hooks-- guarded the pre-push hook's--verifyinvocation with a plugin-name check, mirroringvalidate-commit.sh's existing guard. Without this, third-party repos shipping their ownscripts/bump-version.sh(e.g., SWT v2.3.5's in-tree version bumper) had everypackage.jsonversionfield silently overwritten with the literal string"--verify"on every push. The new guard verifies.claude-plugin/plugin.jsonexists and reportsname == "vbw"before invoking--verify; non-VBW repos now skip the check silently. (Issue #635, PR #636)
debug-- added deterministic selected-todo startup, duplicate completed-session repair, and named foreground non-team subagent support for debug/fix routing. (PR #627, PR #633)
uat-- strengthened remediation and verification lifecycle routing: active current UAT now resumes Verify, archive is blocked by non-terminal UAT, round 08/09 parsing is safe, re-verification transitions finalize through the deterministic owner, and accepted summary deviations can be tracked as VBW todos without reopening remediation loops. (PR #596, PR #600, PR #601, PR #606, PR #608, PR #611, PR #612, PR #615, PR #618, PR #620)state-- preserved ROADMAP phase numbering across prefix gaps and compacted machine-managed TODO insertion for accepted UAT deviations and known-issue promotion. (PR #604, PR #632)qa-- optimized known-issue comparisons for large registries and kept active QA remediation dormant after UAT cutover. (PR #622, PR #624)docs-- clarified contributor PR automation gates plus README debug routing, command coverage, and hook-handler counts. (PR #603, PR #623)
guards-- scoped active-agent guard state by session and allowed foreground non-team agentnamelabels without treating them as team membership. (PR #594, PR #633)debug-- honored accepted exception remediation requests, enforced inline UAT AskUserQuestion boundaries, and kept active UAT prompts self-contained. (PR #618, PR #622, PR #631)uat-- made issue capture synthesize durable remediation descriptions, kept product checkpoint modals self-contained, and made status parsing robust under parallel test load. (PR #611, PR #618, PR #625)
gitnexus-- added comprehensive GitNexus skills for CLI operations, debugging, code exploration, impact analysis, and refactoring. (PR #591)agents-- exposed local VBW development agents for GitHub Copilot and added thefix-planner-vbwagent to support plan-audit workflows.
qa-- hardened QA remediation plan validation, UAT remediation orchestration, process-exception evidence gates, known-issue comparisons, and pending-QA explanations. (PR #566, PR #568, PR #574, PR #582, PR #583)state-- fixed linked ROADMAP checklist drift and archive state consistency during UAT/archive reconciliation. (PR #573)dev-- hardened Dev subagent spawns against worktree isolation failures. (PR #579)release-- refreshed GitNexus index metadata and added Serena onboarding memories.
hooks-- scoped agent-spawn-guard named-agent blocking to active VBW workflows and stripped unsupported isolation/cwd fields from hallucinated subagent spawn parameters instead of blocking. (PR #585, PR #590)scout-- allowed read-only Bash live validation for Scout workflows. (PR #588)hooks-- guarded non-PDF Read pages and reverted the sanitizer after runtime evidence showed validation happens before hook execution. (PR #562, PR #563)
guards-- resolved Claude Code internal sidechains at.claude/worktrees/agent-*to host repository root for hook state, blocking accidental writes into unmanaged sidechain copies. (PR #558)state-- added deterministic shell-sideSTATE.mdreconciliation for UAT/remediation artifacts, fixing drift caused by Bash scripts mutating authoritative files outside PostToolUse hooks; also suppressed invalid SubagentStart/SubagentStophookSpecificOutputJSON. (PR #560)hooks-- reverted non-PDFRead.pagessanitizer addition after runtime evidence showed Claude Code validatesRead.pages: ""before firingPreToolUse:Read, making the hook-seam fix insufficient for the observed validation-before-hook path. (PR #562, PR #563)
rtk-- added explicit opt-in RTK management for install, setup, verification, update, uninstall, and runtime smoke proof capture. (PR #540)research-- added/vbw:research Nso numbered todo snapshots can launch context-only Scout research without removing the todo. (PR #543)skills-- nudged agents to read skill-referenced follow-up material afterSkill()activation when relevant to the active task. (PR #534)
config-- replaced the typed pseudo-menu with bounded structured prompts for top-level and core settings choices. (PR #528)debug-- split competing-hypothesis investigation from implementation ownership so report-only investigators tear down before a fresh implementation debugger is spawned. (PR #530)init-- aligned init and discussion prompts with the shared AskUserQuestion interaction contract. (PR #541)testing-- expanded interactive prompt contract coverage to keep structured prompts and freeform handoffs on the intended paths. (PR #539)
debug-- improved sparse/vbw:debugskill selection by enriching sparse todo context and logging runtime skill decisions. (PR #526)execute-- routed true teams only when dependency graphs have real parallel delegate width, preserving serial execution for linear work. (PR #555)rtk-- made install complete setup, added runtime smoke proof recording, and hardened verification when history tails stall or mismatch. (PR #545, PR #547, PR #549, PR #551)skills-- restricted follow-up file resolution to Claude skill roots so lookalike directories are ignored. (PR #538)testing-- hardened verifier pipelines and archive contract matching to avoid pipefail and broken-pipe false failures. (PR #532, PR #537, PR #552)
All notable changes to VBW will be documented in this file.
fix-issue-- Reduce agent verbosity by ~22% through inline bash extraction and pattern consolidation. (PR #523)workflow-- Support user-authored plan reuse and in-place audits in agent and planner. (PR #497)docs-- Revise smoke test guidelines for consumer repos. (ffc7f50c)agents-- Revise plan audit process and add loop procedure for fix-planner. (932e04c9)
debug-- Improve sparse todo skill selection with bounded enrichment and orchestrator/runtime logging. (PR #526)debug-session-- Normalize completed no-verification results toskipped_no_fix_required. (PR #522)debug-- Preserve completed explicit session state instead of silently reopening. (PR #520)skills-- Clarify high-cardinality selection behavior for bounded vs. freeform prompts. (PR #519)todo-- Stabilize list-with-snapshot flake with deterministic age metadata. (PR #518)planning-- Sync transient planning ignores in all tracking modes. (PR #517)debug-- Commit debug completion artifacts at planning-artifact boundary. (PR #516)readme-- Sync config defaults reference for caveman and discussion_mode settings. (PR #515)debug-session-- Rename metadata export tosession_statusand add stranded session recovery. (PR #514)debug-- Clean up already-fixed outcome state. (PR #509)skills-- Make spawned-agent skill activation additive. (PR #505)todo-debug-- Harden todo-to-debug pickup with deterministic normalization. (PR #504)todo-- Honor planning git boundary for backlog writes. (PR #503)todo-- Restore canonical detail storage. (PR #500)todo-- Make pickup immediate and snapshot-safe across commands. (PR #498)compact-- Harden unauthenticated context cache handling. (PR #495)debug-- Normalize debug session lifecycle spacing. (PR #493)
lifecycle-- Add post-archive lifecycle hook for cleanup after plan archival. (PR #481)caveman-- Integrate caveman token compression language for reduced context usage. (PR #456)run-all-- Add execution contract verification script. (PR #454)fix-- Add fix-commit marker and context for inline QA/UAT. (PR #446)commands-- Deprecate standalone /qa and /verify commands in favor of integrated flow. (PR #439)state-consistency-- Add cross-file state consistency checks at lifecycle boundaries. (PR #420)debug-- Make /debug self-contained by absorbing QA and UAT inline. (PR #415)summary-- Add ac_results frontmatter for acceptance criteria reconciliation. (PR #403)research-- Backfill all frontmatter fields during migration. (PR #401)research-- Add standalone research storage with staleness tracking. (PR #397)vibe-- Add skill evaluation at remediation recovery spawn sites. (PR #393)testing-- Add AskUserQuestion maxItems contract test. (PR #389)debug-- Separate active and completed debug sessions into subdirectories. (PR #387)debug-- Add standalone debug session lifecycle with integrated QA/UAT flow. (PR #385)skills-- Add visible skill evaluation reporting in agent output. (PR #384)scripts-- Add todo detail registry for rich context preservation. (PR #375)vibe-- Auto-continue UAT remediation when re-verification finds issues. (PR #366)map-- Add advisory MCP tool guidance for Scout agents. (PR #355)lifecycle-- Auto-promote known issues to STATE.md Todos at lifecycle boundaries. (PR #354)report-- Add LLM-based classification and template matching for issue filing. (PR #345)report-- Add /vbw:report command for self-diagnosis and bug reporting. (PR #336)discussion-engine-- Add codebase-first assumptions mode. (PR #333)discussion-engine-- Add recommendation-led questions. (PR #331)
ghcp-- Add custom agentic development flow for GitHub Copilot users. (PR #476)testing-- Sanitize README path example and add docs contract checks. (PR #473)known-issues-- Fail closed on degraded known-issues status probes. (PR #465)skills-- Add Step 5b global path contract test. (PR #463)testing-- Auto-throttle local run-all suites across worktrees. (PR #449)gitignore-- Update .gitignore entries. (PR #442)report-- Add contract test for temp-file diagnostic handoff. (PR #417)cleanup-- Remove accidentally committed fix-issue agent file. (PR #408)report-- Add contract test for template alignment and classification. (PR #399)copilot-- Add Copilot cloud-agent onboarding instructions for VBW. (PR #377)docs-- Clarify UAT remediation round-cap semantics. (PR #372)ghcp-- Expose local VBW development agents for GitHub Copilot. (e1a878b)gitignore-- Add presentation source doc to .gitignore. (2718d3b)
list-todos-- Replace AskUserQuestion with display-and-stop pattern. (PR #490)vibe-- Extract AskUserQuestion contract into shared reference. (PR #489)agents-- Prevent stop-hook recovery ambient context and cross-thread contamination. (PR #479)context-- Stabilize cache-context target resolution. (PR #477)testing-- Stabilize run-all overlap auto-tune test. (PR #475)commands-- Consolidate agent settings resolution. (PR #474)debug-target-- Add configurable debug target and root CLAUDE symlink. (PR #472)tests-- Stabilize hook-wrapper cache env isolation. (PR #471)commands-- Eliminate embedded inline plugin-root spans. (PR #468)run-all-- Harden command contracts against local drift. (PR #464)phase-detect-- Harden flaky QA routing under parallel load. (PR #462)testing-- Eliminate statusline SIGPIPE flake. (PR #461)report-- Consolidate diagnostic helpers into /vbw:report. (PR #460)state-updater-- Harden temp file operations against parallel BATS flakiness. (PR #457)init-- Align CLAUDE_CONFIG_DIR prose with resolve-claude-dir.sh semantics. (PR #455)skills-- Unify global skills path resolution around CLAUDE_DIR cascade. (PR #445)phase-detect-- Harden QA pending detection under parallel BATS. (PR #441)tests-- Stabilize live PID fixtures. (PR #440)session-start-- Replace inline sed parser with shared extract_summary_status(). (PR #438)skills-- Correct global install path display to ~/.claude/skills/. (PR #437)scripts-- Remove extract_summary_status() override in recover-state.sh. (PR #432)tests-- Replace hardcoded dead PIDs with get_dead_pid helper. (PR #430)scripts-- Stabilize phase-detect QA pending status under parallel BATS. (PR #429)vibe-- Self-heal retryable phase-detect cache with sentinel and blank handling. (PR #427)qa-gate-- Skip change-evidence checks for plan-amendment-only remediation rounds. (PR #425)hooks-- Bypass task-verify for non-commit tasks. (PR #423)tests-- Harden run_phase_detect retry with exponential backoff. (PR #419)suggest-next-- Add debug-session awareness to fix case. (PR #416)security-filter-- Anchor directory patterns as path components. (PR #412)tests-- Use _ORIG_HOME for real HOME capture after setup_temp_dir. (PR #410)tests-- Fix flaky phase-detect.bats under parallel execution. (PR #409)report-- Persist diagnostics to temp file for reliable issue filing. (PR #400)qa-result-gate-- Support gitignored planning directory evidence in corroboration. (PR #395)track-known-issues-- Aggregate accepted-process-exception outcomes across remediation rounds. (PR #394)config-- Sync config.md settings table with defaults.json. (PR #392)debug-- Handle legacy migration failures in find_latest_unresolved and list. (PR #391)commands-- Address AskUserQuestion maxItems:4 violations. (PR #383)scripts-- Route to needs_uat_remediation when round UAT shows issues_found. (PR #370)commands-- Disambiguate step 4 task list from TaskCreate delegation. (PR #368)skills-- Require explicit skill outcome blocks for subagent spawns. (PR #364)qa-remediation-- Remove hardcoded MAX_ROUNDS=3 cap on remediation loop. (PR #362)hooks-- Fix CC 2.1.94 hook breakage (zsh nomatch, BSD awk locale, test isolation). (PR #359)qa-remediation-- Stop false remediation loops from docs-only rounds and carried known issues. (PR #358)known-issues-- Persist phase known-issues lifecycle through QA remediation. (PR #352)execute-- Enforce real team semantics in /vbw:vibe. (PR #351)research-- Use phase-wide RESEARCH.md naming for Plan mode. (PR #350)report-- Always auto-file issues; remove --file-issue flag. (PR #346)report-- Add scope guardrails to prevent unauthorized side effects. (PR #343)vibe-- Clarify routing prohibition to require AskUserQuestion tool call. (PR #342)testing-- Ungate verify-vibe.sh and add to CI. (PR #334)
hooks-- Prefer nativeagent_id/agent_typefields for subagent detection across start, stop, health, and compaction hooks, with backward-compatible fallback for legacy payload shapes. (PR #321)ci-- Add Discord release notification workflow that posts rich embeds with release notes when a GitHub release is published. (PR #314)
config-- Canonicalizeprefer_teamsaliases by removingwhen_parallelas a distinct mode, normalizing legacy values toalways|auto|nevervia a shared helper, and adding contract test coverage. (PR #320)ci-- Add.github/referencesto.gitignore.
vibe-- Fold Milestone UAT Recovery's two sequential AskUserQuestion calls into a single three-option question, eliminating the redundant yes/no entry gate. (PR #327)vibe-- Add explicitAskUserQuestiontool-call directives at all confirmation gates in the routing table, co-locating tool instructions with routing decisions. (PR #326)commands-- Add missing tools (AskUserQuestion, Agent, team management) toallowed-toolsfrontmatter across 14 command files, and add contract test for consistency. (PR #323)archive-- Use the Write tool instead of Bash for SHIPPED.md creation so PostToolUse hooks fire for artifact tracking. (PR #319)verify-- Keep execute-time UAT generation human-only by explicitly excluding programmatic and UI-automation checks from UAT inexecute-protocol.md. (PR #318)verify-- Add--remediation-kindfilter tocompile-verify-context.shso UAT remediation re-verification scans only the UAT directory when both QA and UAT remediation exist. (PR #317)
qa-- Add QA gate enforcement that blocks UAT until QA passes, treats deviations as defects, and auto-loops dev agents through a remediation state machine before UAT can start. (PR #296)qa-- Write per-roundR{RR}-VERIFICATION.mdinto round directories, making each QA remediation round a self-contained artifact set while preserving phase-level verification as frozen QA history. (PR #299)scope-- Persist milestone-level scope decisions in a newCONTEXT.mdartifact capturing decomposition rationale, requirement mapping, key decisions, and deferred ideas. Route project-level decisions into STATE.md. (PR #292)testing-- Reduce full test-suite latency from ~4m to ~1m via parallelized BATS execution, CI matrix sharding, auto-sized worker counts, and workspace-scoped cache/temp isolation. (PR #303)
vibe-- Route active UAT via metadata-only routing fromphase-detect.shand direct file read, eliminating the fragile precomputed extraction payload that caused intermittent zero-issue counts during template expansion. (PR #311)statusline-- Use-geinstead of-gtfor cache mtime comparison so lifecycle artifacts written in the same second correctly invalidate stale statusline values. (PR #309)uat-- Add consistency guard toextract-uat-issues.shthat detects when the awk parser produces zero issues but frontmatter declares a non-zero count, outputtinguat_extract_error=trueinstead of misleading zeros. (PR #307)qa-- Prevent deviation laundering through metadata-only remediation rounds by hardening classification rules, original FAIL carry-forward, and deterministic gate evidence requirements. Also fix first-UAT scope selection, fail-closed UAT finalization, and statusline lifecycle labeling. (PR #301)debugger-- Enable MCP tool access for Debugger and QA agents by switching fromtools:allowlists todisallowedTools:denylists, and add skill pre-evaluation at spawn points. (PR #289)plan-- Remove team creation from Plan mode so Scout and Lead are always spawned as plain sequential subagents, fixing incorrect concurrent execution that broke Scout-before-Lead dependency. (PR #287)statusline-- Anchor root resolution to script location instead of agent CWD, fixing monorepo scenarios where agents navigate to sub-packages. (PR #267)
scripts-- Addresolve-artifact-path.shfor deterministic artifact filename resolution, replacing inline LLM-computed filenames across plan, summary, research, context, UAT, and verification artifacts. (PR #277)agents-- Enable MCP tool access for VBW subagents by switching Scout from a restrictivetools:allowlist to adisallowedTools:denylist pattern, and adding MCP evaluation guidance to orchestrator commands. (PR #280)
hooks-- Exempt remediation round summaries (R{RR}-SUMMARY.md) from the file-guard terminal-status check that blocks SUMMARY writes with non-terminal status values. (PR #255)state-- Restore canonical## Current Phasesection in STATE.md after milestone archive, introduce shared phase-state helpers, and harden brownfield/archive recovery for consistent phase status across all downstream readers. (PR #256)statusline-- Resolve.vbw-planning/config.jsonby walking up from CWD instead of using bare relative paths, fixing monorepo subdirectory hook execution. (PR #259)vibe-- Re-route after milestone UAT "Start fresh" instead of dead-ending, by re-running phase-detect after marking milestones as remediated. (PR #269)phase-detect-- Addneeds_verificationas a first-class state to eliminate ambiguousauto_uatrouting, and prohibit UAT delegation to subagents. (PR #271)uat-- Fixadvancefromverifystage to start a new remediation round instead of dead-ending as a no-op. (PR #282)uat-- Normalize LLM-generated UAT status synonyms (all_pass,passed,verified, etc.) to canonical forms so phase-detect recognizes them as terminal statuses. (PR #284)
statusline-- Normalize context percentages and X/Y totals to Claude Code's usable autocompact window, including override-aware calculations for larger context windows. (PR #250)
docs-- Clarify the v1.33.0 release wording across the changelog, Discord draft, and published GitHub release notes. (PR #245)docs-- Remove the temporary v1.33.0 Discord draft file from the repository. (PR #247)
statusline-- Treat HTTP 429 usage responses as stale OAuth tokens, back off retries to 5 minutes, and skip usage fetches when nonessential traffic is disabled. (PR #251)
remediation-- Store UAT remediation artifacts in dedicatedremediation/round-XX/directories withRXX-*filenames, while keeping legacy flat layouts readable via fallback logic. (PR #236)remediation-- Add dedicated remediation research, plan, and summary templates so remediation rounds use their own structured artifacts instead of overloading phase-level files. (PR #236)statusline-- Add four display controls to hide the limits line, hide it only for API-key sessions, suppress the build/agent line in tmux, or collapse the statusline to a single line in tmux. (PR #118)statusline-- Color the[VBW]badge dynamically based on whether VBW context is active in the current session. (PR #213)dx-- Addscripts/dev-setup.shplus an optionalclaude-vbwlauncher to automate local VBW setup, teardown, status checks, and launching Claude Code against a local clone. (PR #212)ci-- Require pull requests to link an issue, recognizing closing keywords, bare issue refs, full issue URLs, and sidebar-linked issues. (PR #194)
qa-- QA now persistsVERIFICATION.mdthroughwrite-verification.shdirectly instead of relying on a heredoc escape hatch or parent-command passthrough. (PR #231)skills-- Fix forced skill evaluation for spawned agents by moving skill injection toSubagentStart, so team and subagent workflows reliably receive the skills they need; also stop accidental YAML turn ceilings from capping agents that should be unlimited. (PR #197)session-start-- Warn when GSD is co-installed to reduce cross-wired/gsd:*versus/vbw:*workflows in VBW sessions. (PR #197)remediation-- Make remediation planning and execution self-contained and sequential so remediation no longer inherits normal execute-mode wave/team behavior or creates unnecessary worktrees. (PR #236)todo-- Make/vbw:todoand/vbw:list-todosfail cleanly in restricted permission modes with actionable guidance. (PR #227)release-- Make release-audit changelog insertion non-interactive: generated entries are written immediately and completed before the separate README prompt. (PR #243)prompts-- Add and clarify AskUserQuestion spacing guidance so dialogs do not obscure surrounding output. (PR #241)config-- Change the defaultprefer_teamssetting fromalwaystoautoso VBW only spins up teams when parallelism is useful. (PR #236)
verify-- Limit post-remediation re-verification to the latest remediation round, carry prior UAT issues forward into targeted tests, and write remediation UAT files into the round directory instead of the phase root. (PR #236)compact-- Clear stale.context-usagedata at session boundaries and stamp it with session IDs so new sessions do not inherit false high-context warnings from old sessions. (PR #239)statusline-- Fix stale completion counts after reset/undo, count only valid completed summaries, and render remediation/UAT lifecycle states more accurately. (PRs #197, #236)statusline-- Fix first-render reliability by seeding caches on session start and avoiding blocking stdin reads. (PR #236)agents-- Add explicit mechanicalshutdown_responsetool-call instructions so shutdown requests are acknowledged through the team protocol. (PR #226)teams-- Clean stale or orphaned VBW team directories beforeTeamCreateand afterTeamDeleteto prevent ghost agent labels in Claude Code's status bar. (PR #230)worktree-- Fully remove residual.vbw-worktrees/directories and stale git worktree metadata during cleanup, including locked worktrees. (PR #229)watchdog-- Kill tmux panes that get stuck indefinitely during compaction. (PR #190)agents-- Prune dead PIDs from.agent-pidson session start. (PR #228)ci-- Harden linked-issue enforcement withsynchronizecoverage, tighter matching rules, clearer failure messages, explicit permissions, and updated docs/templates. (PR #196)
release-- Finalize Step 4 now checks remote branch existence before attempting deletion, preventing spurious error output when the branch was already removed by PR merge. (PR #184)
release-- Finalize now auto-merges the release PR (marks ready, waits for checks, merges, pulls) instead of requiring manual GitHub merge. Prepare returns to main after pushing release branch. Changelog extraction uses awk instead of sed for macOS compatibility. (PR #181)
release-- Version bump now accounts for pending release branches. Computes bump base as highest of local, remote, and pending semver values with dry-run/no-push safety contracts preserved. (PR #178)plugin-- Exclude release command from marketplace installs via.claude-plugin/excludepatterns. (PR #161)
gitignore-- Ignore.claudedirectory to keep user-local Claude Code state out of version control. (PR #158)
release-- Correct git date format placeholders (%cd --date=shortinstead of%Y-%m-%d) and clarify tag existence checks to test stdout content rather than exit code. (PR #175)release-- Robust grep pattern eliminates false positives from[Unreleased]section indirection; changelog entries now target versioned headers directly. (PR #173)release-- Auto-create versioned changelog section and populate from merged PRs and commits when section is missing. (PR #170)release-- Two-phase prepare/finalize workflow respects branch protection rules by using draft PRs instead of direct pushes to main. (PR #163)commands-- Add symlink fallback and robust grep pattern for plugin root resolution whenCLAUDE_PLUGIN_ROOTis unset. (PR #168)commands-- Add phase-detect fallback to prevent template-block race condition when phase-detect.sh runs concurrently. (PR #165)commands-- Revert session-key from SHA1 to deterministic value and injectsession_idfrom hook stdin for reliable session tracking. (PR #154)planning-- Enforce deterministic plan filename convention (PLAN.md) to prevent naming drift across phases. (PR #164)auto-uat-- Trigger verification mid-milestone before continuing execution, ensuring UAT gates are not bypassed. (PR #149)
auto-uat-- Auto UAT detection and routing for unverified phases. Newauto_uatconfig flag runs/vbw:verifyinline after QA instead of only suggesting it. Config controls whether UAT runs inline or is merely suggested. (PR #135)require-phase-discussion-- Newrequire_phase_discussionconfig flag gates plan execution behind a mandatory discussion step per phase. Undiscussed phases route toneeds_discussioninstead of jumping toneeds_plan_and_execute. (PR #123)deterministic-verification-- Deterministic VERIFICATION.md generation viawrite-verification.sh. Per-category 6-col verification tables derived from §4–§7 semantics. Replaces ad-hoc QA agent output with structured, reproducible verification artifacts. (PR #138)remediation-- Seeding of ROADMAP.md and STATE.md for remediation phases. Pre-seed CONTEXT.md with UAT report for in-phase remediation. Pre-seeded discussion annotations in Next Up hints. Skip discuss step for in-phase UAT remediation.config-- Multi-locationCLAUDE_CONFIG_DIRfallback across all scripts and commands. Removes hardcoded$HOME/.claudeassumptions. (PR #116 by @halindrome)context-- Pre-flight context guard to prevent mid-workflow compaction from disrupting agent execution.brand-- Horizontal bars format for LLM-generated output, replacing box-drawing characters.suggest-next-- Annotate pre-seeded discussion in Next Up hints for remediation phases.
docs-- Source-validated GSD comparison document added. Consolidated 6 config sections into unified Configuration section in README. Execution Model section added. Contributing guide rewritten with QA review process and local dev setup. Default settings reference and runtime feature flags details added. (PRs #133, #125, #119)verify-- Added "Skip" option for checkpoint testing. CHECKPOINT loop now blocking with explicit wait-for-input via AskUserQuestion. Clarified UAT test requirements and banned automated checks. Updated code blocks to specify language for formatting clarity.suggest-compact-- Replacedeval jqwith safe direct assignment. Dynamic token cost calculated from actual file sizes instead of hardcoded estimates.suggest-next-- Removed misleading--resumefrom verifyissues_foundoutput.execute-protocol-- Enhanced user interaction for first test results with CHECKPOINT and AskUserQuestion.commands-- Replaced shared temp file with deterministic symlink path. Updated inline script execution syntax for clarity. Used fenced code blocks for template processor execution. Removed deadCLAUDE_PLUGIN_ROOTfallback in!expressions.codeowners-- Added dpearson2699 as code owner.ci-- Improved ShellCheck and shell syntax check steps. Removed branch restriction for pull requests.
phase-detect-- Prioritize earlier incomplete phases over later UAT issues (fixes #145). Addrequire_phase_discussiongate to earlier-incomplete scan. Collect all phases with UAT issues. Exclude SOURCE-UAT from UAT globs and cross-reference remediation phases. Implementlatest_non_source_uatfunction for UAT file retrieval. Replacels-pipe patterns with glob for-loops. Usefindfor file counting instead ofls|wc -l. Exclude SOURCE-UAT.md from UAT scanning globs. (PRs #146, #142)archive-- Detect unresolved UAT in shipped milestones, block archive on UAT issues. Deterministic milestone slug derivation via script. (PR #121)plugin-root-- Replace model-executedCLAUDE_PLUGIN_ROOTwith inline load-time resolution. Content validation and process-tree fallback for--plugin-dir. Hardened resolver policy and runtime callsites. Replacelsglob withfindfor zsh compatibility. Use temp file for deterministic plugin root resolution. Added preamble for plugin root resolution in skills. (PR #137)symlink-- Canonicalize symlink target viapwd -Pto survive cache deletion. Spin-wait guard for symlink race condition. (PR #144)hooks-- Add quad-resolution fallback for local dev mode (--plugin-dir).uat-- Flow discovered issues into remediation pipeline. Enhanced UAT issue detection and remediation suggestions. Make CHECKPOINT loop blocking with explicit wait-for-input. Multiple QA rounds: pad normalization, orphan guard, mid-execution guard, numeric sort,sort -Vportability. (PR #142)vibe-- Read phase-detect live to avoid stale routing state. Fallback to preamble phase-detect on transient live-read failures. Avoid shared default session key collisions. Standardize deterministic session-key fallback. Run phase-detect.sh atomically in preamble.resume-- Use phase-detect.sh for deterministic state detection instead of inline logic.recovery-- Automatic state recovery wired into session-start viarecover-state.sh. 5 rounds of QA hardening.remediation-- Pre-seed CONTEXT.md with UAT report for in-phase remediation. Skip discuss step for in-phase UAT remediation. Update sed commands to create backup files during ROADMAP.md and STATE.md edits. Seed ROADMAP.md and STATE.md for remediation phases.guards-- Narrow milestone path guards to prevent false positives. Harden milestone path guards in bash scripts.routing-- Harden milestone path guards in bash scripts to prevent cross-milestone pollution.cache-- Handle empty plugin cache globs under pipefail. Bridge local dev plugin root into marketplace cache. 5 rounds of QA hardening on cache operations.cache-nuke-- Handle empty plugin cache globs underpipefail.config-- Remove dir-existence check fromresolve-claude-dir.sh. Fix bareHOME/.claudereferences in scripts. Auto-UAT runs UAT inline, not just suggests.session-start-- Bridge local dev plugin root into marketplace cache. Remove unusedPHASES_DIRassignment.verify-- Add trailing newline to CHECKPOINT expected line. Use bash builtins for cross-platform portability. Enhance user interaction by adding AskUserQuestion for checkpoint responses.lint-- Resolve SC2034 warnings inextract-verified-items.sh.tests-- Hardened for BATS 1.10 compatibility — avoid backtick escaping incompatible with bats 1.10. Harden stale cache test against SIGINT from bats runner. Plugin root resolver safety and callsite checks. 6 rounds of QA on test suite. Regression tests for deterministic path pattern.docs-- Update defaults for lease locks and event recovery settings. Fix all markdown lint errors in README. Remove.markdownlint.jsonconfiguration file. Update plugin root resolution documentation.
session-start-- Shipped milestones detection and QA verification script for post-ship validation.context-- Pre-flight context guard to prevent mid-workflow compaction from disrupting agent execution.
milestone-- Removed ACTIVE-file milestone indirection. Slug derivation and decision/todo extraction enhanced in milestone scripts.flags-- Partially reverted flag graduation — restored 9 flags as configurable. Legacy key fallback added to all consumer scripts. Runtime defaults aligned with legacy fallback behavior.suggest-compact-- Replaced eval jq with safe direct assignment. Dynamic token cost from actual file sizes.worktree-- Defaults to off. Fixed merge args, normalized boundary paths, added branch-exists fallback.
crash-recovery-- Avoid false fallback and preserve last words when SUMMARY.md missing.uat-- 4 rounds of QA fixes: pad normalization, orphan guard, mid-execution guard, numeric sort, sort -V portability, false remediation fallback.vibe-- Hardened UAT remediation routing and chain enforcement. Auto-route UAT remediation from /vbw:vibe.milestone-- Fixed rename-default-milestone slug derivation, post-ship state, collision guard, phases/ guard, generic tag strip, contract scope.scripts-- Null-safe jq patterns for false-default booleans. Tightened artifact globs. Removed unused variables (SC2034). Removed unreachable exit in validate-contract.sh.config-- Graduated v2_token_budgets. Stripped graduated V2/V3 flags from brownfield configs. Synced EXPECTED_FLAG_COUNT after worktree_isolation addition.tests-- Updated 90+ tests for graduated flags and portable paths. Fixed BATS_TEST_TMPDIR usage. Updated rollout-stage counts.
crash-recovery-- Agent crash recovery vialast_assistant_messagefrom CC 2.1.47 SubagentStop hook.agent-stop.shcaptures agent final output to.agent-last-words/when SUMMARY.md missing.validate-summary.shuses crash recovery files as fallback (60s freshness window).session-start.shauto-cleans stale files (7-day TTL). Event log tracksagent_shutdownevents withlast_message_lengthmetric.bash-classifier-tests-- 55 new BATS tests validating all 26 VBW hook bash patterns against CC 2.1.47 stricter permission classifier. Tests cover hook-wrapper.sh resolution, individual script invocations, bash-guard.sh pattern matching, and end-to-end integration. All patterns confirmed classifier-safe with zero changes needed to hooks.json.agent-memory-- Nativememoryfrontmatter added to all 7 agents. Core agents (lead, dev, qa, debugger, architect) usememory: projectfor cross-session learning. Ephemeral agents (scout, docs) usememory: local.agent-restrictions-- NativeTask(agent_type)spawn restrictions added to all 7 agents. Lead can only spawn Dev. Debugger can only spawn Debugger (competing hypotheses). Dev, architect, docs cannot spawn agents. Scout and QA already lacked Task capability.cc-version-table-- Claude Code version requirements table added to README with minimum versions for key features (hooks, teams, classifier, memory).token-analysis-- v1.30.0 token analysis document covering CC Alignment + Worktree Isolation impact: 85 scripts, 57 test files, 825 BATS tests. Per-request +12% (enforcement content), 73% reduction vs stock teams.worktree-isolation— Git worktree-per-plan isolation for Dev agents (6 scripts, enabled by default viaworktree_isolationconfig, set"off"to disable).
feature-flags-- Graduated 20 always-true feature flags (6 v2_, 14 v3_). Removed dead else-branches from 18+ scripts. Consolidated duplicate flags (v3_lock_lite→v3_lease_locks,v3_contract_lite→v2_hard_contracts). Graduated v3 observability flags (metrics, schema_validation, smart_routing, monorepo_routing, validation_gates, delta_context).config--v2_token_budgetsnow defaults tofalsein config/defaults.json. CC manages context natively.compaction-- Plan mode compaction workaround removed (CC 2.1.47 handles natively). Instructions updated to clarify CC-native vs VBW-specific behavior.update--/vbw:updateno longer displays restart requirement. Changes active immediately (CC 2.1.45+).token-budgets-- Added inline documentation to config/token-budgets.json explaining budget sizing rationale (1-4% of agent context windows) and per-task complexity scaling (0.6x-1.6x multipliers).
file-guard--scripts/file-guard.shmarked as worktree-deprecation candidate with header comment.
lock-lite— Deprecated script removed (replaced by lease-lock.sh).
skills-- Installation scope selection step (project vs global) when installing skills. (PR #91 by @Aperrix)discovered-issues-- Pre-existing test failures surfaced as "Discovered Issues" across/vbw:fix,/vbw:debug,/vbw:qa, and/vbw:verifyinstead of being silently dropped. New DEVN-05 deviation code for dev agent. (PR #99 by @dpearson2699)debugger-report-- Newdebugger_reportschema type replacing incorrectly-shapedblocker_reportusage by the Debugger agent. Role Authorization Matrix updated. (PR #99 by @dpearson2699)state-persistence-- Root STATE.md content (todos, key decisions, project-level notes) now persists across milestone shipping. Newmigrate-orphaned-state.shandpersist-state-after-ship.shscripts. (PR #103 by @dpearson2699)
bootstrap-- Codebase mapping bootstrap across all agents now uses META.md gating and "whichever exist" qualification to avoid wasted tool calls. Compaction re-reads codebase files. (PR #99 by @dpearson2699)bootstrap-- Key Decisions section removed from CLAUDE.md template. Existing decision rows migrated to STATE.md. Deprecated section handling added. (PR #93 by @dpearson2699)requirements-- Added minimum Claude Code version table to README documenting CC 2.1.47+ requirement for agent teams model routing and plan mode native support.
help-- CLAUDE_CONFIG_DIR fallback added to backtick expansion in help.md for non-standard config paths. (PR #101 by @halindrome)hooks-- Circuit breaker added totask-verify.shpreventing infinite loop when failed verification re-triggers itself. Expanded stop words in session-stop.sh. (PR #95 by @dpearson2699)
- Agent Teams Model Routing -- Requires Claude Code >= 2.1.47. Earlier versions had silently broken model routing for team teammates, causing all agents to use the default model instead of role-specific model profiles. If you experience agents using incorrect models (e.g., all agents using Sonnet when configured for Opus/Haiku), upgrade to Claude Code 2.1.47 or later.
rolling-summary-- Newscripts/compile-rolling-summary.shcompiles prior SUMMARY.md files into a condensedROLLING-CONTEXT.md(<200 lines). Integrated intocompile-context.shandcache-context.shhash. Feature-flagged viav3_rolling_summaryconfig. Archive and execute modes call it automatically.correlation-id--scripts/log-event.shauto-resolvescorrelation_idfromVBW_CORRELATION_IDenv var or.execution-state.json. Execute protocol generates UUID at phase start. Zero caller changes required — all shell callers get correlation threading for free.
readme-- Updated token efficiency section with v1.21.30 analysis (17% per-request reduction). Consolidated comparison table. Linked all 5 analysis reports. Fixed agent count (6→7). Trimmed Manifesto section.statusline-- Halved progress bar widths for compact display.agent-start-- Added tmux pane detection for PID→pane_id tracking.agent-stop-- Added tmux pane auto-close on agent stop.session-stop-- Added.agent-panesto cleanup list.
shutdown-protocol-- Agent shutdown request/response schema and handlers for all agents. Message typesshutdown_request/shutdown_responseadded tomessage-schemas.json. Agents gracefully terminate on orchestrator request. 4 rounds of QA hardening. (PR #90 by @dpearson2699)codebase-mapping-- Bootstrap codebase mapping for dev, qa, lead, and architect agents. Agents now reference PATTERNS.md, DEPENDENCIES.md, ARCHITECTURE.md, and CONCERNS.md from.vbw-planning/codebase/when available. (#78, #79, #80, #81 by @dpearson2699)list-todos-- New/vbw:list-todoscommand to browse and act on pending todos with milestone context resolution.config-- Effort-aware agentmaxTurnsresolution. Unlimited turns via0/false. Documented in README.
todos-- Migrated from### Pending Todosto flat## Todosstructure. Todo extraction delegated tolist-todos.shfor token efficiency.vibe-- Scout returns findings to orchestrator instead of writing files directly. Add/Insert Phase modes now load codebase mapping. Strengthened Lead shutdown HARD GATE with compaction-resilient verification and pre-chain checks.execute-- Post-summary improvisation blocked with STOP + Discovered Issues pattern. Post-shutdown verification gate added for Pure-Vibe loop safety.readme-- Updated hook diagram for SubagentStart/Stop hardening, added Agent Turn Limits section.
hooks-- Isolation marker lifecycle causing false blocks. Compaction and session-stop marker races. CLAUDE_PLUGIN_ROOT fallback for--plugin-dirinstalls. Role alias normalization and concurrency hardening for agent-start/stop. YAML frontmatter context required for expanded VBW command detection. Security-filter path-based root resolution with stale TTL. Commit keyword matching skipped for role-only task subjects. Docs role added to lifecycle hook matchers and token budgets. 14 regression tests added.hooks-- MissinghookEventNamein session-start.sh and blocker-notify.sh. Skip commit gate for[analysis-only]tasks.agents-- Lead shutdown gates strengthened to prevent tmux pane accumulation. PR review diff-vs-repo rule added to CLAUDE.md.commands-- CLAUDE_PLUGIN_ROOT resolution in model-executed contexts. Slug validation and milestone resolution alignment in pause.md.
debugger-- Codebase mapping bootstrap now references PATTERNS.md and DEPENDENCIES.md in addition to ARCHITECTURE.md and CONCERNS.md. Compiled context includes dynamic "Codebase Map Available" hint section for debugger role. Cache invalidation when mapping files change. (PR #77 by @dpearson2699)debug-- Codebase bootstrap instruction added to both Path A (team) and Path B (single) debugger task prompts. (PR #77 by @dpearson2699)
statusline-- Replacetrwithsedfor UTF-8 progress bar characters on Linux. GNUtrcorrupts multi-byte characters (█ ░ ▓) by processing byte-by-byte.
discussion-engine-- Unified discussion engine replacing three competing subsystems (bootstrap discovery, phase discovery, phase discussion). One engine, three entry points (/vbw:vibe,/vbw:vibe --discuss N,/vbw:discuss N). Auto-calibrates between Builder and Architect modes from conversation signals. Generates phase-specific gray areas instead of keyword-matched templates. ~1095 lines removed, ~180 lines added.discuss-- New/vbw:discuss [N]standalone command for explicit phase discussions.agent-health-- New agent health monitoring system with start/stop/idle/cleanup/orphan-recovery subcommands, lifecycle hooks (SubagentStart, SubagentStop, TeammateIdle, Stop), and integration tests.circuit-breaker-- Circuit breaker protocol added to all agent definitions for resilient failure handling.compaction-- Enhanced compaction recovery with TaskGet reminder for better context restoration.debugger-- Bootstrap investigation from codebase mapping when available.
vibe-- Bootstrap B2 now delegates to the discussion engine instead of 6-round rigid questioning. Plan mode no longer auto-asks 1-3 questions; users run/vbw:discussexplicitly if they want context. Discuss mode replaced with a 5-line engine delegation.bootstrap-requirements-- Removed tier-based annotation (table_stakes/differentiators/anti_features) and research cross-referencing. Reads simplifieddiscovery.jsonschema directly.readme-- Updated all counts (31k lines, 81 scripts, 23 commands, 7 agents, 461 tests). Added/vbw:discusscommand, Docs agent, agent health monitoring, discussion engine references.
discovery-protocol-- Deleted 800-linereferences/discovery-protocol.md. Replaced byreferences/discussion-engine.md(~150 lines).
teams-- Remove stale tmux forced in-process workaround, restore split-pane mode.token-budget-- Switch from line-based to character-based budgets, default to head truncation, remove cross-task escalation.ci-- Add executable bit toagent-health.sh. Remove unused variables intoken-baseline.sh(SC2034). UseCLAUDE_CONFIG_DIRfallback inagent-health.sh.
hooks-- Skip commit gate for[analysis-only]tasks. Debugger agents in team mode no longer get blocked by the TaskCompleted hook when investigating hypotheses without applying fixes. Contributed by @dpearson2699 (#71, fixes #70).hooks-- Add missinghookEventNameto session-start.sh (3 outputs) and blocker-notify.sh (1 output). Without this field, Claude Code's schema validator silently dropped all VBW context injection on session start. Also hardens compaction marker handling for clock skew and corrupted markers. Contributed by @dpearson2699 (#73, fixes #72).
planning-git-- Add 24 missing transient runtime artifacts toensure_transient_ignore()— session tracking, metrics, caching, events, snapshots, logging, watchdog, and codebase mapping files are now excluded fromplanning_tracking=commitcommits. Contributed by @dpearson2699 (#69, fixes #66).
max-turns-- Agent turn-budget resolver with effort-based scaling (thorough=1.5x, balanced=1x, fast=0.8x, turbo=0.6x), per-agent defaults, and config override support. Contributed by @dpearson2699 (#64).
hooks-- Prevent false isolation blocks by tightening SubagentStart/Stop/TeammateIdle/TaskCompleted matchers from 6 to 24 patterns for correct VBW agent detection. Contributed by @dpearson2699 (#64).compact-- Role-matched snapshot restore prefers snapshots from the same agent role after compaction, with fallback to latest. Task-level resume hints from event log provide in-progress task and resume candidate context. Contributed by @Solvely-Colin (#65).agents-- Add docs agent to max-turns resolver and fix missing EOF newline in resolve-agent-max-turns.sh.
agents-- New docs agent (vbw-docs) for documentation tasks with Sonnet as default model across all profiles, plusqa_skip_agentsconfig to auto-skip QA for documentation work.
prompt-preflight-- Stop deleting.vbw-sessionmarker on non-/vbw:prompts. The marker was being removed mid-workflow when users sent follow-up messages (plan approvals, answers), causingsecurity-filter.shto block Write/Edit calls to.vbw-planning/. Cleanup now only happens at session end viasession-stop.sh.hook-wrapper-- AddCLAUDE_CONFIG_DIRfallback in SIGHUP handler. Previously used bare$HOME/.claudewithout respecting custom config directory.shellcheck-- Fix SC2155 warnings (declare and assign separately) intmux-watchdog.sh,session-start.sh,clean-stale-teams.sh,doctor-cleanup.sh. Fix SC2034 unused variable indoctor-cleanup.sh.security-filter-- Remove.vbw-planning/self-blocking. The marker-based isolation (.gsd-isolation+.active-agent+.vbw-session) caused false blocks in too many scenarios: orchestrator after team deletion, agents before markers set, Read calls before prompt-preflight runs. GSD isolation is enforced by CLAUDE.md instructions +.planning/block (VBW→GSD direction).
help-- Dynamic help output from command frontmatter. Addedcategory:field to all 22 commands andscripts/help-output.shthat reads frontmatter to generate grouped, formatted help. Zero tokens, zero drift.tmux-- PID tracking utility for agent process registration and cleanup on SubagentStop.tmux-- Watchdog script for tmux detach detection, launched on session start.tmux-- SIGHUP trap fallback in hook-wrapper for tmux disconnect resilience.tmux-- Forced in-process mode auto-patch when tmux is detected.session-- Orphan claude process cleanup on SessionStart.session-- Stale team cleanup script with paired tasks directory cleanup, integrated into SessionStart.compaction-- Pre-compact agent state snapshots with agent metadata, restore in post-compact hook.doctor-- Runtime health checks: cleanup preview, --cleanup flag, cleanup logging with summary counts.
doctor-- Removelocalkeyword from top-level case block (bash compatibility).
gitignore-- Runtime files (snapshots, watchdog, doctor logs) documented in .gitignore.
commands-- Plugin root preamble dual-fallback. WhenCLAUDE_PLUGIN_ROOTis empty (observed in some marketplace installs), the backtick preamble now resolves via the plugin cache using the samesort -V | tail -1pattern established by PR #54 for hooks. Zero regression whenCLAUDE_PLUGIN_ROOTis set — the fallback never fires. 15 command preambles + 4 backtick bash calls updated. Extends PR #50's preamble architecture with PR #54's dual-fallback philosophy.testing--verify-plugin-root-resolution.shupdated to recognize both simple and dual-fallback preamble forms.
security-- Database safety guard. PreToolUse hook (bash-guard.sh) intercepts every Bash command and blocks 40+ destructive patterns across Laravel, Rails, Django, Prisma, Knex, Sequelize, TypeORM, Drizzle, Diesel, SQLx, Ecto, raw SQL clients, Redis, MongoDB, and Docker volumes. Fires on all agents with Bash access (Dev, QA, Lead, Debugger). Override withVBW_ALLOW_DESTRUCTIVE=1orbash_guard=falseconfig.security--config/destructive-commands.txtblocklist file, extensible per-project via.vbw-planning/destructive-commands.local.txt.agents--## Database Safetyprompt sections added to vbw-qa, vbw-dev, vbw-debugger, and vbw-lead agents.contracts--forbidden_commandsgate type inhard-gate.shfor per-plan command restrictions.templates--forbidden_commands: []field added to PLAN.md template.docs--docs/database-safety-guard.md— full design document with flowchart, pattern table, override guide, and architecture decisions.readme-- Database safety guard in Features section, Settings Reference (bash_guard), Security hook diagram updated.
readme-- Feature Flags Reference documenting all 20 v2/v3 flags with descriptions, dependencies, and toggle instructions. Settings Reference for all config keys. Planning & Git section forplanning_trackingandauto_push. (PR #62, @dpearson2699)readme-- Contributors section now uses contrib.rocks auto-updating image.
config--/vbw:configfeature flags display uses human-friendly labels instead of rawv2_/v3_prefixed keys. (PR #62, @dpearson2699)
planning_tracking+auto_pushconfig — Control how VBW planning artifacts are tracked in git. Three modes:manual(default),ignore(auto-gitignore.vbw-planning/),commit(auto-commit at lifecycle boundaries: bootstrap, plan, execute, archive). Push behavior:never/after_phase/always. Newscripts/planning-git.shwith sync-ignore, commit-boundary, and push-after-phase subcommands./vbw:initasks preference during setup. (PR #59, @dpearson2699)- Contributors section in README — Added @dpearson2699, @halindrome, and @navin-moorthy.
scripts/migrate-config.sh— Brownfield config migration extracted into shared script. SessionStart hook and/vbw:configboth run migration, ensuring missing v2/v3 flags are backfilled even when hooks didn't fire. Legacyagent_teamskey renamed toprefer_teamsand cleaned up. (PR #57, @dpearson2699)
scripts/migrate-config.sh --print-added— Generic defaults merge fromconfig/defaults.json. Future config keys auto-backfill without code changes./vbw:confignow shows a notice when settings were added. (PR #58, @dpearson2699)
hooks.json+hook-wrapper.sh— Hooks now resolveCLAUDE_PLUGIN_ROOTvia dual fallback: cache path first, then--plugin-dirpath. Previously,--plugin-dirinstalls had noCLAUDE_PLUGIN_ROOTset, causing all hook scripts to silently fail. Newresolve-claude-dir.batsandsessionstart-compact-hooks.batstests added. (PR #54, @dpearson2699)
statusline-cache-isolation.bats— CI runners now have git identity configured in test setup(), fixing "Author identity unknown" failures on GitHub Actions. Default branch detection is now dynamic instead of hardcoding "main". Statusline also renders repo label on detached HEAD. (PR #46, @dpearson2699)commands/*.md— 15 commands now includePlugin root: !backtick preamble so${CLAUDE_PLUGIN_ROOT}resolves correctly in model-executed Bash blocks. Previously the variable expanded to empty string, silently breaking all path references. Regression test added. (PR #50, @dpearson2699)research-persistence.bats— Tests now use trackedtemplates/RESEARCH.mdas fixture instead of gitignored.vbw-planning/runtime state. CI no longer depends on local development artifacts existing at repo root. (PR #56, @dpearson2699)
.github/CODEOWNERS—* @yidakeeensures all PRs auto-request code owner review.- Branch protection — main branch now requires passing CI and PR review before merge.
session-start.sh— Config migration now backfillsprefer_teamsfor users who initialized before v1.20.8, defaulting to"always". Previously these users had the old booleanagent_teamsbut noprefer_teamsenum, causing silent config drift.session-start.sh— jq migration failures now log a warning to stderr instead of failing silently. Users see a clear message when config migration encounters malformed JSON.
- 16 new tests across 2 new test files:
tests/config-migration.bats(9 tests) — empty config, partial config, full config no-op, idempotent migration, malformed JSON handling, EXPECTED_FLAG_COUNT sync validation, prefer_teams migration, prefer_teams preservation, count=23 validationtests/research-persistence.bats(7 tests) — Phase 1 RESEARCH.md section format, research-warn.sh JSON schema (4 cases), compile-context RESEARCH.md inclusion, flag-disabled path verification
- Test suite: 383 → 397 total tests (zero regressions)
research-persistence— Documented Scout write path, test coverage, and known hook isolation limitation for RESEARCH.md creation
/vbw:config— Feature flags (v3_*,v2_*) now displayed in config output. Previously these flags existed inconfig.jsonbut were invisible to users — the only way to discover or toggle them was manual JSON editing. Now shown as a dedicated "Feature Flags" section with descriptions and toggle hint. (Reported by @dpearson2699)
vbw-debugger— Increased maxTurns from 40 to 80. The scientific method protocol (reproduce + hypothesize + evidence + diagnose + fix + verify) legitimately needs more turns for complex bugs in large codebases. (Reported by @dpearson2699, #43)
vbw-debugger— Turn budget awareness: debugger now proactively checkpoints its state (hypotheses, evidence, files examined, next steps) when running long, preventing context loss on turn exhaustion. (Proposed by @dpearson2699, #44)
- PR #41 (@halindrome) — Statusline cache isolation: per-repo cache key with cross-platform hash
- PR #36 (@dpearson2699) — CI consolidation, contributor workflow improvements, ShellCheck fixes
vbw-statusline.sh— Statusline cache key now includes a hash ofgit rev-parse --show-toplevel, preventing cross-repo cache leakage when running Claude Code in multiple repos. Cross-platform hash:md5sum(Linux) ->md5(macOS) ->cksum(fallback). Local-only repos (no remote) display directory name instead of leaking another repo's data. (Thanks @halindrome)pre-push-hook.sh— Relaxed to consistency-only checks: verifies 4 version files match but no longer requires VERSION to change on every push. Removes ~30 lines of false-positive-prone enforcement. (Thanks @dpearson2699)resolve-claude-dir.sh— AddedexporttoCLAUDE_DIRassignment so sourcing scripts can access it. (Thanks @dpearson2699)bootstrap-project.sh— Removed unusedCREATEDvariable (ShellCheck SC2034). (Thanks @dpearson2699)bootstrap-requirements.sh— Removed unusedCOMPETITORSandANSWERED_COUNTvariables (ShellCheck SC2034). (Thanks @dpearson2699)
- CI workflows — Consolidated
.github/workflows/verification.ymlintoci.yml: shell syntax check + 4 contract verification steps now run as part of the test job. One workflow instead of two overlapping ones. (Thanks @dpearson2699) CLAUDE.md— Removed version bump instruction (contributor concern, not end-user relevant); push guard retained. (Thanks @dpearson2699)CONTRIBUTING.md— Rewritten Version Management section: describes manual merge-time bumping, no release-please references. (Thanks @dpearson2699)
- 10 new tests in
tests/statusline-cache-isolation.batsfor cache key isolation, cross-repo leakage prevention, no-remote directory name display, stale cleanup, and non-git directory handling. (Thanks @halindrome)
- PR #35 (@dpearson2699) — Security filter bypass fix:
hook-wrapper.shexit code 2 passthrough - PR #32 (@dpearson2699) — Progress dashboard fix:
state-updater.shmilestone-aware path resolution
hook-wrapper.sh— exit code 2 (Claude Code's "block tool call" signal) was silently converted to exit 0, disabling the security filter for.env,.pem,.keyfiles. Now passes exit 2 through correctly. (Thanks @dpearson2699)session-start.sh— compaction-marker check prevents error output during compaction cycles. (Thanks @dpearson2699)map-staleness.sh— compaction-marker skip and_diag()stderr helper for hook mode. (Thanks @dpearson2699)post-compact.sh—.compaction-markeradded to cleanup. (Thanks @dpearson2699)state-updater.sh— three bugs fixed: missing.execution-state.jsonno longer skips STATE.md/ROADMAP.md updates; reads.plans[](current schema) instead of only.phases{}(old schema);planning_root_from_phase_dir()resolves milestone-aware paths instead of hardcoded.vbw-planning/. (Thanks @dpearson2699)
- 14 new tests in
tests/sessionstart-compact-hooks.batsfor compaction marker, hook wrapper exit codes, and session start behavior. (Thanks @dpearson2699) - 4 new tests in
tests/state-updater.batsfor milestone-aware path resolution and schema support. (Thanks @dpearson2699)
config-- Newprefer_teamsconfig option (always|when_parallel|auto) replaces booleanagent_teams. Defaultalwayscreates Agent Teams for all operations, maximizing color-coded UI visibilityvibe.md-- Plan mode respectsprefer_teams— creates team even for Lead-only when set toalwaysdebug.md-- Debug mode respectsprefer_teams— uses team path for all bugs when set toalways
init.md-- Config scaffold createsprefer_teamsinstead of deprecatedagent_teamssession-start.sh-- Readsprefer_teamsconfig instead ofagent_teamsphase-detect.sh-- Readsprefer_teamsconfig instead ofagent_teams
config.md-- Settings reference table updated forprefer_teamsenumtest_helper.bash-- Test fixtures updated forprefer_teamsconfig
vbw-statusline.sh-- OAuth token lookup now detects keychain access denial vs API key usage. Users with OAuth (Pro/Max) whose keychain blocks terminal access now see an actionable diagnostic message instead of misleading "N/A (using API key)". AddedVBW_OAUTH_TOKENenv var as escape hatch. Usesclaude auth statusto distinguish auth methods when credential store fails.
- PR #38 (@navin-moorthy) -- Human UAT verification gate with CHECKPOINT UX
commands/verify.md-- New/vbw:verifycommand for human acceptance testing with per-test CHECKPOINT loop, resume support, and severity inferencetemplates/UAT.md-- UAT result template with YAML frontmatter for structured pass/fail/partial trackingexecute-protocol-- Step 4.5 UAT gate after QA pass (autonomy-gated: cautious + standard only)vibe.md----verifyflag, Verify Mode section, and NL keyword detection (verify, uat, acceptance test, etc.)execute-protocol-- TeamCreate for multi-agent execution (2+ plans get colored labels, status bar entries, peer messaging)vibe.md-- TeamCreate for Plan mode when Scout + Lead co-spawn (research + planning as coordinated team)
suggest-next.sh-- UAT suggestions surfaced after QA passes (cautious + standard autonomy)execute-protocol-- Step 5 shutdown now conditional on team existence (skip for single plan/turbo)
- PR #27 (@halindrome) -- CLAUDE_CONFIG_DIR support across all hooks and scripts
- PR #29 (@dpearson2699) -- Fix 6 command names missing
vbw:prefix for discovery
resolve-claude-dir.sh-- Central helper for Claude config directory resolution
hooks-- All 21 hook commands now respectCLAUDE_CONFIG_DIRenvironment variablecommands-- 6 commands missingvbw:prefix now discoverable under/vbw:*doctor-- Plugin cache check respects custom config directoryverify-vibe-- Removed unusedGLOBAL_MIRRORvariable
scripts-- 8 scripts source central resolver instead of inline fallback pattern
resolve-claude-dir-- 19 new bats tests for CLAUDE_CONFIG_DIR resolution
shellcheck-- resolved all shellcheck warnings across scripts. Removed unused variables, quoted command substitutions, added targeted disables for intentional patterns (ls|grep for zsh compat, git@{u}syntax, read-consumed vars).ci-- bats tests now pass on GitHub Actions Ubuntu runner. Added git user config for phase-detect tests, fixed cross-platformstatflag order (GNU first, BSD fallback) in resolve-agent-model.scripts-- added executable bit to 6 scripts missing chmod +x: generate-incidents.sh, lease-lock.sh, recover-state.sh, research-warn.sh, route-monorepo.sh, smart-route.sh.testing-- corrected command name expectation in verify-commands-contract.sh. Test now accepts both bare names (map) and prefixed names (vbw:map) since the plugin system auto-prefixes.
discovery-protocol-- complete rewrite ofreferences/discovery-protocol.mdfor coherence and completeness. B2 bootstrap and Discuss mode logic fully specified with gap fixes from research. Removed brittle line number references from Integration Points.vibe-- updated Discuss mode and Bootstrap B2 to align with rewritten discovery protocol.
Merges 10 pull requests from @dpearson2699 (Derek Pearson). These contributions identified bugs, proposed fixes, and directly influenced the v1.20.0 architecture. Previously closed without proper merge credit — now properly merged and attributed.
- #10 --
fix(update): identifiedCLAUDE_PLUGIN_ROOTbreakage when commands copied to user directory. - #11 --
fix(compile-context): fixed unpadded phase number resolution incompile-context.sh. - #12 --
fix(stack): identified nested manifest scanning gap indetect-stack.sh, added iOS/Swift mappings. - #13 --
fix(map): identified zshnomatchglob crash, triggering a repo-wide zsh compatibility audit. - #14 --
fix(todo): fixed STATE.md/todo.md heading mismatch causing unreliable insertion after/vbw:init. - #15 --
test(verification): built repo-wide verification harness (228+ tests, GitHub Actions CI, command frontmatter validation). - #17 --
fix(bootstrap): hardened CLAUDE.md bootstrap with centralized isolation, brownfield stripping, input guardrails. - #19 --
refactor(isolation): designed the two-layer defense model and auto-migration that ships as the canonical isolation architecture. - #22 --
fix(vibe): identified and fixed scope mode writing lifecycle actions into Todos section. - #24 --
fix(hooks): systematic audit ofhookEventNamecompliance across 8 hook scripts.
ci-- GitHub Actions CI workflow (.github/workflows/verification.yml) for automated PR and push checks.testing-- repo-wide test harness:testing/run-all.sh,verify-bash-scripts-contract.sh,verify-commands-contract.sh.bootstrap--scripts/verify-claude-bootstrap.shwith 27 contract tests.hooks--hookEventNamecompliance added to 7 additional hook scripts (post-compact,map-staleness,prompt-preflight,validate-commit,validate-frontmatter,validate-summary).stack-- expandedconfig/stack-mappings.jsonwith iOS/Swift and recursive detection entries.
update-- prefix allclaude plugincommands withunset CLAUDECODE &&to prevent "cannot be launched inside another Claude Code session" error when running/vbw:updatefrom within an active session.statusline-- remove misleading agent count that counted all system-wideclaudeprocesses instead of VBW-managed agents.
doctor--/vbw:doctorhealth check command with 10 diagnostic checks: jq installed, VERSION file, version sync, plugin cache, hooks.json validity, agent files, config validation, script permissions, gh CLI, sort -V support.disable-model-invocation: true.templates-- CONTEXT.md and RESEARCH.md templates for agent context compilation and research output structure.blocker-notify-- TaskCompleted hook auto-notifies blocked agents when their blockers resolve, preventing teammate deadlocks.control-plane-- lightweight Control Plane dispatcher (scripts/control-plane.sh, 328 lines) that sequences all enforcement scripts into a unified flow. Four actions: pre-task (contract → lease → gate), post-task (gate → release), compile (context compilation), full (all-in-one). Fail-open on script errors, JSON result output, lease conflict retry with 2s wait. 15 unit tests + 3 integration tests.rollout-stage-- 3-stage progressive flag rollout automation (scripts/rollout-stage.sh). Stages: observability (threshold 0), optimization (threshold 2), full (threshold 5). Actions: check prerequisites, advance flags atomically, status report with all 14 v3_ flags. Stage definitions inconfig/rollout-stages.json. Supports--dry-run. 10 tests.token-baseline-- per-phase token usage measurement and comparison (scripts/token-baseline.sh). Actions: measure (aggregate from event log), compare (delta with direction indicators), report (markdown with budget utilization by role). Baselines saved to.baselines/token-baseline.json. 10 tests.token-intelligence-- per-task token budgets computed from contract metadata. Complexity scoring (must_haves weight 1, files weight 2, dependencies weight 3) maps to 4 tier multipliers. Fallback chain: per-task → per-role → config defaults. Token cap escalation emitstoken_cap_escalatedevent and reduces remaining budget for subsequent tasks. 12 tests.context-index--context-index.jsonmanifest generated in.cache/with key-to-path mapping per role/phase. Atomic writes via mktemp+mv. Updated on every cache miss, timestamps refreshed on cache hits. 6 tests.execute-protocol-- Control Plane orchestration block in Step 3, context compilation and token budget guards in Steps 3-4, cleanup in Step 5. Individual scripts (generate-contract.sh, hard-gate.sh, compile-context.sh, lock-lite.sh) preserved as independent fallbacks.
isolation-- consolidated to single root CLAUDE.md with context isolation rules for both VBW and GSD plugins.agents-- removed deadmemory: projectfrom all 6 agent frontmatters. Clarified standalone vs teammate session scope in debugger.references-- fixed internal references in verification-protocol.md (S5→§5/VRFY-06). Added per-model cost basis to model-profiles.md methodology note.README-- token efficiency section updated with v1.20.0 numbers (8,807 lines bash, 63 scripts, 21 commands, 11 references). Command/hook counts updated to 21. Typo and incomplete sentence fixes.compile-context-- ROADMAP metadata parser fixed (### Phase→## Phaseto match actual format). Scout, Debugger, and Architect roles extended with conventions, research, and delta files. Code slices added to Debugger and Dev contexts.token-budget-- extended argument parsing for contract path and task number. Per-task budget computation with complexity scoring. Escalation config added toconfig/token-budgets.json.detect-stack-- expanded coverage for Python, Rust, Go, Elixir, Java, .NET, Rails, Laravel, Spring. 4 new manifest file detections.control-plane--context_compilerdefault harmonized fromfalsetotrueto match phase-detect.sh and defaults.json.config-- all 20 V2/V3 feature flags available in project config (default: off, enable via/vbw:config). 15 flags added to migration: lock_lite, validation_gates, smart_routing, event_log, schema_validation, snapshot_resume, lease_locks, event_recovery, monorepo_routing, hard_contracts, hard_gates, typed_protocol, role_isolation, two_phase_completion, token_budgets.
task-verify-- bash 3.2 compatibility: replacedcase...esacinside piped command substitution withgrep -Evstop word filter. macOS bash 3.2.57 parsing bug.bump-version-- added--offlineflag to skip remote GitHub fetch for CI/air-gapped environments.phase-detect-- compaction threshold now configurable viacompaction_thresholdin config.json (default: 130000).scope-- prevent lifecycle actions from polluting Todos.init-- remove*.slnglob that crashes zsh on macOS.teams-- auto-notify blocked agents when blockers clear.defaults-- harmonize model_profile fallback to quality across all scripts.migration-- comprehensive flag migration and jq boolean bug fix.release-- resolve 8 findings from 6-agent pre-release verification.validate-commit-- heredoc commit messages no longer overwritten by-mflag extraction. macOS sed compatibility fix.session-start-- zsh glob compatibility across session-start, snapshot-resume, lock-lite, and file-guard scripts.security-filter-- stale marker detection (24h threshold) prevents false positive blocks on old markers.
tokens-- v1.20.0 Full Spec Token Analysis (664 lines): 258 commits, 6 milestones, per-request -7.3%, ~85% coordination overhead reduction maintained despite 33% codebase growth.
- 86 new tests across 5 new test files: hooks-isolation-lifecycle.bats (10), phase0-bugfix-verify.bats (16), token-budgets.bats (12), context-index.bats (6), control-plane.bats (18), rollout-stage.bats (10), token-baseline.bats (10). Plus 4 context metadata tests. Test suite: 237 → 323 (zero regressions).
isolation-- context isolation to prevent GSD insight leakage into VBW sessions. New### Context Isolationsubsection in Plugin Isolation instructs Claude to ignore<codebase-intelligence>tags and use VBW's own codebase mapping. bootstrap-claude.sh now strips 8 known GSD section headers when regenerating CLAUDE.md from existing files.
config-- interactive granular model configuration. Second menu in/vbw:configModel Profile flow offers "Use preset profile" or "Configure each agent individually". Individual path presents 6 agent questions across 2 rounds (4+2 split), writes model_overrides to config.json, and displays before/after cost estimate. Status display marks overridden agents with asterisk (*). Feature implemented in commits 1ac752b through 91da54f (Phase 1, Plan 01-01).init-- GSD project detection and import. Step 0.5 detects existing.planning/directory before scaffold, prompts for import consent, copies to.vbw-planning/gsd-archive/(preserves original), generates INDEX.json with phase metadata and quick paths. Enables seamless migration from GSD to VBW with zero-risk import.scripts-- generate-gsd-index.sh for lightweight JSON index generation (<5s performance). Creates INDEX.json with imported_at, gsd_version, phases_total, phases_complete, milestones, quick_paths, and phases array for fast agent reference without full archive scan.help-- GSD Import section documenting detection flow during /vbw:init, archive structure (.planning/ → gsd-archive/), INDEX.json generation, and isolation options.docs-- migration-gsd-to-vbw.md comprehensive migration guide (273 lines, 9 sections) covering import process, archive structure, version control best practices, INDEX.json format, usage patterns, GSD isolation, migration strategies (full/incremental/archive-only), troubleshooting scenarios, and FAQ.bootstrap-- 5 reusable bootstrap scripts in scripts/bootstrap/ (project, requirements, roadmap, state, claude). Each accepts arguments only, outputs to specified path, uses set -euo pipefail. Enables shared file generation between /vbw:init and /vbw:vibe.inference-- brownfield intelligence engine. infer-project-context.sh (247 lines) reads codebase mapping to extract project name, tech stack, architecture, purpose, and features with source attribution. infer-gsd-summary.sh (163 lines) reads GSD archives for latest milestone, recent phases, key decisions, and current work.init-- auto-bootstrap flow (Steps 5-8). After infrastructure setup, init detects scenario (greenfield/brownfield/GSD migration/hybrid), runs inference engine, presents always-show confirmation UX with 3 options (accept/adjust/define from scratch), field-level correction, then calls bootstrap scripts to generate all project files. Seamless flow with no pause between mapping and project definition.
vibe-- Bootstrap mode (B1-B6) refactored to call extracted bootstrap scripts via ${CLAUDE_PLUGIN_ROOT}/scripts/bootstrap/. Discovery logic stays inline, file generation delegated to shared scripts. Verified standalone mode, config compliance, and zero regression across all 11 modes.
statusline-- model profile display on L1 after Effort field. Shows current active profile (quality/balanced/budget) dynamically from config.json.
defaults,init,session-start,suggest-next,statusline-- default model profile changed from "balanced" to "quality". New VBW installations and auto-migrations now use Opus for Lead/Dev/Architect/Debugger by default for better output quality.
pre-push-hook-- hook now skips enforcement in non-VBW repos. Added early exit guard that checks for VERSION and scripts/bump-version.sh files. If both absent, hook exits cleanly without blocking pushes. Fixes issue where installing VBW plugin blocked git pushes in existing brownfield repositories.
model-profiles-- cost control via model profile configuration. Three preset profiles (quality/balanced/budget) with per-agent model assignments, plus per-agent override support for advanced users.config-- model profile selection in/vbw:configinteractive menu with quality/balanced/budget options. Settings table displays current profile and all 6 agent model assignments.config-- CLI arguments/vbw:config model_profile <profile>for direct profile switching and/vbw:config model_override <agent> <model>for per-agent overrides.vibe-- Phase Banner displays active model profile during Plan and Execute modes.execute-protocol-- agent spawn messages include model name in parentheses format: "◆ Spawning {agent} ({model})...".model-routing-- agent model resolution helper script (scripts/resolve-agent-model.sh) with hybrid architecture pattern: commands read config, helper handles merge/override logic.init-- new projects seeded withmodel_profile: "quality"in config.json.session-start,suggest-next,statusline-- auto-migration addsmodel_profile: "quality"to existing projects without the field.
vibe,execute-protocol,debug,research,qa-- all agent-spawning commands now pass explicitmodelparameter to Task tool based on active profile and overrides.references-- newmodel-profiles.mdwith complete preset definitions, cost comparison table, and override syntax documentation.effort-profiles-- all 4 effort profile files updated to clarify effort controls planning depth while model profile controls cost.help-- Model Profiles section added with command examples and cost estimates.README-- Cost Optimization section added with 3-profile comparison table and usage guidance.
statusline-- distinguish auth expired from network failure in usage limits. Previously, both a stale OAuth token (401/403) and a network timeout showed the same "fetch failed" message. Now shows "auth expired (run /login)" for auth failures, keeping "fetch failed (retry in 60s)" for actual network issues.
init-- preserve existingCLAUDE.mdin brownfield projects./vbw:initStep 3.5 was blindly overwriting the user's rootCLAUDE.md. Now reads first — if it exists, appends VBW sections to the end instead of clobbering. Same brownfield-awareness added to/vbw:vibeBootstrap B6 and Archive Step 8.
config-- respectCLAUDE_CONFIG_DIRenv var across all scripts and commands. Users who setCLAUDE_CONFIG_DIRto relocate their Claude config directory were hitting hardcoded~/.claude/paths. All 9 affected files now resolve via${CLAUDE_CONFIG_DIR:-$HOME/.claude}fallback pattern — zero breakage for existing users.
refs-- purge all stale/vbw:implementand standalone command references from shipped codemeta-- correct stale counts in README (9→10 reference files, 11→15 disable-model-invocation commands) and agent memory (27→20 commands, 18→20 hooks, 8→11 event types)
hooks--qa-gate.shtiered SUMMARY.md gate: commit format match now only grants a 1-plan grace period; 2+ missing summaries block regardless. Replaces||logic where format match bypassed missing summaries entirely.references--execute-protocol.mdStep 3b hardened to mandatory 4-step verification gate. No plan marked complete without verified SUMMARY.md.docs-- README TeammateIdle hook description updated to reflect tiered gate. CHANGELOG execute.md references corrected to execute-protocol.md. qa-gate.sh comments fixed (GSD → conventional).
/vbw:vibe-- single intelligent lifecycle command replacing 10 absorbed commands (implement, plan, execute, discuss, assumptions, add-phase, insert-phase, remove-phase, archive, audit). 293 lines, 11 modes, 3 input paths (state detection, NL intent parsing, flags), mandatory confirmation gates. 76/76 automated verification checks PASS.references/execute-protocol.md-- execution orchestration logic (Steps 2-5) extracted from execute.md for on-demand loading by vibe.md Execute mode. Zero per-request cost.scripts/verify-vibe.sh-- 241-line automated verification script validating all 25 vibe command requirements across 6 groups.- Context compiler milestone -- 3-phase optimization reducing agent context loading by 25-35% across all project sizes. 14 feat/refactor commits, 65/65 QA checks (3 phases, all PASS). Agents now receive deterministic, role-specific context instead of loading full project state.
scripts/compile-context.sh-- new script producing.context-lead.md(filtered requirements + decisions),.context-dev.md(phase goal + conventions + bundled skills),.context-qa.md(verification targets). Config-gated withcontext_compilertoggle.config--context_compilertoggle (default:true) inconfig/defaults.json. Setting tofalsereverts all compilation to direct file reads.compiler-- skill bundling readsskills_usedfrom PLAN.md frontmatter, resolves from~/.claude/skills/, bundles into Dev context. No-op when no skills referenced.hooks-- compaction marker system:compaction-instructions.shwrites.compaction-markerwith timestamp on PreCompact.session-start.shcleans marker at session start for fresh-session guarantee.
commands-- all 6 commands (execute, plan, qa, discuss, assumptions, implement) now use pre-computedphase-detect.shoutput instead of loading 89-linephase-detection.mdreference doc.commands-- plan.md, execute.md, implement.md callcompile-context.shbefore agent spawn with config-gated fallback to direct file reads.agents/vbw-dev.md-- removed STATE.md from Stage 1 (never used). Added marker-based conditional re-read with "when in doubt, re-read" conservative default.agents/vbw-qa.md-- replacedverification-protocol.mdruntime reference with inline 12-line format spec. Tier provided in task description instead of loaded from full protocol.
hooks--pre-push-hook.shrestored to actual validation logic (was replaced by delegator wrapper causing infinite recursion).hooks--qa-gate.shtightened to tiered SUMMARY.md gate. Commit format match now only grants a 1-plan grace period; 2+ missing summaries block regardless. Replaces previous||logic where format match could bypass missing summaries entirely.references--execute-protocol.mdStep 3b hardened: SUMMARY.md verification gate is now a mandatory 4-step checkpoint after Dev completion. No plan marked complete without verified SUMMARY.md.
commands-- 29 commands consolidated to 20. Ten lifecycle commands hard-deleted:implement,plan,execute,discuss,assumptions,add-phase,insert-phase,remove-phase,archive,audit. All absorbed into/vbw:vibe(single intelligent router with 11 modes). Global commands mirror cleaned. No aliases, no deprecation shims.
hooks--state-updater.shnow updates ROADMAP.md progress table and phase checkboxes when PLAN.md or SUMMARY.md files are written. Previously only STATE.md was updated, leaving ROADMAP.md permanently stale after bootstrap.hooks--pre-push-hook.shrestored to actual validation logic. Commitda97928accidentally replaced it with the.git/hooks/pre-pushdelegator wrapper, causing infinite recursion that hung everygit push.
discovery-- intelligent questioning system for/vbw:implement. Discovery protocol reference (references/discovery-protocol.md) with profile-gated depth (yolo=skip, prototype=1-2, default=3-5, production=thorough), scenario-then-checklist format, and example questions.config--discovery_questionstoggle (default:true) inconfig/defaults.json. Disabling skips all discovery prompts.implement-- bootstrap discovery (State 1) rewrites static B2 questions with intelligent scenario+checklist flow, feeding answers into REQUIREMENTS.md.implement-- phase-level discovery (States 3-4) asks 1-3 lightweight questions scoped to the phase goal before planning. Checksdiscovery.jsonto avoid re-asking.discuss-- answers now written todiscovery.jsonfor cross-command memory, so/vbw:discussand implement share the same question history.
hooks--state-updater.shauto-advances STATE.md to the next incomplete phase when all plans in a phase have summaries. Sets status to "active" on plan writes.hooks--pre-push-hook.shsimplified to a thin delegator routing to the latest cached plugin script viasort -V | tail -1.profile-- added Discovery depth column to built-in effort profiles table.config-- addeddiscovery_questionsto settings reference table.
statusline-- removed Cost field from Line 4. Moved Prompt Cache from usage line (L3) back to context line (L2) after Tokens field.README-- promoted "What You Get Versus Raw Claude Code" comparison table from subsection to top-level section, moved above Project Structure for better visibility.
statusline-- removed Cost field from Line 4. Status line now shows Model, Time, Agents, and VBW/CC versions only.
security-filter.sh--.planning/block now conditional on VBW markers, so GSD can write to its own directory when VBW is not the active caller. Previously blocked GSD unconditionally in every project./vbw:init-- creates.vbw-sessionmarker after enabling GSD isolation, so the security filter allows VBW writes during the remainder of the init flow (codebase mapping).
- Autonomy level rename --
dangerously-viberenamed topure-vibeacross all commands, references, README, and changelog. Tone adjusted to be informative without scare language. - Statusline economy line -- renamed "Cache" to "Prompt Cache" for clarity.
- Token economy engine -- per-agent cost attribution in the statusline. Each render cycle computes cost delta and attributes it to the active agent (Dev, Lead, QA, Scout, Debugger, Architect, or Other). Accumulated in
.vbw-planning/.cost-ledger.json. DisplaysCost: $X.XXon Line 4 and a full economy breakdown on Line 5 (per-agent costs sorted descending, percentages, cache hit rate, $/line metric). Economy line suppressed when total cost is $0.00. - Agent lifecycle hooks --
SubagentStarthook writes active agent type to.vbw-planning/.active-agentviascripts/agent-start.sh.SubagentStophook clears the marker viascripts/agent-stop.sh. Enables cost attribution to know which agent incurred each cost delta. /vbw:statuseconomy section -- status command reads.cost-ledger.jsonand displays per-agent cost breakdown when cost data is available. Guarded on file existence and non-zero total.- GSD isolation -- two-layer defense preventing GSD from accessing
.vbw-planning/. Layer 1: rootCLAUDE.mdPlugin Isolation section (advisory). Layer 2:security-filter.shPreToolUse hard block (exit 2) when.gsd-isolationflag exists and no VBW markers present. Two marker files (.active-agentfor subagents,.vbw-sessionfor commands) prevent false positives. Opt-in during/vbw:initwith automatic GSD detection.
- Statusline cache consolidation -- 6 cache files (
-ctx,-api,-git,-agents,-branch,-model) reduced to 3 (-fast,-slow,-cost). Grouped by update frequency to reduce file I/O. - Pure shell formatting --
awkreplaced with shell functions (fmt_tok,fmt_cost,fmt_dur) for token, cost, and duration formatting. Eliminates 3 subprocesses per render cycle. session-stop.shcost persistence -- session stop hook now reads.cost-ledger.jsonand appends a cost summary line to.session-log.jsonlbefore cleanup.post-compact.shcost cleanup -- compaction hook resets cost-tracking temp files (.active-agent, stale cache entries) to prevent attribution drift after context compaction.- README statusline documentation -- updated hook counts (18/10 to 20/11), added SubagentStart to hook diagram, documented economy line in statusline description.
/vbw:initGSD detection -- Step 1.7 checks~/.claude/commands/gsd/and.planning/to detect GSD. Prompts for isolation consent only when GSD is present; silent skip otherwise.
suggest-next.sh-- context-aware Next Up suggestions (ADP-03). New script reads project state (phases, QA results, map existence, milestone context) and returns ranked suggestions. 12 commands updated to call it instead of hardcoded static blocks. After QA fail, suggests/vbw:fixinstead of/vbw:archive; when codebase map is missing, injects/vbw:maphint.
templates/SUMMARY.md-- slimmed frontmatter to consumed fields only (TAU-05). Removedduration,subsystem,tags,dependency_graph,tech_stack,key_files(never read by any command). Addedtasks_completed,tasks_total,commit_hashes(actually consumed by status and QA). Saves ~80-120 output tokens per SUMMARY write.references/shared-patterns.md-- added Command Context Budget tiers (TAU-02 formalization). Documents Minimal/Standard/Full context injection convention so future commands don't cargo-cult STATE.md injections.
/vbw:status --metrics-- removed broken flag that referencedtokens_consumedandcompaction_countfields which never existed in the SUMMARY template.
/vbw:update-- version display now uses the actual cached version after install, not the GitHub CDN estimate. Fixes misleading "Updating to vX.Y.Z" and false version mismatch warnings when CDN lags behind the marketplace.
hooks-- all 18 hook commands now exit 0 when the plugin cache is missing, preventing "PostToolUse:Bash hook error" spam during/vbw:update. Previously,cache-nuke.shdeleted the cache but hooks kept firing and failing until the cache was re-populated.
config-- default autonomy level changed frompure-vibetostandard. New installations now require plan approval and stop after each phase for review, giving users guardrails by default.
commands-- lazy reference loading (TAU-01). Cross-command@-references inimplement.mdandinit.mdreplaced with deferredReadinstructions soplan.md,execute.md, andmap.mdare only loaded when the model reaches the state that needs them. Removed unused STATE.md injections fromfix,todo, anddebugcommands. Saves 200-500 tokens per invocation for states that don't use the deferred files.
/vbw:update-- always runs full cache refresh even when already on latest version. Fixes corrupted caches or stale hook schemas without requiring a version bump.
statusline-- hourly update check with visual indicator. When a newer VBW version is available, Line 4 turns yellow bold showingVBW {current} → {latest} /vbw:update. Cached 1 hour, single curl, zero overhead otherwise.
hooks.jsoninvalid event types --PostCompact(not a valid Claude Code event) replaced withSessionStartmatcher"compact".NotificationReceivedrenamed toNotification. Fixes fresh install validation error on newer Claude Code versions.notification-log.shfield mismatch -- script was reading.sender/.summary(non-existent fields). Now reads.notification_type,.message, and.titleper theNotificationevent schema.- README event type count -- corrected "11 event types" to "10 event types" after PostCompact was merged into SessionStart.
README Quick Start-- prominent warning against using/clear, explaining Opus 4.6 auto-compaction and directing users to/vbw:resumefor context recovery.
install-hooks.shresolves .git from project root -- scripts useddirname "$0"to find.git, which resolves to the plugin cache directory (~/.claude/plugins/cache/...) for marketplace users instead of the user's project. Now usesgit rev-parse --show-toplevel. Also replaces symlink-based hook install with a standalone wrapper script that delegates to the latest cached plugin version viasort -V | tail -1.pre-push-hook.shuses git for repo root -- replaceddirname "$0"+ relative path navigation (../../) withgit rev-parse --show-toplevel. Works regardless of invocation method (symlink, direct call, or delegated from hook wrapper).session-start.shhook install guard -- auto-install check now usesgit rev-parse --show-toplevelto find the project's.git/hooks/instead of checking relative to$PWD.
/vbw:releaseGitHub auth —gh release createnow extractsGH_TOKENfrom the git remote URL whengh authis not configured, instead of failing silently.- Statusline layout — moved Diff (
+N -M) from Line 4 to Line 1 after repo:branch. AddedFiles:andCommits:labels to the staged/modified and ahead-of-upstream indicators.
- Context Diet:
disable-model-invocationon 13 commands — manual-only commands (add-phase, assumptions, audit, discuss, insert-phase, map, pause, qa, release, resume, skills, todo, whats-new) no longer load descriptions into always-on context. ~7,500+ tokens/session savings. - Context Diet: brand reference consolidation —
vbw-brand-essentials.mdmade self-contained (~50 lines), removing 329-linevbw-brand.mdinjection from 27 command references. - Context Diet: effort profile lazy-loading — monolithic
effort-profiles.mdsplit into index + 4 individual profile files. Commands load only the active profile (~270 tokens/execution savings). - Context Diet: initialization guard consolidation —
plan.mdguard deduplicated to shared-patterns reference. - Script Offloading:
phase-detect.sh— new script pre-computes 22 key=value pairs for project state, replacing 7 inline bash substitutions inimplement.md(~800 tokens/invocation savings). - Script Offloading: SessionStart rich state injection —
session-start.shnow injects milestone, phase position, config values, and next-action hint viaadditionalContext(~100-200 tokens/command savings). - Script Offloading: compaction instructions — CLAUDE.md Compact Instructions section + enhanced
compaction-instructions.shwith main session detection guide context preservation during auto-compact. - Script Offloading: inline substitution cleanup — 10 inline
config.jsoncats removed from 6 commands (plan, execute, status, qa, fix, implement). Config pre-injected by SessionStart. - Agent Cost Controls: model routing — Scout→haiku, QA→sonnet (40-60% cost reduction). Lead/Dev/Debugger/Architect inherit session model.
- Agent Cost Controls:
maxTurnscaps — all 6 agents capped (Scout: 15, QA: 25, Lead: 50, Dev: 50, Debugger: 75, Architect: 30). Prevents runaway spending. - Agent Cost Controls: reference deduplication — 3 redundant
@references removed from agent files (~1,600 tokens/agent spawn savings). - Agent Cost Controls:
state-updater.shenhancement — auto-updates STATE.md plan counts when PLAN.md or SUMMARY.md files are written (PostToolUse hook, no LLM involvement). - Agent Cost Controls: effort-profiles and model-cost docs — updated for consistency with new Scout/QA frontmatter model fields.
/vbw:releasepre-release audit — new audit section runs after guards but before mutations. Finds commits since last release, checks changelog coverage against them, detects stale README counts (command count, hook count), presents branded findings with✓/⚠symbols, and offers to generate missing changelog entries or fix README numbers. Skippable with--skip-audit. Respects--dry-run./vbw:releasegit tagging — creates annotated git tagv{version}on the release commit./vbw:releaseGitHub release — creates a GitHub release viagh release createwith changelog notes extracted from the versioned section. Gracefully warns ifghis unavailable. Skipped when--no-push.- Statusline local commit count —
↑Nindicator (cyan) on Line 1 shows commits ahead of upstream.
- Statusline Line 1 consolidates all git/GitHub info — clickable
repo:branchlink moved from Line 4 to Line 1, replacing the duplicateBranch: Xfield. Staged, modified, and ahead-of-upstream indicators all on Line 1. - Statusline Line 4 cleaned up — removed duplicate GitHub link (now on Line 1).
- Statusline progress bars fixed and unified — all usage bars (Session, Weekly, Sonnet, Extra) now width 20, matching the Context bar. Previously Sonnet was width 10 and Extra was width 5, causing bars to render empty at low percentages (e.g., 7% × 10 = 0 filled blocks). Added minimum-1-block guarantee for any non-zero percentage.