Rta-Smriti Brain v1.0.4-alpha

Rta-Smriti Brain v1.0.4-alpha

Alpha prerelease. Back up an existing brain before upgrading.
Rta-Smriti remains local-first and does not send project data to a hosted
service.

v1.0.4-alpha is a narrow launcher-integrity patch for the v1 Project Reality
line. It preserves the v11 SQLite schema and all stable v1 interfaces while
ensuring an installed Rta-Smriti command cannot be shadowed by an older source
checkout in the operator's current working directory.

What Changed

• Installed-package CLI and MCP wrappers now invoke Python in isolated mode before loading rta_brain.cli or rta_brain.mcp_server.
• A global launcher therefore resolves its verified installed runtime even when invoked from an older checkout or another folder containing an rta_brain package.
• Source-checkout script launchers and standalone native binaries retain their existing launch paths.
• Windows and POSIX wrapper regressions verify the isolated command shape, including generated MCP host configuration.
• The installed-upgrade gate now proves the immutable public v1.0.3-alpha package can upgrade to this candidate without database downgrade or rewrite.

Compatibility And Migration

• Python package metadata: 1.0.4a1
• Display/tag target: v1.0.4-alpha
• SQLite schema: v11; no schema change from v1.0.3-alpha
• Upgrade baseline: immutable public v1.0.3-alpha commit 76961d475905cb528d7959fa3b0166afe8606d0a
• Existing brains: upgrade the launcher; never downgrade or rewrite a brain database
• Existing MCP hosts: restart the host after reinstalling so it starts the updated isolated wrapper

Evidence Boundary

Focused launcher regressions prove the generated installed-package command is
isolated from the working directory. Full source, installed-package, native
artifact, cross-platform CI, browser, privacy, and security qualification will
be recorded in the release verification ledger.
Artifacts are built from the immutable annotated tag; checksums verify download
integrity but do not provide platform code signing.

Honest Boundaries

• Isolated mode protects installed Python launchers from local module shadowing; it is not an operating-system sandbox.
• Routine cognition uses the latest completed index; consequential work still requires a live or deep freshness check.
• Call and impact edges remain bounded hints, not compiler-perfect analysis.
• The packaged benchmark is a synthetic regression harness, not external proof of superiority.
• User-level workers are explicit local processes, not privileged operating-system services.

Build Provenance

Rta-Smriti Brain was conceived, researched, and product-directed by Sulabh
Dubey. It was built with OpenAI Codex as the primary
design, engineering, testing, and documentation agent under Sulabh's review and
release approval. This attribution does not imply OpenAI endorsement.

Rta-Smriti Brain v1.0.3-alpha

Rta-Smriti Brain v1.0.3-alpha

Alpha prerelease. Back up an existing brain before upgrading.
Rta-Smriti remains local-first and does not send project data to a hosted
service.

v1.0.3-alpha is a narrow maintenance patch for the v1 Project Reality line.
It preserves the v11 SQLite schema and all stable v1 interfaces while making
launcher/database mismatches and expired console capabilities safe and
understandable during real operator use.

What Changed

• A console opened without its current one-session capability now shows a dedicated authorization-recovery screen instead of a misleading empty-brain/bootstrap state.
• The recovery screen explains that no projects were removed and provides a copyable rta-brain console open --brain-dir ... command.
• API authorization failures cancel stale project discovery and clear session-only capability state before recovery.
• A launcher that encounters a newer brain schema now names both versions, instructs the operator to upgrade the active launcher, and explicitly refuses database downgrade or rewriting guidance.
• Installation guidance now documents safe upgrades, active-command discovery, MCP-host restart requirements, and stale-console recovery across Windows, macOS, and Linux.
• Regression coverage proves future-schema rejection is non-mutating and exercises expired-console recovery in the packaged mobile dashboard.

Compatibility And Migration

• Python package metadata: 1.0.3a1
• Display/tag target: v1.0.3-alpha
• SQLite schema: v11; no schema change from v1.0.2-alpha
• Upgrade baseline: immutable public v1.0.2-alpha commit 272674cca094447a35307c93ceb05863b84a1b50
• Existing brains: upgrade the launcher; never downgrade or rewrite a brain database
• Existing console tabs: run rta-brain console open --brain-dir <brain-directory> and continue in the fresh authorized tab

Evidence Boundary

The focused schema and rendered authorization regressions pass locally against
the packaged dashboard. Full source, installed-package, native artifact,
cross-platform CI, browser, privacy, and security qualification is recorded in
the release verification ledger. Artifacts are built
from the immutable annotated tag; checksums verify download integrity but do
not provide platform code signing.

Honest Boundaries

• The capability-bearing console URL is intentionally session-scoped; a plain loopback URL is not an alternate login path.
• Routine cognition uses the latest completed index; consequential work still requires a live or deep freshness check.
• Call and impact edges remain bounded hints, not compiler-perfect analysis.
• The packaged benchmark is a synthetic regression harness, not external proof of superiority.
• User-level workers are explicit local processes, not privileged operating-system services.

Build Provenance

Rta-Smriti Brain was conceived, researched, and product-directed by Sulabh
Dubey. It was built with OpenAI Codex as the primary
design, engineering, testing, and documentation agent under Sulabh's review and
release approval. This attribution does not imply OpenAI endorsement.

Rta-Smriti Brain v1.0.2-alpha

Rta-Smriti Brain v1.0.2-alpha

Alpha prerelease. Back up an existing brain before upgrading.
Rta-Smriti remains local-first and does not send project data to a hosted
service.

v1.0.2-alpha is the operator-hardening patch for the v1 Project Reality
release. It preserves the v11 SQLite schema, stable interfaces, and product
boundary while correcting Windows onboarding, background lifecycle, large-repo
sync fallback, and one dashboard accessibility defect found in real local use.

What Changed

• Windows login startup now uses a hidden direct Win32_Process.Create launcher instead of a visible Startup-folder .cmd; re-enabling startup migrates the legacy entry.
• Console, watcher, capture, and continuity workers share the same terminal-independent detached-process primitive, including explicit no-window flags on Windows.
• Standard packages and native binaries now include Watchdog for event-driven repository sync by default.
• The emergency polling fallback backs off to 30 seconds for repositories with at least 10,000 indexed files and 60 seconds at 50,000 files, while retaining five-minute deep verification.
• Windows private-directory hardening no longer reclaims an already-correct owner, closing onboarding failures on owner-controlled roots without weakening foreign-owner checks.
• Re-running start or bootstrap-project now preserves an existing brain's retrieval provider unless --embedding-provider is supplied explicitly; new brains continue to default to the built-in hash provider.
• The active-project secondary label now meets WCAG AA contrast in the dark dashboard theme.

Compatibility And Migration

• Python package metadata: 1.0.2a1
• Display/tag target: v1.0.2-alpha
• SQLite schema: v11; no schema change from v1.0.1-alpha
• Upgrade baseline: immutable public v1.0.1-alpha commit c2dff01b368bdb4d2b759e7a077d07ae0985a966
• Windows login startup: run rta-brain console login-enable --brain-dir <brain-directory> once after upgrade to replace a legacy .cmd registration
• Rollback: restore a pre-upgrade backup or reinstall the v1.0.1 public asset

Evidence Boundary

Focused Windows startup, detached-worker, Watchdog packaging, polling fallback,
ACL, onboarding, and install-local regressions pass locally. The final main
commit also passed the hosted five-lane Windows, macOS, and Ubuntu matrix.
Installed-package, native artifact, browser, privacy, security, and publication
evidence is tracked in the release verification ledger.
Release artifacts are built and smoke-tested from the immutable annotated tag;
checksums verify download integrity but do not provide platform code signing.

Honest Boundaries

• Routine cognition uses the latest completed index; consequential work still requires a live or deep freshness check.
• Call and impact edges remain bounded hints, not compiler-perfect analysis.
• Media descriptions remain unverified until explicitly promoted with provenance.
• The packaged benchmark is a synthetic regression harness, not external proof of superiority.
• User-level workers are explicit local processes, not privileged operating-system services.

Build Provenance

Rta-Smriti Brain was conceived, researched, and product-directed by Sulabh
Dubey. It was built with OpenAI Codex as the primary
design, engineering, testing, and documentation agent under Sulabh's review and
release approval. This attribution does not imply OpenAI endorsement.

Rta-Smriti Brain v1.0.1-alpha

Rta-Smriti Brain v1.0.1-alpha

Alpha prerelease. Back up an existing brain before upgrading. Rta-Smriti
remains local-first and does not send project data to a hosted service.

v1.0.1-alpha is the operator-readiness patch for the v1 Project Reality
release. It preserves the v1 schema, interfaces, and product boundary while
including two lifecycle corrections found during post-publication proof.

What Changed

• Direct dashboard feedback now remains authoritative over delayed background refreshes.
• Temporal-truth API handlers close database state before sending the HTTP response, preventing follow-up requests from overlapping cleanup on slower runners.
• Release metadata, installation guidance, launch surfaces, and upgrade tests now target the corrected patch source.

What Remains In v1

• Deterministic Project Cognition over repository evidence, bitemporal truth, observations, work state, decisions, and media
• Project Reality readiness, project-twin conflicts, knowledge coverage, decision debt, and bounded change-impact hints
• Governed local multimodal evidence with provenance, verification, redaction, retention, deletion, and metadata-only export
• Stable Python SDK plus CLI, authenticated loopback console, and read-only MCP parity
• Opt-in local capture, governed context compilation, canonical project integrity, and inspectable evidence classes

Compatibility And Migration

• Python package metadata: 1.0.1a1
• Display/tag target: v1.0.1-alpha
• SQLite schema: v11; no schema change from v1.0.0-alpha
• Upgrade baseline: immutable public v1.0.0-alpha commit a1b05022aff6df3a066ae5abcad3877f6407eafb
• Rollback: restore a pre-upgrade backup or reinstall the v1.0.0 public asset

Evidence Boundary

Publication requires the corrected source to pass the full hosted Windows,
macOS, and Ubuntu matrix, installed upgrade/uninstall proof, rendered operator
and launch-site journeys, native binary smoke tests, dependency and workflow
audits, privacy and secret scans, SBOM and checksum verification, and anonymous
download acceptance. Exact runs and artifact hashes are recorded in the
release verification ledger; no pending gate is
represented as passed.

Honest Boundaries

• Routine cognition uses the latest completed index; consequential work still requires a live/deep freshness check.
• Call and impact edges remain bounded hints, not compiler-perfect analysis.
• Media descriptions remain unverified until explicitly promoted with provenance.
• The packaged benchmark is a synthetic regression harness, not external proof of superiority.
• User-level workers are explicit local processes, not privileged operating-system services.

Build Provenance

Rta-Smriti Brain was conceived, researched, and product-directed by Sulabh
Dubey. It was built with OpenAI Codex as the primary
design, engineering, testing, and documentation agent under Sulabh's review and
release approval. This attribution does not imply OpenAI endorsement.

Rta-Smriti Brain v1.0.0-alpha

Rta-Smriti Brain v1.0.0-alpha

Alpha prerelease. Back up an existing brain before upgrading. Rta-Smriti
remains local-first and does not send project data to a hosted service.

Rta-Smriti v1 completes the shift from a searchable project index to a
sovereign, evidence-aware project-reality layer for AI-assisted software work.
It remains local-first and agent-neutral: it prepares trusted context and
continuity, but does not plan, execute, route models, or train systems.

What Is New

• Deterministic Project Cognition over repository evidence, bitemporal truth, observations, work state, decisions, and media
• Readiness that distinguishes database health from continuation readiness
• Project Twin reconciliation with bounded conflicts and omission accounting
• Knowledge coverage, decision debt, and approximate change-impact views
• Governed local multimodal evidence with provenance, verification, redaction, retention, deletion, and metadata-only export
• Stable Python SDK plus CLI, authenticated loopback console, and read-only MCP parity
• Synthetic quality gates for continuation, contradiction, debt, abstention, governance, and stale rejection
• Project Reality operator cockpit with accessible native interactions and explicit loading/error/empty states

Evidence Boundary

The frozen v1 source passed 817 Python tests on hosted Windows (12
explicit platform skips), dashboard unit and production builds, 8 rendered
operator journeys, launch-site desktop/mobile QA, the packaged synthetic quality
gates, a 10,000-source performance probe, clean upgrade/uninstall, and native
binary smoke. The five-job hosted matrix passed on Windows, macOS, and Ubuntu
across Python 3.11, 3.12, and 3.13.

Privacy, Gitleaks source/history/artifact scans, dependency audits, actionlint,
and sealed Codex Security scan 8ab0e2aa-a366-4c4a-b60c-8fbccd36e7e2 passed
with no findings in their tested scopes. Live continuity-daemon capture and the
generated 29-tool MCP configuration were exercised successfully. Native release
artifacts, their CycloneDX SBOMs, checksums, and anonymous download acceptance
are recorded in the release verification ledger.

Compatibility And Migration

• Python package metadata: 1.0.0a1
• Display/tag target: v1.0.0-alpha
• SQLite schema: v11, migrated transactionally by the existing database layer
• Upgrade baseline: immutable public v0.9.1-alpha commit 721bd2ec98395f2be36a3b7ebb60c14bfa63c882
• Rollback: restore the pre-upgrade backup or reinstall the v0.9.1 public asset; do not reuse a schema-v11 database with an older binary without restoring its compatible backup

Honest Boundaries

• Routine cognition uses the latest completed index; consequential work still requires a live/deep freshness check.
• Call and impact edges remain bounded hints, not compiler-perfect analysis.
• Media descriptions remain unverified until explicitly promoted by an operator with provenance.
• The packaged benchmark is a small synthetic regression harness, not external proof of superiority.
• User-level workers are explicit local processes, not privileged operating-system services.

Build Provenance

Rta-Smriti Brain was conceived, researched, and product-directed by Sulabh
Dubey. It was built with OpenAI Codex as the primary
design, engineering, testing, and documentation agent under Sulabh's review and
release approval. This attribution does not imply OpenAI endorsement.

Rta-Smriti Brain v0.9.1-alpha

Rta-Smriti Brain v0.9.1-alpha

v0.9.1-alpha is an operator-readiness patch for the Universal Capture
release. It preserves the v0.9 data model and trust boundaries while making the
multi-project dashboard and MCP gateway more predictable under real local load.

Build Provenance

Conceived and researched by Sulabh Dubey. Built with OpenAI Codex as the primary design, engineering, testing, and documentation agent under the maintainer's product direction and release approval. See Contributors And Build Provenance.

Operator Reliability

• Starts the dashboard from a lightweight authenticated registry snapshot, then
  loads repository, continuity, capture, checkpoint, and graph details
  progressively.
• Keeps usable project information visible while expensive freshness or
  continuity checks are still running.
• Uses bounded request timeouts and explicit checking, stopped, and
  not configured states instead of blank or misleading health indicators.
• Prevents delayed responses from a previously selected project from replacing
  the current graph, files, preview, capture, retrieval, governance, or truth
  state.
• Preserves the operator's newer project selection when a slower background
  registry refresh completes.
• Requires an explicit project argument in every advertised multi-project MCP
  gateway tool schema. Single-project MCP configurations remain project-bound.

Local Qualification

The frozen Windows candidate passed:

• 788 Python tests, with 23 explicit platform or optional-capability skips;
• five dashboard unit tests;
• four adversarial progressive-loading and project-switch isolation journeys;
• seven complete rendered operator journeys covering files, graphs, canvas,
  capture, context packs, governance, MCP diagnostics, snapshots, workspaces,
  accessibility, mobile layout, and fault states;
• a real multi-project local audit with no browser console errors, failed API
  requests, persistent loading states, false integrity alerts, or horizontal
  mobile overflow;
• package build, installed-package dependency, npm dependency, privacy,
  Gitleaks, actionlint, and patch-integrity checks; and
• frozen Codex Security diff scans covering 13 of 13 operator-readiness
  implementation surfaces and 12 of 12 release/website code-bearing surfaces,
  with zero findings in either scan.

The frozen security scans covered the original operator-readiness and
release/website surfaces, not an independent audit of every historical line in
the repository. The later seven-entry concurrency repair was manually reviewed
and separately passed regression, privacy, secrets, dependency, workflow, and
patch-integrity checks; it is not misrepresented as part of those frozen scans.
Temporary local reports and private operator data are not release artifacts.

Compatibility And Upgrade

The package version is 0.9.1a1. The installed-package qualification upgrades
from the immutable v0.9.0-alpha tag. This patch introduces no database schema
migration and does not change Universal Capture retention, authority, or
promotion semantics.

Managed watchers, continuity workers, and capture daemons remain explicit,
local, opt-in lifecycle services. A stopped service is reported as stopped; the
dashboard does not silently enable persistent collection.

Publication Evidence

• PR CI run 32719412677
  passed all five Windows, macOS, and Ubuntu jobs.
• Post-merge CI run 32722109549
  passed all five jobs. Its first Windows attempt reached rendered acceptance
  after all Python and installation checks, then hit runner-level
  ERR_NO_BUFFER_SPACE; one bounded rerun passed the complete job without a
  code change.
• Native workflow run 32724105024
  built, audited, smoke-tested, packaged, privacy-scanned, and uploaded all
  Windows, Linux, and macOS artifacts from the exact annotated tag.
• Pages workflow run 32722110481
  deployed the current website. Public desktop and 390 px mobile checks found
  no console errors, WCAG A/AA violations, broken media, or horizontal overflow.
• All eight public release files were downloaded without authentication. Every
  manifested SHA-256 matched; the public Windows binary reported 0.9.1a1, and
  the public wheel installed in a clean environment and passed doctor.

The existing v0.9.0-alpha release notes, screenshots, demo, hashes, and
verification record remain historical evidence and are not rewritten.

Rta-Smriti Brain v0.9.0-alpha

Rta-Smriti Brain v0.9.0-alpha

v0.9.0-alpha is the Universal Capture prerelease. It joins canonical project
identity, bitemporal truth, governed context compilation, and private
agent-session capture into one local, inspectable continuity system.

Universal Capture

• Opt-in adapters normalize supported Codex, Claude Code, Cursor, GitHub
  Copilot, Gemini CLI, and generic local events into one versioned schema.
• Private bounded spools isolate producers from SQLite and apply explicit
  backpressure, recovery, quarantine, and receipt policies.
• One managed normalizer daemon per brain records append-only, hash-chained
  events with project, checkout, session, causal, privacy, and policy anchors.
• Metadata-only, continuity, and explicitly granted encrypted forensic profiles
  keep raw payload retention off by default.

Continuity And Governance

• Causal replay reports interruption, incomplete spans, gaps, late events, and
  the latest trusted cursor without executing captured actions.
• Interactive retention, redaction, and deletion are preview-first, require a
  separate destructive capability, and bind confirmation to the exact policy,
  cutoff or scope, affected state, and operator. Background retention enforces
  only the already-authorized immutable policy.
• Agent-hook installation is preview-first, reversible, canonical-root-bound,
  and separated from read-only MCP and dashboard capabilities.
• Captured text remains untrusted evidence. It cannot promote itself to project
  truth, policy, approval, or verified memory.
• The multi-project MCP gateway is project-scoped and read-only; all mutation
  classes require an explicit single-project binding.
• Replay reads event content and deletion tombstones from one database snapshot;
  default MCP diagnostics verify a bounded journal prefix and report whether the
  verification is complete. Routine capture-write grants can pause or resume a
  source but cannot irreversibly remove it.
• Capture redaction covers root-level Windows and POSIX absolute paths before
  default read-only MCP responses are emitted.
• Project-bound status reports include only that project's spool occupancy and
  omit database-wide daemon activity counters and global session-binding
  diagnostics.
• Login supervision revalidates every private enrollment receipt against the
  brain's current canonical project root before starting any enrolled worker.
• Shared free-text redaction recognizes provider credentials including Google
  API keys and Stripe secret or webhook keys before spool, journal, replay, and
  export verification.
• Default MCP continuity and readiness reads use a strict path-free lifecycle
  projection: local roots, database paths, process identifiers, launch material,
  and raw errors remain private while a bounded has_error signal preserves
  fail-closed readiness.
• Managed continuity workers bind liveness to both PID and cross-platform
  process-birth identity. A confirmed mismatch is stale and recoverable without
  signaling the unrelated process; an unavailable identity fails closed rather
  than risking a duplicate worker.
• Continuity start/stop is an independent MCP process-control capability rather
  than an implicit side effect of memory-write permission, and control responses
  use the same path-free lifecycle projection as status and readiness reads.
• Capture export and replay verify normalized content, event envelopes, every
  hidden privacy-filtered chain link, and the pagination anchor before returning
  a page. Receipts label the bounded verification scope instead of implying a
  full-journal proof.
• Rendered operator acceptance parses the downloaded export and verifies its
  schema, journal proof, redaction proof, payload exclusion, canonical-path
  exclusion, and provider-credential exclusion rather than checking only its
  filename.
• The package README, CLI inventory, installation guide, verification record,
  and website identify v0.9 consistently and link to the same formal prerelease.

Qualification Contract

The frozen source candidate passed:

• 760 Python tests, with 23 explicit optional-dependency or privilege skips and
  649 subtests on Windows;
• dashboard unit tests plus the complete rendered Playwright operator journey;
• the dedicated Universal Capture operator journey, including parsed export
  integrity and privacy assertions;
• installed-wheel upgrade and uninstall checks from 0.8.0a1 to 0.9.0a1;
• Windows native CLI, SQLite/FTS, MCP, benchmark, bundled Tree-sitter,
  Universal Capture, encrypted snapshot, Ed25519, sync, and console smoke tests;
• the 10,000-event capture probe at 170.445 events per second, with 43.071 ms
  p99 replay-page latency and verified fail-closed backpressure;
• repository and artifact privacy scans, Gitleaks history and working-tree
  scans, actionlint, npm audit, Python dependency audit, wheel inspection, and
  checksum verification; and
• hosted CI on Windows, macOS, and Ubuntu across Python 3.11, 3.12, and 3.13.

A later Codex Deep Security Scan coordinator attempt did not inspect the
workspace because both workers were blocked, and its completion call rejected a
non-UUID scan identifier. It therefore produced no usable sealed report and is
not represented as security coverage. Release confidence instead rests on the
completed code-specific security tests, threat-model controls, Gitleaks,
dependency audits, privacy scans, and cross-platform hosted matrix documented in
the verification record.

Tag-generated binaries, wheel, SBOMs, and checksums are accepted only after the
native workflow passes and downloaded assets match the combined manifest. Any
code or release-relevant content change invalidates the affected evidence and
requires proportionate reruns.

Rta-Smriti Brain v0.6.0-alpha

Rta-Smriti Brain v0.6.0-alpha

Release commit: 6c086f5e421f8ec5506e7ee6e6cb0296ca43fed3

Hosted CI: https://github.com/sulabhdubey/rta-smriti-brain/actions/runs/32484754948
Native binaries: https://github.com/sulabhdubey/rta-smriti-brain/actions/runs/32487134222

What changed

• Initial repository ingestion now warms the persistent SHA-256 cache, reducing the cost of later deep verification.
• Common Tree-sitter grammars and Ed25519 snapshot support are included in standard installs and native binaries.
• Oversized eligible sources default to metadata-only isolation with fresh_with_warnings; strict fail-closed blocking remains selectable.
• LSP mode can discover trusted system-installed language servers, uses bounded shell-free JSON-RPC, rejects project-local discovery, and falls back conservatively.
• Optional Ollama continuity compaction is loopback-only, bounded, redacted, append-only, and always marked unverified.
• Codex transcript capture can rebind at a verified repository turn_context, excluding earlier content from unrelated roots.
• mcp-doctor verifies the generated MCP command through initialization, tool discovery, and ping before configuration is copied.
• Encrypted v3 snapshots add scrypt, streaming AES-256-GCM, bounded verification, optional Ed25519 signing, and atomic restore.
• Multi-project workspaces now expose member health, degraded partial search, member removal, and safe workspace deletion without deleting project brains.
• Retrieval benchmarks gain bounded historical comparisons; Tree-sitter call extraction avoids comments and strings while retaining conservative fallback behavior.

Install

Download the binary for your operating system and SHA256SUMS.txt, verify its SHA-256 value, then run --version before onboarding a non-sensitive repository.

Source installation remains available:

git clone --branch v0.6.0-alpha https://github.com/sulabhdubey/rta-smriti-brain.git
cd rta-smriti-brain
python -m venv .venv
python -m pip install --upgrade pip
python -m pip install .

Verification

• Main CI passed Windows/Python 3.11, macOS/Python 3.11, and Ubuntu/Python 3.11, 3.12, and 3.13.
• The native workflow built and smoke-tested separate Windows x64, Linux x64, and macOS binaries from this tag.
• The release candidate passed 251 Python tests, rendered operator and launch-site QA, clean install/upgrade/uninstall checks, bounded performance probes, Gitleaks, actionlint, the 174-file privacy scan, and npm/Python runtime dependency audits.

Trust boundaries

Rta-Smriti remains local-first alpha software: no hosted account, telemetry, or cloud sync is added. Managed processes are user-level. Discovered language servers and Ollama are explicit local integrations that Rta-Smriti cannot independently certify. Metadata-only large files are never represented as content-indexed. Call edges remain impact hints, and the public benchmark is a transparent synthetic regression harness rather than external proof of superiority. Native alpha binaries are not platform code-signed; checksums prove artifact integrity against this release manifest, not publisher identity.

Rta-Smriti Brain v0.5.0-alpha

Rta-Smriti Brain v0.5.0-alpha

Release commit: be534d9

Hosted CI: https://github.com/sulabhdubey/rta-smriti-brain/actions/runs/32407147824
Native binaries: https://github.com/sulabhdubey/rta-smriti-brain/actions/runs/32417096347

Theme

v0.5.0-alpha turns the v0.4 continuity foundation into a more trustworthy
operator loop:

• actions get checked before they are taken;
• retrieval explains why each result was selected;
• the dashboard opens the intended project brain instead of falling back to the
  first available database;
• onboarding starts repository sync, the managed console, and Codex continuity
  capture together when the local sessions folder is present.

New And Improved

• Pre-action governance can include operational context: checkpoint readiness,
  continuity lifecycle, dirty worktree state, canonical-root risk, and freshness.
• Retrieval diagnostics now include normalized query terms and per-result
  selection reasons for lexical, structural, and hybrid matches.
• Managed console startup preserves root-level --db for dashboard and
  console start, so a copied or custom brain directory opens the requested
  database instead of defaulting to the first brain.
• Dashboard health defaults are converted into an exact project/database
  identity, preserving fail-closed protection for duplicate project names.
• One-command start onboarding starts Codex continuity capture when the
  sessions folder exists, skips it explicitly when unavailable, and exposes
  controls for sessions root, interval, inactivity, lookback, and backlog tail.
• The multi-project MCP gateway and native MCP tools remain read-oriented and
  project-scoped; existing Codex tasks still need a fresh task after MCP
  registration changes.

Fixes

• Fixed the copied-repo/local-console path where rta-brain --db ... console start launched successfully but the React dashboard did not auto-select the
  requested brain.
• Fixed the dashboard state flow so /api/health default_db and
  default_project select only the exact matching brain.
• Kept context-pack buttons observable: New Task Prompt and Copy Command now
  surface visible copied-state feedback in the operator console.

Boundaries

• No hosted accounts, telemetry, cloud sync, privileged service, or automatic
  reboot persistence is added.
• Optional Sentence Transformers and LSP adapters remain operator-installed and
  operator-selected.
• Gitleaks is not claimed unless it is available and run in the release
  environment.
• Public release evidence must use synthetic data only. Local brain databases,
  private project paths, capability tokens, context packs, daemon state, and
  scratch artifacts are not release assets.

Candidate Gates

Before tagging v0.5.0-alpha, run:

npm audit --audit-level=high
npm run build
npm run build:launch
npm run test:unit
npm run test:operator
npm run test:launch
python -m unittest discover -s tests -v
python -m compileall -q rta_brain tests scripts
pip install -e . --dry-run --no-deps
python scripts/build_installed_smoke.py
python -m rta_brain.cli --json benchmark
python scripts/privacy_scan.py --root .
python -m rta_brain.cli publish-readiness --json
git diff --check

Then commit the reviewed tree, wait for hosted Windows/macOS/Linux CI, create
the annotated v0.5.0-alpha tag, build/upload release assets, verify
SHA256SUMS.txt, and perform a clean install from the public release page.

Rta-Smriti Brain v0.4.0-alpha

Rta-Smriti Brain v0.4.0-alpha

This alpha turns repository memory from a manual snapshot into an operator-controlled living index while preserving local-first defaults.

New

• Foreground watch-repo plus managed-background watcher lifecycle commands using incremental manifests, event-scoped content hashing, and periodic polling verification

• Managed Codex continuity lifecycle with canonical-project session discovery, resumable cursors, heartbeat-validated start/status/stop controls, and conservative automatic checkpoints

• Bounded 30-day discovery plus explicit recent-tail recovery for oversized transcript backlogs, preventing silent disk growth while recording provenance-bearing truncation events

• Append-only session events with common credential redaction, oversized-output bounds, partial-write recovery, and verification status

• Structured work-state records and filesystem reconciliation for assets, jobs, approvals, blockers, QA decisions, fallbacks, and next actions

• Operational readiness that distinguishes database health from checkpoint, reconciliation, capture-backlog, lifecycle, and error readiness

• One fail-closed multi-project MCP gateway for native context, checkpoint, event, work-state, reconciliation, continuity, and doctor tools

• Dashboard lifecycle health, backlog, checkpoint provenance, and safe continuity controls

• Owner-only POSIX database/daemon modes, linked-artifact rejection, bounded MCP frames, protocol negotiation, and notification-safe JSON-RPC handling

• Foreground watch-repo plus managed-background watcher lifecycle commands using incremental manifests, event-scoped content hashing, and periodic polling verification

• Dashboard repository-sync controls with heartbeat, backend, and error visibility

• Persistent SHA-256 cache for repeated deep freshness checks

• Optional local hybrid FTS and vector retrieval

• Dependency-free deterministic hash embedding provider

• Lazy Sentence Transformers adapter for separately installed local models

• Pluggable regex, Tree-sitter, LSP-command, and entry-point parser registry

• Per-project source-size, parser, embedding, model, and ranking policy

• Dashboard controls with blocked-source and optional-provider warnings

• Canonical-root binding with explicit rebind protection and duplicate-root dashboard warnings

• Repository root, branch, HEAD, and dirty-file diagnostics

• Structured continuation checkpoints and one-click new-task prompts

• Claim provenance fields for source path, hash, verification command, timestamp, and status

• Compact anomaly-first deep freshness output with bounded details

• Default exclusions for worktrees, bundled browser runtimes, test scratch folders, and generated tool caches

• Managed console start/open/status/restart/stop with stale-process, authorization, and occupied-port recovery

• One-command start onboarding for canonical-root detection, migration, indexing, watcher startup, console startup, and readiness proof

• Optional owner-controlled login startup on Windows, macOS, and Linux

• Evidence-aware Action Gate with typed policies, trust thresholds, expiry, path/action scope, required checks, owner overrides, and short-lived action/policy digest receipts

• Approximate call and test links plus bounded dependency, dependent, impact, evidence, and relevance graph queries with explicit relation semantics

• Retrieval diagnostics for provider, embedding coverage, parser fallback, freshness, rank components, latency, and source hashes

• Packaged public synthetic benchmark for no-memory, lexical, hash-hybrid, and explicitly optional Sentence Transformers retrieval plus stale, contradiction, continuation, and governance gates

• Explicit query-only local workspaces that search existing unlinked project-brain databases without merging or mutating them

• Selective redacted memory/checkpoint/policy export and integrity-checked import with rename, merge, or fail conflict handling

• Non-mutating selective bundle previews, bounded schema validation, staged all-or-nothing imports, and restrictive atomic private writes

• Authenticated HMAC-SHA256 local snapshots with consistent SQLite backup and tamper verification

• Cross-platform standalone CI artifacts with versioned filenames, full runtime smoke tests, and SHA-256 manifests

• Worktree-aware, frozen-binary-safe Git hooks with linked-file protection and atomic installation

• Opt-in managed post-commit checkpoint hooks that refuse to replace unknown hooks

• Operator memory feedback and conservative decay restricted to eligible old, unverified inference and hypothesis records

• Read-only-by-default, project-bound MCP capabilities with explicit write/ingestion flags, allowed thread roots, bounded frames/tasks, and downgraded agent provenance

• Descriptor-bound thread, bundle, snapshot, and repository reads; private SQLite modes and trusted-schema disablement

• Rendered Chromium operator acceptance configured for Windows, macOS, and Linux CI

• Isolated install, forced-upgrade/reinstall, and uninstall lifecycle proof

• Sanitized 100, 1,000, and 10,000-file indexing and retrieval resource evidence

• Destination-wide axe WCAG checks, accessible Intelligence and Bases tabs, explicit active-state
  semantics, live status announcements, modal focus containment/restoration, normal-mode contrast,
  reduced motion, forced-colors structure, and a zero-overlap mobile Canvas

Fixed On Main

• Installed wheels now generate valid Python-module commands instead of pointing to repository-only source wrappers.
• install-local emits native .cmd launchers on Windows and executable POSIX launchers on macOS and Linux.
• Generated agent, MCP, dashboard, bootstrap, and context-pack commands use the active operating system's quoting and invocation rules.
• Setup no longer assumes the wrapper directory is already on PATH.
• Dashboard recovery guidance now distinguishes the foreground console from opt-in background repository sync.
• Windows, Ubuntu, and macOS CI build and smoke-test an installed wheel through bootstrap, retrieval, wrappers, MCP config, dashboard assets, and API authorization.
• Bootstrap now writes agent bridge files before the final index pass, so a newly created brain starts fresh instead of immediately reporting those generated files as added.
• Unsigned selective-bundle imports now downgrade memories and quarantine checkpoints and policies instead of inheriting authority.
• The release privacy gate now suppresses only exact hashed detector-definition lines; unrelated private paths in those modules fail the scan.
• Generated PowerShell and POSIX commands now keep substitutions, variables, backticks, quotes,
  apostrophes, and embedded newlines literal instead of allowing a project path to alter the command.
• Disposable operator QA uses a random per-launch capability token and proves that the retired fixed
  token is rejected.
• Post-bootstrap project selection now binds the exact project/database tuple, rejects ambiguous
  same-name roots, and clears a previous selection when the identity verification refresh fails.

Defaults And Boundaries

• SQLite remains local.
• No hosted account, cloud sync service, telemetry, privileged service, login item, or automatic reboot persistence is added.
• FTS5 and regex remain available deterministic defaults.
• The recommended bootstrap flow enables the dependency-free hash provider; lexical-only retrieval remains selectable and external embedding packages are never installed automatically.
• Sources above the selected cap remain blocked and keep freshness fail-closed.
• Call relationships are approximate impact hints, not compiler-perfect call graphs.
• Workspaces reference local brain database paths and do not combine project stores.
• Selective bundles exclude source code by default; snapshots contain the complete local brain and should remain private.
• Selective bundles provide SHA-256 integrity but no sender authentication; inputs are capped at 25 MB.
• HMAC snapshots use a local shared secret and are authenticated, not encrypted or public-key signed.
• Snapshot payloads are capped at 64 MiB, with legacy envelopes capped at 16 MiB.
• The bundled benchmark is a transparent regression harness, not independent evidence of competitive superiority.

See Architecture and Usage Guide.

Refresh Provenance

This prerelease was refreshed in place after the continuity hardening batch.

• Tag target: $target
• Hosted CI: https://github.com/sulabhdubey/rta-smriti-brain/actions/runs/32302463544, completed successfully on Windows, macOS, Ubuntu Python 3.11, Ubuntu Python 3.12, and Ubuntu Python 3.13
• Artifact source: CI artifacts from the green run above
• Release payload: universal wheel, Windows x64 binary, Linux x64 binary, macOS ARM64 binary, public benchmark JSON, and combined SHA256SUMS.txt
• Local pre-upload checks: combined manifest verified, wheel metadata/static assets inspected, staged Windows binary --version and doctor passed, bundled privacy scan passed

Note: Gitleaks was not run in the current shell and is not claimed as fresh release evidence.