CHANGELOG.

What's Changed

Features

• feat(workflow): Cognithor Resilient Workflow Engine (CRWE) — JSONL-streaming task runner with crash-recovery + audit chain (#499) (a6622ddd)

Fixes

Other

• chore: bump version to v0.99.0 — CRWE Resilient Workflow Engine (93bd9332)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.98.0...v0.99.0

What's Changed

Features

• feat(operational-trust): close remaining audit sinks (PR-D/4) — Compliance-Spring closer (#497) (bdf7778f)
• feat(operational-trust): WeightOptimizer encrypted snapshots + channel_contributions upstream fix (PR-B/3) (#496) (8798f9bd)
• feat(operational-trust): Reflector audit-event helper + causal atomic wiring (PR-A/3) (#494) (f593cddc)
• feat(pack-4): community-registry Ed25519 signing (TUF-Light) (#478) (6f586308)
• feat(extensions/vscode): cost-budget gutter decoration (Sprint-27 PR-I) (#456) (c641f5ed)
• feat(extensions/vscode): Decision-Explanation hover provider (Sprint-27 PR-H) (#455) (7b202d95)
• feat(extensions/vscode): TRUST-1 receipt sidebar webview (Sprint-27 PR-G) (#454) (1b8aa6f9)
• feat(mcp): TRUST-9 wiring for video_render — provenance + cost (Sprint-27 HF-4) (#451) (a544048f)
• feat(extensions/vscode): Cognithor Run Plan palette + WebSocket client (Sprint-27 PR-F) (#453) (d6d4649e)
• feat(mcp): video-composition skills layer (Sprint-27 HF-5) (#452) (3c4ab175)
• feat(mcp): video_compose + video_render MCP tools (Sprint-27 HF-3) (#450) (60b387f6)
• feat(extensions/vscode): MCP-stdio bridge auto-spawn (Sprint-27 PR-E) (#448) (6212288b)
• feat(core): multimodal helpers + VISION_TOKENS CostKind (Sprint-27 VLM-3) (#449) (518604cd)
• feat(video): RendererABC + HyperFramesRenderer (Sprint-27 HF-2) (#444) (a259c071)
• feat(vllm): VLMConfig + orchestrator launch flags for Qwen3-VL on RTX 5090 (Sprint-27 VLM-2) (#445) (b8c8432c)
• feat(extensions/vscode): scaffold (TypeScript + esbuild + vsce) — Sprint-27 PR-D (#442) (eb3d6ca1)
• feat(streaming+cli): WebSocketSink + cognithor agent ws (Sprint-27 PR-C) (#440) (29c03b3f)
• feat(streaming+cli): JsonlSink + cognithor agent run --stream (Sprint-27 PR-B) (#439) (1d08d873)
• feat(streaming): EventEmitter foundation + JSON Schema v1 (Sprint-27 PR-A) (#438) (d8d25cd5)

Fixes

• fix(pass-3): CB-wide bug sweep — 3 CRITs + 6 HIGHs + 3 MEDs + PACK-4 follow-up (#479) (82493e52)
• fix(deep-pr7): refuse ARC corpus task paths that escape corpus_root (#476) (1ba32dd4)
• fix(deep-pr6): warn when pack registers tools without explicit risk-level (#475) (8caa2a71)
• fix(deep-pr3): channel hardening (Telegram voice / Mattermost allow-list / Signal fail-closed) (#472) (b6657c22)
• fix(deep-pr5): stop logging user-text content + promote i18n locale-miss to WARNING (#474) (fa39ec33)
• fix(deep-pr4): MCP hardening (substitution, timeout cap, ReDoS, SSRF, format allowlist) (#473) (d6aa3f57)
• fix(deep-pr2): TRUST receipt builder best-effort + hashline reanchor on break (#471) (b8481a50)
• fix(deep-pr1): gateway task-leak + sync-in-async + unbounded set (3 deep-audit findings) (#470) (e24fcfbc)
• fix(audit-pr9): Feishu encrypt_key wiring + ReDoS guard on user regex (#468) (89e9649d)
• fix(audit-pr10): SQL inject_check bypass + --arc null-deref + list_directory depth cap (#469) (18eb0d2d)
• fix(audit-pr8): pack-loader sys.modules registration + richer error logging (#467) (2adeccd2)
• fix(audit-pr7): bootstrap OLLAMA_HOST + preflight env-var rebrand + Dart format CI gate (#466) (325c6897)
• fix(audit-pr6): i18n duplicate-key shadowing + Dart Map type-cast bugs (#465) (8e0cd47c)
• fix(audit-pr5): always clear ContextVars in message-handler finally (#464) (06528f92)
• fix(audit-pr4): Gatekeeper risk-classification batch — 14 tools + Kanban write demotion (#463) (50e66006)
• fix(audit-pr3): GDPR endpoint hardening — SQL injection + IDOR + path traversal (#462) (b97d6748)
• fix(audit-pr2): AuditLogger thread-safe persist + canonical disk write (#461) (ac011359)
• fix(audit-pr1): Sprint-27 fresh-code hardening (5 audit findings) (#460) (90bebd43)
• fix(cli): ruff format src/cognithor/main.py — Sprint-27 agent-cmd wiring (#441) (caa08075)

Other

• chore: bump version to v0.98.0 — Compliance-Spring (6b5c93f9)
• tools(burn-in): nightly Compliance-Spring smoke-test workflow + script (#498) (3e13de0d)
• test(operational-trust): Hypothesis property tests for audit-completeness (PR-C/3) (#495) (1933eca1)
• test(flutter): cover last 6 untested top-level screens (~15 tests) (#491) (a14c5bb4)
• test(backend): deep coverage for mcp/browser + mcp/database_tools (~50 tests) (#493) (a3a7501a)
• test(flutter): cover last 4 untested providers (~39 tests) (#492) (cb6a6b53)
• test(flutter): deep interaction tests for top-5 daily-driver screens (~26 tests) (#490) (7199f939)
• test(backend): deep coverage for message_handler + pge_loop (~40 tests) (#488) (1692c0fd)
• test(flutter): smoke for 4 daily-driver screens with timer-mocks (#489) (c385b1db)
• test(vscode): receipt_view + ws_client + mcp_bridge unit suites (~19 tests) (#487) (0d78708b)
• test(vscode): test infra + cost_gutter + decision_hover suites (~12 tests) (#485) (8148b06c)
• test(backend): deep coverage for agent_vault + post_processing (~99 tests) (#486) (02e4e821)
• test(flutter): cover 8 stateful config sub-pages (+8 tests) (#484) (71bd20bf)
• test(flutter): cover 14 deferred screens via silent-provider helpers (+38 tests) (#483) (58292b5c)
• test(flutter): 8 screen smoke tests + 2 helper modules (+32 tests) (#482) (c4ca8f08)
• test(flutter): cover 4 previously untested providers (+57 tests) (#481) (76f2c44d)
• refactor(gateway): extract _sync_core_inventory + reload_components + execute_delegation (#480) (af29a70b)
• chore(deps-dev): bump esbuild (#447) (875b6d3f)
• test(video): VLM-4 integration smoke + hardware-gated full pipeline (Sprint-27) (#459) (e5c6b518)
• test(streaming): end-to-end smoke for Plan→Gate→Execute roundtrip (Sprint-27 PR-K) (#458) (ea299332)
• docs(extensions/vscode): IDE compat guide + smoke (Sprint-27 PR-J) (#457) (1e592c8a)
• docs(spikes): HyperFrames + Qwen3-VL on RTX 5090 — research output for HF-1 + VLM-1 (#443) (872cbf37)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.97.0...v0.98.0

What's Changed

Features

• feat(memory): TRUST-9 wiring into MemoryIndex.upsert_chunk (#435) (898c54fe)
• feat(security): TRUST-10 self-audit backfills for cost / escalation / scope ledgers (#434) (d2ccf101)
• feat(learning): TRUST-9 wiring into KnowledgeIngestService.ingest_file (#433) (0dadea88)
• feat(security): TRUST-10 self-audit backfills for provenance + fingerprint ledgers (#432) (bc2e4e10)
• feat(cli): cognithor receipt diff — surface trust-ledger deltas between two receipts (#431) (003ca45c)
• feat(cli): cognithor receipt export-all — bulk-export per-session receipts (#430) (6e23b96e)
• feat(gatekeeper): TRUST-3 — derive failure_mode for new policy_names (#429) (6c9423b0)
• feat(gatekeeper): TRUST-2 explanation on capability-matrix block path (#428) (7c8c95e6)
• feat(packs): TRUST-7 SCHEMA-kind fingerprint of PackManifest schema (#427) (49f13ded)
• feat(memory): TRUST-9 wiring into CoreMemory.save (#426) (51e70247)
• feat(memory): TRUST-9 wiring into ProceduralMemory.save_procedure (#425) (75577f4f)
• feat(cli): cognithor receipt list — enumerate session_ids in audit log (#424) (ae2dfb59)
• feat(gatekeeper): TRUST-2 explanation on disabled-tool block path (#423) (55005e12)
• feat(gatekeeper): TRUST-3 — explicit failure_mode on gatekeeper audit entries (#422) (5c51010e)
• feat(audit): TRUST-3 — classify TRUST-5..10 failure modes (#421) (53d9b4d1)
• feat(security): TRUST-9 wiring into IsolatedSessionStore.create_session (#420) (bfb80898)
• feat(gatekeeper): TRUST-2 explanation on OFFLINE-mode network blocks (#419) (9b2a5117)
• feat(gatekeeper): TRUST-2 explanation on credential-mask path (#418) (89dbfb19)
• feat(packs): TRUST-10 backfill — record pack_manifest schema lineage (#417) (94c5db0e)
• feat(audit): TRUST-10 backfill — record SEC-HIGH-5 schema migration (#416) (947c9d87)
• feat(gatekeeper): TRUST-2 explanations on Python-code + path-validation blocks (#415) (ad65f7a3)
• feat(backends): TRUST-7 auto-wire fingerprint_model on first chat() (#414) (cb612fdc)
• feat(backends): TRUST-7 MODEL-kind fingerprint capture in OllamaBackend (#413) (e41ed4a5)
• feat(security): TRUST-9 wiring into AgentVault.store (#412) (04d182fc)
• feat(memory): TRUST-9 wiring into SemanticMemory.add_relation (#411) (641a1c25)
• feat(memory): TRUST-9 wiring into EpisodicMemory.append_entry (#410) (19f3ef1a)
• feat(memory): TRUST-9 wiring into SemanticMemory.add_entity (#409) (d8502d52)
• feat(packs): TRUST-7 fingerprint capture in PackLoader (#408) (e11fc19f)
• feat(mcp): TRUST-7 fingerprint capture in register_tool (#407) (047eeb0d)
• feat(security): TRUST-6/8 cross-wiring helper (#406) (2adaf466)
• feat(api): GET /api/crew/trace/{trace_id}/receipt (#405) (bfcc858f)
• feat(cli): cognithor receipt show / verify (TRUST-1 + trust-bundle) (#404) (6a1ecd74)
• feat(audit): opt-in include_trust= folds TRUST-5..10 ledgers into run_receipt (#403) (c1311923)
• feat(security): TRUST-1 trust-bundle composer (#402) (48e2dfc0)
• feat(security): TRUST-6 cost-ledger foundation (#401) (bbbb8149)
• feat(security): TRUST-10 migration-ledger foundation (#400) (13a94cbd)
• feat(security): TRUST-8 cloud-escalation ledger foundation (#399) (990c4c01)
• feat(security): TRUST-7 ToolFingerprint foundation (#398) (f6c57a8f)
• feat(memory): TRUST-9 ProvenanceTag foundation (#397) (21054b20)
• feat(security): TRUST-5 production wiring — Gatekeeper integrates ScopeRegistry (#396) (78cc9038)
• feat(security): TRUST-5-Lite PermissionScope foundation (#395) (77262ca9)
• feat(pse): wire Sprint-26 domains into gateway boot (#394) (583cf612)
• feat(pse): Sprint-26 polish — register_all_sprint26_domains() central wiring (#393) (7e874f39)
• feat(pse): Sprint-26.4 — BinaryData + Image V2 + Cross-Domain Demo (Sprint-26 final) (#392) (66945df5)
• feat(pse): Sprint-26.3 — Datetime + AST + Float domains (Woche 3) (#391) (7d7981df)
• feat(pse): Sprint-26.2 — SQL + JSON + Bridge whitelist (Cut-Off 2) (#390) (2dd84a7d)
• feat(pse): Sprint-26.1 — Domain-Expansion Foundation + Cost-Tracker (#389) (5d8386fa)
• feat(audit): TRUST-3 FailureMode enum + classifier + aggregator (#376) (e479b7a2)
• feat(packs): TRUST-4 pack rollback with automatic pre-upgrade snapshots (#377) (9bf733d2)
• feat(gatekeeper): TRUST-2 structured DecisionExplanation for "why" reconstruction (#375) (cb7a4ef1)
• feat(audit): TRUST-1 Run-Receipts API + signed bundle for post-mortem reconstruction (#374) (8b247497)
• feat(pse): Sprint-25 — Planner-Refinement-Loop nach PSE-Synthese (#365) (930ed72e)
• feat(pse): Sprint-24 — PSE auto-switch in ContextPipeline produktiv schalten (#364) (31233650)
• feat: Sprint-23 PR#J — context-profile REST endpoints + Flutter picker (#356) (56cf2754)
• feat(backend): Sprint-23 PR#E — num_ctx through LLMBackend (vLLM + Ollama + others) (#351) (a68e4dac)
• feat(router): Sprint-23 PR#D — pipe num_ctx end-to-end into Ollama wire (#350) (dc59f3ab)
• feat(router): Sprint-23 PR#C — context-manager helper for profile scoping (#349) (540a60af)
• feat(router): Sprint-23 PR#A — pure context-profile selection heuristic (#348) (d0aba201)
• feat(router): Sprint-23 — task-aware context profile selection (#347) (b7588474)
• feat(pse): Sprint-22 Track B PR#4 — List/Sequence-DSL family + IntList type (#346) (441c6898)
• feat(pse): Sprint-22 Track B PR#3 — Number/Int-DSL family + Int as first-class (#345) (3f6843fb)
• feat(pse): Sprint-22 Track B PR#2 — Regex/Pattern-DSL family (#344) (5e1be776)
• feat(pse): Sprint-22 Track B PR#1 — String-DSL family + generic input typing (#343) (71f374dd)
• feat(pse): Sprint-22 — JSONL persistence for PSECache (Infrastruktur 9→10) (#341) (0a902721)
• feat(planner): Sprint-22 A.3 — PSE routing nudge in Planner system prompt (#340) (ac4368c4)
• feat: Sprint-22 A.4+A.5 — PSE open-world hardening (held_out + auto-escalate) (#339) (290b9293)
• feat(integration): Sprint-22 Track A — PSE engine as first-class MCP tool surface (#337) (7ad8b0e0)
• feat(integration): Sprint-22 Track A.2 — PSE tools as GREEN in Gatekeeper (#338) (a4b0ab64)
• feat(integration): Sprint-21 — arc_tools prefers central VLLMBackend HTTP path (#336) (68cdcfb7)
• feat(pse): Sprint-20 Hebel V — few-shot win-demo store + auto-record (#335) (800a2600)
• feat(driver): Sprint-19 Hebel U — Phase-A driver multi-game default expansion (#334) (4b343531)
• feat(pse): Sprint-19 Hebel T — first-time-action exploration bonus (#333) (49301833)
• feat(pse): Sprint-19 Hebel S — per-action pixΔ histogram in vision prompt (#332) (282144ca)
• feat(pse): Sprint-19 Hebel R — RESET-bonus when stalled+destructive (#331) (0e09d270)
• feat(pse): Sprint-19 Hebel Q — single-pixΔ-spike>1000 plan-scorer trigger (#330) (408b2330)
• feat(pse): Sprint-19 Hebel P — raw LLM reasoning in audit JSONL (#329) (e1a48822)
• feat(pse): Sprint-19 Hebel O — stalled-progress warning in vision prompt (#328) (2c67d99f)
• feat(pse): Sprint-19 Hebel N — pixΔ-safety multiplicative gate in plan_scorer (#326) (431e5ab5)
• feat(pse): Sprint-19 Hebel M — absolute pixΔ in trajectory + concrete prompt rule (#325) (85f17511)
• feat(pse): COGNITHOR_PSE_DEBUG_LLM env-flag — surface choice_fn exceptions (#324) (d91eb758)
• feat(pse): Sprint-19 — full-stack score-attack (Hebels A+B+C+D+E+L) (#323) (968e95a3)
• feat(pse): Sprint-18 — add llm_planning agent variant for plan-horizon A/B (#321) (6f0203ec)
• feat(pse): Sprint-17 — bp35 observed-rules + max_tokens + pixΔ in history (score-lift hebels) (#320) (94eacc35)
• feat(pse): Sprint-16 Hebel 2 — action-streak detector + vLLM perf tuning (#316) (6b6af4af)
• feat(pse): Sprint-16 Hebel 1 — anti-loop in LLMActionDecoder (ff1cf973)
• feat(pse): Sprint-15 — MTP stats + Hashline-Guard audit-chain integration (#312) (dcdff467)
• feat(pse): Sprint-15 — LLM call telemetry (evidence for tuning) (#311) (0cc8282f)
• feat(pse): Sprint-15 — vLLM tuning knobs (FP8 KV cache + MTP speculative) (#310) (1a37723e)
• feat(pse): Sprint-15 — training-free GraphExplorerAgent (no LLM) (#309) (74fa5add)
• feat(pse): Sprint-14 — multi-step planning + vision-mode for Qwen3.6 multimodal (#308) (7f1110de)
• feat(pse): Sprint-13 — LLM prompt perfection (game-rules + GoalInferer) (#307) (8766f2c2)
• feat(pse): Sprint-12 PR-13 — wire ClickTargetSampler into Sprint10DSLAgent (#304) (47e76a7d)
• feat(mcp): rewire arc_play/arc_replay onto the new program_synthesis stack (#303) (020be108)
• feat(pse): Sprint-12 PR-6+7 — wire ArcAuditTrail + GameProfile into Sprint10DSLAgent (#297) (3b20f434)
• feat(pse): Sprint-12 PR-5+11 — wire FrameAnalyzer + level-transition resets (#299) (40f2fafd)
• feat(pse): Sprint-12 PR-12 — feed FrameAnalyzer summary into LLMActionDecoder prompt (#300) (8ce2d276)
• feat(pse): Sprint-12 — in-process EpisodeRunner for ad-hoc episodes (#302) (95a74ff8)
• feat(pse): Sprint-12 PR-9 — salience-based ClickTargetSampler for click-target games (#295) (ec08b9cf)
• feat(pse): Sprint-12 PR-8 — wire plan_click_solution() as Sprint10DSLAgent fast-path (#294) (6294ba7f)
• feat(pse): Sprint-12 PR-4 — FrameAnalyzer (per-action movement signatures) (#293) (0532d0c2)
• feat(pse): Sprint-12 PR-2 — persistent observation log + GameProfile (#291) (6b4b2488)
• feat(pse): Sprint-12 PR-3 — pure-NumPy fast-path planner for click-toggle games (#292) (3278b5e3)
• feat(pse): Sprint-12 — StateGraph + state-keyed counts + per-game prompts (#290) (e1d19bf3)
• feat(pse): Sprint-12 — NVFP4 model + InProcess vLLM backend (RTX 5090 validated) (#289) (190fc496)
• feat(pse): Sprint-11 Wave-6 — arcengine harness shim + scorecard parser (Sprint-11 closed) (#287) (442f5ba0)
• feat(pse): Sprint-11 Wave-5 — LLMReasoningAgent over vLLM/qwen3.6:27b (#286) (55782fc6)
• feat(pse): Sprint-11 Wave-4 — Sprint10DSLAgent + DSLActionDecoder (#285) (066d4205)
• feat(pse): Sprint-11 Wave-3 — EpisodeMemory + ChangeDetector + StuckDetector (#284) (a46eb4e8)
• feat(pse): Sprint-11 Wave-2 — FrameBridge + ActionDecoder (#283) (294ee106)
• feat(pse): Sprint-11 Wave-1 — ARC-AGI-3 Game-Agent foundation (#282) (38e1ebee)
• feat(pse): Sprint-10 Track B — vLLM/qwen3.6:27b LLM-Prior wiring (infrastructure) (#281) (5d99560a)
• feat(pse): Sprint-10 Wave-4 — crop_to_least_common_color_cells (+2 solves) (#280) (a86be18c)
• feat(pse): Sprint-10 Wave-3 — crop_smallest_component + neighbor_count_grid (#279) (7e461ebb)
• feat(pse): Sprint-10 Wave-2 PR-4 — crop_largest_component (solves 1f85a75f, be94b721) (#277) (cded1b7d)
• feat(pse): Sprint-10 Wave-1 PR-3 — fill_with_most_common_color (solves 5582e5ca) (#276) (24278884)
• feat(pse): Sprint-10 Wave-1 PR-2 — 4 symmetry-completion primitives (#275) (6b4eafa6)
• feat(pse): Sprint-10 Wave-1 — self_tile_by_mask primitive (solves real ARC 007bbfb7) (#274) (f1272740)
• feat(pse): Phase-2 Sprint-9 — real ARC-AGI corpus reality-check baseline (#273) (ca254ef5)

Fixes

• fix(test): _CapturingClient ignores TRUST-7 /api/show probe + accepts kwargs (9962044c)
• fix(test): seed _fingerprinted_models in test_chat_basic — TRUST-7 cache hit avoids /api/show probe (8f0c3443)
• chore(release): v0.97.0 — Operational Trust (e2bc9f1a)
• feat(api): GET /api/crew/trace/{trace_id}/receipt (#405) (bfcc858f)
• fix(workflow): pack update stub exits 2 and requires qualified_id (WF-6) (#380) (d9b7b8de)
• fix(workflow): iMessage on Win/Linux reads BlueBubbles env or skips with warning (WF-4) (#379) (781ba81b)
• fix(workflow): research_screen skips API calls when deep-research pack absent (WF-1) (#378) (6eb53499)
• fix(security): add SHA-256 hash chain to audit log JSONL (#373) (5d1a1a0f)
• fix(security): wrap web content in + drop trust directive (#372) (b8678980)
• fix(workflow): robot_office polls correct kanban endpoint with bare-list parser (WF-3) (#371) (de6b3f7b)
• fix(security): route start_background through destructive-command guard (#369) (ad006044)
• fix(security): SQLCipher fail-closed when encryption requested but unavailable (#368) (45185602)
• fix(security): pack manifest can no longer downgrade built-in tool risk (#367) (8a3c94d3)
• fix(workflow): drop double /api/v1 prefix in 3 Flutter providers (#366) (262c62e8)
• fix(types): mypy --strict zero across entire src/cognithor/ + CI gate (10 rounds) (#363) (6d1a1061)
• chore(types): Sprint-23 PR#M — final mass mypy round (970 → 965) (#360) (f90c4649)
• fix(tests): import-sort I001 in test_context_profile_routes (main CI hotfix) (#358) (5f382050)
• fix(pse): Run #27 — broaden Hebel N + bump max_tokens 4096→6144 (#327) (e1b3ddbd)
• fix(pse): revert max_tokens 1024→2048 — Sprint-17 caused 40→17 tok/s drop (#322) (8dbaf289)
• feat(pse): Sprint-17 — bp35 observed-rules + max_tokens + pixΔ in history (score-lift hebels) (#320) (94eacc35)
• fix(pse): Sprint-16 — VRAM 0.93 + max_num_seqs=8 + disable fast_path on llm_full (#319) (14ee645b)
• fix(pse): roll back vLLM defaults — VRAM 0.90 + ctx 32K (Run #12 init crash) (#317) (ac3a6942)
• fix(pse): TTFT capture — pipe ttft_s through audit chain into JSONL (e8fb17a1)
• fix(pse): num_speculative_tokens=1 — match MTP head's training (37fb8893)
• fix(pse): switch llm_full to greedy decoding for MTP (0e2b1b4d)
• fix(pse): re-enable vLLM log_stats when telemetry/MTP-stats wired (2649acbf)
• fix(pse): vLLM 0.20-v1 telemetry adapters — TTFT field + MTP delta-poll (d769601d)
• fix(pse): correct MTP checkpoint id — sakamakismile/Qwen3.6-27B-Text-NVFP4-MTP (4db7de1f)
• fix(pse): Qwen3.6-27B has no instruct variant — switch defaults to FP8 base + add live validation runbook (#288) (e8491108)
• fix(test_security): replace openssl subprocess with cryptography library (#272) (30a8681b)

Other

• docs(sprint-27): IDE-Integration plan-doc — VS-Code + headless CLI (#437) (8f59a2a3)
• docs: operational_trust reference document (#436) (1ca7dd34)
• docs(integrations): sevdesk is in development, not v1.0 Launch (A-2) (#383) (8550e0ac)
• chore(integrations): catalog now lists 136 tools incl. pse_synthesize family (A-6) (#382) (6d94b73e)
• docs(skills): mark 4 marketplace-scaffolding modules NOT WIRED (A-8) (#385) (8628a42a)
• refactor(arc): mark cognithor.arc.classic deprecated for one minor cycle (A-4) (#384) (ad3acec7)
• refactor(core): delete dead collaboration module (A-7) (#381) (4ba02d59)
• test(flutter): add 28 unit tests for KanbanProvider (#388) (ec549dd9)
• test(flutter): add 24 unit tests for EvolutionProvider (#387) (0c991f56)
• test(flutter): add 24 unit tests for SessionsProvider (#386) (53746849)
• chore(gateway): delete dead _idle_loop class attribute (#370) (41a0a709)
• chore(types): Sprint-23 PR#O — mypy --strict zero across the entire codebase (#362) (2144d5a3)
• chore(types): Sprint-23 PR#N — surgical mypy round (965 → 737) (#361) (f8d17f93)
• chore(types): Sprint-23 PR#L — mass mypy --strict cleanup (1547 → 970, -37 %) (#359) (1380b68a)
• chore(config_routes): Sprint-23 PR#K — Protocol-based RoutableApp typing (#357) (70ee66ff)
• chore(mypy): exclude cookiecutter templates + bundled packs from strict scan (#355) (d1652d75)
• chore(core): Sprint-23 PR#H — finish core/ mypy --strict (63 → 0) (#354) (6edc55b7)
• chore: Sprint-23 PR#G — delete 10 orphan modules + fix legacy asyncio test (#353) (70e85900)
• chore(core): Sprint-23 PR#F — sweep pre-existing mypy issues across core/ (#352) (8072596c)
• perf(pse): Sprint-22 — per-execute memoization in InProcessExecutor (Engine-Performance 9→10) (#342) (cbe14a95)
• Update README.md (9967f654)
• docs(pse): Sprint-15 Phase-A report + finalise driver to MTP=1 (b332b527)
• diag(pse): Phase-A run #9 — MTP-variant as MAIN model (53ee9f9d)
• diag(pse): MTP-off baseline run for Phase-A (dc196453)
• scripts: Sprint-12 Phase-A live validation driver (#305) (1a95e50f)
• docs(pse): comprehensive README for the arc_agi3 package (#301) (c87eac46)
• docs: Sprint-12 Phase-A live validation runbook (#298) (83f35d75)
• test(pse): Sprint-12 — end-to-end smoke tests + level-transition fix (#296) (73b1818a)
• docs(pse): Sprint-10 Wave-1+2 real-ARC validation — +7 training / +2 eval solves (#278) (4062333c)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.96.0...v0.97.0

What's Changed

Features

• feat(pse): Phase-2 Sprint-8 Track B — 100 % on every ARC corpus subset (#271) (1a85c361)
• feat(pse): Phase-2 Sprint-6 Track A — first real Score-Lift via Symbolic-Repair cascade (#269) (30acee61)
• feat(pse): Phase-2 Sprint-5 Track 1 — A/B-Test on hard subset reveals partial activation (#268) (c968b276)
• feat(pse): Phase-2 Sprint-4 — ARC-AGI-3 corpus + Phase-1 reality-check (#267) (2aca8e85)
• feat(pse): Phase-2 Sprint-3 Track 1 — A/B benchmark with --phase2 flag (#266) (56368990)
• feat(pse): Phase-2 Sprint-2 Track D — Benchmark report + nightly CI (#264) (3e743587)
• feat(pse): Phase-2 Sprint-2 Track C — 20-Task Leak-Free fixture set (#263) (605a0a46)
• feat(pse): Phase-2 Sprint-2 Track E — MCTS extensions (parallel + restart + diversity) (#265) (ce870906)
• feat(pse): Phase-2 Sprint-2 Track B — Production wiring (#262) (546c72da)
• feat(pse): Phase-2 Sprint-2 — End-to-end Verifier evaluator (§7.3.3, Sprint-2 Track A) (#261) (b0444cca)
• feat(pse): Phase-2 Sprint-1 — Benchmark runner (plan task 12, §12.3 / §12.4) (#260) (be8bfd61)
• feat(pse): Phase-2 Sprint-1 — Top-level synthesis engine (plan task 10, §3.2) (#259) (3dfd2bc5)
• feat(pse): Phase-2 Sprint-1 — Module B MCTS controller (plan task 7, §5) (#258) (696d5490)
• feat(pse): Phase-2 Sprint-1 — Symbolic-Prior heuristic catalog (plan task 4, §4.4) (#257) (89ad2348)
• feat(pse): Phase-2 Sprint-1 — Refiner Escalation logic (plan task 9 final slice, §6.6) (#256) (80fb6154)
• feat(pse): Phase-2 Sprint-1 — Hybrid-Repair (plan task 9 slice, §6.5.2 Zone-2) (#255) (366d122b)
• feat(pse): Phase-2 Sprint-1 — Trace-Replay (plan task 9 slice, §6.4) (#253) (fa89ae12)
• feat(pse): Phase-2 Sprint-1 — LLM-Repair Two-Stage (plan task 9 slice, §6.5.2 Zone-1) (#254) (33da64d8)
• feat(pse): Phase-2 Sprint-1 — Symbolic-Repair Advisor (plan task 9 slice, §6.5.2) (#252) (18476703)
• feat(pse): Phase-2 Sprint-1 — Diff-Analyzer (plan task 9 slice, §6.3) (#251) (3f05639d)
• feat(pse): Phase-2 Sprint-1 — CEGIS Refiner stage (plan task 9 slice, §6.5.3) (#250) (6dbf420c)
• feat(pse): Phase-2 Sprint-1 — Local-Edit Repair (plan task 9 slice, §6.5.1) (#249) (f702f27b)
• feat(pse): Phase-2 Sprint-1 — triviality rules (plan task 8 slice, §7.3.1) (#248) (bf2b3273)
• feat(pse): Phase-2 Sprint-1 — partial_pixel_match metric (plan task 8 slice) (#247) (2852c7a9)
• feat(pse): Phase-2 Sprint-1 — α-Controller + PriorPerformanceTracker (plan task 6) (#246) (61ba82ff)
• feat(pse): Phase-2 Sprint-1 — verifier final-score aggregation (plan task 8, §7.2) (#245) (402776d8)
• feat(pse): Phase-2 Sprint-1 — typed datatypes (plan task 2, spec §9) (#244) (39bb2512)
• feat(pse): Phase-2 Sprint-1 — YAML heuristics loader (plan task 1) (#243) (0cb3fb3e)
• feat(pse): Phase-2 Sprint-1 — F3 interaction tests (spec v1.4 §12.2) (#242) (387ef6c4)
• feat(pse): Phase-2 Sprint-1 — Dual-Prior mixer + Symbolic-Prior interface (#241) (142034d2)
• feat(pse): Phase-2 Sprint-1 — Module A LLM-Prior over vLLM/Qwen 3.6 27B (#240) (46c74195)
• feat(pse): Phase-2 Sprint-1 — telemetry counters (spec v1.4 §11) (#239) (289ff74f)
• feat(pse): Phase-2 Sprint-1 — α-mixer + sample-size-dampening (§4.4 / §4.4.4) (#238) (39f7460e)
• feat(pse): Phase-2 Sprint-1 — compute_suspicion helper (spec v1.4 §7.3.2) (#237) (4035a7ae)
• feat(pse): Phase-2 Sprint-1 — auto-classify primitives via @primitive decorator (#236) (68969607)
• feat(pse): Phase-2 Sprint-1 — F1 (DSL classification) + F2 (Refiner mode controller) (#235) (0a37de82)
• feat(pse): Phase-1 final close — K4 subprocess sandbox + D5/D7 + D13 (#234) (d6a4b742)
• feat(pse): D15 — mypy --strict clean + pre-commit pytest hook (#233) (d4081830)
• feat(pse): K4 — flip on 6 type-system-blocked adversarial cases (#232) (2a2fd534)
• feat(pse): ARC-AGI-3 eval-suite scaffold (D5 prep) (#231) (de2902bb)
• feat(pse): seed Predicate/Lambda bank so HO primitives are reachable (#228) (c4e5d3ec)
• feat(pse): channel-side telemetry + audit-trail wiring — Week 7 day 5 (#225) (11dfa4cf)
• feat(pse): telemetry counters + audit trail — Week 7 day 4-5 (D10 + D11) (#224) (f9192c2b)
• feat(pse): cognithor pse CLI — Week 7 day 4 (D8 — spec §19.2) (#223) (c728bfea)
• feat(pse): deterministic Cost-Auto-Tuner — Week 7 day 1 (spec §7.6) (#222) (34cd4e95)
• feat(pse): branch (H5) — conditional-Lambda combinator — Week 6 day 6 (v1.2 spec) (#221) (ae745a92)
• feat(pse): sort_objects (H4) + SortKey enum — Week 6 day 5,5 (v1.2 spec addition) (#220) (49321e6d)
• feat(pse): align_to (H3) + AlignMode enum — Week 6 day 4-5 (#219) (ee820501)
• feat(pse): Lambda type + map_objects + filter_objects — Week 6 day 3-4 (Phase 1.5 H1+H2) (#218) (dc6de659)
• feat(pse): closed-set Predicate type + 13 constructors — Week 6 day 1-2 (Phase 1.5 prerequisite) (#217) (360cfcba)
• feat(pse): trace + replay system — Week 5 day 4-5 — K9/K10 HARD GATES (#215) (745db7f1)
• feat(pse): PGE adapter + state-graph bridge + numpy fast-path — Week 5 day 1-2 (#214) (35918d23)
• feat(pse): sandbox strategy + platform detection — Week 4 day 4-5 (spec §11.6) (#213) (76b7b249)
• feat(pse): tactical-memory cache + capability-token constants — Week 4 day 3 (#212) (9585e551)
• feat(pse): five-stage Verifier pipeline — Week 4 day 1-2 (spec §10) (#211) (0c69f99e)
• feat(pse): bottom-up enumerative search engine — Week 3 day 8 — FIRST END-TO-END SYNTHESIS (#210) (a525d5a3)
• feat(pse): in-process executor + observational-equivalence pruner — Week 3 day 3-7 (#209) (ce76a082)
• feat(pse): search candidate-tree types — Week 3 day 1-2 (Program/InputRef/Const) (#208) (2edcc088)
• feat(pse): add mask/logic + construction + color-constant primitives — Week 2 part D (21/56) (#207) (7d5a19a7)
• feat(pse): add object-detection DSL primitives — Week 2 part C (8/56) (#206) (de374b51)
• feat(pse): add size/scale + spatial DSL primitives — Week 2 part B (12/56) (#205) (ed5c6e71)
• feat(pse): add geometric + color DSL primitives — Week 2 part A (15/56) (#203) (5d7beea2)
• feat(pse): scaffold Program Synthesis Engine — Week 1 of Phase 1 spec (#202) (d355a847)
• feat(bridge): risk caching, Planner-driven goal evaluator, HITL-resume fallback (#160) (075015b7)

Fixes

• feat(pse): Phase-2 Sprint-8 Track B — 100 % on every ARC corpus subset (#271) (1a85c361)
• fix(audit): batch 3 — hashline prev_hash chain, canvas-static-registration, catalog drift trigger (#186) (7ecb84de)
• fix(audit): batch 2 — owner-gate, paths cleanup, CI hygiene, pre-commit hook (#185) (906c160a)
• fix(audit): software-side fixes for #2 #3 #4 from 2026-04-29 workflow audit (#184) (62591d17)
• fix(identity): close two latent observations from identity-test sessions (442 total) (#173) (dd7711da)

Other

• chore(release): v0.96.0 — ARC-AGI-3 Pass (9a5f2c68)
• docs(pse): record D3/D9/D12 + fix HO-count drift (#230) (075ecc90)
• docs(pse): benchmarks.md scaffold + drift gate (D9) (#229) (40529ac8)
• docs(pse): hello-world tutorial + drift-test (D12) (#227) (d99bcbc5)
• docs(pse): auto-generated DSL reference + overview + architecture (D9) (#226) (626d1c98)
• test(pse): adversarial security test scaffold — Week 5 day 3 (K4 hard gate) (#216) (ba97d80f)
• test(flutter): unit tests for SecurityProvider (+12 tests) (#201) (f4728da4)
• test(flutter): unit tests for SkillsProvider (+17 tests) (#200) (76b6977b)
• perf(flutter): migrate workflows-categories tab to CustomScrollView+SliverList.builder (#199) (5c30a3e3)
• test(flutter): unit tests for 4 more ChangeNotifier providers (+22 tests) (#197) (bf893af0)
• chore(tooling): pre-commit auto-install + verify_readme_claims in CI + JARVIS_HOME fallback fix (#194) (f91a75f4)
• test(flutter): unit tests for 4 more ChangeNotifier providers (+45 tests) (#198) (19338a9a)
• test(flutter): add unit tests for 4 simple ChangeNotifier providers (#196) (1d2b1c8b)
• perf(flutter): migrate chat-history search results + models catalog to lazy slivers (#195) (640554e6)
• refactor(gateway): extract lifecycle.py from gateway.py — split COMPLETE (6/6) (#193) (31c9770c)
• refactor(gateway): extract message_handler.py from gateway.py (split 5/6) (#192) (fdf4650b)
• refactor(gateway): extract post_processing.py from gateway.py (split 3/6) (#190) (596b6b6c)
• refactor(gateway): extract pge_loop.py from gateway.py (split 4/6) (#191) (4a0102bf)
• refactor(gateway): extract session_mgmt.py from gateway.py (split 2/6) (#189) (fcdd0fe2)
• refactor(gateway): extract message_utils.py from gateway.py — split step 1/6 (#188) (87a63497)
• perf(flutter): migrate leads + monitoring eager ListView to CustomScrollView+SliverList.builder (#187) (e2d1c05e)
• refactor(config_routes): extract learning.py + autonomous.py + social.py — split complete (8/8) (#183) (7e8c0797)
• refactor(config_routes): extract evolution.py + infrastructure.py + workflows.py from _factory.py (7/8) (#182) (89efb340)
• refactor(config_routes): extract ui.py from _factory.py (6/8) (#181) (6488076a)
• refactor(config_routes): extract skills.py from _factory.py (5/8) (#180) (a150ea1c)
• refactor(config_routes): extract security.py + governance.py from _factory.py (4/8) (#179) (d1d3dbf8)
• refactor(config_routes): extract session.py + monitoring.py from _factory.py (3/8) (#178) (cf6769cb)
• refactor(config_routes): extract system.py + config.py from _factory.py (2/8) (#177) (dfeb7498)
• refactor(config_routes): backbone — convert single-file module to package + route-inventory drift test (1/8) (#176) (3fde3965)
• chore(benchmark): archive in-process benchmark/ module — no callers, duplicates cognithor_bench/ (#175) (d49ac3a0)
• docs(audit): benchmark/ archive recommendation (closes last VERIFY-WIRING item) (#174) (7fdbd6d8)
• test(identity): Session 4 Phase 4 (FINAL) — state mgmt + kill switch + admin (28 tests, 438 total) (#172) (79a62ed7)
• test(identity): Session 4 Phase 3 — retrieve_memories + build_context_for_llm (21 tests, 410 total) (#171) (2f0c4e8f)
• test(identity): Session 4 Phase 2 — process_interaction flow (27 tests, 389 total) (#170) (37344bfe)
• test(identity): Session 4 Phase 1 — engine init + Genesis Anchors (20 tests, 362 total) (#169) (dbdc1ff1)
• test(identity): Session 5 — attention + temporal + local_store + merkle_batcher (147 new tests, 342 total) (#168) (3cd6208d)
• test(identity): Session 3 — vector_store + working_memory (50 new tests, 195 total) (#167) (42436ac5)
• test(identity): Session 2 — adapter + llm_bridge (32 new tests, 145 total) (#166) (e5f1fe1f)
• test(identity): Session 1 — memory + epistemic + input_sanitizer (81 tests) (#165) (a93283dc)
• docs(audit): scope identity/ test-coverage gap before writing tests (#164) (f141e308)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.95.0...v0.96.0

What's Changed

Features

• feat(flutter): WP4 Trace-Screen + WP5 AutoGen-Shim verification (v0.95.0 PR 4) (#159) (6741650f)
• feat(channels): WP3 WebSocket Live-Stream for Crew Traces (v0.95.0 PR 3) (#158) (b0d696ee)
• feat(api): WP2 REST API for Crew-Traces (v0.95.0 PR 2) (#156) (7418bd98)
• feat(trace): WP1 TraceBus + Owner-Gating (v0.95.0 PR 1) (#155) (2e881f0d)
• Merge pull request #153 from Alex8791-cyber/feat/claude-code-hook-bridge (c1eedb62)
• feat(core): register CLAUDE_CODE_SUPERVISED backend + fix parallel tool pairing (05608d9e)
• feat(gateway): route Claude Code APPROVE through ApprovalManager (HITL) (fdf007c5)
• feat(core): ClaudeCodeSupervisor — autonomous outer-loop driver for the claude CLI (e63a8141)
• feat(gateway): Claude Code hook bridge — Gatekeeper + Observer over VS Code (6f4aebab)

Fixes

• fix(ci): unblock main on Win-py3.12 — lxml-html-clean dep + perf budget bump (#154) (74033cf9)

Other

• docs(changelog): roll [Unreleased] to [0.95.0] (0e72c1eb)
• docs(readme): add v0.95.0 Trace-UI highlights (5bd7e767)
• build: bump version to 0.95.0 (22b2b0ab)
• chore(format): apply ruff format to bridge + supervisor files (03c6f846)
• chore(lint): clear all 14 Ruff findings on the hook bridge + supervisor (d7455b1d)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.94.1...v0.95.0

Cognithor v0.94.1 — Hotfix for v0.94.0

Released: 2026-04-26

Post-v0.94.0 dry-run audit surfaced three fixable wiring concerns. All addressed in this hotfix.

Was wurde gefixt

[autogen] Extra ist jetzt vollständig (HIGH)

Bei pip install cognithor[autogen] && cognithor-bench --adapter autogen kam vorher eine irreführende ImportError. Grund: OpenAIChatCompletionClient lebt im PyPI-Paket autogen-ext, nicht in autogen-agentchat. Beide [autogen]-Extras (root + cognithor_bench) ziehen jetzt zusätzlich autogen-ext[openai]==0.7.5. Adapter-Error-Message aktualisiert.

pip install "cognithor[autogen]==0.94.1"
# zieht jetzt: autogen-agentchat==0.7.5 + autogen-ext[openai]==0.7.5

docs/integrations/catalog.json ist sauber (MEDIUM)

Die zwei sevDesk-Tools (sevdesk_get_invoice, sevdesk_list_contacts) waren im Catalog gelistet, aber NICHT im laufenden MCP-Server registriert (@mcp_tool-Decorator ist ein No-op-Marker; das Modul wird nirgends importiert). Die cognithor.ai/integrations-Seite hat damit Funktionalität versprochen, die der Server nicht liefert.

Der Generator-Script scripts/generate_integrations_catalog.py hat jetzt eine explizite NOT_YET_REGISTERED_PREFIXES-Filter. Die sevDesk-Einträge werden automatisch wieder im Catalog auftauchen, sobald der Connector live verdrahtet wird.

Windows-Konsole zeigt §-Zeichen korrekt (LOW)

insurance-agent-pack run --interview rendert §34d nicht mehr als �34d Mojibake auf Windows-Konsolen mit cp1252-Default. CLI rekonfiguriert sys.stdout/sys.stderr auf UTF-8 beim Entry.

Compatibility

Pures Bug-Fix-Release, keine Breaking-Changes vs v0.94.0. Apache 2.0 unverändert. Migration einfach:

pip install --upgrade cognithor==0.94.1

Verweise

• Audit-Report: docs/superpowers/reports/2026-04-26-v094-dry-run-audit.md
• CHANGELOG: CHANGELOG.md
• Hotfix PR: #152
• Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.94.0...v0.94.1

Cognithor v0.94.0 — AutoGen Strategy Adoption

Released: 2026-04-25

Microsoft hat AutoGen offiziell in den Maintenance-Modus versetzt und verweist auf das Microsoft Agent Framework (MAF) 1.0 als Nachfolger — mit einem harten Migrations-Pfad von conversation-zentriert zu graph-basiert. Cognithor v0.94.0 positioniert sich in diesem Window als lock-in-freie EU-Alternative und AutoGen-Migrations-Onramp, ohne unsere PGE-Trinity-Architektur zu verbiegen.

Highlights

cognithor.compat.autogen — Source-Compat Shim für AutoGen-AgentChat 0.7.5

Search-and-replace-Migration vom AutoGen-AgentChat-Code auf Cognithor, ohne PGE-Trinity zu kompromittieren.

- from autogen_agentchat.agents import AssistantAgent
- from autogen_agentchat.teams import RoundRobinGroupChat
+ from cognithor.compat.autogen import AssistantAgent, RoundRobinGroupChat

• 17-Feld-Signatur-Parität für AssistantAgent (verifiziert gegen echtes autogen-agentchat==0.7.5)
• 1-Shot-Pfad delegiert an cognithor.crew.Crew(...).kickoff_async()
• Multi-Round via eigenem _RoundRobinAdapter (~250 LOC)
• Komponierbare Terminations: MaxMessageTermination, TextMentionTermination, __and__ / __or__
• OpenAIChatCompletionClient Wrapper über Cognithor's Model-Router (16 Provider)
• Bewusst NICHT unterstützt (Sicherheitsmodell): SelectorGroupChat, Swarm, MagenticOneGroupChat — siehe ADR 0001.
• Migration-Guide: src/cognithor/compat/autogen/README.md
• Install: pip install cognithor[autogen] (zieht autogen-agentchat==0.7.5 für Signatur-Tests)

cognithor_bench/ — Reproduzierbares Multi-Agent-Benchmark-Scaffold

In-Monorepo-Submodul, eigenes pyproject.toml, console-script cognithor-bench.

cognithor-bench run scenarios/smoke_test.jsonl
cognithor-bench tabulate results/
cognithor-bench run scenarios/foo.jsonl --adapter autogen   # opt-in

• Adapter Protocol mit Cognithor-Default + opt-in AutoGen-Adapter (lazy ImportError-safe)
• JSONL-Scenario-Format, --repeat / --subsample / --seed / --native (default) / --docker (post-v0.94.0)
• Markdown-Reporter mit Pass-Rate-Aggregation
• Bundled smoke_test.jsonl (3 trivial DACH-Szenarien für CI)
• 94.79% Test-Coverage auf das Submodul

examples/insurance-agent-pack/ — DACH-Versicherungs-Pre-Beratung Reference Pack

Standalone-installierbares Apache-2.0-Beispielpaket — KEINE cognithor.packs-Registrierung (das ist für private Commerce-Packs reserviert).

pip install ./examples/insurance-agent-pack
insurance-agent-pack run --interview

• 4 Agenten: NeedsAssessor, PolicyAnalyst (NEU: PDF-Tool-Use via PyMuPDF), ComplianceGatekeeper (NEU: sichtbarer PGE-Demo mit regelbasiertem Klassifikator), ReportGenerator
• Synthetische Knowledge-Seeds (PKV/GGF/bAV/BU)
• §34d-NEUTRAL: blockt Rechtsberatungs-Anfragen + konkrete Produkt-Empfehlungs-Demands
• DISCLAIMER persönlich verfasst — siehe docs/DISCLAIMER.md
• Konzeptionell verwandt mit dem versicherungs-vergleich Template aus v0.93.0; fokussiert auf Tool-Use + Visible-PGE-Demo

ADR 0001 — PGE Trinity vs Group Chat

Erste Architecture Decision Record im Repo. Dokumentiert, warum Cognithor bewusst keine SelectorGroupChat / Swarm Patterns adoptiert (LLM als Sicherheitsgrenze nicht tragfähig; Gatekeeper-Regeln sind inspectable und testbar).

docs/adr/0001-pge-trinity-vs-group-chat.md

Competitive-Analysis-Dokumente

Vergleich mit AutoGen, Microsoft Agent Framework, LangGraph, CrewAI — sachlich, mit Quellen-Zitaten.

docs/competitive-analysis/

Install

pip install cognithor==0.94.0                       # core
pip install "cognithor[autogen]==0.94.0"            # mit AutoGen-Compat-Tests
pip install "cognithor[full]==0.94.0"               # alles inkl. optionaler Features

Platform-Artefakte unten (Windows Installer, Linux .deb, Android APK, iOS IPA, Flutter Web).

Compatibility

• Keine Breaking-Changes vs v0.93.0. Reine Erweiterung.
• cognithor.crew (v0.93.0) unverändert.
• PGE-Trinity unverändert.
• Apache 2.0 unverändert. NOTICE bekommt MIT-Attribution für AutoGen-Inspiration (kein verbatim AutoGen-Code).

Verweise

• AutoGen-Migration-Guide: src/cognithor/compat/autogen/README.md
• CHANGELOG: CHANGELOG.md
• Spec: docs/superpowers/specs/2026-04-25-cognithor-autogen-strategy-design.md
• Implementation Plan: docs/superpowers/plans/2026-04-25-cognithor-autogen-strategy.md
• Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.93.0...v0.94.0

What's Changed

Features

• feat(crew): Feature 2 Quickstart Documentation (v0.93.0) (#145) (dff283a8)
• feat(crew): Feature 7 Integrations + sevDesk DACH connector (#144) (b6e92304)
• feat(crew): Feature 3 CLI + Templates (#143) (30d26dfb)
• feat(crew): Feature 4 Guardrails (#142) (d32d3807)
• feat(crew): Feature 1 Crew-Layer Core (#141) (45abe90b)

Fixes

• fix(build): remove duplicate force-include for crew templates (8d9c8e8f)

Other

• chore(release): bump pyproject.toml to 0.93.0 (was missed in initial bump) (d22be571)
• chore(release): v0.93.0 — Crew-Layer + Guardrails + CLI + Integrations + Quickstart (db41cde5)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.92.7...v0.93.0

What's Changed

Features

• feat(vllm-video): native video input via Qwen3.6-27B + 5 review rounds (#140) (7eadc600)
• feat(flutter): implement URL-einfügen dialog for video URL paste (d2ede0b9)
• feat(installer): bundle LGPL ffmpeg for Windows + CI verification step (fc4ce15a)
• feat(flutter): chat bubble renders video kind with thumbnail + long-video banner (a84f60af)
• feat(flutter): ChatProvider.sendVideo + URL-paste video detection (b7ce496d)
• feat(flutter): chat-input paperclip becomes PopupMenuButton with Video entry (02927090)
• feat(gateway): video attachment classification + MediaServer/Cleanup lifecycle (55a9dcab)
• feat(api): /api/media/upload + /api/media/thumb endpoints (363a14f9)
• feat(llm): UnifiedLLMClient hard-errors on video when vLLM DEGRADED (6fbf130f)
• feat(planner): route video_attachment to vision_model_detail (97db0411)
• feat(models): WorkingMemory.video_attachment per-turn field (3f95e883)
• feat(vllm): orchestrator threads spike flags + media-url env into docker run (082f50cc)
• feat(vllm): VLLMBackend.chat(video=) with mm_processor_kwargs payload (96b3f2c4)
• feat(video): VideoCleanupWorker with session registry + mtime TTL sweep (5aaf4605)
• feat(media): MediaUploadServer FastAPI app + ephemeral-port lifecycle (5791048d)
• feat(media): MediaUploadServer save_upload with quota + LRU eviction (ee7c76d9)
• feat(config): video + launcher fields on VLLMConfig (97f6276e)
• feat(llm): MediaUploadError hierarchy for video upload failures (f5c03845)
• feat(video): ffprobe-based resolve_sampling with adaptive + override modes (b92244b6)
• feat(video): VideoSampling dataclass + duration bucket rules (b65108e7)

Fixes

• fix(test): browser-agent not_started_error on Python 3.12+ (00496ba7)
• fix(installer): also flatten iss ffmpeg paths + ruff wrap (f6a9ef00)
• fix(installer): flatten ffmpeg archive to fixed bin/ path (5098df4e)
• fix(flutter): promote KeyboardListener FocusNode to state field (3c218c84)
• fix(vllm): _attach_images_to_last_user preserves list content symmetrically (4ef4c36c)
• fix(api): surface quota-exceeded recovery hint to client (338e93bd)
• fix(vllm): thread live VLLMConfig through orchestrator construction (fff7cb62)
• fix(vllm): log docker run command + failure details for observability (e07d48d7)
• fix(flutter): dispose URL-dialog TextEditingController to prevent leak (bbf36202)
• fix(video): VideoCleanupWorker.start() is idempotent (9c790a91)
• fix(api): share VLLMOrchestrator between Gateway and backends_api endpoints (aad95066)
• fix(flutter): defer _isUploading reset to post-frame callback (e42828e1)
• fix(vllm): VLLMBackend.chat_stream accepts + forwards video kwarg (558c92ca)
• fix(media): threading.Lock around save_upload to prevent quota TOCTOU (c9726288)
• fix(media): harden path-traversal guards via resolved is_relative_to check (2cc64ccb)
• fix(gateway): wire VideoCleanupWorker.on_session_close into stale-session sweep (74bfe548)
• fix(flutter): strip query + fragment from URL-paste filename derivation (e51f3e97)
• fix(gateway): offload _build_video_attachment to thread pool (2db74443)
• fix(api): media upload handler offloads ffprobe + ffmpeg to thread pool (05a033bd)
• fix(flutter): block send button while video upload is in progress (b8b38c96)
• fix(vllm): _attach_video_to_last_user preserves pre-existing list content (90dcfeaa)
• fix(config): default vLLM image to cu130-nightly per spike (e082d6ee)

Other

• chore(release): bump Flutter kFrontendVersion to 0.92.7 (f38408eb)
• chore(release): sync version + flutter pubspec version to 0.92.7 (45f0cb28)
• chore(release): bump to 0.92.7 for video-input feature (a2cecb2d)
• docs(video): document cu130-nightly rolling-tag risk + mitigation (5bba52a7)
• style: clear all repo-wide ruff errors + format (5cab95aa)
• chore(changelog): document video-input via vLLM feature (ba108cda)
• docs(video): manual smoke-test recipes 9-15 for video input (8defc553)
• docs(video): user guide video-input section with troubleshooting + config (b3381ae7)
• test(vllm): fake server round-trip for video_url + mm_processor_kwargs (ddfe5f13)
• refactor(video): pin resolve_sampling mocks to module-scoped subprocess (3323f1a3)
• refactor(video): tighten VideoSampling docstring + error-message test (ca7a38e1)
• docs(spike): Day-1 video-input spike complete, plan/spec amended (cc1e21c1)
• docs(plan): video-input via vLLM implementation plan (0a6d6c6f)
• docs(spec): fix inconsistencies introduced by single-video refactor (b70bf3b0)
• docs(spec): revise video-input spec with honest risk review (0afac632)
• docs(spec): video-input via vLLM for Qwen3.6-27B (5c5ae89f)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.92.6...v0.92.7

What's Changed

Features

• feat(vllm): opt-in vLLM backend with Flutter-driven lifecycle (#137) (3185d224)

Fixes

• fix(config): reload planner prompts on language change (#139) (adf44e69)
• feat(vllm): opt-in vLLM backend with Flutter-driven lifecycle (#137) (3185d224)

Other

• chore(release): v0.92.6 (23302e6d)
• chore(release): rename launcher asset to CognithorLauncher-.exe (d879a214)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.92.5...v0.92.6

What's Changed

Features

• feat(qwen3.6): add VLM as first-class option end-to-end (#132) (d0aca3e4)

Fixes

• fix(flutter): bump frontend version to 0.92.4 + CI sync-check (#135) (6a162dfc)
• fix(config): tolerate legacy/unknown keys in user config.yaml (#134) (61074cf7)

Other

• chore(release): v0.92.5 (0d9ef840)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.92.4...v0.92.5

What's Changed

Features

• feat(security): local PII redactor for outbound LLM messages (closes #122) (#130) (217f76f9)
• feat(security): local PII redactor for outbound LLM messages (1d3028c1)

Fixes

Other

• chore(release): v0.92.4 (954d2c43)
• ci(publish): use workflow_dispatch tag input for version validation (a18d2afb)
• ci(publish): add workflow_dispatch trigger for manual PyPI publish (6dac6510)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.92.3...v0.92.4

What's Changed

Features

Fixes

• fix(flutter): update widget_test.dart JarvisApp → CognithorApp (6d902288)
• fix(install): platform-aware sqlcipher dev package name (closes #116) (#127) (cd947d65)
• fix(install): platform-aware sqlcipher dev package name (307c7961)

Other

• chore(release): v0.92.3 (c5325997)
• docs: deep rebrand sweep (DATABASE, DEVELOPER, IDENTITY, CONFIG_REFERENCE) (#129) (e20a0c48)
• docs: deep rebrand sweep over DATABASE/DEVELOPER/IDENTITY/CONFIG_REFERENCE (8da427cc)
• chore(flutter): update ARB + regenerate l10n for Cognithor branding (55b75e5f)
• refactor(flutter): rename JarvisTheme/JarvisApp/Jarvis* widgets to Cognithor* (a65fc119)
• refactor(flutter): rename jarvis_*.dart files to cognithor_*.dart (d5c49c7c)
• docs: purge stale Jarvis references + add Observer Audit Layer mentions (#126) (6733ef13)
• refactor(config): remove backward-compat JarvisConfig/jarvis_home aliases (#125) (e0183759)
• style(config): ruff format after alias removal (0ff64b29)
• docs: purge stale Jarvis references + add Observer Audit Layer mentions (4a416a20)
• refactor(config): remove backward-compat JarvisConfig/jarvis_home aliases (d4ac6b99)
• docs(skills): port SKILL.md improvements from beta to main (#124) (d08fdde0)
• docs(skills): port SKILL.md improvements from beta branch (2e426fef)

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.27.7-beta...v0.92.3

Cognithor v0.27.7-beta

Beta Release — automatically generated from the beta branch.

Changes since v0.27.6-beta

Features

• feat: improve skill scores for cognithor (#117) (e41019ce)
• feat: improve skill scores for cognithor (cebfa1cc)

Fixes

• fix(tests): flaky test_full_lifecycle (Windows timer resolution) (1d5ca91e)
• fix: DDG cache TTL=0 not expiring on Windows (3addadb1)

Other

• chore(skills): complete Jarvis → Cognithor rebrand in SKILL.md files (#123) (c6c4b994)
• chore(skills): complete Jarvis → Cognithor rebrand in SKILL.md files (a8c64260)

---

This is a pre-release. It may contain bugs or incomplete features. For stable releases, use the main branch.

What's Changed

Fixes

• fix(ci): split pip install into own step so args reach python (12e48e6)
• fix(ci): add shell: bash to installer step + retry logic for downloads (38df59d)
• fix(ci): use CLI args instead of env vars for installer skip flags (dbf2bc7)
• fix(ci): use shell: bash for installer step so env vars propagate (03b5cf8)
• fix(ci): set SKIP_FLUTTER_BUILD via PowerShell env syntax (3fb98da)
• fix(ci): use --no-pub to prevent registrant regeneration (cd2d6da)

Other

• debug: print sys.argv in installer to trace missing --skip-flutter (63cc608)

Note — this release was published automatically when the v0.92.2 tag was pushed. Version bump in pyproject.toml / __init__.py and matching changelog entry land in the follow-up commits. Installer artifacts attached to this release are correct.

---

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.92.1...v0.92.2

Deep Research Analyst (second paid Agent Pack) + pack architecture hardening.

Highlights

• Deep Research Analyst (EUR 65/159) — multi-hop web research with citations, source triangulation, Markdown/PDF export, SQLite history, Flutter Research tab
• Pack Architecture — fixed relative imports, added /api/v1/packs/loaded, PacksProvider for Flutter tab gating
• cognithor.ai live on Vercel with auto-deploy, /promise page, 7 packs in catalog

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.92.0...v0.92.1

No release notes provided.

Highlights

• i18n Fix (#109) — Language selection now applied to agents, prompts, and voice STT
• Approval Flow — Real WebSocket approval with 30min timeout (no more auto-approve stub)
• Conversation Tree — session-aware, persists conversation_id in session store
• Live Logs Tab — real-time monitoring in MonitoringScreen
• Invisible Backend — pythonw.exe, no console window
• ConnectionGuard — overlay when backend unreachable
• DB Migration — auto-copies ~/.jarvis/ data to ~/.cognithor/

All 6 GitHub Issues Fixed

• #104 first_run.py KeyError
• #105 model_post_init auto-swap override
• #106 upgrade corruption
• #107 connection-lost overlay
• #108 invisible backend + live logs
• #109 i18n language selection

Infrastructure

• Kanban sqlcipher cursor fix
• Evolution goal progress endpoint now reads from GoalManager
• Duplicate plan prevention (cleaned up 323 duplicates)
• Planner retries on action refusal

See CHANGELOG.md for full details.

Highlights

• Package rename complete: jarvis → cognithor (both CLI commands preserved)
• Release pipeline hardened: Pre-release smoke test validates built wheel on 2 OS before PyPI upload
• All 5 GitHub issues fixed: #104, #105, #106, #107, #108

Bug Fixes

• #104 first_run.py KeyError — config structure fixed from nested to flat
• #105 Model switching — model_post_init no longer overrides explicit llm_backend_type when API keys exist in keyring
• #106 Upgrade corruption — auto-migrate ~/.jarvis/ → ~/.cognithor/, keyring fallback, corrupt DB recovery
• #107 Frontend disconnect — ConnectionGuard overlay blocks UI when backend unreachable, 15s health polling
• #108 Invisible backend — pythonw.exe in --ui mode, Live Logs tab in MonitoringScreen

Infrastructure

• Environment variable overrides: JARVIS_* prefix supported for backward compatibility
• Version consistency check in CI (pyproject.toml vs init.py)
• scripts/prepare_release.py for local pre-release validation
• scripts/verify_readme_claims.py for README claims audit
• README: Pre-v1.0 Beta disclaimer, corrected numeric claims

New Tests

29 new tests: release smoke (4), env overrides (18), security regression (2), migration (5)

Full Changelog

See CHANGELOG.md for complete details.

Two Major Memory Subsystems

Both features from the Retrieval Extensions prompt are now complete.

CAG Layer — KV-Cache Preloading

• Deterministic prefix generation ensures LLM backends (Ollama, llama.cpp, vLLM) automatically reuse their internal KV-cache for Core Memory
• Content normalizer with BOM strip, whitespace collapse, sha256 hashing
• Cache store with atomic writes, JSON + txt sidecar
• Selector determines which memory tiers qualify (Core always, Semantic conditional)
• Metrics track prefix hit/miss rate
• PrefixCacheBuilder for any HTTP-based backend
• NativeLlamaCppBuilder for explicit state save/load (llama-cpp-python)
• 2 integration hooks: Gateway prepares prefix, Planner uses it instead of raw core_memory_text
• Config: opt-in (cag.enabled=false), auto-rebuild on change
• 71 tests

Hierarchical Document Reasoning — 4th Retrieval Channel

• 5 document parsers: Markdown, PDF, DOCX, HTML, Plaintext (German legal heuristics)
• Tree builder: hierarchy construction, content splitting, bottom-up LLM summaries
• SQLite tree store with CASCADE delete
• LLM node selector: top-down traversal with JSON parsing + regex fallback
• 4-channel score fusion: auto-normalizing weights (vector + BM25 + graph + hierarchical)
• 4 new MemoryManager methods, HierarchicalConfig
• 136 tests, 14 edge cases

Combined: 207 new tests, ~35 new files

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.88.0...v0.89.0

New Feature: 4th Retrieval Channel

Hierarchical Document Reasoning — index long structured documents (contracts, manuals, legal texts) as hierarchical trees and let the LLM navigate to relevant sections via top-down semantic reasoning. No vector embeddings needed.

Components

• 5 Document Parsers: Markdown (ATX+Setext), PDF (TOC+font-size), DOCX (heading styles), HTML (h1-h6+filtering), Plaintext (German legal heuristics)
• Tree Builder: Flat sections -> hierarchy -> content splitting -> bottom-up LLM summaries
• SQLite Tree Store: Same DB as existing indexer, CASCADE delete, transactional writes
• LLM Node Selector: Top-down traversal with JSON parsing + regex fallback
• 4-Channel Score Fusion: Auto-normalizing weights (vector + BM25 + graph + hierarchical)

Integration

• 4 new MemoryManager methods: index/remove/list/reindex hierarchical documents
• HierarchicalConfig in config.yaml (enabled by default, score_weight=0.25)
• Existing 3-channel search unchanged when no hierarchical docs exist

Quality

• 136 new tests (models, 5 parsers, store, builder, selector, retrieval, integration, 14 edge cases)
• All existing memory tests pass unchanged
• 14 edge cases covered (no headings, heading jumps, corrupt files, concurrent indexing, encoding issues, etc.)

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.87.2...v0.88.0

v0.87.2 — Zero Silent Failures + Install Overhaul

Silent Exception Audit

• 79 except Exception: pass blocks replaced with log.debug(..., exc_info=True) across 31 files
• Zero silent exception swallowing remaining in codebase

install.bat Overhaul (#100)

• PATH append (not overwrite), exact model matching, VRAM validation
• Network retry (3 attempts), disk space check, AMD GPU detection
• --silent mode, install log, COGNITHOR_OLLAMA_URL env var
• All 13 audit points addressed

install.sh Overhaul (#101)

• set -eo pipefail (removed -u), retry() for all network ops
• Multi-distro: apt-get, dnf, pacman, brew
• Portable: no GNU timeout, nproc/sysctl fallback, POSIX sort
• --silent mode, disk space check, install log
• macOS compatible (sysctl, no systemd assumption)

Bugfixes

• WindowsPath concatenation crash in encrypted_db.py (#99)
• deploy/cognithor.service for systemd (#CI fix)
• PREREQUISITES.md updated to cognithor
• Remaining python -m jarvis in tests fixed
• Graceful shutdown log event restored

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.87.1...v0.87.2

New Features

• cognithor config — Interactive CLI config TUI with dynamic model discovery (#95)
  • cognithor config list / get KEY / set KEY VALUE
  • Live model lists from Ollama, OpenAI, Gemini, LM Studio
  • Cached fallback registry (CI-updated weekly)
• contrib/ directory for community modules (Item 4 refactor)

Fixes

• Graceful CLI shutdown — no more CancelledError traceback on /quit (#96)
• 7 remaining from jarvis inline imports fixed (CI was failing)
• Ruff formatting for 17 files post-rename
• Version-pinned upgrade example in README (#94)
• Installer/bootstrap scripts updated for cognithor rename

Upgrade

pip install cognithor[all]==0.87.1

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.87.0...v0.87.1

BREAKING CHANGE: Package renamed from jarvis to cognithor

All imports change: from jarvis.X -> from cognithor.X

Migration

# Before
from jarvis.core.planner import Planner

# After
from cognithor.core.planner import Planner

Scope

• 1,265 files changed
• 4,474 Python import replacements
• 358 env var renames (JARVIS_ -> COGNITHOR_, both still work)
• 339 path renames (~/.jarvis -> ~/.cognithor, fallback to old path)
• 599 Flutter package renames (jarvis_ui -> cognithor_ui)

Backwards Compatibility

• jarvis CLI command still works (alias)
• Both COGNITHOR_* and JARVIS_* env vars accepted
• ~/.cognithor preferred, ~/.jarvis fallback
• "Jarvis" personality name unchanged

Also included (from v0.86.4)

• AST-based Python/Shell security analysis
• _safe_call subsystem failure tracking
• README accuracy overhaul
• Evolution Engine research optimization

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.86.4...v0.87.0

v0.86.4 — Security AST, _safe_call, README overhaul, Evolution fixes

Security (Item 5 — Reddit Refactor)

• Python AST Guard: Detects getattr/setattr bypass, chr() concat, __import__, exec/eval, dunder escape (40 tests)
• Shell AST Guard: bashlex parsing with regex fallback, 60+ blocked commands, substitution/chaining detection (48 tests)
• Gatekeeper: AST Layer 1 + Regex Layer 2 dual-defense for both Python and Shell

Observability (Item 3 — Reddit Refactor)

• _safe_call: All 22 subsystem inits in advanced.py tracked with failure registry
• Health endpoint: Shows degraded status + failed subsystem names
• 11 unit tests for safe_call module

README Accuracy (Item 2 — Reddit Refactor)

• Live CI + PyPI badges replace static ones
• Sandbox claims corrected (platform-adaptive, not "4-level")
• Tool count 120+, installation extras documented
• Development Status downgraded to Beta

Evolution Engine

• LLM-generated search queries instead of static templates
• search_and_read instead of web_search for source discovery
• Coverage thresholds increased, proportional scoring (1% granularity)
• Back-to-back cycles (5s pause instead of 5min cooldown)
• Journal API endpoint, Evolution config page in Flutter UI
• Learning goals persist across config saves

Reddit Lead Hunter

• Hard routing bypasses Planner for reddit_scan
• Product parameter added to reddit_scan
• LLM scoring uses qwen3:32b (was missing model parameter)

Kanban

• pending_review status for human-in-the-loop
• Scheduled tasks panel with cron job toggle
• Drag-and-drop reorder within columns

Fixes

• Bootstrap PEP 563 annotation resolution
• Config save: social/kanban/vault sections now editable
• ModelsConfig field names corrected
• DB corruption from force-kills handled
• Auto-upgrade syncs code + UI + procedures

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.86.3...v0.86.4

v0.86.3 — Config fixes, Evolution goals, Reddit scoring

Fixes

• Reddit scanner LLM call: Was missing model parameter — all 500 posts scored 0. Now uses qwen3:32b
• ModelsConfig crash: models.default.name doesn't exist, changed to models.planner.name
• Config save: social, kanban, vault sections were not in editable sections list — toggles didn't persist
• Evolution Goals API: Added POST (create single goal) and PATCH (update by ID) endpoints. GET now returns structured objects

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.86.2...v0.86.3

v0.86.2

Reddit Lead Hunter

• Hard routing: Gateway bypasses Planner, directly calls reddit_scan
• Product parameter added to reddit_scan
• JSON Schema fix for all 6 tools

Kanban Board

• New pending_review status (human-in-the-loop)
• Scheduled Tasks panel with pause/resume
• Drag-and-drop reorder within columns

Fixes

• Bootstrap endpoint PEP 563 fix
• Identity tools registered (4 new)
• Auto-upgrade syncs code + UI + procedures
• requirements.txt synced
• F821 active_skill scope fix

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.86.1...v0.86.2

What's Changed

Bugfixes

• Installer runtime error — Fixed Inno Setup {userprofile} constant (now {%USERPROFILE}) (#90)
• Logo not centered — Centered logo in collapsed sidebar rail (#91)
• Stale version banner — Updated hardcoded v0.80.0 to v0.86.x in start_cognithor.bat (#92)
• Reddit Skill unavailable — RedditLeadService now always initializes so MCP tools are registered even without product config
• Ruff formatting — Fixed lint errors in 4 files

New Features

• Kanban Scheduled Tasks — New "Scheduled" tab showing cron jobs with live next_run countdown, pause/resume toggle
• Drag-and-drop reorder — Reorder tasks within Kanban columns (sort_order persisted)
• Auto-upgrade from source — Installer now detects newer source trees and auto-upgrades on launch
• Bootstrap download progress — Verbose progress bar during Flutter Web UI download (#93)
• Morning Briefing expansion — Enhanced Day Planner procedure
• System architecture docs — Complete ASCII schematic of entire system
• CI workflow_dispatch — Flutter Web + Windows installer builds can now be triggered manually

API Changes

• — Pause/resume individual cron jobs
• — Jobs with next_run times
• — Batch sort_order updates

Full Changelog: https://github.com/Alex8791-cyber/cognithor/compare/v0.86.0...v0.86.1

Cognithor v0.86.0 -- Lead Wizard, Playwright Auto-Post, Deep Audit Fixes

Lead Wizard (Flutter v2)

• Full-screen sequential lead processor with progress bar
• Keyboard shortcuts: A=Archive, S=Skip, R=Reply, I=Improve (disabled inside TextField)
• Integrated RefinePanel: LLM improves draft with hint + 3 style variants
• TemplatePicker: select from saved reply templates
• Summary screen at completion: replied/skipped/archived counts
• "Process Queue" button on Leads tab (filters new leads, sorts by score)

Playwright Auto-Post

• Real browser automation with Reddit session cookie persistence
• Multiple selector strategies for Reddit UI changes
• Graceful clipboard fallback on any failure
• BrowserAgent wired in gateway post-init

Performance + Feedback in Detail Sheet

• PerformanceBadge: color-coded engagement score for replied leads
• FeedbackDialog: 5 tags (converted/conversation/ignored/negative/deleted)
• RefinePanel integrated into LeadDetailSheet

Deep Audit Fixes (2 Critical + 5 Important)

1. CRITICAL: ReplyRefiner LLM never wired -- refine was completely broken
2. CRITICAL: PerformanceTracker author_replied always False
3. Wizard keyboard shortcuts fired inside TextField
4. R=Reply shortcut documented but not implemented
5. Flutter scanInterval default 60 != backend 30
6. Skill file tools_required expanded to all 6 MCP tools

Pre-Release Verification

• Ruff lint: 0 errors
• Flutter analyze: 0 issues
• Python tests: 13,117 passed, 0 failed
• Flutter web build: successful

Cognithor v0.85.0 -- Intelligence Layer: Self-Improving Reddit Lead Hunter

LLM Reply Refinement

• refine() improves drafts with style profile + few-shot context + user hints
• generate_variants() creates 3 distinct reply styles (technical/casual/question)
• MCP tool: reddit_refine, REST: POST /leads/{id}/refine

Performance Tracking

• Automatic re-scanning every 6h: finds our comment, counts upvotes + replies
• Manual feedback tags: converted, conversation, deleted, ignored, negative
• Engagement score 0-100: upvotes3 + replies5 + author_replied10 + converted20

Subreddit Discovery

• LLM generates 20 candidates from product description
• Reddit JSON validation: subscribers, activity, sample posts
• Ranked by relevance x activity x log(subscribers), top 10 returned

Reply Templates

• Auto-save high-performing replies (engagement > 85) as templates
• Template variables: {product_name}, {subreddit}, {author}
• CRUD via MCP + REST

Feedback Learning Loop

• Weekly cron: LLM analyzes top-5 vs bottom-5 replies per subreddit
• Generates SubredditStyleProfile: what works, what fails, optimal tone/length
• Few-shot injection: top 3 replies as examples in every new draft
• Auto-adjusts reply_tone based on engagement trends

Scanner Intelligence

• REPLY_PROMPT now includes style context + few-shot examples
• Drafts improve with every learning cycle

New Infrastructure

• 5 new modules: refiner.py, tracker.py, discovery.py, templates.py, learner.py
• 3 new SQLite tables: reply_performance, reply_templates, subreddit_profiles
• 6 MCP tools total (3 new: refine, discover, templates)
• 14 REST endpoints total (8 new)
• 3 cron jobs (scan 30min, tracker 6h, learner weekly)

Stats

• 13,117 tests passing (34 new social tests)
• 0 ruff lint errors / 0 flutter analyze issues