Changelog
Announced releases of the Ninchi platform, newest first. Routine fixes and infrastructure work ship continuously between these entries and are not listed.
The API contract and ninchi-cli are versioned separately. See the documentation.
Ninchi 1.2.0
To be tagged platform-v1.2.0 at 4270a6c, the commit running in
production (EKS and Vercel) on the release date. 296 pull requests merged
between 1.1 and this release.
Ninchi 1.2.0 brings education to individuals and teams: a learning loop for self-directed learners, and an Education module for instructors and their cohorts. It also adds the My Data hub, the Review Guide for code reviewers, Japanese across the product, and many quality-of-life improvements and fixes.
Education for teams and classrooms
The Education module is enabled per organization by Ninchi.
- Cohorts and rosters (EDU-001, EDU-002A, EDU-019, EDU-020, EDU-014, EDU-024). Instructors create cohorts, manage rosters with names and pending invites, rename, archive or hand over a cohort, import cohorts from GitHub Teams, and invite students by email or CSV.
- Comprehension Checks (EDU-002, EDU-003, EDU-011, EDU-011B, EDU-012,
EDU-015, EDU-018, EDU-022, EDU-023). Instructor-launched question sets
generated from students' repositories or their pull-request branches,
with named per-repository perspectives, substantive file selection, a
/ninchicommand on student pull requests, "next question" navigation, and edit, archive and delete. Launches stay within one installation token and resume if issuance times out. Large GitHub organizations get paged repository lists, repositories off by default with name-pattern enablement, and one-request bulk enablement (EDU-013, EDU-017, EDU-021). - Classroom dashboard and evidence (EDU-004). Per-student and per-perspective evidence with drill-down, an audited evaluation export, and a student view of their own evidence.
- Assessment mode (EDU-005). A per-answer timer, copy-paste off, an automated AI-assist check and Teach Me held until the check closes. It makes it harder to paste in an answer and records signals an instructor can look at. The AI-assist result is a signal, not proof, and it is shown only to instructors.
- Student privacy (EDU-025, EDU-027). Students see their own surfaces only; organization-level reads are refused for them at the API.
- Japanese rubric validation (EDU-016). Generated rubrics and key points are checked in the language the question was asked in.
Self-directed learning
- Learning loop (UX-022, UX-017, UX-016). Each result explains why the question was asked and leads with the gap, links to a re-verification, and offers practice. Teach Me is easier to find, and a challenge can be rerolled before answering.
- Questions scaled to the change (ENT-042, ENT-043). Small changes get intent questions, ordinary changes get current-behavior and edge-case questions, and large or contract-touching changes get design-rationale questions.
- Per-member difficulty (UX-044, UX-045). Members can raise their own challenge difficulty above the organization's setting, never below it, within the plan's difficulty cap.
My Data hub
- One place for personal evidence (UX-026A, UX-041, UX-037). Challenge history, narrative, profile and achievements, with an organization switcher. Onboarding samples are kept out of personal evidence (UX-029).
- My Settings (UX-026B). Personal preferences in one page.
Review Guide (Pro and above, per-organization opt-in)
- A guide for the human reviewer (REV-001 to REV-005, REV-007, REV-010). A projection of the pull request for its reviewer, with a unified diff, reviewer attestation, links from the Ninchi PR comment, and attention counts for the organization. Pull requests from authors who are not organization members are covered too (REV-009).
- Linked-ticket requirements (REV-013, REV-014), beta. An AI-summarised section that quotes the linked ticket from a trusted ticket-bot comment and points to related changed files. It never judges whether the pull request meets the requirements. Enterprise on GitHub, off by default and enabled per organization by Ninchi.
Japanese across the product
- The console, onboarding, organization settings, dashboards, knowledge map and transactional emails in Japanese (I18N-001A to I18N-001F).
- Customer API errors carry stable codes and localized messages (I18N-001G).
- Consoles follow the organization's language, and each challenge has a language toggle (I18N-002, I18N-003).
Quality of life
- Unified app and organization navigation and tabbed organization settings (UX-019, UX-020, UX-021, UX-027).
- One Ninchi comment per pull request, reused when a challenge is superseded (UX-036).
- An organization policy for pull requests from non-members, with a silent option (PMF-006A).
- Role-aware onboarding checklist and rotating unscored samples (UX-028, UX-034); plan-aware presets (UX-033); clearer exits from challenges and revisitable results (UX-030); challenge links open in a new tab (UX-039).
- Datadog Developer Platform listing (DD-004).
- Fixes include renamed repositories still receiving challenges, App reinstalls re-linking the existing organization, stale Stripe invoice events ignored, and several dashboard and admin 500s.
Platform and security
- Production runs on EKS with Fargate; the EC2 host is decommissioned (ZTNA-01).
- Egress is default-deny behind AWS Network Firewall (SEC-11, SEC-21).
- Operator access is MFA-gated and audited, with IAM-authenticated database access and a read-only diagnostics role (ZTNA-03).
- A synthetic canary customer and PagerDuty paging (SRE-001).
- A measured backup restore drill (SEC-08).
- VPC flow logs, database and load-balancer alarms, an IAM password policy and 365-day log retention (OPS-055).
Not included in this release
- Ninchi Direct. The concept previews at
/demo(DEMO-001 to DEMO-010) are demonstrations, not product. - Weekly email digest (GTM-009). Remediated (opt-in, Pro-gated) but
still disabled:
digest_sweeper_enableddefaults false. - Executive report export (GTM-007). The shared report payload exists; there is no export yet.
Ninchi 1.1
Tagged platform-v1.1. The last functional commit in this release is
05ed87c, which disarms the weekly digest; everything after it under the tag
is documentation.
The announced bundle is AI-spend accountability: connecting vendor usage data to recorded evidence of demonstrated understanding, so an organization can see what its AI spend is producing in evidence terms and where AI-assisted change lands without verified understanding. Enterprise plans only.
AI-spend accountability
- Vendor usage connections (GTM-012, GTM-019, GTM-022). Organizations can connect GitHub Copilot and Anthropic with read-only credentials and have Ninchi pull organization-level usage on a daily sweep. The two are not equivalent: Copilot reports seats and aggregate utilization but no cost, while Anthropic reports tokens and cost. Vendor data is aggregated before it is stored — no per-user rows are persisted. Connection health is re-validated daily, and failures name their real cause (missing Copilot permission, absent subscription, rejected credential) rather than reporting a generic error. OpenAI and Cursor are not connections.
- Observed AI footprint and subscription floor (GTM-018). Detected per-tool developer activity, and an optional spend floor computed from admin-confirmed seat prices. The floor does not render until an admin confirms the price book, and it is always labelled as an estimate.
- Accountability analytics (GTM-020). Six-month trends joining spend to recorded verified-understanding events, weighted coverage, and seat utilization, plus a provider and model spend breakdown. Team-scoped rows deliberately carry no spend attribution.
- Executive panel visuals (GTM-021). Trend and delta rendering for the AI Investment panel, including explicit states for estimated-floor-only and unattributed spend.
- AI-authorship detection for programmatic checks (GTM-013). Client attribution from the CLI and MCP surfaces feeds the AI-assisted signal.
Cost per verified-understanding event is presented as supporting context —
spend divided by recorded events. It is not a return, productivity, or
outcome measure. See docs/claim-boundaries.md.
Enterprise and analytics
- Manager subtree analytics access and bulk member invites.
- Snapshot history correctness (GTM-008, GTM-016): repaired the sweeper
failures that froze
score_snapshotsfrom 2026-08-03, and removed two presentation faults — repos with zero mapped areas and a phantom prior-week row for brand-new organizations. - Datadog technology partner integration (DD-001, DD-002, DD-003): OAuth connect, a metrics exporter with a claim-safe taxonomy, and listing assets.
- Product and marketing analytics behind opt-in consent (AN-002 through AN-005), and SOC 2 Type 1 attestation on marketing surfaces (UX-015).
Security and platform
- Fail-closed row-level security on tenant tables, with tenant GUC plumbing and a CI guard that fails any fail-closed read which does not bind a tenant (SEC-13, OPS-008).
- Private subnets with per-AZ NAT, Multi-AZ RDS, managed master credential rotation, and distroless container images (SEC-10, SEC-12, SEC-14, SEC-18).
- Opaque organization
public_idin URLs and API responses, with both identifier forms accepted indefinitely (SEC-16). - Admin session hardening: 8-hour TTL, new-IP alerts, staff allowlist.
Not included in this release
- Weekly email digest (GTM-009). The implementation merged in #331 but
is disabled (
digest_sweeper_enableddefaults false) pending the remediation its ticket requires: an opt-in default, an archived-org filter on the sweep, a plan gate, andList-Unsubscribewith a per-recipient preference. No digest email is sent by this release. - Executive report export (GTM-007). Planned; not implemented.
Ninchi 1.0
Tagged platform-v1.0 at 3b7d312. The tag was applied retroactively on
2026-08-24; 3b7d312 is the commit that was running in production through
the soft-launch announcement, and the next deploy did not land until
2026-07-28.
Soft launch. Ninchi records evidence that a named actor answered a generated challenge tied to a specific artifact, and rolls that evidence up for individuals, repositories, and organizations.
- GitHub App installation, pull-request challenge generation, and grading against hidden rubrics.
- The Ninchi Score (individual) and the Organizational Intelligence Index (organization and team) — difficulty-weighted evidence ratios over recorded verified-understanding events.
- Knowledge maps, baselines, org and repository analytics, and narrative summaries over deterministic facts.
- Programmatic checks as a separate, lower-weight evidence class via the API, CLI, and MCP surfaces.
- Stripe billing with plan-gated capabilities, and the Enterprise tier.
ninchi-cli0.2.0 published to PyPI.