SwarmAI v1.24.0

SwarmAI v1.24.0

86 commits since v1.23.0 — Eval OS, Radar, Pollinate & Knowledge-lifecycle release.

Eval System

• Eval OS reframed as the agent-era assert successor; recovery is now a first-class 6th dimension
• Load-time fail-loud guard for off-canonical case dimension/category; Guide tab shows 5 canonical dimensions
• Judge fail-fast throwaway client — one hung judge can't blow the wall; throttle-retry preserved
• Admit safety/memory/runtime_health categories to report buckets

Radar — Jobs & Runs

• New Jobs & Runs section restores job + run-status visibility; shows only ACTIVE pipeline runs
• Crash-residue split from decision pauses in NEEDS YOU (and dropped from the roster)
• 🔔 job cards surface last_error — show WHY a job failed; disabled jobs excluded from the queue
• Needs You shows top 3 + "See more" fold; category tags + click-processing state

Pollinate

• HTML-deck build layer — zero-network render of 34 design systems
• Cross-format gate given teeth (3 hard invariants); verb-anchor PPT-to-web conversion detection
• Latin font backfill; html-deck style-select + PPT import fixes

Knowledge Lifecycle

• Agent-directed retire_entry + ddd-retire CLI (the "out" side); tightened decay (60/150)
• _ENTRY_RE widened to parse emoji-prefix prose entries; prose-matching made opt-in (fixed 2 HIGH data-corruption vectors)

Pipeline / Jobs / Health

• Terminal runs never treated as crash-zombies; publish back-fills artifact_id correctly
• Permission: deny streams through continue so the agent responds
• Health-probe: revived 2 silently-dead runtime checks
• Release skill: dropped stale commit-count scope gate — R6/R11 quality gate is the real readiness bar

Full changelog: v1.23.0...v1.24.0

Full Changelog: v1.23.0...v1.24.0

SwarmAI v1.23.0

Reliability + self-evaluation release. 76 commits since v1.22.0 (36 fix · 11 feat · 3 test · 1 perf · docs/chore). Preview: #92

🔁 Session resilience — resume that actually resumes

• DEAD-state resume recovery + _cleanup_dead TOCTOU guard — a send() hitting a dead subprocess self-recovers instead of waiting out the 60s lifecycle loop.
• First-send zombie fix — _last_turn_clean fail-closed guard recycles a soft-interrupt-poisoned warm subprocess before reuse.
• Bounded tool-call-leak recovery — 1st leak injects a correction prefix + resumes; 2nd does a clean cold restart. No self-reinforcing retry loops.
• Serialized subprocess IO — single SDK channel per session, concurrency-safe _force_kill.

🖥️ Streaming spinner — honest & backend-authoritative (OT01)

• Spinner self-heals within ~3s via streaming-aware reconcile cadence.
• Render-freeze on long streaming turns fixed at the render-source layer.
• Terminal backend states clear the spinner immediately; hard-cap backstop clock guarantees a churn-immune force-clear bound.
• Spinner stays honest while the backend flushes post-disconnect.

🧪 Eval subsystem — golden cases you can trust

• gate_refs anti-drift gate — every case reference must resolve to live content or the case fails.
• hard_delete_cases — physical, resurrection-proof golden-case removal.
• validate-corpus sweep closes the resting-corpus drift hole; score-drift hard gate + weekly scheduled cadence.
• Judge-infra coverage-collapse alerting — an errored case no longer masquerades as a clean pass.

🧠 Governance & memory

• AGENT R30 — context files as cognitive organs (verify-before-quote, touch-it-fix-it) + s_persist Step-0 admission gate.
• perf(memory): index titles capped to pointers, recall keywords recovered.
• Credential-expiry banner driven by /health auth.

🔧 Build/deps

• Tauri 2.11.3 to 2.11.4; time pinned to 0.3.51 (cookie 0.18.1 break).

---

No breaking changes to the desktop app contract.

Full Changelog: v1.22.0...v1.23.0

SwarmAI v1.22.0

SwarmAI v1.22.0 — 212 commits since v1.21.0 (89 fix · 61 feat · 18 test · 13 docs · 7 governance · 5 refactor · 4 harden · 3 perf · 3 ddd · 1 revert)

Highlights

Recall & Context (READ path — pure-filesystem rebuild)

• Recall returns to pure-filesystem: vector/Titan leg physically removed from all recall paths; keyword/FTS5 + entry-level BM25 slice
• Entry-level slice+rank replaces whole-section front-truncation — query-matching entries surface regardless of position
• Archives FTS5 + no-truncate read line + [RECALLED] provenance header
• recall#5 opener un-latch + keyword-miss cap; recall#3 degraded counter (loud-on-degradation)
• Token estimator recalibrated to real tokenizer + unified CJK range; self-report drift check

System Prompt / Briefing

• Cognition-admission: CUT 8 feed sections from system-prompt briefing (700→290 tok) — system prompt is cognition real-estate, not a dashboard
• Ephemeral-budget WARN observability (never truncates load-bearing context)
• EVOLUTION.md loads unconditionally on desktop (O2 coding-gate removed)

Memory

• Per-section dormant threshold — MEMORY ages at 45d, others 90d
• Single-source write-path dedup; caps reconciled to live counts
• SessionRecall resurrected (dead import had killed it in prod)

Eval (decoupled system-level subsystem)

• eval_command_guard — structurally bans running eval inside coding pipeline (DEC05/PIT179)
• Runtime canary teeth for spine probes; memory/resume/assembly-floor probes
• Judge self-heals on transient credential error

Security & Hooks (defense outside the agent)

• bash_syntax_guard — prevents headless hang on syntactically-broken commands
• eval_command_guard, persist-skill routing guard, irreversible-external-op gate (C041 fix)

Chat / Frontend

• Bridge store-phase→isStreaming — kills frozen-tab-no-indicator desync
• Live per-tool elapsed timer on pending tool card
• 10s grace on desync guard

Pipeline

• In-band L2 escalation via AskUserQuestion (true human-in-the-loop)

Critical fix

• spawn: pass allowed_tools=[] not None — SDK 0.2.109 was crashing all sessions

Full Changelog: v1.21.0...v1.22.0

SwarmAI v1.21.0

SwarmAI v1.21.0 — 294 commits since v1.20.1 (8 days). The biggest release to date: a full session-lifecycle resilience overhaul, the #1-recurring frontend reconcile bug fixed at its root, a self-evolution closed loop, and a universal pipeline understanding gate.

🧠 Session Lifecycle & Resource Arbitration (R3–R6)

• Unified RecoveryCoordinator — RSS kill, stuck-WAITING, streaming-timeout, tool-hang, and OOM all route through one coordinator (R3b–R3e). RecoveryTransaction closes the _crash_to_cold TOCTOU.
• R6 multi-tab isolation — spawn/resume decoupled from the UX tab ceiling; pure spawn_budget RAM arbitration; independent concurrent-streaming cap; orphan-only eviction makes cross-tab eviction structurally impossible; orphan-session reaper; slot release on tab close.
• recovery_exhausted now surfaces to the user when the self-heal breaker trips.

🖥️ Frontend Streaming & Reconcile (OT01 — #1 recurring bug)

• Render-source split-brain fixed at the root — store is now the sole render authority; stale-prop fallback deleted. 33 prior patches all hit innocent layers; the real defect was the React render-source.
• Foreground-resume reconcile kills the App-Nap stale-store freeze; background-tab re-render gated on isActive (cross-tab render saturation).
• Streaming text rendered as plaintext (kills O(n²) markdown re-parse jank); [DONE] survives mid-stream gen churn; blue active-tab underline; LeftSidebar nav redesign.

🔬 Self-Evolution Closed Loop

• Autonomous cognitive recording + growth report (OT07); meta-cognitive intake + active maintenance escalation.
• M0–M5 DDD knowledge governance: honest per-entry noise metric, physical reclaim of stale noise, keep-class protection, auto-cultivation quality gate, authoritative-zone write-protect, self-evolving eval with Goodhart guard.
• M3 SELF.md resident self-portrait + Recurrence Radar.

⚙️ Pipeline

• Universal Understanding Gate at the EVALUATE→THINK boundary — diagnosis before build.
• REPRO gate + diagnostic-challenge sub-agent; artifact_cli publish --quiet (parse-proof orchestrator output); schema subcommand.

🛡️ Safety & Hardening

• PreToolUse guard default-denies Bash backgrounding; default 120s Bash timeout (15min opt-in); connection-phase reconnect budget extended to ~65s; auto-resend of swallowed question on backend recovery.

📊 Tally

150 fix · 77 feat · 23 test · 24 docs · 5 refactor · 2 governance — all CI-green.

---

🤖 Generated with Claude Code

Full Changelog: v1.20.1...v1.21.0

SwarmAI v1.20.0

SwarmAI v1.20.0

Major release: 142 commits delivering 6 major subsystem upgrades.

Highlights

🧠 Session Unit Extraction — God module split (4167→2802 lines, −33%) via strangler-fig pattern into SessionUnit + StreamingOrchestrator + RetryManager.

📦 MessageStore Single-Writer — Structural protection for chat messages. Phase-gated writes, rAF-gated subscriptions, cross-tab isolation.

🔄 Streaming State Machine — 11-mode explicit reducer replaces 18 implicit booleans. Per-tab on UnifiedTab. 300 matrix tests.

🏥 Session Lifecycle Resilience — HealthSensor (5 triggers) + HealingLoop (max 3 attempts) + unified _arm_recovery_checkpoint. Self-heal re-enabled with user-stop guard.

📊 OS Eval System — 115 cases, 97.4% score. Two-tier: session canary (31 programmatic, 0.07s) + bi-weekly LLM judge (84 cases, ~$0.12). Dashboard live.

🧬 Evolution L1: Governance Mining — Weekly evolution cycle now mines EVOLUTION.md for recurring judgment patterns (correction classes + GC candidates) → governance proposals surfaced in session briefing.

Other Notable Changes

• Cross-tab strict isolation (subscription guard fix)
• Unified recovery primitive (_arm_recovery_checkpoint)
• RSS kill thresholds raised (3.5GB/7GB for 36GB machine)
• Frontend viewport lock for Settings/Eval pages
• Security sanitization for v1.20.0 public release
• Stall/watchdog timeout raised to 90s for cold-start
• Queued message ops routed through MessageStore

Stats

• 142 commits | 6 major subsystems | 9/9 smoke tests pass
• Backend: session_unit 4167→2802 LOC (−33%)
• Frontend: 300+ new streaming state tests
• Eval: 115 golden set cases operational

Full Changelog: v0.0.86...v1.20.0

SwarmAI v1.19.0

What's New

Sub-Agent Progress Observability

• Tiered elapsed-time banners show when sub-agents (pipeline, deep-research) are running longer than expected
• New /api/chat/sessions/streaming-state endpoint for real-time agent status
• Cross-turn bleed protection + concurrent agent tracking

DDD Runtime Read-Side Activation (F1-F5)

• 5 new callers wired into DDD documents for real-time project context during coding

UX Improvements

• R17: Mandatory prompt suggestions after every task completion
• Amazon Narrative Format added to s_docx style standard
• Workspace explorer: race condition fix, path-escape, portability improvements

Quality

• 3 adversarial review rounds: 2 HIGH + 1 MEDIUM bugs caught and fixed pre-ship
• Document metadata convention with lint + auto-update

---

Full Changelog: v1.18.2...v1.19.0

Full Changelog: v1.18.2...v1.19.0

SwarmAI v1.18.2

What's New

Bug Fixes

• fix(chat): message swallowing during queue drain — P1 UX bug where user messages were lost (246dd0d3)
• fix(chat): reconcile race — streaming indicator stuck + layout overflow (9f0821ef)
• fix(layout): viewport-lock — structural fix for page scroll (71f3675f)
• fix(ui): ChatInput spacing — softer separator + tighter action bar (bff7785f, c1ce43c3)
• fix(ui): memoize PDF data decoding — prevent re-load on every render (f24cbd10)

Features

• feat: max_turns 200→400 — prevent deep pipeline sessions from being truncated (a786b9e8)

Pipeline Quality

• Iron Law enforcement on R1 (05a49d75)
• Anti-rationalization tables at BUILD/REVIEW/TEST/DELIVER decision points (171abbd0)
• Spec-compliance enforced as fresh sub-agent (1c29ff48)
• Graphviz decision trees for profile selection + checkpoint logic (06097681)
• Iron Law framing + P5 gate sequence + "zero-context" plan specs (088e856c)

Tests

• Regression guards for drain lifecycle, viewport lock, and streaming fixes

---

Full Changelog: v1.18.1...v1.18.2

Full Changelog: v1.18.1...v1.18.2

SwarmAI v1.18.1

Highlights

🔬 Pipeline Meta-Intelligence (v5)

• 5-layer self-learning system: Observe → Analyze → Adapt → Recover → Specialize
• All 6 profiles now include THINK stage — no profile skips reasoning
• 9 code gates (non-bypassable) — including goal adversarial, REPORT.md mandatory, profile immutability
• Non-skippable stages enforced (goal_cycle, deliver, reflect)
• Mandatory completion summary — no more silent pipeline endings

📝 File Editor Enhancements

• Selection-based review: select code → send feedback with full context
• SSE-driven auto-refresh: real-time reload after agent edits (replaces polling)
• Diff highlight on reload: changed lines visually highlighted after edit
• Comment popover fix: positioning + race condition resolved

⏱️ Streaming UX

• Elapsed time shown during tool execution (no more "is it stuck?" uncertainty)

🛡️ Pipeline Quality Gates

• Goal profile gets full stage coverage + aggressive detection heuristic
• REPORT.md enforced as mandatory deliver output
• Pipeline reports use requirement as title (not batch H1)
• run-update --status completed emits reminder to output summary
• Docs compressed -122 lines (formatting templates → descriptions)

Stats

• 27 commits | 24 files changed | +2,825 / -425 lines
• Focus: Pipeline hardening + File Editor UX + Streaming polish

Full Changelog

v1.18.0...v1.18.1

Full Changelog: v1.18.0...v1.18.1

v1.18.0 — Streaming Spinner-Hang Root Cause Fix

What's New

Streaming Fixes (P0)

• Root cause fix: spinner-hang on background tabs eliminated via readonly isStreaming + Omit<> type enforcement — direct mutation now impossible at compile time
• Atomic clear primitive: all 9 streaming-clear paths now go through setIsStreaming() — no more scattered .isStreaming = false
• L2 reconciliation poll: 15s background check against backend streaming-state endpoint catches any residual stale flags

Workspace & Editor

• Bare-filename resolution via Projects + workspace root walk (Stage 3/4)
• 39 adversarial security tests (symlink traversal, path injection, dot-segment attacks)

Signals

• Weibo trending adapter (China social)
• East Money (东方财富) market adapter (replaced Xueqiu which required auth)

Infrastructure

• streaming-state endpoint added to post-build smoke test
• Co-Authored-By identity fixed to Swarm <swarm@swarmai.dev>

Upgrade Notes

• Frontend-only changes — no backend migration needed
• Requires app relaunch after DMG install to load new JS bundle

---

Full Changelog: v1.17.7...v1.18.0

SwarmAI v1.17.7

Bug Fixes

• Chat text duplication — merge consecutive assistant DB rows on read, eliminating duplicate content (#63)
• Cross-turn response bleed — prevent interrupted sends from leaking into next response (#62)
• DailyActivity checkpoint noise — reduce redundant checkpoint entries (#66)
• Radar Todo noise flooding — dedup gate + confidence filter + auto-purge (#67)
• Scheduler event race — close race condition in emit_event_atomic + pagination merge seam
• Build verification deadlock — drain stdout pipe to prevent 64KB buffer deadlock in verify_build
• Build timeout — increase startup timeout 90s → 180s for PyInstaller cold extraction

Enhancements

• Memory decay scoring — Ebbinghaus + Hebbian decay integration into distillation hook
• Governance audit — P5 proactive propose, R15 any-API verification, stale data cleanup
• Context refinement — convergence metrics, session_unit timeout docs

Stats

• 21 commits, 16 bug fixes, 2 new features
• All fixes verified by adversarial review sub-agents

Full Changelog: v1.17.6...v1.17.7

Full Changelog: v1.17.6...v1.17.7