What’s New

In progress (not yet versioned): /silver:multi-ai-task and /silver:deep-research-multi-ai — multi-ai-task-v2 spine, v4 manifests, capability spike, dual HTML reports.

Workflows

• feat(skills): APO catalog registers WF-SILVER-MULTI-AI-TASK and WF-SILVER-DEEP-RESEARCH-MULTI-AI
• docs(site): multi-ai-task and deep-research-multi-ai help pages

Minor release: five-tool routed compression stack, multi-AI orchestration primitive, and a batch of hook gate-integrity fixes — several gates were failing open or blocking Stop permanently.

Features

• feat(tools): activate five-tool routed stack — graphify, leanctx, context-mode, rtk, agentmemory
• feat(tools): five-tool reconciler and Cursor install integration
• feat(multi-ai): multi-AI primitive, deep-research composer, Excel matrix, and personas
• feat(multi-ai): Codex/OpenCode host adapters for live deep-research dispatch
• feat(clarify): per-plan CLARIFY filenames and planning path helper

Fixes

• fix(hooks): stop instruction-ledger self-enrollment from fenced agent reports
• fix(hooks): stop instruction-ledger and review-fix-ladder gates failing open
• fix(hooks): add a real claude arm to sb_leanctx_hooks_present()
• fix(hooks): site regression gate no longer wedges on a BSD mktemp template
• fix(hooks): stop shell-hook duplication and orphan bash leaks
• fix(hooks): site session gates skip agent-delegation false positives (#244, #245)
• fix(leanctx): resolve the claude host MCP artifact to ~/.claude/settings.json
• fix(leanctx): dedupe the lean-ctx MCP server entry on SB install
• fix(doctor): report D10-routes cross_tool unsupported as a platform-limitation warning instead of a FAIL recommending --fix=host
• fix(tests): repair a BSD-first stat probe that wedged CI
• fix(ci): repair the graphify artifact schema failing validation on main
• fix(agents): implement agent_delegate_append_workdir_evidence for the delegate log floor

Docs

• LeanCTX operating guide, public site reconcile, shared Graphify repository artifacts

Patch release: Cursor desktop /silver commands fix — hyphen-safe command filenames and install identity lock.

Fixes

• fix(cursor): desktop-safe command filenames (silver-<route>.md stubs with silver: frontmatter) — Cursor desktop rejects colon-bearing command filenames
• fix(cursor): install identity lock — VERSION-aligned cache/current/local/marketplace pins prevent registry drift claiming version N against installPath N-1
• fix(tests): install-cursor and host command naming regression coverage

Patch release: Cursor desktop local-plugin materialization — fixes external-symlink security rejection under ~/.cursor/plugins/local/ so /silver commands register after public reinstall.

Fixes

• fix(cursor): materialize commands-only local plugin surface into ~/.cursor/plugins/local/silver-bullet instead of symlinking outside plugins/local — Cursor desktop rejects external symlink targets and silently skips command registration
• fix(plugin-mirror): restore plugins/silver-bullet/scripts symlink to repo scripts/
• fix(tests): install-cursor regression coverage for materialized local plugin directory and commands-only backend caches

Patch release: permanent Cursor desktop/TUI picker fix — local plugin discovery symlink, commands-only backend caches, marketplace SHA alignment at install, D14 commands-only contract, stale custom-agent quarantine, and tagged archive/public install regression coverage.

Fixes

• fix(cursor): commands-only picker install path for desktop and TUI — symlink ~/.cursor/plugins/local/silver-bullet, sync marketplace SHA, prune skill-source from backend caches
• fix(cursor): align manifest SHA with install before verify — registry/primary/backend SHA narrowing, local plugin link as desktop discovery fallback
• fix(cursor): enforce D14 commands-only cache validation; stop malformed /silverui and internal/subagent route drift in TUI
• fix(cursor): stale SB custom-agent cleanup and quarantine during install
• fix(tests): archive install regression tests for tagged public release paths

Patch release: Cursor slash-picker deduplication and custom subagents (install, doctor D21, RFL ladder), stack double-compression recovery hardening, MultAI deep-research port (/silver:compare), Codex user-scope registration, and tri-host post-release verify orchestrator.

Features

• feat(cursor): SB custom subagents installer — install-cursor.sh, doctor D21, silver-init §3.2.3 tri-state consent
• feat(cursor): RFL custom subagents — 6-rung review-fix ladder, hooks, tests, skill
• feat(deep-research): port MultAI solution-landscape and solution-compare — need-profile interview, SCR extraction, weighted matrix SPA, /silver:compare
• feat(agents): silver-agent-opencode and silver-agent-pi delegation surfaces
• feat(release): mandatory tri-host post-release verify orchestrator

Fixes

• fix(cursor): slash picker deduplication — prune backend plugin caches, honor .cursorignore, commands-only registration
• fix(cursor): harden D21 doctor probe and host-agnostic allowlists
• fix(stack): double-compression mutex recovery, SEC-01/SEC-02 hardening, doctor D20
• fix(codex): register Silver Bullet plugin with user scope
• fix(commands): align Cursor and Codex slash routes with silver: colon convention
• fix(deep-research): v2 audit follow-ups and Thermos hardening

Patch release: tri-host RC matrix green (cursor/codex/claude × fresh/upgrade), Cursor hook generator repairs, Claude home-path install alignment, Codex path-rewrite scoping, LeanCTX five-tool stack foundation, and silver-deep-research v2 engine (retroactively documented).

Fixes

• fix(cursor): hook generation, merge dedupe, marketplace gitPath, and /silver slash command materialization
• fix(hooks): restore canonical manifest; harden Claude/Codex install surfaces
• fix(install): scope Codex path-rewrite to plugin surfaces (prevents test fixture corruption)
• fix(rc): unblock cursor RC delegate smoke; Claude first-class RC host
• fix(tests): install-claude/codex/doctor gate fixtures; record-leanctx-usage coverage

Features

• feat(leanctx): Phase 1–2 five-tool stack config, install scripts, and hook gates
• feat(testing): five-tool pre-release gate via agent-cursor harness
• feat(deep-research): v2 engine — search_orchestrator with SB-owned skill_portals, capability_score eval harness, validate_report language gates, adversarial/report-block references, and July 2026 landscape replay benchmark

Patch release: semantic-compress hook test isolation, bundle mirror sync, and Stage 4a site scan completion with pre-release gate site freshness wiring.

Fixes

• fix(test): isolate test-semantic-compress-hook.sh from active repo .planning/ E2E state

Chores

• chore(sync): regenerate agent bundle mirrors after create-release skill doc sync
• docs(site): complete Stage 4a public site scan; wire site freshness into pre-release-gate.sh

Patch release: CI green gate enforcement, scheduler test isolation, completion-audit enterprise policy alignment, and CI-safe E2E harness fixtures.

Fixes

• fix(ci): scheduler test isolation and orchestrator join-gate stdout leak
• fix(ci): completion-audit enterprise policy shadowing + scripts/pre-release-gate.sh CI gate
• fix(ci): regen atomic-flow-index.json, agent preflight dry-run, matrix grep pipefail
• fix(ci): CI-safe mktemp work-dir for minimal-intent/tri-criteria dry-runs

Documentation

• CI policy: never tag/release on red CI (docs/RELEASE.md, AGENTS.md)

Tri-criteria E2E validation harness: live/cold matrix across Cursor, Claude, and Codex; multi-workflow orchestrator scheduler (composer_chain); greenfield 6/6 PASS.

Features

• feat(tri-criteria): live/cold validation harness, tri-host live matrix, greenfield 6/6 PASS
• feat(autonomous-enterprise): agent-claude track and orchestrator scheduler with composer_chain multi-WF advance
• feat(audit): AF/FS compliance script; Audit mode for catalog compliance
• feat(skills): silver-deep-research replaces silver-research/multi-ai

Fixes

• fix(tri-criteria): E2E scorer false fails and runtime composition wiring
• fix(agent-codex/agent-claude): graphify and agentmemory preflight before delegate
• fix(agent-claude): folder trust seeding and plugin cache alias resolution
• fix(catalog): remediate WF/AF audit failures to 157/157 PASS

Verification

• Tri-criteria greenfield matrix — 6/6 PASS (cursor+claude+codex)
• Release version alignment, template parity, render freshness — PASS

Tri-host pre-release smoke gate: mandatory 12/12 isolated host smoke before tag. Agent-codex harness improvements and Cursor CLI smoke timeout fix.

Features

• feat(agent-codex): v0.2.0 harness dir, R9 learnings, pilot PASS

Fixes

• fix(agent-codex): sibling-prompt compliance gaps and unified delegate runtime env
• fix(pre-release): HOST_API_KEY alias, operator key-file fallback, Cursor CLI smoke 300s timeout

Verification

• run-pre-release-host-smoke.sh — 12/12 PASS (tri-host isolated)
• run-all-tests.sh — 5068 passed, 0 failed

Pre-release test hardening: full test suite must be green before patch release. Fixes delegation fallout and version-alignment drift from v0.50.3.

Fixes

• fix(tests): outcomes-check stop-coalesce teardown + jq-less Stop state-dir preservation
• fix(hooks): SEC-04 rm -f -- hardening in delegation and stop-coalesce libs
• fix(validate): host-agnostic allowlist for AF-AGENT-DELEGATE surfaces
• fix(release): config/site/sentinel version alignment; agent skill scenarios + SENTINEL audits

Verification

• run-all-tests.sh — 5349 passed, 0 failed (6/6 suites green)
• run-apo-authoring-compliance.sh — 26/26 PASS

Official Cursor isolation for pre-release smoke and shellcheck fixes in host smoke scripts.

Features

• feat(release): official Cursor isolation for pre-release smoke

Bug Fixes

• fix(release): satisfy shellcheck SC2155 in pre-release host smoke

Enterprise E2E matrix hardening, WBS meta-supervision MVP, and orchestrator read-only parent Bash.

Features

• feat(wbs): cross-cutting WBS meta-supervision MVP

Bug Fixes

• fix(e2e): matrix session isolation, E2E-011 five-tool preflight, /compact recovery, rows 21–22 gate patterns
• fix(orchestrator): allow read-only parent Bash; block SB state mutations only

Documentation

• docs(issues): enterprise E2E issues log with fix commit SHA traceability

Instruction-following reduction closure, silver-doctor coverage, subagent model gates, and enterprise E2E harness hardening.

Features

• feat(if-reduction): complete Waves 1–4 and subagent engagement audit closure
• feat(validation): enterprise E2E claims overlay on 22-row matrix with pre-matrix dry-run gate

Bug Fixes

• fix(if): close remaining test, hook coverage, and documentation gaps
• fix(ladder): enforce Composer 2.5 only for all Cursor subagent rungs
• fix(doctor): host-aware checks — D8 orchestrator rule Cursor-only; D2/D3/D13 use active host plugin paths via runtime-paths.sh
• fix(e2e): fail fast when Claude token gateway credentials are missing

Documentation

• docs(test): silver-doctor catalog, scenario, sentinel manifest coverage, and mocked-host unit tests for Claude/Codex/Cursor doctor paths

Scoped RTK coexistence: full agent RTK when opted in; verbatim harness mode only.

Bug Fixes

• fix(rtk): scope RTK_DISABLED to harness scripts (SB_RTK_COMPAT_MODE=verbatim) and opted-out projects — hook bridge omits disable when recommended_tools.rtk.enabled_by_user is true
• fix(rtk): unwrap rtk / RTK_DISABLED=1 prefixes in completion-audit.sh gate regexes (sb_shell_command_unwrap_rtk)

Tests

• test(rtk): opted-in / opted-out / verbatim modes in test-rtk-compat.sh
• test(hooks): RTK-wrapped git push classification in test-completion-audit.sh

Enterprise E2E live test suite, silver:multi-ai-task model alignment, and silver-prefix skill exposure.

Features

• feat(e2e): opt-in enterprise-e2e-live-test suite (SB_ENTERPRISE_E2E_LIVE=1) with runbook docs/ENTERPRISE-E2E-LIVE-TEST.md and entrypoint scripts/run-enterprise-e2e-live-test.sh
• feat(skills): expose silver:multi-ai-task (renamed from multi-ai-task) with scripts/multi-ai-task-models.py — review-fix-ladder model set @ medium reasoning; OCG plan when OpenCode agents configured

Tests

• Structural enterprise E2E live wiring validation (tests/enterprise-e2e-live/test-enterprise-e2e-live-suite.sh)
• tests/scripts/test-multi-ai-task-models.sh for resolver + Claude bundle prefix

Enterprise E2E hardening patch: friction fixes #2–#13 (hooks, orchestrator Bash in Task workers, RTK compat, session-start branch scope), turn-level TUI watcher + matrix monitor, interactive 22/22 matrix harness, multi-ai-task v2.1–2.6 doc/sentinel fixes, and test isolation for semantic-compress / plugin-surface gates. run-all-tests 4672 passed, 0 failed at release SHA.

Bug Fixes

• fix(hooks): resolve SB frictions #2–#13 for enterprise E2E
• fix(hooks): allow Bash in orchestrator-spawned Task workers
• fix(rtk): export RTK_DISABLED for SB shell entry points
• fix(install-claude): rsync full hooks tree to plugin cache
• fix(hooks): add hookEventName to high-frequency PostToolUse output
• fix(matrix): monitor empty-array crash when all rows complete; retry on API 429

Tests

• test(hooks): accept two-line branch scope file in session-start
• Semantic-compress + Claude plugin-surface test isolation; orchestrator parent-guard coverage

Recommended-tools expansion and global wiring: RTK and Context Mode opt-in gates, Alumnium opt-in, Graphify + agentmemory stack optimizer (synergy_max) across hosts, multi-agent global RTK/Context Mode optimizer, and Context Mode read-deny when enforced. Documentation and site workflow SDLC ordering; enterprise E2E fixture shift; CI/shellcheck and diagnostics fixes. Integrates doc-scheme Graphify exclusion test fix; restores opt-in Graphify gate behavior after mandatory-graphify branch merge.

Enterprise readiness ladder run 2: FLOW-16/17 worker split (DESIGN-HANDOFF.md / DOCUMENT.md), SECURITY.md skill-dispatched template, catalog skill_worker_templates parity, depth-aware review-loop docs, core-rules hash pin enforcement, Cursor GPT-5.5 ladder slug substitutions, and review-loop safety-cap non-convergence guard.

Bug Fixes

• Split silver-handoff (FLOW 16) from silver-ensure-docs (FLOW 17) worker templates; added FLOW-DOCUMENT runtime queue token
• Split security vs silver-secure worker templates with distinct flow-log labels
• Fixed RELEASE.md mandatory skill, ROUTER.md colon invoke form, orchestrator flow-label CSV mappings
• Lazy-load worker template mapper in orchestrator-directive.sh when sourced before orchestrator-parent.sh
• Core-rules injection requires verified hash pin; missing pin injects warning instead of unpinned content

Tests

• Extended orchestrator handoff, workflow CSV, instruction-flow parity, worker-template parity, and review-fix-ladder coverage

Post-release APO runtime alignment, orchestrator queue fixes, and help-center refresh for the v0.48 catalog model.

Bug Fixes

• Fixed silver-devops orchestrator post-exec queue token (FLOW-QUALITY-GATE-PRESHIP instead of stale FLOW-DEVOPS-QUALITY-GATE-PRESHIP)
• Registered silver-deep-research and silver-ensure-docs as flow atoms and extended the silver-deep-research default queue

Tests

• Added orchestrator queue-order and worker-template parity coverage for research and devops flows
• Tightened APO schema test for subagent-only dispatch_mode and composition triple alignment gates
• Added stale 18-flow phrase guard across public docs in check-apo-invariants.py

Documentation

• Refreshed README, help center, OG card, and workflow pages for the canonical AF-* catalog (27 atomic flows)
• Added AGENTS.md rule requiring Composer 2.5 subagents for website and help-center authoring

Atomic-flow APO redesign.

Features

• Added docs/apo-catalog.json as the authoritative Agentic Process Orchestrator catalog
• Replaced the legacy FLOW 1-18 mirror with 27 canonical AF-* atomic flows, 22 catalog-backed workflows, and 85 skill-backed flow steps with local V-loops
• Added generated catalog views: docs/composable-flows-contracts.md, docs/workflow-composition-matrix.md, and docs/generated/atomic-flow-index.json

Tests

• Added blocking APO gates for catalog schema/SOT, atomic-flow deduplication, per-flow and per-step V-loops, composition SOT, evidence/intent models, subagent execution metadata, runtime alignment, worker parity, tool policies, router coverage, and site/doc freshness

Documentation

• Updated silver-bullet.md, templates, README/help/reference/workflow docs, and search metadata for the APO catalog authority and migration from FLOW 1-18 aliases