The router counted Slack as ``dispatched`` and a broken stack quietly dropped every Slack notification — operators only saw it in the warn-rate metric, with no rejected-channel signal. Post-fix: throw a plain Error so the rejection propagates as an infra rejection through ``Promise.allSettled``. The router pushes the record into ``batchItemFailures``, Lambda retries the batch, the ``fanout.dispatcher.rejected`` warn fires per record, and operators get a distinct alarm. Also bumps the existing log line from ``warn`` to ``error`` and attaches an ``error_id: FANOUT_SLACK_MISSING_TASK_TABLE`` so the deployment-bug case can be distinguished from per-record failures. Test: ``throws when TASK_TABLE_NAME env var is missing`` deletes the env var, asserts the throw, asserts no DDB call was attempted (env-var guard fires first). * fix(fanout): match SlackApiError by name as well as instanceof (#79 review #7) When a bundler ever duplicates the slack-notify module (rare with NodejsFunction tree-shaking but possible if dual-bundled), two distinct SlackApiError classes coexist and ``instanceof`` against one fails for instances of the other. The dispatcher would see a foreign-class SlackApiError, fall through to the rethrow branch, and the router would treat it as an infra rejection — flipping a channel-terminal swallow into infinite Lambda retries. Add an ``err.name === 'SlackApiError'`` fallback so the swallow branch fires either way. Mirrors the duck-typed ``GitHubCommentError`` check used elsewhere in the same handler. Test: synthesise a plain Error with name === 'SlackApiError' (NOT an instance of the mock's SlackApiError class) and assert batchItemFailures stays empty — proving the swallow path catches both shapes. * fix(fanout): extend TERMINAL_SLACK_API_ERRORS with permission codes (#79 review #8) Original set omitted documented Slack permission/scope failures. Codes outside the set fall to the retryable branch, so a misconfiguration like ``ekm_access_denied`` or ``missing_scope`` would burn 3 Lambda retries before DLQ on every event — even though the failure is fundamentally a configuration bug that no retry can clear. Adds: - Permission/scope: missing_scope, ekm_access_denied, team_access_not_granted, posting_to_general_channel_denied - Payload shape: invalid_arguments Reorganized the set into commented blocks (channel-shape, auth, permission/scope, payload-shape) so future additions go in the right bucket and the rationale stays visible. Test coverage: parametrized over the full TERMINAL_SLACK_API_ERRORS set (21 codes) — every one must throw SlackApiError so the router swallows it. The existing retryable test.each remains intact and covers the negative-class case (codes outside the set throw a plain Error and escalate to retry). * fix(fanout): promote Slack reaction/delete network errors to error logs (#79 review #5) The reaction / delete helpers (``addReaction``, ``removeReaction``, ``deleteMessage``) used to log every catch at warn with a single generic event key, lumping API-level rejections (e.g. ``no_reaction``) together with infrastructure failures (DNS lookup, TLS handshake, fetch timeout, JSON parse error from a hostile gateway). Operators who alarmed on the warn rate saw a flat signal that masked genuine infra problems. Split the boundary: - API-level (``!result.ok`` after a successful HTTP call) stays at warn with channel-specific event keys (``fanout.slack.reaction_add_api_error``, ``fanout.slack.reaction_remove_api_error``, ``fanout.slack.message_delete_api_error``). These are per-message UX problems; operators don't page. - Network errors (the outer ``catch (err)`` after ``fetch``) promote to ``logger.error`` with dedicated event keys (``fanout.slack.reaction_add_network_error``, ``fanout.slack.reaction_remove_network_error``, ``fanout.slack.message_delete_network_error``) and ``error_id``s (``FANOUT_SLACK_REACTION_NETWORK``, ``FANOUT_SLACK_DELETE_NETWORK``) so each has its own alarmable signal. User-visible symptoms when these fire silently: stale emoji reactions (hourglass never swaps to ✅) and orphaned intermediate messages. Behaviour unchanged: errors are still swallowed (per-message reactions and intermediate cleanup are best-effort by design; they must not fail the batch), but operators now get distinct metrics for each failure class. * fix(fanout): emit fanout.slack.dup_delete_failed on ghost-message accumulation (#79 review #6) The conditional UpdateItem dup-delete path (``task_created`` / ``session_started`` lifecycle persists) calls ``deleteMessage`` to clean up the duplicate Slack message that landed when a sibling retry won the race. The delete is inherently best-effort — but if it fails, the duplicate becomes a permanent ghost in the thread and operators had no way to alarm on the rate. Refactor ``deleteMessage`` to return a boolean (``true`` on success or ``message_not_found``-as-already-gone, ``false`` otherwise) and emit a dedicated ``fanout.slack.dup_delete_failed`` event with an ``error_id: FANOUT_SLACK_DUP_DELETE_FAILED`` from the dup-delete callsites when the cleanup couldn't complete. The terminal-event cleanup paths (``slack_session_msg_ts``, ``slack_created_msg_ts``) intentionally don't fire this event — those paths target genuinely-stale UX cleanup, not retry-driven duplicates, so an alarm there would be noise. No new tests beyond the existing dup-delete coverage; the ``deleteMessage`` return value isn't yet asserted at the unit level, but the behavior is fully exercised by the existing ``dup-delete`` integration paths (test gap #31 will add an explicit failure-path assertion when it lands). * chore(fanout): tighten RouteOutcome arrays to ReadonlyArray (#79 review #9) ``RouteOutcome.dispatched`` and ``infraRejections`` were typed as plain ``NotificationChannel[]`` — which made ``readonly`` on the property prevent reassignment but still allow callers to mutate the underlying array via ``.push``, ``.splice``, or ``.sort``. Inconsistent with the ``ReadonlySet<string>`` used for ``CHANNEL_DEFAULTS`` in the same file. Tightening to ``ReadonlyArray<NotificationChannel>`` makes the contract honest: the router owns the arrays, callers read them. Test suite updated to use ``[...outcome.dispatched].sort()`` where it previously called ``.sort()`` directly — the explicit copy makes the intent clear and would have surfaced any silent test-side mutation. * refactor(fanout): make SlackDispatchEvent a type alias of FanOutEvent (#79 review #10) The two interfaces were structurally identical: same five fields, same readonly modifiers, same metadata shape. The decoupling was purely nominal and a silent-drift footgun — adding a field to ``FanOutEvent`` (e.g. when the router starts plumbing an ``approval_required`` ID through) would not flow into ``SlackDispatchEvent``, leaving the dispatcher unaware until a downstream test happened to fail. Replace with a one-line type alias: export type SlackDispatchEvent = FanOutEvent; The slack-notify module now type-imports ``FanOutEvent`` from fanout-task-events. ``import type`` is erased at compile time, so the runtime bundle still has the one-way dep (fanout-task-events → slack-notify) — no module-cycle hazard. Reviewer-suggested ``Pick<FanOutEvent, 'task_id' | …>`` was considered and rejected: the dispatcher uses every field of ``FanOutEvent``, so the Pick would just enumerate the same five fields with extra noise. A direct alias keeps the intent obvious and prevents drift identically. * fix(fanout): generalize Slack dedup to cover agent_error + log Retry-After (#79 review #4) PR #79 review #4 surfaced a sibling-channel-failure hazard: when GitHub or Email rate-limits, the record lands in ``batchItemFailures``. On the Lambda retry, every Slack-subscribed event for that record runs again. Terminal events were already guarded by ``slack_notified_terminal``; ``agent_error`` was not — operators would page twice on a single agent failure if a sibling channel happened to fail. Generalize the dedup mechanism. ``TERMINAL_EVENTS`` is replaced by a ``SLACK_DEDUP_ATTRIBUTE`` map that marks each event type with the ``channel_metadata`` attribute that should guard the post: - 5 terminals share ``slack_notified_terminal`` (any first-arriving terminal claims the right; subsequent terminals dedup against it) - ``agent_error`` gets its own ``slack_dispatched_agent_error`` so a duplicate agent_error doesn't reuse the terminal slot - ``task_created`` / ``session_started`` map to ``null`` because they already use the per-event ``slack_*_msg_ts`` conditional persists from review #1 — the conditional already provides full idempotency (a separate marker would be redundant) Also surfaces Slack's ``Retry-After`` header on rate-limited responses through a dedicated ``fanout.slack.retryable_api_error`` warn so operators reading CloudWatch can see the recovery window instead of guessing from sustained warn rate. Tests: - logs Retry-After header on rate-limited Slack responses (new): asserts ``retry_after_seconds`` propagates from Slack's response header into the warn metadata - existing terminal-codes parametrized test untouched (terminal branch doesn't read headers) - existing retryable test gains a ``headers: { get: () => null }`` stub on the fetch mock so the headers.get call doesn't crash Reviewer suggested a per-channel dispatch bitmap as the alternative. Rejected as premature: the duplicate-GitHub-PATCH is harmless (idempotent), Email is still a stub, and the dedup map covers the specific agent_error pain identified above. A bitmap would add a new table + IAM grants + per-dispatch DDB cost for a hypothetical problem (Slack rate-limiting AND a sibling channel failure). * test(fanout): conditional UpdateItem race + dup-delete coverage (#79 test gap) Adds 4 tests covering the lifecycle-persist conditional path that review fix #1 introduced and review fix #6 hardened. Pre-PR-#79 the only ConditionalCheckFailed coverage was the terminal-dedup path; the new lifecycle-persist + dup-delete code lacked direct assertions and was flagged 9/10 criticality by the reviewer. - task_created persist ConditionalCheckFailed → posts duplicate then deletes it: pins the cleanup behaviour that prevents ghost task_created posts in the channel - session_started persist ConditionalCheckFailed → posts duplicate then deletes it: parallel coverage for the other lifecycle attribute (slack_session_msg_ts) - dup-delete failure emits fanout.slack.dup_delete_failed with error_id: pins the operator-alarm signal added in review fix #6; asserts both the event key and the FANOUT_SLACK_DUP_DELETE_FAILED error_id propagate - chat.delete returning message_not_found is treated as success (no dup_delete_failed): negative-class assertion. Prevents false-positive alarms when the race resolves cleanly (the duplicate was already deleted by a prior retry). The ghost / message_not_found tests use ``fetchMock.mockImplementation`` URL-routing rather than ``.mockResolvedValueOnce`` chains because ``updateReaction`` issues 2-3 reaction-API fetches between chat.postMessage and chat.delete; routing by URL keeps the test focused on the load-bearing chat.delete behaviour without coupling to reaction call order. * test(fanout): cover task_stranded + agent_error renderers (#79 test gap #32) Pre-PR-#79 the new ``taskStrandedMessage`` and ``agentErrorMessage`` helpers in slack-blocks.ts had no direct unit tests. Reviewer flagged this as a 7/10 gap because the renderers carry the prior_status / error_type / message_preview metadata threaded through from the event source — silent drift in the metadata field names would produce ugly fallback messages in production. Adds 5 tests: - task_stranded WITH metadata renders the prior_status parenthetical (``Task stranded for org/repo (last status: RUNNING)``) so operators can tell at a glance whether the task hung in HYDRATING vs RUNNING — without the parenthetical the reviewer's "generic Event: ..." UX regression would resurface. - task_stranded WITHOUT metadata still renders cleanly (legacy events written before the reconciler started stamping metadata must not crash or leak ``undefined``). - agent_error with full metadata (error_type + message_preview) renders the rotating_light, type, and preview. - agent_error WITHOUT metadata stays sensible — no leaked ``undefined`` strings or empty ``_Type:_`` line. - agent_error truncates a 500-char message_preview to keep Slack channel UX readable. * test(fanout): cover agent_error dedup + dedup-slot isolation (#79 test gap #33) Pre-PR-#79 review-fix #4 there was no direct test for the ``slack_dispatched_agent_error`` dedup attribute or its interaction with the existing ``slack_notified_terminal`` slot. A future refactor that collapsed the two slots — or renamed one of them — would silently break the sibling-channel-failure-retry guarantee that fix #4 added. Adds 4 tests: - ``agent_error claims its own dedup attribute``: pins the UpdateExpression and ConditionExpression strings so a refactor that renames the attribute breaks loudly. - ``agent_error retry hits the dedup guard``: end-to-end scenario matching review #4 — task already has ``slack_dispatched_agent_error: true``, retry must short-circuit before chat.postMessage. Without the guard, a second rotating_light fires. - ``terminal dedup attribute is per-class``: a flaky task_completed-then-task_failed sequence dedups against the same ``slack_notified_terminal`` slot. Catches the regression where the orchestrator emits both terminal types and we'd otherwise post both ✅ and ❌. - ``agent_error and terminals use distinct dedup slots``: the important negative — having ``slack_dispatched_agent_error`` set must NOT shadow a subsequent ``task_completed``. Pins the slot separation so a future merge into a single slot can't silently drop terminals after an agent_error. * test(fanout): add construct-level tests for FanOutConsumer (#79 test gap #34) The construct shipped on issue #64 with no unit-level coverage of its IAM contract. The only synth-level signal lived inside ``slack-integration.test.ts`` ("0 EventSourceMapping") which proved the migration didn't regress the OTHER construct. Reviewer flagged this 6/10 — and the gap is what allowed review #2 (unconditional Slack secret grant) to slip through in the first place. Adds 6 tests: - ``attaches a single DynamoEventSource on the TaskEventsTable stream``: pins the architectural invariant — issue #64 was fundamentally about reaching exactly-one stream reader. Adding a second consumer must fail this test loudly. - ``creates a DLQ for the fanout Lambda``: pins retention period + presence; a DLQ-less deployment would silently drop poison-pill records past retryAttempts. - ``omits the bgagent/slack/* grant when slackSecretArnPattern is not provided``: the review #2 invariant. Iterates every IAM::Policy and asserts NONE of them grant secretsmanager:* on a bgagent/slack/* ARN. A regression that re-introduces the unconditional grant breaks this test. - ``attaches the bgagent/slack/* grant only when slackSecretArnPattern is provided``: the positive case. Pins the grant shape (action, effect, resource pattern). - ``passes TASK_TABLE_NAME env var when taskTable is provided``: review #3 dependency — the dispatcher throws on missing env. - ``omits TASK_TABLE_NAME env var when taskTable is not provided``: graceful degrade for dev stacks that haven't onboarded the TaskTable yet (matches the construct's documented contract). * test(fanout): cover task_stranded through terminal dedup (#79 test gap #35) The reconciler at handlers/reconcile-stranded-tasks.ts:170 emits BOTH ``task_stranded`` and ``task_failed`` for a heartbeat-expired task — one for the operator signal, one to drive the FAILED status transition. Pre-PR-#79 this pair had no test coverage; reviewer flagged this 8/10 because the visible failure mode (a paired "Task stranded" + "Task failed" double-page in Slack) would surface in production but be silent in CI. Adds 2 tests: - ``task_stranded posts and writes the terminal dedup marker on first arrival``: pins that task_stranded participates in the shared terminal slot and renders the warning message with metadata. Catches a regression that omits task_stranded from the dedup map entirely. - ``task_stranded after a sibling task_failed dedups``: the operational scenario — task_failed already claimed ``slack_notified_terminal``; the subsequent task_stranded must short-circuit before chat.postMessage. Without this guard, operators get the double-page the reviewer warned about. * fix(fanout): re-read TaskRecord before terminal cleanup to close orphan-message race Live observation during PR #79 review verification: the same Slack @mention happy path sometimes leaves the 🚀 task_created message in the thread (orphaned beside the ✅ task_completed) and sometimes deletes it cleanly. The race window: 1. ``task_created`` stream batch posts the rocket message and persists ``slack_created_msg_ts`` via the conditional UpdateItem introduced in PR #79 review fix #1. 2. ``task_completed`` stream batch fires ~30s later. Its initial GetItem races the prior UpdateItem and sees a stale ``channel_metadata`` WITHOUT ``slack_created_msg_ts``. 3. The terminal cleanup branch checks ``channelMeta.slack_created_msg_ts`` — undefined — silently skips the chat.delete. The rocket message stays in the thread. Add a fresh GetItem inside the TERMINAL_EVENTS cleanup branch, after the dedup UpdateItem has linearized our view of the table. Any prior ``slack_*_msg_ts`` writes are visible by then, so the cleanup fires correctly. On a re-read failure (DDB throttle / transient blip) we fall back to the dispatch-entry snapshot and emit ``fanout.slack.cleanup_reread_failed`` so operators can alarm on the rate. Pre-existing race (the unconditional UpdateItem in pre-PR-#79 was the same shape — wrote, GetItem on the next batch could miss it). PR #79 doesn't introduce it but doesn't fix it either; this commit does, since the live screenshot evidence appeared during review verification. Tests: - ``terminal cleanup re-reads TaskRecord``: scripts a stale dispatch-entry GetItem followed by a fresh re-read GetItem with ``slack_created_msg_ts`` present; asserts chat.delete fires against the freshly-read ts. - ``terminal cleanup falls back to dispatch-entry snapshot when re-read fails``: defense-in-depth — DDB throttle on the re-read must not break terminal delivery; cleanup uses the entry snapshot and emits the fallback warn. --------- Co-authored-by: bgagent <bgagent@noreply.github.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: Alain Krok <alkrok@amazon.com>

Addresses the blocker + critical items from PR review: - Refresh-token race (review blocker). Linear rotates refresh_tokens on every use; concurrent Lambdas/agents racing the same secret will all read the same expiring token and one's refresh will succeed while the others get `invalid_grant`. On `invalid_grant`, re-read the secret from Secrets Manager (bypassing cache). If the refresh_token has changed, another caller already rotated; use the freshly-read token (or retry refresh once if it's also expiring). If unchanged, the refresh_token is permanently rejected and the workspace needs re-onboarding. Implemented in both the TS resolver (linear-oauth-resolver.ts) and Python resolver (config.py). - Unguarded bedrock_agentcore import in agent/src/server.py (review critical). The bare `from bedrock_agentcore.runtime.context import BedrockAgentCoreContext` inside `_run_task_background` killed the entire pipeline thread with no diagnostic if the SDK was missing or its module structure changed. Wrap in try/except (ImportError, AttributeError) and log via _warn_cw — the Linear token resolver has its own SM fallback, so the agent can proceed without the workload-token bridge. - Cache invalidation on fetch-level refresh failure (review high). The TS resolver's `invalidateLinearOauthCache()` only ran in the `!resp.ok` branch; if `fetch()` itself threw (timeout, DNS), the catch returned null without invalidating, leaving the stale expiring token cached for 60s and hammering Linear's token endpoint. Move invalidate into the fetch-level catch too. - Malformed expires_at log (review medium). The Python `_is_expiring` caught `ValueError` and silently returned True, masking consistently-bad writes. Add a WARN log so operators see the bad data instead of just an unexplained refresh on every task. - Positive-path refresh log (review non-blocking aws-samples#5). Added INFO-level breadcrumb on successful refresh in both resolvers so operators diagnosing intermittent 401s have a trace of which workspace refreshed and to what expiry. 11/11 existing resolver unit tests still pass; will add tests for the new race-recovery branch in a followup commit.

theagenticguy requested a review from a team as a code owner April 2, 2026 19:50

theagenticguy changed the title ~~Add Iteration 3e: Memory security and integrity (OWASP ASI06)~~ docs: Memory security and integrity Apr 2, 2026

theagenticguy changed the title ~~docs: Memory security and integrity~~ docs: add Iteration 3e for memory security and integrity (OWASP ASI06) Apr 2, 2026

theagenticguy changed the title ~~docs: add Iteration 3e for memory security and integrity (OWASP ASI06)~~ feat: add Iteration 3e for memory security and integrity (OWASP ASI06) Apr 2, 2026

krokoko approved these changes Apr 2, 2026

View reviewed changes

krokoko merged commit af6dafe into main Apr 2, 2026
7 of 12 checks passed

krokoko deleted the feature/memory-security-roadmap branch April 6, 2026 23:07

leandrodamascena mentioned this pull request Apr 12, 2026

Agent / Orchestrator: Handle merge conflicts in pr_iteration workflow - NEED DISCUSSION #22

Open

2 tasks

scottschreckengaust mentioned this pull request Apr 24, 2026

docs: add least-privilege deployment roles and deployment guide #46

Merged

13 tasks

isadeks mentioned this pull request May 13, 2026

feat(fanout): migrate SlackNotifyFn to FanOutConsumer subscriber (#64) #79

Merged

8 tasks

scoropeza mentioned this pull request May 19, 2026

Migrate dashboard LogQueryWidget widgets to EMF metrics #115

Open

3 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: add Iteration 3e for memory security and integrity (OWASP ASI06)#5

feat: add Iteration 3e for memory security and integrity (OWASP ASI06)#5
krokoko merged 1 commit into
mainfrom
feature/memory-security-roadmap

theagenticguy commented Apr 2, 2026

Uh oh!

Uh oh!

theagenticguy commented Apr 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

theagenticguy commented Apr 2, 2026

Summary

Key findings from research

9 identified gaps

4-phase implementation roadmap

Test plan

Uh oh!

Uh oh!

theagenticguy commented Apr 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants