v1.26.0 — Code Intelligence v3, dual-mode auth, first-run trust [preview]

[xg-gh-25]

Preview — heads-up on what's landing in v1.26.0 (currently on main, not yet tagged). Tracking issue for the release; feedback welcome before it ships.

Since v1.25.0 (2026-07-15), main has accumulated 133 commits (32 code-intel · 11 auth · 6 onboarding · 11 fix · 5 eval · plus tests/docs/chore). The theme of this cycle is understanding + first-run trust: making SwarmAI genuinely understand any codebase, get new users in the door without friction, and never crash on boot.

---

Highlights

🧠 Code Intelligence v3 — AI-Ready-Repo engine (the big one, ~32 commits)

Turn any repo into something an AI agent truly understands — not just greps.

• Tree-sitter AST parsing revived — real structural extraction (functions, classes, edges) across many languages.
• Value-reference edges — module-scope constants, config, and cross-module data consumers now show up in the graph (Python, PHP, Ruby, Swift, Kotlin, C/C++).
• Domain layer + spec-details — code-intel now emits domain grouping and an 8-section rich-understanding spec per domain, with [human] edits preserved on re-generation.
• Confidence-aware edges — EXTRACTED (explicit import) vs INFERRED (call-graph guess), plus a god-node guard and a topology layer (god-nodes / surprising deps / open questions).
• Multi-package / monorepo wired end-to-end.
• Graded incremental re-analysis — file-fingerprint diffing so re-index is fast, and auto-refresh when the tree changes.
• Coverage that can't lie — fail-closed anchor-accounting + mermaid-node-anchoring gates; "100%" now means 100%.

🔐 Auth — Bedrock and direct Anthropic API-key mode

• Bedrock API-key (bearer-token) mode added alongside the existing SSO/ADA path — pick what fits your setup.
• Method-aware remediation — expiry/401 errors now tell you the right fix for your auth method, not a generic banner.
• Durable secret store — API keys survive restart and never leak into config.json or backups (atomic write).
• Runtime Anthropic 401 → actionable Settings message.

🚪 Onboarding — new users get in, first try

• First-launch wizard no longer blocked by a backend-connectivity gate; partial-init new users reach the wizard, not a broken ChatPage.
• Step-1 failure surfaces cleanly (no infinite spinner); desktop "Configure later" escape hatch.
• SSO read-only account display, verify-then-persist, real Step-4 values, double-click guard.

🛡️ Boot hardening — a 6-axis startup audit, all findings closed

• Cold-start ~47s stall removed — dropped an O(db-size) quick_check probe.
• ErrorBoundary no longer renders a blank screen on an explicit fallback={undefined}.
• CredentialBanner boot crash fixed (was mounting useLayout outside its provider).
• Backup restore partial-DB + freshness-lockout data hazard fixed (adversarial-reviewed); restore-step SSE deadlock resolved.

🔎 Recall & Eval — first real numbers

• Mechanical recall@K / MRR benchmark — the first quantified recall score (ddd@5 = 0.71), with an anti-circularity guard that caught 7 fake-hard queries.
• Entry-level BM25 leg + reserved domain slot so buried lessons actually surface.
• Section-name is now a matchable recall signal (MEMORY context_files recall 0.00 → 0.25).
• DDD access-decay hit-log — entries that get used stay alive; unused ones decay.

🧹 Governance & hygiene

• Scheduled per-project DDD semantic-drift review (self-audit job).
• Removed per-call $ budget controls from handler logic (cost is governed centrally, not buried per-call).
• AGENT R30 extended to session-local jargon; assorted design-doc drift sweeps.

---

Notes

• No breaking changes to the desktop app contract.
• Full changelog will accompany the tagged release.

Comment below if there's anything you'd like verified or expanded before v1.26.0 ships.