CHANGELOG.md

Changelog

0.8.5

• Add trellis-exec login command for interactive OAuth inside Docker with credential persistence in a named volume
• Forward host plugins and settings into containers, run as non-root claude user
• Fix container auth failure caused by --bare flag skipping OAuth credential loading in newer Claude Code versions

0.8.4

• Fix container mode default network: change from none to bridge so the Claude CLI can reach the Anthropic API

0.8.3

• Fix duplicate binary name in Docker inner command that caused "Unknown command: trellis-exec" when using --container
• Auto-build Docker image when missing instead of failing with a cryptic pull error
• Add docker:slim and docker:browser npm scripts for easier image builds

0.8.2

• Reject plans with phase headings but no task items, falling back to LLM decomposition instead of silently producing empty phases

0.8.1

• Implement Docker container mode (Layer 4): --container launches docker run with the project mounted, re-invoking trellis-exec inside the container with full tool access and OS-level isolation
• Add container launcher module with pure buildDockerArgs and buildInnerCliArgs functions
• Add multi-stage Dockerfile (slim ~200MB, browser ~1.5GB with Playwright)
• Add container e2e tests with graceful skip when Docker is unavailable
• Add docs/container-mode.md with full documentation on mounts, networking, resource limits, and troubleshooting

0.8.0

• BREAKING: Safe mode is now the default. Agents run with granular permission controls instead of --dangerously-skip-permissions. Use --unsafe for legacy unrestricted access.
• Add permission controls: buildPermissionArgs() with safe, unsafe, and container modes. Judge and reporter agents are read-only in all modes.
• Add git checkpoints: automatic commit + tag before each phase for recovery on failure
• Add budget enforcement: per-phase cap via --max-phase-budget and cumulative run-level caps via --max-run-budget and --max-run-tokens
• Add container mode plumbing (--container flag and related options) for Docker-based isolation
• Add init-safety subcommand to generate reference safety config for interactive Claude Code sessions
• Strip tools: from all agent frontmatter; tool permissions are now controlled entirely by the execution mode via CLI flags
• Add default 30-minute phase timeout in safe mode when no explicit timeout is set
• Display budget usage in summary report when limits are configured

0.7.21

• Fix browser acceptance tester output parsing when CLI returns content block arrays instead of plain strings
• Add verbose logging of raw browser-tester output for debugging parse failures
• Strengthen browser-tester agent output format instructions to improve JSON compliance
• Track sub-agent token usage (judge, fix) in phase summary reports for accurate cost reporting
• Fix outdated model defaults and add --timeout documentation
• Fix ASCII diagram alignment in README

0.7.20

• Change default orchestrator model from sonnet to opus
• Add default values to CLI reference flags in README

0.7.19

• Make test auto-detection language-agnostic: support Python (pytest), Go, Rust, Ruby (rspec), Java (Maven/Gradle), Elixir, and Makefile projects
• Extend web app detection for Django, Rails, and Phoenix frameworks (two-signal heuristic to avoid false positives on API-only projects)
• Replace JS-specific prompt examples with language-neutral alternatives across prompts, agents, and skills

0.7.18

• Fix runSinglePhase missing orchestrator correction pre-application (--phase runs now apply corrections correctly)
• Split phaseRunner.ts into focused modules: judgeRunner.ts, browserRunner.ts, testDetector.ts
• Move RunContext type from cli.ts to src/types/runner.ts (fixes inverted dependency)
• Remove backward-compatibility re-exports from phaseRunner.ts
• Fix tryParseAssessment mutating input object in-place (spread into new object)
• Cache spec/guidelines content in RunContext to avoid repeated disk reads during prompt building
• Wrap realpathSync in try/catch to handle broken symlinks gracefully
• Add KNOWN_AGENT_TYPES enum for subAgentType validation
• Log SHA fallback in getChangedFiles instead of silently falling back
• Validate fallback JudgeAssessment through Zod schema
• Fix immutability violations in phaseRunner corrective task injection and planEnricher mergeResolvedField
• Enable noUnusedLocals, noUnusedParameters, noImplicitReturns in tsconfig
• Replace greedy regex in parseJudgeResult with iterative JSON.parse scan
• Add truncation notice to reporter prompt when diff exceeds 50k chars
• Invert context authority: learnings (Current Understanding) positioned before spec with anti-hack instructions
• Add orchestrator self-correction via corrections field in PhaseReport

0.7.17

• Fix markdown lint errors for blank lines around lists in README and docs

0.7.16

• Update README with browser testing architecture, auto-detection, verification pipeline, new CLI flags, and new agents
• Sync all docs with current implementation: fix orchestrator timeouts, remove stale REPL/worktree references, add browser and verification coverage to phase-runner and harness-comparison docs

0.7.15

• Fix browser acceptance empty-results loop: break on unparseable tester output instead of dispatching fixer on nothing
• Reinforce JSON output requirement in browser-tester prompt and agent definition
• Fix judge attempt numbering off-by-one (passed on attempt 0 → passed on attempt 1)
• Add judgeFixCycles to PhaseReport; combine with phaseRetries in summary report Retries column
• Rewrite summary table with minimal box-drawing (│ ─ ┼ delimiters)
• Show explicit message for empty browser acceptance results instead of 0/0 criteria passed
• Add constraint generalization instruction to orchestrator prompt for better cross-phase learning
• Strengthen plain-text output instruction in phase-orchestrator agent to reduce markdown leakage

0.7.14

• Remove dead phaseReport field from SharedState schema; make exitCode required in CheckResultSchema
• Extract shared judge/rejudge prompt helpers and applyJudgeOutcome() to reduce duplication
• Inline single-use buildSubAgentPrompt/buildSubAgentArgs into dispatchSubAgent()
• Consolidate getChangedFiles() to use single git status --porcelain call (halves git spawns for no-ref path)
• Batch applyReportToTasks with Map lookup instead of per-task O(n) scans
• Add prompts.test.ts covering buildRejudgePrompt, formatIssue, normalizeReport, parseJudgeResult, collectLearnings
• Add tests for reviewPhaseContract, detectTestCommand, selectJudgeModel
• Add edge case tests for devServer, completionVerifier, and stateManager (350 total tests, up from 309)

0.7.13

• Add judge corrections mechanism: judge can return corrections (e.g., targetPath renames) that update tasks.json before the completion verifier runs, eliminating false-positive failures for .module.css/.css and .jsx/.js mismatches
• Reorder phase runner flow: completion verifier now runs after judge (not before) so corrections are applied first
• Remove EXTENSION_VARIANTS hack from completionVerifier — path reconciliation is now handled by the judge
• Auto-inject CLAUDE.md scaffolding task into phase-1 during compilation, giving all agents persistent project orientation that survives context compaction
• Add browser smoke and dev server integration tests with Playwright fixtures
• Add foundational CLAUDE.md for the trellis-exec project itself

0.7.12

• Consolidate getChangedFiles/getChangedFilesRange and getDiffContent/getDiffContentRange into single functions with optional fromSha parameter, eliminating duplication and simplifying callers
• Extract 13 prompt-building and normalization functions from phaseRunner.ts into new src/runner/prompts.ts, reducing phaseRunner from 2,202 to 1,510 lines

0.7.11

• Add subAgentType-aware execution guidance to phase orchestrator (scaffold/implement/test-writer strategies)
• Add Task Type Summary section to phase context for orchestrator awareness
• Clarify classifySubAgentType() role as orchestrator hint, not automatic dispatch trigger

0.7.10

• Add project-level web app detection (detectWebApp) that checks for frontend framework deps, build-tool configs, and HTML entry points
• Fix requiresBrowserTest heuristic to propagate across phases for web app projects: sticky propagation once any phase has UI output, and last-phase guarantee so end-of-build acceptance tests always run

0.7.9

• Fix token/cost extraction to read from nested usage.input_tokens (actual CLI format) instead of non-existent top-level num_input_tokens, with fallback for legacy format
• Add extension-variant tolerance to completion verifier so .js target paths resolve when .jsx exists on disk, preventing infinite retry loops
• Skip corrective tasks in contract review to suppress false warnings about missing acceptance criteria and target paths

0.7.8

• Fix sub-agent CLI calls (judge, fix, etc.) failing silently due to missing --verbose flag required by Claude CLI for stream-json with piped stdin
• Skip fix-judge retry loop when judge sub-agent process itself fails, avoiding wasted corrective task cycles on infrastructure errors
• Add tests for sub-agent CLI failure handling and stream-json --verbose invariant

0.7.7

• Add two-tier browser testing with Playwright (optional peer dependency)
• Tier 1: per-phase deterministic smoke check (console errors, blank page detection, interactive element click test) runs before the judge on UI phases
• Tier 2: end-of-build LLM-driven acceptance tests generated from spec criteria, with browser-fixer retry loop (default 3 retries)
• Add requiresBrowserTest flag to Phase schema, set by compiler prompt and deterministic heuristic
• Add --dev-server, --save-e2e-tests, --browser-test-retries CLI flags
• Add language-agnostic dev server autodiscovery (Node, Python, Rails, Go, Docker Compose)
• Add browser-tester (Opus) and browser-fixer (Sonnet) agent definitions
• Feed browser smoke results to judge prompt as additional evidence
• Include browser acceptance results in end-of-run summary report

0.7.6

• Upgrade phase learnings to authoritative "Spec Amendments" that take precedence over spec assumptions
• Add constraint decision tier for runtime/toolchain facts discovered during implementation (never evicted)
• Reorder phase context so amendments appear after spec/guidelines, giving discovered constraints last-word authority
• Add structured handoff template (Architecture State / Deviations from Spec / Watch List) to orchestrator prompt
• Add consistent output style instructions to phase-orchestrator for [task-id] progress format

0.7.5

• Add end-of-run summary report showing per-phase time, task completion, judge results, retries, token usage, and cost
• Switch CLI subprocess output from --print to --output-format stream-json to capture token usage from Claude CLI result events
• Extract and accumulate token usage (input/output tokens, cost) per phase

0.7.4

• Randomize orchestrator spinner message from 10 fun labels instead of always showing "Orchestrating…"
• Reframe README tagline as phased execution harness

0.7.3

• Start judge attempt count at 1 instead of 0 for user-friendly output
• Use descriptive log message when orchestrator starts, keep spinner as "Orchestrating…"
• Adjust spinner frames so bounce endpoints show 1 bar instead of 0

0.7.2

• Fix false projectRoot/git-root mismatch warning on case-insensitive filesystems (macOS) by using realpathSync for path canonicalization
• Add success log for completion verifier so users can confirm it ran

0.7.1

• Fix projectRoot resolution: auto-detect git root instead of defaulting to tasks.json directory, preventing infinite retry loops when tasks.json lives in a subdirectory like .specs/feature/
• Add "all paths missing" diagnostic in completion verifier to fail fast with a clear error instead of snowballing corrective tasks
• Add early projectRoot sanity warning when resolved path is inside .specs/ or differs from the git root

0.7.0

• Relax sub-agent tool restrictions: implement agent gains Glob/Grep, scaffold agent gains Edit
• Increase default orchestrator timeout to 30 minutes, add --long-run flag for 2-hour timeout
• Add long-running phase protocol with intermediate commit reminders for reporter fallback safety
• Add lightweight completion verification pass (target path existence, TODO/FIXME scan) before judge
• Replace flat 20-entry decisionsLog cap with tiered learnings: architectural decisions never evicted
• Add pre-phase contract review that flags missing or vague acceptance criteria before execution
• Add long-running harness comparison doc analyzing Trellis against Anthropic research

0.6.2

• Fix caller mutation: deep-clone tasksJson in runPhases and shallow-copy ctx to prevent side effects
• Add stdin error handler in execClaude to surface backpressure/early close as a proper rejection
• Fail fast on unreadable spec/guidelines files instead of silently injecting error text into prompts
• Add guidelinesRef to LLM decompose fallback path in compilePlan
• Wrap loadState JSON.parse in try/catch with file path in error message
• Increase default orchestrator timeout to 15m, add --timeout CLI flag and reporter fallback sub-agent
• Allow judge to upgrade timed-out phases when committed work passes review
• Sync task statuses back to tasks.json after each phase completes
• Fix check command auto-detection to use range-based diff for committed test files

0.6.1

• Validate and rewrite all docs to match current native-tools architecture
• Archive 7 obsolete docs (REPL, worktree, old spec) to docs/_archive/
• Update README with all current CLI flags, compile options, env vars, and adaptive judge model

0.6.0

• Add adaptive judge model selection: Sonnet for small diffs (<150 lines, <3 tasks), Opus for larger work
• Add --judge flag (always|on-failure|never) to control when the judge runs
• Add --judge-model flag to explicitly override the judge model
• Use targeted re-judge prompt after fix (fix-only diff instead of full phase diff)
• Switch orchestrator sub-agent dispatch from CLI subprocess to native Agent tool
• Remove unused modifiedFiles and schemaChanges state fields
• Consolidate duplicate stripCodeFences utility

0.5.11

• Fix spinner bounce to show zero bars at both endpoints for symmetric animation

0.5.10

• Polish orchestrator spinner: add trailing ellipsis to label and smooth ping-pong with 1-frame dwell at bounce endpoints
• Fix spinner animation to bounce back and forth instead of jumping to start
• Normalize orchestrator spinner label to action word "Orchestrating…"

0.5.9

• Fix spinner leak: stop spinner in executePhase catch block so process exits after orchestrator timeout
• Add cross-phase learnings: surface decisionsLog entries from prior phases in orchestrator context
• Cap accumulated learnings at 20 entries with phase ID prefix for provenance
• Update phase-orchestrator agent to reference learnings and write forward-looking decisionsLog entries

0.5.8

• Add test coverage for streamParser, agentLauncher, and spinner (3 new test files)
• Expand phaseRunner tests with range-based judging and startSha tracking coverage
• Expand prompts tests with buildEnrichmentPrompt coverage
• Export formatElapsed from spinner for testability
• Fix markdown lint errors and migrate agnix config to tools array

0.5.7

• Add per-task git commits via orchestrator prompt (conventional commit format with scope and bullet summary)
• Add per-phase git commit in runner after judge passes, summarizing all completed tasks
• Track startSha/endSha per phase in PhaseReport for commit range tracking
• Fix judge silently auto-passing by using range-based diffs (startSha..HEAD) instead of working tree vs HEAD
• Add git helpers: getCurrentSha, ensureInitialCommit, commitAll, getChangedFilesRange, getDiffContentRange
• Add getPhaseCommitRange convenience function to state manager

0.5.6

• Use absolute path for phase report file in completion protocol so subagents resolve the correct location

0.5.5

• Stream orchestrator output in real-time using claude --output-format stream-json instead of buffered --print
• Add pause/resume to spinner so real-time output and elapsed-time indicator coexist cleanly
• Add NDJSON stream parser (src/ui/streamParser.ts) for extracting assistant text and result events

0.5.4

• Stream orchestrator stdout/stderr in real-time with --verbose flag
• Show actual recommendation (retry/halt/advance) in interactive prompt instead of misleading "continue" label
• Run judge even when phase fails if files were changed, preventing silent skip of judge/fix loop
• Default to "retry" instead of "halt" for transient failures like missing report files
• Store current phase report in state.json (phaseReport field) and clean up temp file after reading
• Auto-detect new test files after each phase and set check command from package.json or test runner configs
• Refactor execClaude to use options object instead of positional params

0.5.3

• Use Opus instead of Haiku for plan decomposition to handle ambiguous, architectural inputs
• Add stage-based progress messages during compile (parsing, decomposing, enriching, validating)
• Show elapsed time in spinner during long-running LLM calls
• Increase default compile timeout to 10 minutes with --timeout CLI flag for user override
• Add onStderr callback to execClaude() for real-time subprocess progress streaming
• Add simplified orchestration design document and native tools architecture doc

0.5.2

• Remove unused imports in phaseRunner test to fix lint warnings

0.5.1

• Fix executable bit on dist/cli.js lost during tsc rebuild, causing npx entrypoint to fail with "Permission denied"

0.5.0

• Replace REPL-mediated orchestrator with single claude --print invocation per phase using native tools (Read, Write, Edit, Bash, Glob, Grep)
• Orchestrator signals completion by writing .trellis-phase-report.json to disk instead of calling a REPL helper
• Remove worktree isolation (--isolation flag) — phases run directly in project root
• Remove turnLimit and maxConsecutiveErrors CLI options (no longer applicable without REPL turn loop)
• Add previous-attempt context to phase prompt for retries (last report, judge issues, corrective tasks)
• Extract git-diff helpers into standalone src/git.ts module
• Replace AgentLauncher dependency in compilePlan with simple query callback
• Delete replManager, replHelpers, worktreeManager and all associated tests (~8,700 lines removed)

0.4.10

• Add input validation to dispatchSubAgent() — returns descriptive error instead of crashing on wrong argument format
• Add writeFile(path, content) REPL helper for simple file creation without spawning a sub-agent
• Add task completion gate on writePhaseReport() — rejects reports missing any task from tasksCompleted or tasksFailed
• Add stuck-loop detection — intervenes after 4 identical REPL outputs with alternative approach guidance
• Update orchestrator prompts with writeFile docs, dispatchSubAgent worked example, and stricter completion rules
• Fix safePath ancestor resolution for symlinked tmpdir paths

0.4.9

• Fix REPL async IIFE wrapper silently dropping sub-agent return values, causing orchestrator to skip tasks and prematurely complete phases
• Wrap dispatchSubAgent and runCheck in sandbox with auto-reporting via capturedConsole so results always appear in REPL output
• Fix statement-form IIFE to return the last var declaration's value

0.4.8

• Add adaptive REPL timeout: 30s for sync expressions, 5min for long-running helpers (dispatchSubAgent, runCheck, llmQuery) so sub-agents aren't killed prematurely
• Strengthen post-timeout feedback to prevent orchestrator from hallucinating that timed-out work was completed
• Pre-load spec and guidelines content into orchestrator phase context to eliminate warm-up turns

0.4.7

• Accept structured judge output objects ({task, severity, description}) as the primary format, with plain strings as fallback
• Normalize detail → description field before validation to handle LLM field name variance
• Fix "continue" action incorrectly mapping to "halt" when the judge recommends retry
• Show retry counter (retries used: 1/2) in the interactive phase prompt
• Log a message when max retries are exceeded instead of exiting silently
• Remove habbit-tracker example files

0.4.6

• Add bouncing-bar spinner animation on stderr during LLM wait states (orchestrator launch, REPL turns, judge dispatch, plan compilation)

0.4.5

• Add normalizeReport() to validate and map orchestrator phase reports to the canonical schema, fixing Zod validation errors from LLM-style field names
• Detect and skip comment-only code blocks in the REPL turn loop; log raw orchestrator responses in verbose mode for debugging
• Increase verbose output limits from 200 to 500/1000 chars for code/results
• Add default file-existence check when no --check command is provided, verifying all phase targetPaths exist
• Include untracked files in getChangedFiles() and getDiffContent() so the judge reviews new files created by sub-agents
• Improve timeout error messages to prevent orchestrator from writing false "complete" reports after sub-agent timeouts
• Enforce all-tasks iteration in orchestrator prompts — writePhaseReport() must not be called until every task is attempted

0.4.4

• Document judge → fix correction loop in phase-runner.md
• Update README architecture diagram to show judge/fix as post-orchestration Phase Runner step
• Fix judge model reference (Sonnet → Opus) and remove stale env vars from README

0.4.3

• Fix markdown lint errors in docs (trailing punctuation, blank lines, code fence labels)

0.4.2

• Add comprehensive test coverage for critical execution paths (191 → 273 unit tests)
• New test files for replHelpers, extractCode, and compilePlan covering previously untested modules
• Add mergeWorktree tests for clean merges, conflicts, and missing branches
• Add CLI handler subprocess tests for handleStatus, handleCompile, and handleRun
• Add buildJudgePrompt edge case and runPhases halt action tests
• Export stripCodeFences from compilePlan for direct unit testing

0.4.1

• Fix interactive prompt to accept full words "retry"/"skip"/"quit" (not just single chars)
• Fix REPL variable scoping — var declarations now persist across eval turns for synchronous code
• Fix searchFiles to auto-detect glob patterns in first param instead of treating them as regex
• Make spec and guidelines file references explicit in phase context with readFile() examples

0.4.0

• Move judge invocation from orchestrator to phase runner — judge now runs as a system-controlled gate between phases using git diff for accurate changed-file detection
• Add lightweight fix agent (agents/fix.md) for targeted corrections from judge feedback
• Upgrade judge model from sonnet to opus for better reasoning on tricky issues
• Add bounded judge-fix correction loop (max 2 attempts) before surfacing issues
• Add getChangedFiles() and getDiffContent() git diff helpers to worktree manager
• Add progress logging during phase runner startup

0.3.12

• Fix cross-phase task dependency validation — tasks can now reference IDs from prior phases without being rejected as non-existent

0.3.11

• Store specRef, planRef, and guidelinesRef as relative paths in tasks.json, making it portable across machines

0.3.10

• Add projectRoot as required field in tasks.json, enabling tasks.json to live outside the project directory
• Introduce RunContext type that normalizes CLI flags + tasks.json refs into a single resolved config before execution
• Add --spec, --plan, --guidelines CLI override flags for the run command
• Refactor phaseRunner to accept pre-resolved RunContext instead of resolving paths internally
• Fix markdown lint errors in docs and README

0.3.9

• Format-agnostic plan compiler: accepts any well-structured technical plan, not just phase/task formatted plans
• New buildDecomposePrompt decomposes plans via LLM using full spec + plan + guidelines as context
• Add --guidelines CLI flag for the compile command
• Add optional guidelinesRef field to TasksJson schema
• Copy guidelines file to project root during execution, mirroring spec pattern
• Include guidelines reference in buildPhaseContext orchestrator context

0.3.8

• Replace spec-section-injection with spec file copy into project root for simpler, more reliable spec access
• Remove readSpecSections REPL helper and parseSpecSections — orchestrator now reads spec directly via readFile()
• Update phase-orchestrator and skill docs to reference readFile('spec.md') instead of readSpecSections()

0.3.7

• Fix readSpecSections to accept both array and varargs calling conventions
• Update phase-orchestrator prompt to direct agent to use pre-loaded spec sections
• Add test coverage for parseSpecSections, buildPhaseContext spec embedding, and varargs readSpecSections

0.3.6

• Pre-load spec sections into phase context prompt for --project-root compatibility
• Add graceful error handling to readSpecSections when spec file is missing

0.3.5

• Add --project-root CLI flag to decouple project root from tasks.json location

0.3.4

• Fix invalid git branch names when specRef is an absolute file path in worktree creation

0.3.3

• Add habit-tracker example specs (plan, spec, tasks, guidelines, pitch)

0.3.2

• Auto-fallback to LLM parsing when deterministic plan parser cannot identify phase boundaries

0.3.1

• Fix CLI entrypoint detection for npx and npm link by resolving symlinks with realpathSync
• Fix vitest config resolve option incorrectly nested inside test (ts2769)

0.3.0

• Rewrite orchestrator to use sequential --print --continue calls instead of persistent process
• Fix --agent-file → --agent CLI flag for sub-agents and orchestrator
• Add --dangerously-skip-permissions for headless sub-agent execution
• Disable orchestrator file tools via --disallowedTools to enforce REPL-only interaction
• Add extractCode() to parse JS from Claude responses, filtering natural language
• Add corrective nudge when orchestrator outputs natural language instead of JS
• Add REPL helper function docs to phase context
• Fix dryRun passthrough in executePhase (was hardcoded to false)
• Gate e2e CLI test behind TRELLIS_E2E_CLAUDE env var
• Split vitest config into unit and e2e with dedicated scripts
• Add docs/cli-integration-architecture-changes.md documenting the new architecture
• Update sub-agent prompt and agent files to use Write/Edit tools instead of text output

0.2.14

• Fix command injection in worktreeManager by replacing execSync with execFileSync
• Fix getState() ENOENT crash on first phase turn (returns empty initial state)
• Fix in-place mutation of phase tasks during retry (spread copies + unique IDs)
• Fix listener leak in agentLauncher orchestrator handle
• Add regex validation in searchFiles to prevent ReDoS and SyntaxError
• Add --enrich flag to compile CLI for opt-in LLM enrichment
• Fix cleanupWorktree running from inside the directory being deleted
• Track and clear sandbox timers on session destroy to prevent leaks
• Add 14 new security and edge-case tests
• Add docs/security.md documenting attack surface and mitigations
• Redesign README banner and add architecture diagram drafts

0.2.13

• Add Claude CLI pre-flight check with clear install message on failure
• Remove dead TRELLIS_EXEC_COMPACTION_THRESHOLD env var from help text
• Add tests for llmQuery default model and interactive mode promptForContinuation
• Document sub-agent permission enforcement model (Claude CLI --agent-file)
• Clean up stale TODO reference in skills doc

0.2.12

• Add e2e integration tests verifying §10 success criteria (compile, dry run, state round-trip, parallel scheduling, phase retry, handoff, REPL truncation, architectural validation)
• Add test fixtures: minimal Node.js test project, sample spec, and sample plan
• Add Group 2 claude CLI tests that skip gracefully when CLI is unavailable
• Add e2e integration tests documentation

0.2.11

• Package for dual npm CLI and Claude Code plugin distribution
• Add README with installation, CLI reference, architecture, configuration, and custom agents docs
• Update package.json with files, keywords, author, and prepublishOnly script
• Update plugin.json description and keywords
• Fix Zod 4 z.record() arity in agent linter

0.2.10

• Add eight orchestrator skills: compile, dispatch-agent, explore-codebase, manage-phase, quick-query, run, status, and verify-work
• Add skill architecture documentation explaining the rationale for skill-based orchestrator design
• Add agnix linter configuration
• Fix agent tools frontmatter to use YAML array syntax

0.2.9

• Fix agent frontmatter to use tools instead of allowed-tools per official Claude Code sub-agent docs
• Add Zod-based agent frontmatter linter with strict schema validation to catch unknown fields

0.2.8

• Add agent markdown files: phase-orchestrator, implement, test-writer, scaffold, and judge
• Update agent frontmatter to allowed-tools syntax matching current Claude Code format
• Add agnix and markdownlint-cli2 dev dependencies with lint:code, lint:md, and lint:agents scripts
• Add .markdownlint-cli2.jsonc config and fix all markdown lint errors across docs and agents
• Fix skill allowed-tools to space-delimited format per Agent Skills spec

0.2.7

• Add plan enricher (Stage 2) with targeted Haiku calls for ambiguous fields flagged by the deterministic parser
• Add prompt templates for enrichment and full-plan fallback parsing
• Add compilePlan pipeline wiring Stages 1–3: deterministic parse → targeted enrichment → fallback
• Add plan compiler architecture documentation

0.2.6

• Add CLI entry point (src/cli.ts) with run, compile, and status subcommands
• Argument parsing via Node built-in util.parseArgs() with environment variable fallbacks
• CLI flags override environment variables; environment variables override defaults
• Add bin field and build script to package.json
• Include compiled dist/ output for direct installation
• Add CLI reference documentation (docs/cli.md)

0.2.5

• Add phase runner — the deterministic outer loop that composes all sub-modules into the full execution pipeline
• Implements §6 Phase Runner Logic: phase iteration, orchestrator ↔ REPL turn loop, advance/retry/skip/halt action handling, worktree commits at phase boundaries, and resume from saved state
• Add phase runner documentation

0.2.4

• Add agent launcher module for managing claude CLI subprocesses (sub-agent dispatch, LLM queries, orchestrator sessions)
• Support real, dryRun, and mock operating modes for testability
• Add agent launcher documentation

0.2.3

• Add git worktree manager for isolating execution runs on separate branches (create, commit, merge, cleanup)
• Add check runner for executing user-defined verification commands as a deterministic gate after each task
• Add documentation for worktree manager and check runner modules

0.2.2

• Add REPL manager (replManager.ts) with vm-based sandboxed eval, expression-first async execution, output truncation, scaffold restoration, and consecutive error tracking
• Add REPL helper factory (replHelpers.ts) with filesystem helpers (readFile, listDir, searchFiles, readSpecSections, getState) and stubs for agent/LLM helpers
• Add REPL architecture documentation
• Add JSDoc comments to orchestrator, planParser, and scheduler functions

0.2.1

• Add deterministic plan parser (Stage 1 of plan compiler) that extracts phases, tasks, spec references, file paths, dependencies, sub-agent types, and acceptance criteria from plan.md without LLM calls
• Flag ambiguous fields for Stage 2 enrichment
• Move test fixtures to test/fixtures

0.2.0

• Add task scheduler with dependency resolution and parallel execution grouping
• Implement targetPaths overlap detection for implicit dependency serialization
• Add dependency validation (missing refs, self-refs, circular dependencies)
• Document scheduler grouping vs spec §10 #8 ordering rationale
• Fix invalid --verbose flag in typecheck script

0.1.6

• Add oxlint linter with lint and lint:fix npm scripts
• Fix no-empty-file lint violation in test setup

0.1.5

• Migrate test suite from node:test to vitest with expect-style assertions
• Add vitest.config.ts scoped to *.test.ts files only
• Add test/ directory for shared test concerns
• Add typecheck script with verbose tsc output
• Add tsconfig.test.json for typechecking test files separately from build

0.1.4

• Add state persistence layer (stateManager) with atomic writes and Zod validation
• Add crash-safe trajectory logger with JSONL append
• Enable @types/node in tsconfig and add tsx for test execution
• Add test suites for stateManager and trajectoryLogger (19 tests)
• Add .npmrc with save-exact config

0.1.3

• Add core data model types and Zod schemas for tasks, state, and agents
• Add zod as a dependency
• Switch package type to ESM

0.1.2

• Add Trellis RLM Executor spec v3 documentation
• Add gitkeep files for agents, hooks, and skills directories

0.1.1

• Add project scaffolding (package.json, tsconfig, .gitignore)
• Add plugin manifest and version bump skill
• Add Claude Code settings