Hackathon submission packet

This is a working, copy-ready submission packet for the CockroachDB AI Agents Hackathon. Bracketed fields are release blockers, not claims. Replace them only after the linked artifact is public and verified. The revision-10 AWS/CockroachDB Cloud deployment is verified; the public-relevance receipt is historical revision-7 evidence, the reference-agent and replacement artifacts are historical revision-6 evidence, and the representative Cloud EXPLAIN is separately captured historical plan evidence. The final public video and entrant-controlled Devpost fields remain release blockers. The release-state table is authoritative. The live revision-10 AWS observations establish only the read-only HTTP route surface and its returned CockroachDB-backed data; they do not identify or attest the deployed database principal or grant matrix. The current checked-in PUBLIC-03 reader boundary is supported by local proofs and remains unapplied to that historical live stack.

Submission deadline: August 18, 2026 at 5:00 PM EDT / 4:00 PM CDT. The submitted project must remain available free of charge and without restriction through the end of judging: September 15, 2026 at 5:00 PM EDT / 4:00 PM CDT. Do not tear down or scale the judging deployment to zero before that hold expires.

Devpost fields

Project name

OSTK Fleet Recall

Tagline

Agents are replaceable. Their memory shouldn't be—and when two disagree, memory should say so.

Repository

https://github.com/os-tack/ostk-fleet-recall

Working demo

https://d13zrqfh66r7ub.cloudfront.net

Testing instructions

Open the working-demo URL without signing in. Start with the specification-versus-code example. Confirm its exact open conflict, both sides, relevant evidence cards, and measured server/round-trip time. Try the migration example for the second exact conflict mapping. The CockroachDB, Rust, project-purpose, and datastore-library examples should return relevant evidence without a conflict; nonsense should return no results and no conflicts. Expand “View raw evidence envelope” to inspect the bounded CockroachDB-backed result. Open a repository-backed source card and confirm that its link targets commit 56b577c82b9c5a5c80d73103f7f6b56d51698872 with an exact #Lx-Ly anchor. Inline Markdown, code, and a relative repository link rendered as a code-styled anchor should remain bounded without horizontal overflow at desktop and 390px mobile sizes. Then open https://d13zrqfh66r7ub.cloudfront.net/healthz and confirm {"status":"ready"}. The public surface exposes only read routes: /, /healthz, /api/status, and bounded POST /api/recall; no public mutation route is exposed. That is a route-level observation only and says nothing about the credential below the router. Do not infer from it that the live revision-10 task uses the checked-in fleet_publication principal or fleet_publication_reader grants.

The public URL uses CloudFront's default certificate. Its generated-hostname viewer policy has a TLSv1 minimum and may negotiate newer TLS; CloudFront uses restricted HTTP to the ALB origin. The origin accepts only the CloudFront origin-facing prefix list plus a secret header. Do not describe this topology as end-to-end TLS or as enforcing TLS 1.2 at the viewer.

Demo video (under three minutes)

[PUBLIC_YOUTUBE_OR_VIMEO_URL]

Project thumbnail

docs/assets/devpost-thumbnail-v2.png — 1536×1024 PNG (3:2, under 5 MB).

Built-with tags

Rust, CockroachDB, Vector Search, Agent Skills, AWS, Amazon ECS, AWS Fargate, Amazon S3, Application Load Balancer, Amazon CloudFront, Amazon ECR, AWS Secrets Manager, Amazon CloudWatch, Terraform, MCP, model2vec, Axum, SQLx. Add OSTK only if the optional adapter is actually used in a final captured artifact; it is not part of the default proof path.

The listed AWS services were exercised by the live deployment and evidence capture, not merely declared in Terraform. Remove a tag before submission if its live integration is later removed or cannot be shown safely.

Qualifying checkboxes

CockroachDB: Distributed Vector Indexing and Agent Skills Repo.
AWS: Amazon ECS / EKS and Amazon S3. Mention ALB, ECR, Secrets Manager, and CloudWatch in the narrative or “Other” field if available.
Do not select CockroachDB Cloud Managed MCP Server unless it is separately used against the final cluster and captured.

Subjective form answers to confirm with the entrant

Learning level: suggested Significant.
AI value usable in your career: suggested Yes.
Team members: [CONFIRM_DEVPOST_TEAM].

Gallery plan

Project thumbnail: ready at docs/assets/devpost-thumbnail-v2.png.
Architecture: ready at docs/assets/architecture.png, generated from the validated deployment topology in docs/ARCHITECTURE.md; regenerate it only if the hosted topology changes.
Proof: capture the agent action/conflict result and post-replacement cloud recall without account IDs, secret ARNs, credentials, or tenant-sensitive logs.

Run ./docs/assets/verify-media.sh before uploading the gallery. The script checks the committed image formats, dimensions, and conservative 5 MB limit; the manual visual acceptance checks live in docs/assets/README.md.

Inspiration

A strong agent can remember locally, but a fleet has a different problem: workers restart, run on different hosts, and can simultaneously learn incompatible things. A shared vector database alone finds similar text; it does not establish scope, provenance, correction, replay safety, or honest conflict semantics. We wanted Recall's local-first memory quality to remain intact while giving any agent fleet one durable place to coordinate, with OSTK available as an optional orchestrator rather than a runtime dependency.

What it does

Fleet Recall is a distributed memory backend and MCP service for agent fleets, including OSTK. It combines CockroachDB's scoped C-SPANN vector indexes with lexical retrieval, then hydrates Recall-compatible chunks and fuses candidates. Agents can also record deliberate typed claims with provenance. If two active claims make incompatible assertions about the same key, Fleet Recall marks them disputed and returns an explainable conflict instead of silently overwriting one.

Every Fleet Recall process is bound to a trusted tenant/project scope. Mutations are serializable and idempotent: an identical canonical request retried with the same tenant-wide key produces at most one durable claim, receipt, and audit trail, while independent incompatible writes remain distinct and become an explicit conflict. ECS workers are disposable; the live deployment keeps CockroachDB Cloud as the memory source of truth after every task is replaced. A historical revision-6 forced deployment changed the complete serving task set to a fully disjoint set, while public lexical/dense RRF continued to return the exact action and escalation claims 16 and 18. That replacement proof was not rerun for revision 10.

The submission core does not depend on OSTK or an LLM. Its default AWS agent is a standalone deterministic reference policy implemented in Fleet Recall. Four separately deployment-bound one-off Fargate tasks observe durable memory, apply an explicit rollout-safety policy, and persist cited actions; the wrapper emits a verified summary only when their claim, action, exact conflict, and escalation identifiers form one correlated chain. The policy agent resolves search results through exact get and rereads each durable action before attesting it. That is agentic policy execution, not a claim of autonomous LLM reasoning. An optional OSTK adapter can run the same chain with bounded model sessions through a checked-in Bash bridge; it remains opt-in and unclaimed until an explicitly authorized live run produces verified evidence. Its boundary is documented in docs/OSTK_DEMO.md.

How we built it

Rust 1.94 application and MCP protocol with bounded inputs/outputs, running as immutable ARM64 image git-56b577c82b9c at source commit 56b577c82b9c5a5c80d73103f7f6b56d51698872 on ECS/Fargate.
CockroachDB Cloud 26.2.5 is the live memory plane for corpus chunks, typed claims, passage vectors, conflicts, idempotency receipts, and events. The revision-10 idempotent rich-seed task exited zero and upserted exactly 552 rows: 346 documentation chunks, 2 code chunks, and 204 operations chunks. /api/status reports the vector, lexical, conflict-membership, and claim-support-chunk indexes, working cosine distance, and embedding dimension 512.
Distributed Vector Indexing with project- and source-prefixed VECTOR(512) cosine indexes, plus a stored TSVECTOR inverted index for hybrid recall.
CockroachDB Agent Skills pinned at commit e14e86d: the coding agent invoked the transaction and SQL skills, then used their guidance to shape the fresh 40001 retry loop, embedding-outside-transaction boundary, SQL invariants, connection pooling, concurrency tests, and EXPLAIN plan gates. The full skill-to-code/test trace is in docs/AGENT_SKILLS_AUDIT.md.
Current PUBLIC-03 source and local proof: the externally provisioned fleet_publication principal inherits one NOLOGIN fleet_publication_reader role with only database CONNECT, public-schema USAGE, and SELECT on _sqlx_migrations, memory_corpus_models, memory_chunks, memory_claim_embeddings, memory_claim_support, memory_claims, memory_conflict_members, and memory_conflicts; it has no sequence, DML, or DDL authority. The exact official-v26.2.3 TLS wrapper and secondary Docker RBAC proof passed locally. The accepted LocalStack receipt additionally binds the production image, prefix 1–17, reader status/recall, writer denial, replacement persistence, and zero residue at clean commit cd6ecfca2c1a6d112ba058aad899a21aa34bb0f4. It explicitly proves no TLS, database-password authentication, IAM enforcement, Fargate, or AWS apply.
The checked-in AWS ECS/Fargate, Application Load Balancer, and CloudFront deployment targets a replaceable public demo task, S3 for private content-addressed model delivery, Secrets Manager for Cockroach TLS URLs, ECR for scanned immutable images, and CloudWatch for logs/Container Insights. A dedicated one-off Fargate task definition runs the deterministic reference policy agent under A/B/C deployment identities; deploy/aws/run-reference-agent.sh verifies their correlated durable-memory chain. Historical revision-6 run devpost-final6-20260814T143523Z, using reference-agent task definition revision 6, verified decision/action/incompatible/escalation claims 15/16/17/18 and open conflict 5. The replacement wrapper used serving revision 6 and verified exact claims 16 and 18 after a fully disjoint task-set change. Separately, the historical revision-6 source-conflict self-audit used claims 9/10 and open conflict 3 to prove that a semantic result can carry an exact hash-bound documentation/code disagreement. The publication-safe self-audit, reference-agent, replacement, and validation receipts are checked in. These self-audit/agent/replacement receipts were not rerun for revision 10. The live serving, migration, seed, and reference-agent task-definition families are revision 10, and the service is 1/1 healthy. The ECR Basic OS-package scan completed with an empty finding-severity count; it does not claim language or application-dependency coverage. The current PUBLIC-03 Terraform suite passed all 21 static tests, but those planned execution/task-role changes remain unapplied; the live revision-10 route must not be described as a deployed PUBLIC-03 database boundary.
Backend-neutral corpus traits extracted into upstream ostk-recall, pinned by immutable Git revision so the local and fleet stores share retrieval contracts without making local Recall depend on CockroachDB.

Do not add CockroachDB Managed MCP to this list unless the final team actually uses it against the submission cluster and captures the inspection evidence. The qualifying tool pair is Distributed Vector Indexing plus Agent Skills.

Challenges

CockroachDB vector indexes reward a schema designed around their execution shape. Range predicates cannot become equality prefixes for ANN, so active recall rows are physically separated from history and source-specific dense search gets its own prefixed index. Time filtering is explicitly bounded and diagnosed rather than disguised as exact.

The harder semantic problem was contradictory memory. We made claims, provenance, conflict membership, transitions, and request receipts one atomic serializable unit. Embeddings are computed before the transaction, while the entire database unit is retried only on 40001. That keeps transactions short and makes contention a normal, tested path.

Accomplishments

One model generation per project, pinned by a digest of the exact local files.
Tenant/project predicates and keys on every data path, including adversarial request tests that reject authority smuggling.
Hybrid vector/lexical Recall semantics without linking the local LanceDB implementation into the fleet backend.
Explainable, coverage-aware conflicts and at-most-one durable claim mutation with idempotent replay under concurrency.
Real CockroachDB capability and query-plan tests, including representative C-SPANN and inverted-index evidence.
A dormant-by-default AWS Terraform deployment definition that separates one-off DDL credentials from a non-admin, writer-capable runtime secret. The live seed/serving verification and historical revision-6 migration, four-task-agent, and replacement runs prove that historical topology; the current publication-reader split is locally proven source and remains unapplied to that stack.

What we learned

Distributed memory is primarily a correctness and trust problem, not an ANN API swap. The vector index determines important table boundaries; serializable retries determine application orchestration; and model identity is part of the data contract. CockroachDB's transaction and SQL Agent Skills were particularly useful because they turned those distributed-system concerns into reviewable code and test gates.

What's next

Package Fleet Recall as an optional OSTK Recall plugin/backend.
Add authenticated workload identity and dynamic multi-project routing without trusting MCP parameters.
Add set-based bounded ingestion, changefeed-driven projections, and multi-region locality policies.
Load-test numeric claim-ID access paths and introduce hash sharding or a versioned UUID public contract where needed.
Add private AWS/Cockroach connectivity, WAF/rate limiting, operational SLOs, and contention/hot-range alerts.

Pre-existing work disclosure

OSTK and ostk-recall existed before the hackathon and remain independently useful local-first projects. The Fleet Recall CockroachDB schema/store, claim and conflict persistence, fleet MCP adaptation, deployment, demo, and the backend-neutral upstream extraction used here were built for this project. Git history and pinned revisions make that boundary inspectable.

Tool-use evidence

Requirement	Evidence to show judges	Release state
CockroachDB persistent memory	Local 26.2 schema/round-trip/conflict tests; current local PUBLIC-03 product/RBAC/production-image proofs; live seed/status/API verification; historical public relevance, cloud self-audit, reference-agent, and restart capture	Revision 10 deployed exactly 552 rich-corpus rows (346 documentation, 2 code, 204 operations). Current API results carry the exact release revision and source-line ranges, and the final seven-query smoke gate passed. Those are route-level live observations, not a deployed-reader-grant attestation. The exact official-v26.2.3 TLS wrapper, Docker RBAC proof, and clean-commit LocalStack receipt cover the current reader source locally; the checked-in seven-query receipt remains historical revision-7 evidence, while the agent and replacement receipts remain historical revision-6 evidence and were not rerun for revision 10
Distributed Vector Indexing	DDL and local representative `EXPLAIN`; cloud plan and reference-agent lexical+dense RRF capture	Live capability flags, cosine operation, dimension 512, and lexical/dense RRF verified. The publication-safe Cloud `EXPLAIN` selects both C-SPANN indexes and the lexical inverted index for the exact production SQL shapes on 10,001 disposable rows; all assertions passed
CockroachDB Agent Skills	Pinned coding-agent invocation audit mapping each skill to decisions, code/tests, and accepted deviations	Ready in repository
AWS service	ECS/ALB/CloudFront/S3/Secrets/ECR/CloudWatch Terraform; one-off self-audit/reference-agent wrappers; LocalStack contract harness	Live revision-10 image and all four task-definition families are deployed; the service is 1/1 healthy, the 552-row rich seed succeeded, the ECR Basic OS-package scan completed with empty severity counts, and final desktop/390px mobile QA passed. The seven-query public-relevance receipt is historical revision-7 evidence; the two-task self-audit, four-task reference run, and disjoint serving-task replacement are historical revision-6 evidence. The current PUBLIC-03 Terraform tests passed 21/21 locally, but the changes remain unapplied and are not part of those live claims
Public open source	Source, licenses, locked dependencies, setup and sample data	Published and anonymously verified at https://github.com/os-tack/ostk-fleet-recall
Functional URL	HTTPS landing page, bounded recall, `/healthz`	Live at https://d13zrqfh66r7ub.cloudfront.net; health, status, and source-linked bounded recall verified
Public video <3 min	Live AWS UI plus reviewed cloud agent/replacement receipts by default, then YouTube/Vimeo link and final duration; standalone Fleet Recall is optional local footage and OSTK is an optional alternate	46-second 1600×900 sanitized rehearsal rendered; `[RENDER_LIVE_NARRATE_AND_UPLOAD]`

Judging-criteria evidence map

The official judging criteria are equally weighted. This table keeps the final story balanced instead of treating successful vector search as the whole project.

Criterion	Strongest repository evidence	Final proof still to capture
Agentic Memory Design	Hybrid lexical/vector recall, typed claims, hash-bound source support, correction-aware conflict membership, the checked-in revision-6 claims-9/10 source self-audit, and the historical revision-6 claims-15–18 four-task cloud run	Show the source cards carrying open conflict 3, then label the cited operator escalation from conflict 5 as historical revision-6 evidence
Technological Implementation	Prefix-scoped C-SPANN and inverted indexes, schema-2 claim-support index, backend-neutral Recall extraction, short serializable retries, immutable model identity, conformance and adversarial tests, plus the publication-safe Cloud plan and live receipts	Show the live revision-10 image/task identity and label the historical revision-7 relevance and revision-6 agent/replacement receipts accurately in the final video
Real-World Impact	A recalled schema-migration decision changes the next agent's rollout action; an incompatible memory stops rollout and produces a cited operator handoff	Explain how the same pattern applies to long-running coding, operations, and research fleets
Product Readiness	Trusted deployment scope, bounded input/output and hydration, route-level read-only live AWS surface, current local PUBLIC-03 product/RBAC/production-image proofs, idempotent receipts, historical revision-6 four-task proof, and historical revision-6 exact recall after full task replacement	Capture final publication-safe footage without exposing identifiers or secrets; explain both the CloudFront-to-ALB HTTP boundary and the fact that current reader grants are locally proven but not deployed to the historical live stack
Creativity & Originality	Local Recall stays local-first while Fleet Recall adds a distributed memory plane; disagreement is durable first-class state rather than an overwrite or a hidden rank choice	Make the conflict-aware memory-to-action transition the visual center of the final video

Final video plan (target 2:40)

VIDEO_DEMO.md is the single authoritative recording plan. It opens on the live conflict-first interface and follows three acts: ASK → DISAGREE → SURVIVE, followed by a short local-first-to-fleet coda. Do not use an older architecture-first timeline or replace serving tasks manually while filming.

The current UI runs at the revision-10 release boundary. Final desktop and 390px mobile QA verified safe inline Markdown, immutable exact source-line links, a relative repository link rendered as a code-styled anchor, and no horizontal overflow. The final seven-query smoke gate passed. The checked-in self-audit, reference-agent, replacement, and validation artifacts remain historical revision-6 evidence; the Cloud plan is separately captured historical evidence. None was rerun during the revision-10 cutover. The checked-in seven-query public-relevance receipt is historical revision-7 evidence for the prior 548-row release; it is not the revision-10 smoke receipt, replacement, or reference-agent proof. Local or rehearsal terminal footage is supporting material, never AWS evidence. The final export needs narration or another accessible audio track; the silent rehearsal is not submission-ready. The accepted current LocalStack PUBLIC-03 receipt is also local supporting evidence only; it must not be edited together with the revision-10 browser footage in a way that implies the checked-in reader grant matrix is deployed.

Navigation Menu

Uh oh!

Hackathon submission packet

Devpost fields

Inspiration

What it does

How we built it

Challenges

Accomplishments

What we learned

What's next

Pre-existing work disclosure

Tool-use evidence

Judging-criteria evidence map

Final video plan (target 2:40)

Final release checklist

Uh oh!

FilesExpand file tree

SUBMISSION.md

Latest commit

History

SUBMISSION.md

File metadata and controls

Hackathon submission packet

Devpost fields

Inspiration

What it does

How we built it

Challenges

Accomplishments

What we learned

What's next

Pre-existing work disclosure

Tool-use evidence

Judging-criteria evidence map

Final video plan (target 2:40)

Final release checklist