v1.0.3

Hotfix release containing four code/mechanism fixes since v1.0.2.

Fixed

• Resolver canary no longer auto-starts on boot; synthetic canary telemetry is explicit diagnostic-only and the sample star map demotes existing canary noise.
• Mímir JS pool applies query-aware relevance scoring so query-relevant nodes rise during first-stage pool scoring.
• Restart resume handoff carries interrupted-turn progress context to reduce post-restart repetition loops.
• Consolidation apply paths are transaction-safe for multi-step fuse/supersede updates.

v1.0.2 — Codex OAuth + packaged startup hotfix

Constellation Engine v1.0.2

Hotfix release for Codex-compatible providers, optional Codex OAuth setup, semantic anchor embeddings, and packaged launcher polish.

Added

• Optional Codex OAuth provider path in first-run setup for users who have installed Codex CLI and run codex login.
• Local-only Codex shim on 127.0.0.1:3457; Constellation never reads or stores Codex OAuth tokens.

Fixed

• Packaged Codex shim cold-start now uses Electron's bundled Node runtime instead of assuming a system node command exists.
• Configured Codex shim host/port are passed into the child process, so custom baseUrl ports no longer fall back to 3457.
• Codex-compatible provider budget defaults are raised to avoid Claude-era budget warnings on OpenAI-compatible usage reporting.
• Star-map dense retrieval now has optional semantic_anchor / embedding_text_version support plus dry-run audit/backfill tools.
• Electron launcher now focuses whichever window is open, including the onboarding wizard, and packages canonical icon assets.

Artifacts

• Windows x64 installer
• Linux x64 AppImage
• macOS arm64 zip

v1.0.1 — vec0 + mimir_actions + self_act hotfix

Hotfix release. Three silent-failure bugs surfaced by a 20-round research-cron sweep (Opus, 2026-05-19) that took the post-1.0.0 codebase as fixed input and walked the cognitive pipeline end-to-end looking for orphan producers / consumers.

Fixed

• vec0 lookup column rename (engine.cjs): consolidation's per-candidate embedding pull used WHERE rowid = ? on the sqlite-vec virtual table, but vec0 exposes the primary key as id, not rowid. The SELECT returned nothing, leaving the consolidation judge embedding-blind on every pair. Fix: WHERE id = ?.
• mimir_actions table now actually populated (src/dashboard.js): the autonomy picker's success branch wrote to mimir_outreach_audit but never wrote the canonical row to mimir_actions. Consumer side (Critic gate, demotion-sweep, persona caps, outreach review queue) was reading from an always-empty table — silently no-op'd in production. Added the missing INSERT with try/catch + action guard + meta sanitization.
• self_act consolidation policy (engine.cjs ALLOWED_OPS_BY_TYPE): Mímir-emitted self_act nodes had no entry in the policy map, so the type gate rejected them before reaching the judge. Added 'self_act': ['FUSE', 'TIMELINE_MERGE', 'INDEPENDENT'] so they participate in consolidation like any other autonomous-source type.

Notes

• No schema migration; no env-var changes; no API changes.
• Dashboard bundle rebuilt to bake the mimir_actions INSERT into the obfuscated dist (the build-platform.sh [1.5/6] overlay step picks up the new bundle).
• electron-updater will offer 1.0.1 as a drop-in update for 1.0.0 installs once the release is published.

Downloads

• Windows: Constellation Setup 1.0.1.exe
• macOS (Apple Silicon): Constellation-1.0.1-arm64-mac.zip
• Linux: Constellation-1.0.1.AppImage

v1.0.0

[1.0.0] - 2026-05-18

First stable release. Same engine that's been running in production daily, with a final pre-launch polish pass on top of the 0.3.0r26 line.

Added (1.0.0)

• Single-chunk cron delivery (src/telegram.js, src/cron.js): cron reports now post as one Telegram message instead of being shredded into ~900-char fragments. send() / sendLong() accept a { style: 'single' | 'layered' } override; cron paths force single. User-driven /talk replies still use the layered chunker.
• IR anti-hallucination pass (src/narrative-ir.js): duty-section tail now carries an explicit deep-retrieval trigger ("if a fact isn't in this packet, dig — don't invent") plus an anti-fabrication hard rule. Reduces low-confidence completions when the pool covers a topic but the L0/L1 layers are thin.
• Top-N force-full precision (src/agent-runtime.js): the top-7 dynamic pool nodes by score now render at precision='full' regardless of token budget. Cap raised from the legacy 5-slot heuristic. POOL_TOP_N_FORCE_FULL env override (default 7).
• RESTART_TOUCH pulse hint (src/behavior-logger.js, src/pulse-handlers.js, src/main.js, electron/main.js): fifth pulse marker joins DEBRIEF / TASK_TOUCH / COGNITIVE_TOUCH / ANCHOR_TOUCH. The agent can request an in-place engine restart via this marker; the Electron launcher catches the exit signal and respawns without dropping session state.
• library_read_log table (scripts/mimir-js/diary.js): records every library read with a 24h KNN-dedup window. Surfaces as a digest for the consolidation re-sweep loop.

Changed (1.0.0)

• Consolidation cosine threshold lowered 0.70 → 0.65 (engine.cjs CONSOLIDATION_COSINE_THRESHOLD). Catches more fusion candidates at the cost of more judge calls; offset by the auto-supersede + type-gate paths.
• Reconsolidate batch threshold raised 0.85 → 0.92 (scripts/mimir-js/reconsolidate.js SUPERSEDE_THRESHOLD) plus a SUPERSEDE_BLOCKED_TYPES gate. The pure-cosine batch path no longer bypasses the judge's TIMELINE_MERGE classification for borderline pairs — true verdicts route through the periodic re-sweep instead.
• Lever B re-sweep added to engine.cjs _consolidationResweep(): every 6h, re-scan recent active nodes through the full judge. Kill-switch ENGINE_CONSOLIDATION_RESWEEP=0.

Added (r26)

• /status payload surfaces mechanism toggles: rumination_enabled and novelty_gate_enabled now read out of getStatus() alongside sa and leiden, so dashboards can pull mechanism state from one endpoint without polling /rumination/status and /hebb/status separately. /config GET/POST still authoritative for runtime flips.

Fixed (r26)

• edges.valid_from now defaults to current epoch-ms on INSERT (schema.sql). Bi-temporal reads filter on valid_from, and main-arch installs got the default expression via migrate_phase1c; OSS shipped without it, so rows from any INSERT path that didn't explicitly stamp valid_from would land NULL and silently fall out of the time-window filter. New DEFAULT (CAST((julianday('now')-2440587.5)*86400000 AS INTEGER)) matches main-arch behavior.

Added (r25)

• Predictive priming ported to mimir-js/sa.js (Python parity: SIGNAL_HISTORY_SIZE=20, PRIMING_STRENGTH=0.08, PRIMING_DECAY=0.97). After every /signal injection, current A_fast is snapshotted into a per-instance signal history. Each tick predicts the next-likely activation pattern via cosine-similarity-weighted average of "what came after past signals like this one" and pre-warms those nodes through the same 3-channel fusion weights as the forward pass. Already-active nodes are damped 10× to push priming outward into unexplored space. Kill-switch: MIMIR_PRIMING=0. Default-ON.
• Reverse propagation ported to mimir-js/sa.js (Python parity: REVERSE_PROPAGATION_SCALE=0.15, DIRECTED_EDGE_TYPES = supports/extends/exemplifies/contains/depends_on/enables/causes/inspires/supersedes). Per-channel reverse CSR matrices built alongside forward CSR during _buildState(); each tick adds a fusion-weighted backward diffusion pass so activation flows effect→cause along directed edges (abductive reasoning substrate). Kill-switch: MIMIR_REVERSE_PROP=0. Default-ON.
• Novelty gate ported to mimir-js/hebb.js (Python parity: NOVELTY_GATE_THRESHOLD=0.3, NOVELTY_EMA_ALPHA=0.1, EMA_MAX_SIZE=3000). LTP (+0.02) now only fires when |coact − EMA| exceeds the threshold — strengthening surprising co-activations and skipping over well-predicted ones. EMA dict prunes at 3000 entries (drop ≤0.01 noise floor, then keep top half by EMA value). Kill-switch: MIMIR_NOVELTY_GATE=0. Default-ON.
• Rumination ported as mimir-js/rumination.js (Python parity: rumination_interval_s=120, rumination_strength=0.35, rumination_n_nodes=12, idle ≥60s). When the daemon goes idle, pick a Leiden zone weighted 70% by mean A_slow (recent cognitive afterglow) + 30% by zone size; re-inject 0.35 energy into the top 12 nodes by weight · (1 + log(1+access_count)). Mirrors the brain's default-mode network. Kill-switch: MIMIR_RUMINATION=0. Default-ON. Routes: POST /rumination/run, GET /rumination/status.
• /config GET and POST now expose four mechanism toggles (rumination_enabled, novelty_gate_enabled, reverse_propagation_enabled, priming_enabled) for runtime control without restart. Env-var kill-switches remain the durable off-switch.

Fixed (r25)

• Identity bootstrap on standalone boots: src/main.js now scaffolds identity/, identity/tasks.json (default {"tasks":[]}), identity/COGNITIVE_STATE.md, engine-inbox/, and library/ on every launch (idempotent). Previously only the Electron path (electron/main.js) did this, so node src/main.js could hit ENOENT inside Anamnesis writes if those files were missing from the repo. Defensive mkdirSync({recursive:true}) guard also added to session-debrief.js#backupAndWrite() so any other write path lands the parent directory before writeFileSync.

Changed (r24)

• Sponsor verification is now permanent on first success. The dashboard previously re-queried the Worker every 7 days and would re-show the support banner if the cache aged out. Now, once a sponsor (recurring or one-time) verifies once, the banner stays dismissed permanently. Cancellations on the recurring path still propagate through the GitHub Sponsors webhook → CF Worker → engine.

Fixed (r22)

• engine.cjs _validEdgeEndpointsSql edges:0 regression: helper defaulted edgeAlias to '', producing bare source/target in EXISTS subqueries. Since the nodes table has its own source column (text: diary/knowledge/inbox), SQLite's subquery scoping resolved source to the inner nodes.source instead of the outer edges.source — making EXISTS always false and silently zeroing every edge-counting path (dashboard stats, conn_count writes, etc.). Default now 'edges'; all five callsites use bare FROM edges, so the fix flows through.
• ANCHOR_TOUCH pulse marker leaking into user-visible text: dashboard.js PULSE_MARKER_RE plus two inline strip regexes and both telegram.js send-path replacers covered only TASK_TOUCH|COGNITIVE_TOUCH|DEBRIEF. ANCHOR_TOUCH (added later) now included in all five strip sites.
• logs/anamnesis-parse-failures/ unbounded growth: parse-failure dumps now cap to the 50 most recent files; older entries unlink on write.

Added

• First public OSS release scaffolding
• Anamnesis auto-archive (session-debrief.js): tasks with status completed/expired/failed that have sat ≥7 days move to identity/tasks-archive.json. Triple safety gate — status whitelist, parseable timestamp, 7-day dwell — in_progress/pending/blocked/suspended/code-ready/code-done are never touched. On archive write failure, all tasks remain in the active list.
• Cognitive-state auto-roll (session-debrief.js): when identity/COGNITIVE_STATE.md exceeds 64 KB, a full snapshot is appended to cognitive-state-archive.md and the active file is trimmed to H1 + the last ~32 KB.
• Default memory-hygiene cron at 0 4 * * * ships ON: nightly star-map hygiene (supersedes / dormant / weak-edge / noise / fusion / stale / gardener) + system health check via /api/doctor + event-timeline detection + inbox review. Users can disable in the Dashboard Cron Editor.
• Dashboard Tour Step 6 ("Cron Editor") now explains the default-on memory-hygiene cron and why to keep it enabled.

Changed

• LICENSE now ships verbatim AGPL-3.0 text
• CONTRIBUTING.md rewritten to match the actual npm scripts and runtime layout
• Identity templates updated to drop stale provider/transport references