Zaxy v0.5 to v1.0 Release-Gate Roadmap

Zaxy is moving from a technically strong memory substrate into a stable product: coordinator memory for multi-agent projects.

The v1.0 goal is not to add every possible memory feature. It is to make the existing thesis reliable, legible, and useful enough that external users can adopt Zaxy with confidence:

Zaxy is the leading open-source system for auditable, replayable, and coordinated memory in multi-agent projects.

Current Baseline

This roadmap starts from the current 0.4.x codebase, not from an empty plan. Several capabilities that normally belong in later phases are already present:

README positioning already leads with "Coordinator memory for multi-agent projects".
zaxy init defaults to the local embedded Codex path with no Neo4j sidecar.
Embedded Kuzu is the default projection backend; Neo4j, pgGraph, and LatticeDB remain optional or experimental paths.
Memory Bootstrap and Memory Checkout are the model-facing context contracts.
MCP exposes memory, context, lifecycle, and coordination tools.
Zaxy Coordinate includes mission, worker, assignment, finding, review, promotion, approval packet, handoff, conflict, and benchmark flows.
LangGraph has a dependency-light native-beta adapter; CrewAI remains a dependency-light native-preview adapter.
CoordinationBench, LongMemEval-compatible reports, backend shootout tooling, release smoke checks, beta UAT, and docs validation already exist.
Test coverage is already near the v1.0 bar, so 92% is a ratchet target rather than a future aspiration.

The work from v0.5 to v1.0 should therefore be release-gate driven: each release must make a narrower public promise, prove it with docs, examples, tests, and benchmarks, and avoid broad speculative expansion.

v1.0 Success Criteria

Zaxy v1.0 is ready when these statements are true:

Positioning is clear: Coordinator Memory for Agent Teams.
A new user can install, initialize, inspect, and run a meaningful example in less than five minutes on a normal local development machine.
MCP is polished enough that clients can discover tools, understand errors, and use structured outputs without reading source code.
LangGraph is first-class and documented as the primary native framework path.
At least one direct model-call integration path works outside MCP.
Coordinate supports real mission workflows: isolated workers, cited findings, conflict review, approval, promotion, checkout, audit, and handoff.
Public benchmarks are transparent about methodology, baselines, limitations, latency, citation coverage, and token tradeoffs.
Public APIs, MCP schemas, CLI commands, and event payloads have a documented v1.0 stability contract.
Coverage stays at or above 92%, with release gates preventing regressions in core memory, coordination, and onboarding flows.
External validation collection remains available as optional post-release evidence when an outside user, project, or case study is available.

Release Principles

Minor releases before v1.0 may introduce breaking changes, but every breaking change must be documented in the changelog and migration guide.
v1.0 freezes the public API, MCP schemas, stable CLI surfaces, and durable event payload contracts.
Every release includes a changelog entry, updated docs, release smoke output, and at least one polished example or benchmark artifact.
Eventloom remains the source of truth. Projection backends remain derived views.
MCP remains the primary framework-neutral interface, but native integrations must share the same Memory Bootstrap, Memory Checkout, capture, and feedback contracts.
Backend experiments do not become defaults unless they preserve temporal semantics, citations, inferred-edge metadata, session isolation, dashboard rendering, rebuild behavior, and published benchmark guardrails.

v0.5: Public Positioning and First-Run Trust

Theme: Make the product legible and trustworthy from the first command.

Target: 4 weeks.

Implementation status: completed against docs/superpowers/plans/2026-05-30-v0-5-public-positioning-first-run-trust.md; the public positioning, first-run docs, quickstarts, examples, changelog, and static site coverage are in place.

Ship

Align package metadata, README, docs homepage, site copy, and examples around "Coordinator Memory for Agent Teams".
Keep architectural language available, but stop making "temporal knowledge graph fabric" the first public explanation.
Turn the current quick start into a measured first-run path: install, zaxy init, memory bootstrap, memory checkout, doctor, and one example.
Polish three examples:
single-agent durable memory;
LangGraph memory checkout before model work;
three-worker Coordinate mission with conflict review and handoff.
Improve MCP tool descriptions for model-facing clarity and client discovery.
Publish public docs for "Why Zaxy", "Getting Started", "MCP Quickstart", "Coordinate Quickstart", and "Architecture".
Create a lightweight external validation script for one person outside the project to run and report time-to-first-success.

Gates

Clean install to successful zaxy doctor completes in less than five minutes on a normal local development machine.
zaxy init leaves a user with clear next steps and no required sidecar.
All polished examples run from a clean checkout.
Documentation links validate.
Coverage remains at or above 92%.
Release artifact and release smoke checks pass.

Explicit Non-Goals

Do not expand backend scope.
Do not make hosted or multi-tenant memory a v0.5 deliverable.
Do not add another framework adapter before the first-run story is clean.

v0.6: MCP and Native Runtime DX

Theme: Become the best MCP-native memory backend while preparing native runtime integrations.

Target: 4 to 5 weeks after v0.5.

Implementation status: completed with the canonical MCP tool contract snapshot at docs/examples/mcp-tool-contract.json, a matching schema test in tests/test_mcp.py, representative response snapshots at docs/examples/mcp-response-snapshots.json for bootstrap, checkout, graph retrieval, verbatim retrieval, context assembly, feedback, and coordination checkout, and structured MCP error payloads with stable error codes and remediation hints. zaxy hook-status now reports structured memory activation remediations with runnable checkout commands, zaxy doctor surfaces the same missing or stale checkout state as a memory_activation check, and top-level zaxy status now reports memory activation freshness, latest checkout/capture details, token efficiency, and checkout remediation commands. LangGraph checkout middleware now emits the zaxy.native.v0.6 metadata contract with diagnostics, quality, feedback guidance, and fail-closed error payloads, and zaxy doctor --release-smoke now runs examples/langgraph_memory.py as the LangGraph release-validation smoke. The shared non-MCP adapter lifecycle and payload keys are published at docs/examples/native-integration-contract.json. CrewAI checkout middleware shares that same non-MCP native contract while remaining native-preview. The MCP Quickstart now documents one recommended local setup route each for Codex, Claude Code, Claude Desktop, Cursor, and generic MCP clients. Beta readiness now checks docs/examples/first-run-timing-report.json so clean first-run doctor and example timings remain under the five-minute budget. The configured coverage ratchet and pytest coverage gate now enforce the 92% floor.

Ship

Harden MCP structured outputs for memory, checkout, context assembly, feedback, and coordination tools.
Standardize MCP error payloads with stable error codes, human messages, and remediation hints.
Add MCP contract tests that snapshot tool names, schemas, required fields, and representative responses.
Improve memory_checkout output for client use: stable summary fields, diagnostics, citation coverage, stale-state warnings, required actions, and feedback templates.
Improve zaxy status, zaxy doctor, and zaxy hook-status around last checkout, last capture, stale memory, missing hooks, and degraded projection states.
Promote LangGraph from native-preview toward beta by stabilizing payload keys, examples, and failure behavior. Status: LangGraph now reports native-beta and emits the zaxy.native.v0.6 checkout contract.
Define the shared native integration contract for non-MCP runtimes:
before model/task call: bootstrap or checkout;
after model/task call: capture assistant or task output;
after tool call: capture redacted observation;
after context use: record feedback.

Status: published in docs/examples/native-integration-contract.json.

Document Claude Desktop, Claude Code, Cursor, Codex, and generic MCP setup paths with one recommended local route each. Status: covered in docs/mcp-quickstart.md and rendered to the static docs.

Gates

MCP schema snapshot tests protect the public tool surface.
LangGraph example runs as part of release validation. Status: covered by the langgraph_example check in zaxy doctor --release-smoke.
Status and doctor commands return actionable remediation for common local failures.
Clean first-run time improves or remains below the v0.5 threshold.

Status: covered by the first_run_timing check in zaxy doctor --beta-readiness.

Coverage remains at or above 92%.

Status: enforced by pyproject.toml and scripts/check-coverage.py.

Explicit Non-Goals

Do not promise full API stability yet.
Do not promote CrewAI or AutoGen beyond current maturity until LangGraph and the shared contract are stable.

v0.7: Coordination Workflows

Theme: Make Coordinate production-useful, not only demonstrable.

Target: 5 to 6 weeks after v0.6.

Ship

Add mission templates for common workflows such as software delivery, research review, benchmark investigation, and release validation.

Status: built-in templates are available through zaxy coordinate template list, show, and apply; applying a template creates replayable mission, worker, and assignment events.

Improve approval flows so pending, conflicted, stale, and evidence-poor findings have obvious next actions. Status: approval packets and static review exports now include explicit next-action guidance for normal review, conflict resolution, stale evidence refresh, and missing evidence.
Improve conflict review:
deterministic source-state conflicts remain default;
lexical semantic conflicts remain opt-in;
hosted semantic conflict adapters remain bounded and auditable.

Status: deterministic source-state conflicts are the default, local lexical semantic conflicts are opt-in through --semantic-conflicts lexical, and hosted HTTP semantic conflict adapters remain opt-in with bounded request payloads and locally validated response IDs.

Improve the dashboard or CLI mission viewer so users can inspect mission state, worker ledgers, findings, evidence, decisions, and promoted state without reading Eventloom JSONL. Status: zaxy coordinate inspect provides a replay-only CLI viewer with mission brief, worker ledgers, findings by status, evidence, review decisions, promoted state, conflicts, approval packet, and handoff records.
Add audit report generation for a mission from Eventloom replay.

Status: zaxy coordinate audit-report generates a read-only Markdown or JSON report from mission and worker Eventloom replay with sequence/hash citations for every audited event.

Publish a CoordinationBench report with clear baselines, adapter status, limitations, and reproduction commands. Status: the coordination-real-v1 report is archived at reports/benchmarks/coordination-real-v1/coordination-benchmark.md with Zaxy metrics, local baselines, disclosure-only adapter status, limitations, and reproduction commands.
Add a polished multi-agent example that includes review and approval steps, not just worker reporting. Status: examples/coordinate_three_worker_project.py now demonstrates a three-worker mission with approval packet export, approval application, accepted promotion, conflict/defer decisions, mission inspection, audit reporting, checkout, and handoff.

Gates

A user can complete a full mission workflow from CLI and MCP: start mission, create workers, assign work, report findings, detect conflicts, review findings, promote accepted state, checkout accepted memory, and create handoff. Status: covered across Coordinate CLI tests and MCP coordination tool tests, including conflict materialization, approval application, checkout, and handoff.
The same workflow is demonstrated through LangGraph or a direct native helper.

Status: CoordinationAdapter now covers start, worker creation, assignment, finding report, conflict detection, approval application, checkout, and handoff as JSON-friendly native helper calls.

CoordinationBench report generation is reproducible from tracked inputs.

Status: coordination-real-v1 regenerates from benchmarks/coordination-real-v1/coordination-workload.json through zaxy coordinate benchmark --workload ....

Audit report output cites Eventloom sequence and hash metadata.

Status: covered by CoordinationManager.audit_report and zaxy coordinate audit-report --json tests.

Coverage remains at or above 92%. Status: enforced by the configured pytest coverage gate and coverage ratchet.

Explicit Non-Goals

Zaxy does not spawn or manage workers itself unless a separate orchestrator design is approved.
Zaxy does not infer accepted findings from raw transcripts without explicit evidence and review.

v0.8: Model-Native Integrations and Observability

Theme: Make memory activation work where model calls actually happen, not only through MCP.

Target: 5 to 6 weeks after v0.7.

Ship

Add direct model-call integration modules outside MCP, sharing the same native contract defined in v0.6. Status: dependency-light OpenAI-compatible and Claude-compatible adapters now run Memory Checkout before provider-shaped model calls, inject checkout context, record bounded model-call metadata, capture sanitized assistant turns, and return the shared zaxy.native.v0.6 metadata.
Prioritize two paths:
OpenAI-compatible model-call wrapper for request/response capture, checkout injection, tool observation, and feedback; Status: available in zaxy.adapters.openai_compatible with redacted tool-call observation and direct memory feedback helpers.
Anthropic or Claude-style wrapper if the local usage path and API boundary can be kept stable. Status: available in zaxy.adapters.claude_compatible as a duck-typed client.messages.create(**request) helper with the same capture, observation, feedback, and fail-closed checkout contract.
Keep direct integrations dependency-light and optional. Core install remains small. Status: both direct model-call adapters are duck-typed and import no provider SDK.
Add examples showing model-call memory activation without MCP. Status: examples/openai_compatible_memory.py and examples/claude_compatible_memory.py run with fake provider-shaped clients and no network or provider dependency.
Add trace correlation from mission, checkout, context assembly, model call, tool call, finding, review, and handoff. Status: zaxy.trace now builds provider-neutral zaxy.trace.v0.8 spans and edges from replayed Eventloom events with sequence/hash citations, and zaxy trace export --json exposes correlated mission, checkout, model-call, tool-call, finding, review, promotion, and handoff paths without requiring a tracing provider.
Improve Pathlight integration or add neutral trace hooks that can feed LangSmith, Phoenix, or local JSONL traces without making any one provider a hard dependency. Status: the neutral trace export is JSON-friendly and provider-independent; zaxy trace export --format jsonl --output trace.jsonl writes ingestion records for local JSONL pipelines while Pathlight remains optional.
Improve replay tools for long-running missions, including branch/fork design if it can be implemented without destabilizing Eventloom semantics. Status: zaxy replay now supports inclusive --from-seq/--to-seq windows so operators can inspect bounded ranges in long Eventloom logs without rewriting history.
Continue performance work on compaction, context assembly, projection rebuild, and query latency based on benchmark evidence. Status: beta readiness now includes a named benchmark_no_regression check over the release benchmark commands so performance work stays tied to reproducible smoke, 40-query performance, and 100-query scale reports.

Gates

At least one outside-MCP direct model integration runs in a documented example. Status: zaxy doctor --release-smoke now runs both examples/openai_compatible_memory.py and examples/claude_compatible_memory.py.
Model-call capture is redacted, bounded, and opt-in where provider cost or privacy could surprise users. Status: direct model adapters capture bounded request metadata and sanitized assistant turns only when applications explicitly call the adapter helpers.
Trace output can follow a useful path from mission to model call to promoted finding. Status: covered by trace correlation and CLI tests that link mission spans to model calls, tool observations, reviewed findings, promoted findings, and handoffs.
Benchmarks show no regression in checkout quality, citation coverage, and latency budgets. Status: zaxy doctor --beta-readiness reports benchmark_no_regression and requires checkout quality floors, citation coverage at 1.0, and p95/p99 checkout latency budgets in scripts/release-check.sh.
Coverage remains at or above 92%. Status: enforced by the configured pytest coverage gate and coverage ratchet.

Explicit Non-Goals

Do not turn the packet analyzer into a required router.
Do not make direct provider integrations the only recommended path; MCP remains the primary framework-neutral route.

v0.9: Hardening and API Freeze Candidate

Theme: Prepare for a stable 1.0 contract.

Target: 4 to 5 weeks after v0.8.

Ship

Publish an API inventory covering:
MCP tool names, schemas, and response contracts;
Python SDK public classes and functions;
stable CLI commands and options;
durable Eventloom event types and payload fields;
projection backend contract;
benchmark artifact schemas. Status: docs/api-inventory.md now inventories these surfaces and links each category to its current contract authority.
Mark each surface as stable, beta, experimental, or internal. Status: the API inventory uses explicit Stable, Beta, Experimental, and Internal labels and keeps candidate backends and Skill Memory out of the stable release surface.
Add migration tests and docs for upgrades from 0.4 through 0.9. Status: docs/migration.md now covers each release band from 0.4 to 0.9, the compatibility test expectations, and the non-destructive rollback policy.
Add fuzz tests for Eventloom payload validation, hash-chain replay, and bounded MCP inputs. Status: parametrized validation tests now reject non-object and oversized Eventloom payloads, sequence-tampered hash chains, and unbounded direct memory_append MCP inputs.
Add failure-injection or chaos tests for projection rebuild, corrupted projection artifacts, missing hooks, stale checkout, and degraded backends.

Status: projection rebuild failure recovery is covered by test_reproject_command_closes_pggraph_backend_after_projection_failure; corrupted projection artifacts are handled through integrity-first replay, reset/rebuild, and rollback paths; missing hooks and stale checkout are covered by hook-status activation and capture tests; degraded backends are surfaced through fallback metrics including zaxy_degraded_operations_total and documented alerting guidance.

Expand release gates so the public examples, MCP smoke, LangGraph smoke, Coordinate mission smoke, benchmark comparison, docs validation, and beta UAT all run or fail with explicit skip reasons. Status: scripts/release-check.sh now has named commands for public examples, MCP smoke, LangGraph smoke, Coordinate mission smoke, docs validation, benchmark comparison, and beta UAT; zaxy doctor --beta-readiness reports release_gate_surface_coverage and accepts intentional omissions only as SKIP:<reason>.
Add contributor guide, issue templates, and benchmark contribution guidance.

Status: CONTRIBUTING.md, GitHub issue templates, and docs/benchmark-contributions.md now document the production-code bar, release checks, issue evidence, and reproducible benchmark contribution requirements.

Freeze v1.0 candidate schemas and begin treating changes as migration events.

Status: docs/examples/v1-schema-freeze.json now binds the v1 freeze candidate surfaces, and docs/agent-events.md defines schema.migration.proposed and schema.migration.applied for non-additive stable or beta contract changes.

Gates

Full release gate passes on Python 3.11, 3.12, and 3.13.
Coverage remains at or above 92%.
No public benchmark claim depends on untracked or local-only inputs.
API inventory has no undocumented stable surfaces.
Migration guide covers 0.4 to 0.9.

Status: current evidence for these v0.9 gates is recorded in docs/v09-gate-audit.md.

External-user feedback is tracked as optional post-release evidence. Status: collection packet and evidence shape are documented in docs/v09-gate-audit.md.

Explicit Non-Goals

Do not add major new feature families after v0.9 unless they are required to fix a release blocker.
Do not change event schemas casually after the freeze candidate.

v1.0: Stable Coordinator Memory Release

Theme: Stable, documented, benchmarked, and positioned release.

Target: 2 to 3 weeks after v0.9.

Ship

Final API and data model stability commitment. Status: docs/stability-commitment.md defines the v1.0 public API and data model compatibility promise, migration-event requirement, and non-commitments for experimental/internal surfaces.
Comprehensive changelog from 0.4 to 1.0. Status: CHANGELOG.md now has explicit 0.9.0 - Release Candidate and 1.0.0 - 2026-05-31 sections above the existing 0.4+ release history.
Migration guide from 0.4. Status: docs/migration.md covers upgrades from 0.4 through the v0.9 freeze candidate and links migration events for post-freeze stable or beta schema changes.
Public v1.0 announcement with positioning, examples, benchmark evidence, limitations, and roadmap beyond 1.0. Status: the stable 1.0.0 announcement is in docs/announcements/zaxy-v1.0.md; it keeps external validation as optional post-release evidence and links the public verification request.
Updated website and docs. Status: generated static site pages include the v1.0 announcement, stability commitment, release validation checklist, gate audit, API inventory, and migration guide.
Final release smoke command and release validation checklist. Status: docs/release-validation-checklist.md records the v1.0 gates, aggregate commands, artifact review steps, and optional external validation path.
Final release gate evidence audit. Status: docs/v10-gate-audit.md maps clean-repo UAT, MCP smoke, LangGraph smoke, direct model integration smoke, Coordinate mission smoke, benchmark guardrails, docs validation, release smoke, coverage, public stability tags, and optional external validation to proof or evidence guidance.
External validation note, case study, or public user feedback if available.

Status: collection packet and report template are published in docs/external-validation.md; outside-user evidence is optional for v1.0.

Gates

Clean-repo UAT passes. Status: command-level evidence is recorded in docs/v10-gate-audit.md; final release execution is tracked by docs/release-validation-checklist.md.
MCP smoke passes. Status: command-level evidence is recorded in docs/v10-gate-audit.md.
LangGraph smoke passes. Status: command-level evidence is recorded in docs/v10-gate-audit.md.
At least one direct model integration smoke passes. Status: command-level evidence is recorded in docs/v10-gate-audit.md.
Coordinate mission smoke passes. Status: command-level evidence is recorded in docs/v10-gate-audit.md.
Benchmark guardrails pass. Status: command-level evidence is recorded in docs/v10-gate-audit.md.
Docs validation passes. Status: command-level evidence is recorded in docs/v10-gate-audit.md.
Release smoke passes. Status: command-level evidence is recorded in docs/v10-gate-audit.md.
Coverage remains at or above 92%. Status: command-level evidence is recorded in docs/v10-gate-audit.md.
Public surfaces are tagged with their stability level. Status: command-level evidence is recorded in docs/v10-gate-audit.md.
External validation is optional post-release evidence. Status: the evidence shape is documented in docs/v10-gate-audit.md.

Cross-Cutting Workstreams

Benchmarks

Run and publish benchmark evidence regularly. Zaxy benchmark claims should show the workload, input fingerprint, baseline, metrics, latency, citation coverage, token tradeoffs, and limitations.

Priority benchmark lanes:

LongMemEval-compatible memory retrieval;
CoordinationBench;
backend shootout for embedded, Neo4j, pgGraph, LatticeDB, and BM25;
first-run onboarding time;
activation efficiency for high-context sessions;
audit and replay completeness.

Documentation and Content

Every release needs docs and public communication. Content should focus on auditable coordination, replayable memory, MCP-native adoption, LangGraph, and real examples.

Testing and CI

The v1.0 quality bar is coverage at or above 92%, with stable tests for public contracts. The release process should continue to include ruff, mypy, pytest, coverage ratchet, docs validation, release smoke, and benchmark guardrails.

Community and External Validation

External validation is optional post-release evidence, not a v1.0 release blocker. Start with one or two users who can run the first-run path and one example when available. Capture where they get stuck, and turn that into docs, doctor checks, or CLI improvements.

Competitive Awareness

Track Mem0, Zep, Letta, MemPalace, Agent Memory, ActiveGraph, and the MCP ecosystem. Avoid fake apples-to-apples claims. Keep competitor comparisons tied to reproducible adapter contracts, public disclosures, or clearly labeled limitations.

v1.1: Accepted-State Recovery and Benchmark Guardrails

Theme: Make Zaxy's accepted-state memory thesis falsifiable under stale, distracting, incomplete, and no-safe-answer histories.

Implementation status: in progress on the 1.1.0 release line.

Ship

Promote StateRecoveryBench as an official benchmark lane with a canonical tracked workload, report schema, Markdown report, guardrail checker, and release-check wiring.
Add zaxy state-recovery-benchmark as the public command. Keep zaxy experimental state-recovery as a research/debug path.
Use memory_fabric_checkout as the production guardrail baseline. Treat associative projection rows as diagnostic research baselines only.
Resolve Coordinate accepted state from parent mission replay so checkout and proof packets agree on accepted findings, diagnostic rows, review refs, promotion refs, and worker source refs.
Update public docs, API inventory, migration notes, testing guidance, changelog, and benchmark docs without broadening product claims beyond the measured lane.

Gates

`python scripts/check-state-recovery-benchmark.py reports/benchmarks/state-recovery-v1/state-recovery-benchmark.json --workload reports/benchmarks/state-recovery-v1/state-recovery-workload.json

--require-git-tracked-inputs`

`zaxy state-recovery-benchmark --output-dir /tmp/zaxy-state-recovery

--workload reports/benchmarks/state-recovery-v1/state-recovery-workload.json`

Focused tests for StateRecoveryBench, Coordinate accepted-state resolution, Memory Checkout coordinate suppression, CLI behavior, ruff, mypy, and release smoke.

Explicit Non-Goals

Do not promote astro-associative projection to core product behavior in 1.1.
Do not claim StateRecoveryBench replaces LongMemEval or CoordinationBench.
Do not tune behavior to benchmark labels; only authority, review, promotion, provenance, stale, and citation metadata may determine accepted state.

Primary Risks

Risk	Mitigation
Scope creep	Treat auditable coordination, MCP, LangGraph, native model-call activation, and DX as the only v1.0 pillars.
Low external feedback	Keep v0.5 first-run validation and v0.7 Coordinate example feedback as optional post-release evidence, while local release gates prove the documented paths.
Benchmark credibility	Require tracked inputs, fingerprints, baselines, methodology, and limitation notes.
Breaking changes	Inventory public surfaces in v0.9 and document every migration.
Backend distraction	Keep embedded Kuzu as default unless another backend beats the same gates without sidecar friction or quality loss.
Native integration fragmentation	Force MCP and direct integrations through the same Memory Bootstrap, Memory Checkout, capture, and feedback contracts.

Minimum Viable v1.0

If timeline or feedback pressure requires a narrower release, preserve only these pillars:

Polished positioning and docs.
Excellent MCP experience.
First-class LangGraph path.
Production-useful Coordinate workflow.
One outside-MCP model integration.
Transparent benchmark evidence.
Stable public API and data model contracts.

Everything else can move past v1.0.

Related references: README.md, site/index.html, Coordinate roadmap, and benchmarks.