CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

[0.26.2] — 2026-08-04

Added

• Added .ultra/project-brief.md as the raw owner-intake artifact and optional Research wayfinding.md navigation for unclear multi-lens paths, without adding a fifteenth Skill or another semantic authority.
• Expanded Research alternatives beyond direct competitors and made a single North Star metric optional when an observable outcome plus guardrails is more honest.

Changed

• Narrowed ultra-init to the Project Brief and empty skeleton, made ultra-research own the first accepted North Star, domain language, and specification baseline, and limited ultra-change to touched baseline sections.
• Made the first six Research lenses dependency-aware and kept Grilling, Think, and Domain Modeling as caller-bounded reusable methods rather than Research stages or owner routes.
• Routed ultra-status from semantic file contents instead of a Research completion bit or any single unresolved marker.
• Gave every Change one stable id across active, archive, and abandoned positions; made the task ledger append-only and every current reader active-id scoped.
• Added explicit none, bounded, and required Research Disposition contracts, one-task planning for quick Changes, and two-pass delivery against a fresh semantic plus product snapshot.
• Gave the exact Change Contract explicit Profile rationale and Risk flags fields so its required blast-radius decision has one stable representation.
• Kept ordinary technical seam selection model-owned and made delegation reachable for task execution, scoped Research evidence, and aggregate Change review without manufacturing task rows.

Fixed

• Made session context prefer the accepted North Star, fall back to the Project Brief before Research, retain legacy one-line compatibility, and always append active task acceptance when present.
• Bound Research distillates to the actual git hash-object value of all three baseline specifications so staleness is mechanically observable.
• Prevented archived or abandoned unfinished tasks, invalid Change ids, blocked dependencies, and ambiguous multiple in_progress rows from being injected as the current Hook acceptance.
• Made canonical change_id override a conflicting legacy change_ref, so migration compatibility cannot revive a historical task.
• Bound Test reports to exact Change id, ordered task ids, intent digest, HEAD, and a product-worktree digest that excludes the report and Change-directory metadata while binding current intent separately, so product reconciliation requires retest while delivery metadata and archive finalization do not self-invalidate.
• Prevented a separate request from opening a second active Change, and gave abandoned intents an exact owner decision, reason, reusable-evidence, and recovery closure read by future Change and Status workflows.
• Kept non-publishing package inspection diagnostic while the explicit Deliver gate blocks release-package creation only for an undispositioned changed export with no non-test consumer.
• Made that gate's authorization path use the finding disposition in the current Test report and carry the accepted decision into delivery.md.
• Aligned .npmignore with the explicit package allowlist and locked all eight maintained product documents into the tarball smoke contract.

Verified

• Passed 119 Node tests, 21 Hook tests, 14 Skill Creator validations, five isolated host install/Doctor/uninstall lifecycles, a 115-entry package dry-run inspection, and an audit with zero vulnerabilities.

0.26.1 — 2026-08-03

Added

• Added reframing to ultra-grilling, five bounded framing questions, and checkable resolved, stalled, and unavailable exits so a literal request can be translated into the role and outcome the owner actually means.
• Added HC-<n> hard-constraint identifiers, task-level constraint mapping, executable Change verification, and a supported changes/abandoned/ exit.
• Added owner-authorized cross-model-family review for major or security-relevant Changes, plus optional native fan-out inside independent research and test regions.

Changed

• Made ultra-dev converge on each task's mapped Change acceptance IDs, track the best-ever passing set, and stop after two no-progress rounds or three repair rounds.
• Expanded delivery reconciliation to every repository document and kept its wiring gate local to an explicitly invoked ultra-deliver; ordinary npm lifecycle commands never depend on project .ultra/ evidence.
• Made ultra-change read relevant archive history before opening intent, preferring delivery.md while supporting legacy archive summaries and verification records.
• Made all six review lenses the default and required every skipped lens to preserve its reason in the aggregate result.

Fixed

• Prevented delegated workers from modifying .ultra/, including when . grants the rest of an isolated checkout.
• Made dangerous-command authorization reachable by hashing the protected effect independently of its retry spelling.
• Kept additive protected-branch publication on the host's native authority path instead of hard-blocking an already owner-authorized release, retained exact-digest guards for history rewrites and branch deletion, and stopped treating inert searches as effects.
• Kept expandable interpreter heredocs inside dangerous-effect inspection while continuing to treat quoted commit-message heredocs as data.
• Required explicit owner authority before ultra-review can launch a cross-host model-family recheck.
• Bound ultra-dev convergence to the active Change acceptance IDs, with explicit stalled and unreachable exits instead of an open-ended repair loop.
• Made ultra-change recover archived intent from legacy archive evidence when the preferred delivery record is absent.
• Kept the ultra-deliver wiring gate local to an owner-invoked delivery workflow; ordinary package commands remain independent of project .ultra/ state.
• Added the missing changes/abandoned/ template directory and migrated completed WIP evidence links to their canonical runtime document.

Verified

• Completed an authenticated Claude-to-Codex continuation on one isolated file-first task: Claude banked the failing test and Resume Note without changing implementation, and Codex resumed from Git plus .ultra/, reached green, and closed the ledger and canonical evidence without conversation transfer.
• Verified the release candidate with 111 Node tests, 10 Hook tests, 14 Skill Creator validations, five healthy host Doctor reports, and zero audit vulnerabilities.

0.26.0 — 2026-08-01

Added

• Added ultra-delegate, five source-backed CLI profiles, and a background worker with bounded instruction, permission, worktree, logs, and three terminal result states.
• Added five file-first hooks for session context, acceptance recall, compact recovery, mechanical evidence observation, and exact-command dangerous-effect protection.

Changed

• Rebuilt the product around fourteen portable Skills: eight owner workflows, five model-invoked disciplines, and one router. Repository files and Git are now the complete cross-session and cross-host authority.
• Replaced all five host adapters with managed, provenance-checked file-first plugin artifacts and isolated install/doctor/update/uninstall behavior.
• Moved rule-side research, review, TDD, and autonomy material into focused Skill references and reduced the project skeleton to one canonical .ultra-template/.

Removed

• Removed the MCP server, SQLite runtime dependency, workflow state machine, orchestrator, operational CLI, command projections, custom-agent projections, and their obsolete schemas and tests.

Fixed

• Made --config-dir the isolation home for Codex plugin and personal-marketplace sidecars, and registered the plugin through Codex with the same isolated HOME and CODEX_HOME, preventing sandbox installs from mutating the real home directory.
• Namespaced multi-host --config-dir installs by runtime so one host cannot replace another host's skills/ or plugins/ tree.
• Made failed Codex registration restore the previous managed plugin and marketplace, or remove a fresh partial install.
• Rejected project-escaping task context and spec paths before hooks read them.
• Rejected delegated writable roots outside the named worktree and readable roots outside the current project before launching another CLI.
• Restored independent plan, development-evidence, and final-wiring checks so a locally green layer cannot stand in for an end-to-end product path.

0.25.1 — 2026-07-31

Fixed

• Published state-migration lock ownership atomically from a fully written and synchronized candidate inode, so concurrent Node 22 processes cannot mistake a valid publication or release window for a malformed lock. Stable malformed locks remain fail-closed.

0.25.0 — 2026-07-31

Added

• Added task_outcome / attest_commit as the sole model-facing path for recording checkout-local integrated commit proof after durable Task completion.

Changed

• Published exact Task identifier, numeric, array, object, digest, and runtime-field boundaries in the seven-tool MCP contract while keeping repository-defined Task, Change, priority, risk, research, and slice labels under model ownership.
• Applied the same exact-input boundary to Baseline, Change, Decision, Artifact, Event, Context, Checkpoint, Sync, Session, Archive, and Doctor calls. Unknown fields and wrong scalar, enum, array, object, digest, or scope types are rejected before SQLite affinity or downstream consumers can reinterpret them.
• Made Dev delivery publish durable completion before one integration commit, then attest that commit locally without changing the Git-facing team ledger.

Fixed

• Removed arbitrary Task patches and model-supplied commit hashes from task_outcome / complete, eliminating the remaining self-referential or bookkeeping-commit path.
• Fixed malformed semantic input and idempotency attempts escaping as transport or raw SQLite errors. Correctable requests now return mutable diagnostics and enter bounded rejected-attempt audit history without creating semantic authority.
• Fixed record, checkpoint, team-checkpoint publication, Session lease, Worker Packet, and archive receipt failures leaving database rows, managed files, worktrees, or Git checkpoint bytes partially committed. Each public mutation now commits its durable receipt with the owned state or compensates the external effect before retry.
• Fixed Task deletion reporting a generic draft-state error before identifying the exact durable Session, Artifact, checkpoint, Context, workflow-history, or recovery reference that must be resolved.
• Added a linked active successor path for immutable archived or cancelled Changes instead of requiring history mutation or direct SQLite repair.
• Removed the public Kernel's runtime dependency on the retired workflow supervisor; the legacy Change implementation remains compatibility-only and is absent from all five bundled host runtimes.
• Revoked an assigned Worker Packet when a parallel reservation is unwound, ensuring the abandoned capability cannot later complete a Task without a live Session.

0.24.2 — 2026-07-31

Changed

• Made the production Change facade single-mode and Kernel-only. The pre-v0.24 Change implementation is explicitly isolated for compatibility, while retired workflow and dialogue rows remain read-only history with no public artifact owner or invalidation writer.
• Made review and debug agents evidence-only: delegated workers can write only their assigned report, while the primary host owns source remediation and final judgment.

Fixed

• Preserved rejected semantic calls as bounded ultra_kernel_attempt audit history visible through ultra.context, without creating the rejected semantic authority.
• Renamed misleading _ultra.state_commit response metadata to _ultra.projection_commit; projection processing no longer resembles a semantic acceptance receipt.
• Removed current Change-packet decisions and artifact invalidation from retired workflow/dialogue tables while retaining safe migration and historical reads.

0.24.1 — 2026-07-31

Changed

• Kept semantic completeness, route selection, baseline sufficiency, and delivery acceptance in Skills and the host model. The MCP kernel now records those conclusions as advisory diagnostics while retaining hard failures only for structural authority, current bytes, digests, scope ownership, concurrency, permissions, and recovery.
• Made local delivery follow the evidence required by the actual Change route. Explicit omissions and their rationale are preserved without imposing a fixed Plan/Dev/Test/Review sequence on every archive.
• Made completed Tasks explicitly reopenable without erasing their prior completion commit or accepted history.

Fixed

• Fixed nonexistent or mismatched Task/Change checkpoint scopes creating ghost checkpoint authority; invalid scopes now fail before any Context or checkpoint write.
• Fixed shallow or contradictory evidence declarations being accepted without checking current managed artifact authority and stable file bytes.
• Fixed baseline semantic gaps, missing optional research, and not_run verification being misrepresented as mechanical authority failures.
• Fixed installed Change revision paths depending on source-only semantic supervisor modules excluded from the npm package.
• Fixed Hooks, Skills, templates, and packaged documentation disagreeing about advisory diagnostics, adaptive delivery, and the persistence/safety-kernel boundary.

0.24.0 — 2026-07-30

Added

• Added normalized Decision Records, reversible Stage Checkpoints, bounded Context Envelopes, and immutable Worker Packets as the shared evidence and handoff model across Skills, Hooks, agents, SQLite, and tracked .ultra artifacts.
• Added a native Grok Build plugin so Claude Code, Codex, OpenCode, Kimi Code, and Grok Build share the same eleven explicit workflows and seven-tool MCP kernel.
• Added schema 22 backup-first migration and real-history compatibility for v4.4, v4.5, v0.22, and v0.23 state and team-ledger formats.
• Added packaged better-sqlite3 provenance, platform and Node ABI verification, and durable Doctor and archive workers for every supported host.

Changed

• Completed the MCP boundary reduction: Skills own adaptive semantic procedure; MCP owns persistence, identity, transactions, leases, paths, recovery, and irreversible safety; Hooks only observe or inject bounded context; agents consume immutable Worker Packets without writing SQLite authority.
• Folded Plan artifact publication into the accepted Plan checkpoint so a later export cannot invalidate its own authority.
• Replaced semantic supervisor failures with mutable diagnostics and repairable drafts. Only corruption, unsafe paths, concurrency conflicts, permission failures, and irreversible external effects remain hard blockers.
• Made ultra.context, ultra.sync, and ultra.doctor own their inspection and repair preflight so an older valid ledger cannot block the tools required to migrate it.
• Reduced packaged and documented MCP contracts to the exact seven public tools; removed the retired fine-grained schemas, fixtures, and command bypasses.

Fixed

• Fixed accepted or stale workflow runs becoming permanently self-locked after an artifact changed, and made reversible stages reopenable without rewriting immutable accepted history.
• Fixed v0.23 team ledgers using research_run_id being rejected under the newer research_checkpoint_id digest contract.
• Fixed native SQLite drivers and Doctor workers being absent or unresolved in installed plugins.
• Fixed session acquisition, archive recovery, Context identity, and Worker Packet cleanup paths leaving partial authority after failure.
• Fixed installed interaction guidance failing to render each host's native question surface while keeping canonical Skills host-neutral.

0.23.0 — 2026-07-30

Added

• Added a seven-tool public MCP kernel: ultra.context, ultra.record, ultra.checkpoint, ultra.sync, ultra.session, ultra.archive, and ultra.doctor.
• Added idempotent batched recording, content-addressed Context snapshots, and semantic checkpoints that commit accepted workflow authority in one model round trip.
• Added mutable diagnostic envelopes for rejected draft checkpoints and archive preflights so the host model can correct or abandon the same draft.

Changed

• Moved workflow sequencing and semantic judgment back into the eleven public Skills. MCP now owns persistence, schema, transaction, concurrency, path, recovery, and irreversible-effect boundaries instead of directing the model.
• Reduced the unconditional tools/list surface from 60 fine-grained tools to seven high-level tools. The 0.22 fine-grained contracts remain callable but hidden for one compatibility release.
• Made draft Changes and nonterminal workflows reversible while keeping accepted revisions immutable and replaceable through a new draft.
• Updated Claude Code, Codex, OpenCode, and Kimi assets to use the same narrow capability graph while preserving each host's native question and bounded worker surfaces.

Fixed

• Fixed Plan drafts becoming permanently self-locked when a task or exported artifact changed after validation.
• Fixed semantic validation failures surfacing as MCP transport errors and forcing models into repeated repair calls.
• Fixed duplicate Context compilation creating competing snapshots for the same semantic inputs.
• Fixed ultra.context creating .ultra state merely by inspecting an uninitialized project.

0.22.0 — 2026-07-30

Added

• Added a versioned, digest-chained Git team checkpoint at .ultra/tasks/tasks.json for portable baseline, Change, task-contract, dependency, and durable task-status handoff.
• Added task.ledger_get, task.ledger_publish, and task.ledger_import with per-record revisions, bounded ancestry, transactional fast-forward imports, and typed baseline, Change, task, deletion, and active-session conflicts.
• Added exact-byte backup and deterministic upgrade of matching legacy v4.4 and v4.5 task projections during project resume.

Changed

• Moved generated task and context views under .ultra/.runtime/projections/; SQLite remains checkout-local operational authority while the Git checkpoint is the narrow team handoff.
• Updated all public workflow Skills, project templates, hooks, adapters, and documentation to inspect and synchronize the team checkpoint without turning hooks or host adapters into semantic authorities.
• Made imported ready baselines require checkout-local revalidation before they may publish a descendant checkpoint.

Fixed

• Fixed tracked task projections creating a self-referential BASELINE_HEAD_STALE treadmill after every Ultra operation or metadata-only commit. Baseline freshness now uses scoped source content plus Git ancestry while continuing to detect specification and source drift.
• Fixed task completion requiring a second bookkeeping commit. The completion SHA is now backfilled only into local SQLite and generated views after the single real completion commit.
• Fixed resume, doctor, status, and four-host install paths disagreeing about legacy task projections, team checkpoints, generated views, and protected files.
• Replaced the daemon's machine-speed-dependent 500ms release assertion with a bounded observable-session wait while preserving the immediate-poll behavior requirement.

0.21.1 — 2026-07-29

Fixed

• Increased the packed four-host integration-test timeout to preserve the same release gate on slower clean Linux runners instead of cancelling a healthy tarball installation and MCP round-trip near the previous limit.

0.21.0 — 2026-07-29

Added

• Added a trackable semantic artifact registry with digest, provenance, ownership, freshness, and orphan detection while keeping mutable SQLite, leases, telemetry, backups, and worktrees under the ignored .ultra/.runtime/ boundary.
• Added typed Change deltas, documentation reconciliation records, immutable task and role context snapshots, transactional plan publication, and self-contained delivery archives with crash-safe recovery journals.
• Added workflow revision and supersession contracts so accepted intent can evolve without rewriting history or leaving active documents detached from their owning workflow.
• Added schema 20 migration and backup-first recovery for legacy .ultra/state.db projects, including safe runtime relocation and inode-pinned archive finalization on supported POSIX platforms.

Changed

• Published one exact eleven-capability graph across all four host interaction contracts. Every adapter now presents the same inspect -> suggest -> host-native ask -> normalize -> persist -> apply -> read back flow, owns no semantic authority, and waits for explicit invocation before another public capability starts.
• Updated current workflow, authority, runtime, and project templates for the separate greenfield, brownfield-adoption, and migrated-authority paths.
• Made every Change converge through a current plan, DB-backed task contract, bounded context, risk-selected verification, review, documentation reconciliation, and local delivery archive. Research remains adaptive and may be omitted when current evidence is sufficient.
• Defined .ultra/ as project-local cross-session workflow memory: registered specifications and evidence carry semantic bodies, while MCP and SQLite own lifecycle state, references, digests, legal transitions, and recovery.
• Extended packed-package conformance to cover prompt hygiene, native interaction parity, exact public capabilities, user-handbook byte preservation, and project inertness before ultra-init.

Fixed

• Fixed archive, plan, session-close, worktree, and migration mutations so partial failures recover the prior files and state instead of reporting false success or leaving split authority.
• Fixed context and delivery gates accepting stale, cross-project, symlinked, malformed, unowned, or digest-mismatched artifacts.
• Fixed draft baselines exempting arbitrary files under .ultra/specs/; only the four reserved baseline scaffold files are provisional before baseline readiness, and every additional spec is diagnosed as an orphan immediately.
• Fixed host hooks and adapters resolving legacy state paths or becoming a second semantic authority. Hooks now observe active workflow state, protect projections, and surface recovery without forcing a semantic route.

Removed

• Removed the non-Ultra learn command and Skill, automatic session-to-Skill mining, and both global output-style personas from source, host adapters, and npm distribution. Ultra no longer creates user Skills or controls communication style.

0.20.0 — 2026-07-28

Added

• Added schema 19 and decision.complete so settled normalized intent reaches a terminal lifecycle state without fabricating a second user approval or an artifact checkpoint. Existing settled schema-18 threads migrate safely.

Changed

• Unified semantic selection across Claude Code, Codex, OpenCode, and Kimi Code as inspect -> suggest -> ask if unresolved -> normalize -> persist, using each host's native structured question surface with a direct-question fallback.
• Made research coverage and planning posture owner-selected after model recommendation while preserving autonomous fact finding, synthesis, decomposition, and reversible implementation judgment.
• Stopped requiring a ceremonial decision checkpoint after every normalized answer. Artifact checkpoints remain explicit digest-bound freshness gates.
• Made unanswered non-blocking follow-ups advisory instead of global workflow gates; unanswered blocking choices and blocking deferrals still fail closed.

0.19.1 — 2026-07-26

Fixed

• Made the four-host handbook-isolation release test hermetic so a clean CI runner does not require the Codex CLI merely to prove that installation and uninstallation preserve user-owned instruction files byte for byte.

0.19.0 — 2026-07-26

Changed

• Made every public Ultra workflow explicitly activated. Claude Code and Kimi Code disable model invocation, Codex keeps implicit invocation disabled, and OpenCode commands load private plugin workflow assets instead of exposing public workflows through its model skill catalog.
• Limited lifecycle context, health, checkpoint, stop, and subagent hooks to DB-authoritative active workflows. Initialized idle projects retain only generated-projection protection.
• Changed workflow handoff to return allowed transitions and a host-owned recommendation without launching another public workflow.

Removed

• Removed ubp-handbook, the shared handbook renderer, Kimi session bootstrap, and every package path capable of creating or rewriting user-level CLAUDE.md or AGENTS.md files.

0.18.0 — 2026-07-25

Added

• Added explicit ubp-handbook --full preview and apply modes that render one complete engineering contract with native Claude Code, Codex, OpenCode, and Kimi semantics, preserve supported provider-managed marker blocks, require a content-bound preview confirmation, and back up every replaced user handbook.

Changed

• Replaced the duplicated legacy root prompts with concise host-native repository guides. General engineering doctrine now has one renderer, while workflow detail remains in portable Skills and host wiring remains in adapters.
• Changed research coverage from a catalog-wide checklist to a model-selected evidence set. MCP persists only selected or explicitly recorded exclusions while retaining synthesis and evidence invariants.
• Codex hook manifests now call the stable managed plugin adapter, so new tasks do not retain an evictable cache-version path.
• Corrected Claude Code namespaced invocation, OpenCode Skill loading, and installation-doctor syntax in every rendered handbook.
• Full handbook convergence now preserves repeated and nested external provider regions, normalizes duplicate Ultra regions, follows dotfile symlinks, and retains the target POSIX mode.
• Extended prompt conformance checks so external Skill declarations cannot enter packaged Ultra Skills or worker prompts.
• Source installations now record dirty state and a deterministic worktree digest without inheriting an enclosing consumer repository.
• The npm package now ships a current decision contract and excludes the repository-only historical implementation plan.

Removed

• Removed stale external graph declarations, foreign interaction-tool, old context-budget, outdated version, and invalid path instructions from the active repository handbooks. External graph and memory capabilities remain separately owned.

0.17.1 — 2026-07-24

Fixed

• Fixed the broad templates/ ignore rule excluding the two generated-project report templates from clean npm packages. Both templates are now tracked explicitly, and the release suite rejects any locally present project template that is absent from Git.

0.17.0 — 2026-07-24

Added

• Added deterministic Git bootstrap for initialization and resume, including explicit auto, initialize, and skip modes, persisted unborn-HEAD state, and an owner-authorized first-checkpoint gate.
• Added a source-backed compatibility crosswalk against the original rocky2431/ultra-builder-pro workflow and a shell-free execute-plan consumer for current completed-plan dependency waves.
• Added schema 18 adaptive transition state, migration of active legacy init/change/plan/deliver runs, Context Manifest v3 Change-authority digests, and one host-specific interaction contract for every supported runtime.
• Added research coverage dispositions (execute, verify_existing, reuse, not_applicable, and accepted deferred) and risk-selected verification profiles with rationale for every excluded dimension.

Changed

• Changed plan execution into a resumable wave state machine. A wave pauses until its DB tasks converge, later dependencies remain pending, and reruns skip completed work.
• Changed MCP session spawning to create the single authoritative Git worktree and session record. Session close now preserves work by default, and change-owned auto-merge requires current dev and review evidence.
• Changed session worktrees to bind their ignored .ultra entry to the central authority and pass non-overridable DB, checkout, and authority-root paths to explicit workers.
• Changed initialization to complete after classification, local Git/scaffold setup, and read-back verification. It never starts research implicitly; baseline research begins only through an explicit invocation.
• Changed the workflow from a canonical next-action pipeline to an adaptive capability graph. MCP exposes valid alternatives and unique hard-recovery transitions; the active host owns semantic recommendations and native user interaction.
• Changed change capture to end after the accepted intent contract, made plan approval conditional on a material owner decision, selected test/review work by risk, and limited Ultra delivery to local convergence and archive.

Fixed

• Fixed process exit zero, worker spawn errors, and daemon failures bypassing task gates or losing session-linked failure and circuit-breaker evidence.
• Fixed daemon dispatch of stale tasks, unresolved dependencies, or overlapping declared files, and made explicit takeover terminate the prior worker before replacing its lease.
• Fixed initialization resume leaving corrected metadata or projections behind after a late failure; DB state, generated projections, and newly bootstrapped Git now roll back together.
• Fixed cleanup paths that could remove uncommitted or unintegrated worktrees, and fixed dependency waves that could start before prior task convergence.
• Fixed change-owned auto-merge trusting task status alone; integration now requires the exact completion commit, ready dev evidence, and current task review.
• Fixed execute-plan accepting empty, cyclic, duplicate-task, stale, or incomplete change plans. Change-owned execution now requires the exact healthy completed plan workflow and current DB task graph.
• Fixed direct MCP and daemon session creation bypassing the plan workflow. Admission and spawn now share the same current task-contract, dependency, and staleness gate before takeover or worktree mutation.
• Fixed plan-authority drift between wave selection and spawn being counted as a worker failure, and fixed isolated workers creating a second project DB.
• Fixed an isolated worker recursively spawning, taking over, or prematurely closing its parent-owned session; the supervising process now remains the only transport-settlement authority.
• Fixed late resume failure after a schema upgrade restoring the migrated intermediate DB instead of the exact pre-migration backup.
• Fixed the directory-only .ultra/ Git ignore form failing to protect an authority symlink from git add -A; fresh initialization uses the symlink-safe .ultra rule, while legacy repositories receive a repository-local info/exclude rule without changing their tracked baseline.
• Fixed semantic change.update operations leaving derived tasks and compiled contexts apparently current. A changed intent contract now marks tasks stale, invalidates Context v3, and records the exact invalidation evidence.
• Fixed stale tasks remaining usable through plan completion, dev startup, verification, delivery, or orchestrator admission after semantic invalidation.
• Fixed marker-only task reconciliation bypassing semantic invalidation. Clearing stale now requires a complete execution-contract rebind and records the current Change authority digest.
• Fixed review convergence accepting incomplete or fabricated worker provenance; every specialist is now selected or skipped with rationale, and completed workers must exactly match persisted specialist artifacts.
• Fixed MCP workflow contracts omitting public freshness, recovery, review, testing, and delivery failure codes needed by host adapters.
• Fixed runtime specification templates and workflow prompts retaining rigid-route wording or omitting durable step and decision-tool handoffs.

0.16.0 — 2026-07-23

Added

• Added schema 16 durable decision threads and items with exactly one current question, normalized owner answers, explicit reversible delegation, consequence- bearing deferral, supersession history, and artifact-bound checkpoints.
• Added nine decision.* MCP contracts, CLI mappings, events, valid/invalid contract fixtures, status/breadcrumb recovery, doctor health checks, and workflow gates.
• Added one canonical progressive-disclosure decision protocol shared by init, research, thinking, change, and planning without copying questionnaires into each Skill.

Changed

• Changed research to preserve all seventeen semantic steps as an internal coverage contract while exposing only the current owner decision or compact checkpoint.
• Changed change and plan formation to inspect facts autonomously, present one evidence-backed choice at a time, stop for the owner, and reuse one approved checkpoint instead of requesting equivalent confirmations.
• Changed dev, test, review, deliver, status, doctor, user handbooks, and lifecycle documentation to recover through the same decision authority and exact next route.

Fixed

• Prevented a decision thread for one change from blocking unrelated changes on the same baseline while preserving baseline-only and matching workflow gates.
• Prevented workflow steps or completion from advancing on an open, blocking-deferred, checkpoint-ready, stale, or superseded owner decision.
• Prevented prompt transcripts, static question queues, large research dumps, and repeated owner approvals from becoming project authority.

0.15.0 — 2026-07-22

Added

• Added schema 15 typed semantic records for selected research areas, including step-specific attributes, stable ids, evidence links, source anchors, stored source digests, synthesis trace validation, and current-source health checks.
• Added complete Change Contracts with executable acceptance, non-goals, public seams, recovery, unresolved decisions, profile rationale/risk flags, and explicit bounded research disposition.
• Added plan acceptance coverage and traceability gates, one-task enforcement for quick changes, verified specification-learning application, and a published ultra-baseline-reconciliation-v1 schema.
• Added task/change/plan review modes and machine-validated worker selection or skip provenance while preserving independent specification-fidelity and engineering axes.

Changed

• Unified every greenfield, brownfield, migrated, and daily route through ultra-change before planning. Initial planning can no longer bypass durable intent, risk, research, recovery, or acceptance authority.
• Reworked the workflow Skills into compact host-neutral execution contracts with focused references for research semantics, Change profiles, planning preflight, review modes, and delivery reconciliation.
• Made context budgets advisory while keeping evidence, authority, acceptance, specification freshness, review, and recovery gates deterministic and blocking.

Fixed

• Fixed research completion that previously stored prose and file hashes without a machine-consumable semantic index or stale-source detection.
• Fixed planning that could complete without proving every Change acceptance criterion had an executable owner or that each task traced to accepted evidence.
• Fixed approved learning and delivery reconciliation that could record status without verifying the applied target, before/after content, source anchor, and archive replay provenance.
• Fixed review artifacts that did not distinguish task, aggregate change, and plan scope or preserve why each bounded worker was selected or skipped.
• Fixed host review-prompt adaptation across intentional Markdown line wrapping and excluded tests, test fixtures, and generated Python bytecode from installed and published plugin payloads.

0.14.0 — 2026-07-19

Added

• Added one read-only JavaScript Context Spine reader and renderer for the authoritative .ultra/state.db, plus a packaged breadcrumb CLI shared by Claude Code, Codex, OpenCode, Kimi Code, and the MCP server.
• Added cross-host regression coverage proving stale JSON projections and context manifests cannot override the active database change or task.

Changed

• Changed task.parse_prd and task.expand to validate and atomically persist task objects derived by the active host model. Ultra MCP no longer starts a second model client or requires an Anthropic, OpenAI, or other provider key.
• Reduced Python lifecycle hooks and the generated OpenCode plugin to thin adapters over the canonical breadcrumb reader. Checkpoint and resume data is advisory recovery evidence rather than a second workflow authority.
• Made collaboration and verification Skills host-neutral, kept the current host responsible for delivery, and generated Codex metadata from the source Skill instead of maintaining duplicated prompt bodies.

Fixed

• Fixed authoritative-state and injected-prompt divergence caused by hooks reading .ultra/workflow-state.json or context manifests after .ultra/state.db had advanced.
• Fixed stale or incomplete state routing: missing projects route to ultra-init, prior schemas route to migration, and corrupt current schemas route to ultra-doctor without fabricating active workflow context.
• Fixed installed OpenCode context injection so it reads the bundled runtime with the installer-selected Node executable and remains portable outside the source checkout.
• Fixed Codex collaboration invocation and generated metadata quoting while preserving the bounded read-only advisor contract.

Removed

• Removed the private LLM client, planner prompt builders, and the @anthropic-ai/sdk and openai runtime dependencies.

0.13.0 — 2026-07-18

Added

• Added schema 12 repository snapshots, branch and dirty-worktree evidence, classification metadata, and a categorized baseline gap ledger with explicit blocker, ownership, acceptance, and resolution state.
• Added safe task.init_project resume semantics, bounded monorepo scope detection, projection-only import routing, explicit migrated-baseline replacement, and backup-first recovery for old state databases.
• Added explicit corrupt-state restore and rebaseline commands that quarantine the original database, WAL, SHM, and task projection and restore every moved artifact when recovery fails.
• Added approved incident break-glass records and mandatory post-incident baseline reconciliation gaps, plus durable archive-journal crash recovery.

Changed

• Made ultra-init the single route for empty repositories, existing codebases, healthy installations, incomplete adoptions, prior Ultra state, and corrupt databases. Existing application and baseline files are preserved; resume installs only missing current scaffold assets.
• New ordinary changes now require a healthy ready baseline. Already-active work remains executable with visible drift warnings, while convergence still requires current approved authority.
• Ready-baseline replacement now requires durable owner identity and rationale; a bare replace_ready boolean is rejected and cannot erase provenance.
• Reworked baseline templates and host-neutral Skills around Observed/Verified/Decided/Unknown evidence, characterization results, owner approval, and one deterministic next action.

Fixed

• Prevented prior-version projects from being treated as approved after a schema upgrade; compatibility baselines remain migrated/adopting until an evidence-backed brownfield replacement is approved.
• Made status and doctor report old or corrupt state as structured recovery guidance, retain pre-migration backup paths on failure, and close failed database handles safely.
• Prevented maintenance CLI telemetry from opening corrupt authority before evidence-preserving recovery checks, and made task creation enforce the same baseline boundary as change creation.
• Closed bulk-parse and task-expansion task-creation bypasses; persisted PRD tasks now carry their active change id, and expanded children inherit the parent task's authorized change ownership.
• Rejected task writes against missing or terminal changes even when the baseline is healthy, and enforced parent-child change ownership for generic task creation and reassignment paths.
• Fixed installed npm tarballs so direct ultra-tools task init-project calls can locate the packaged scaffold outside a source checkout.

0.12.1 — 2026-07-18

Fixed

• Fixed the published project scaffold so its task projection and neutral context template are tracked on fresh checkouts instead of being hidden by a broad ignore rule; the seed projection now uses the authoritative v4.5 state contract.

0.12.0 — 2026-07-18

Added

• Added schema 11.0 project baselines and four baseline.* MCP contracts for greenfield initialization, brownfield adoption, evidence recording, explicit approval, re-adoption history, and archive reconciliation.
• Added automatic new/existing repository detection, neutral baseline templates, metadata-only provider references, and cross-host baseline status/doctor output.

Changed

• Reclassified file count, token estimate, and context-share limits as advisory attention warnings. Necessary incident context and active work are no longer rejected or repaired by arbitrarily raising a threshold.
• Made incomplete or stale baseline state advisory during an active bounded change, requiring approved adoption at change.converge and complete revision/spec reconciliation in the atomic archive transaction.
• Changed the Stop hook from a workflow-completion denial into a lifecycle advisory; direct writes to the authoritative task projection remain blocked.

Fixed

• Prevented old projects from being initialized as empty greenfield products or using generated tasks.json as project metadata authority.
• Updated Claude Code, Codex, OpenCode, and Kimi presentation so hooks expose warnings without turning them into edit, stop, or workflow refusal gates.
• Removed the convergence deadlock where normal HEAD or tracked-spec changes made by the active change were rejected before archive could reconcile them; incomplete archive declarations now fail with full state and artifact rollback.

0.11.0 — 2026-07-18

Added

• Added a portable Skill authoring contract aligned with OpenAI Skills guidance and the Agent Skills progressive-disclosure model, plus regression coverage for source prompts, commands, agents, generated runtimes, and package contents.
• Added an independent review-spec worker and the ultra-review-findings-v2 / ultra-review-summary-v2 evidence contracts so specification fidelity and engineering standards remain separate release-gate axes.
• Added focused research and code-review references that are loaded only when the active workflow needs them.

Changed

• Rewrote the nineteen packaged source Skills as concise, English, host-neutral procedures with portable name and description frontmatter. Runtime metadata, invocation policy, and MCP dependencies now belong to host adapters.
• Reduced all eleven command documents to thin workflow launchers and rewrote the bundled workers as bounded, evidence-based roles without arbitrary confidence, coverage, severity, or option-count thresholds.
• Simplified Claude Code, Codex, OpenCode, and Kimi Code adaptation by removing stale prompt rewriting branches while preserving each host's native invocation, MCP, hook, and agent contracts.

Fixed

• Made the review waiter validate exact named v2 artifacts, reject partial or stale schemas, derive severity counts from evidence, and fail inconsistent two-axis summaries closed.
• Isolated read-only collaboration commands from user configuration and repository rules, and passed the review schema by an explicit absolute path instead of guessing an installed plugin root.
• Prevented Python caches and retired prompt/runtime artifacts from entering the npm tarball.

Removed

• Removed the duplicate ai-collab-base prompt framework, the seventeen-step research prompt tree, redundant review checklists, learned-Skill placeholder, and obsolete v1 verdict updater.

0.10.0 — 2026-07-18

Added

• Added Context Manifest v2 and a DB-derived Context Spine across planning, implementation, checking, review, convergence, and recovery. Context snapshots now carry role, gate, readiness blockers, bounded references and digests, fresh-context budgets, a public seam, an exact verification command, and one deterministic next action.
• Added change.breadcrumb, change.learning_propose, and change.learning_resolve, advancing the live MCP contract to 32 tools and .ultra/state.db to schema 10.0 with 16 tables.
• Added approval-gated specification-learning candidates. Proposed or approved discoveries block convergence until rejected or applied to a declared baseline target, and change.get exposes their authoritative state.
• Added independent spec_fidelity and engineering_standards review evidence, plus public-seam and exact red/green signal fields for convergence.

Changed

• Rebuilt the daily harness around one status router, one highest-value alignment question, fresh-context tracer bullets, explicit expand contracts, independent check/review contexts, and convergence before release.
• Reduced the seven core execution workflow prompts from 1,922 to 758 lines and added a 220-line ceiling for every public Ultra skill so host prompts cannot silently accumulate duplicated procedure. All twelve public workflows are now explicitly user-invocable; four rule skills remain agent-only.
• Session-start, pre-edit, and recovery hooks now share one read-only context_spine.py helper and inject only the compact DB breadcrumb. Intent, transcript, memory-provider, and graph-provider payloads never enter hook context.
• OpenCode now consumes the DB-generated Context Manifest v2 projection, strips provider/intent content, and fails context readiness when its compiled HEAD differs from the current repository.
• Updated Claude Code, Codex, OpenCode, and Kimi Code native renderings, managed handbook blocks, MCP mapping, architecture, and compatibility documentation.

Fixed

• Prevented plan-role task context from becoming ready without an execution contract, public seam, and verification command.
• Prevented breadcrumbs from reusing a snapshot after task state or git HEAD changed, and made an explicitly missing change id fail closed.
• Added auditable 9.1 → 10.0 migration history while preserving existing runtime rows, references, projections, and foreign-key integrity.
• Corrected hook conformance to package context_spine.py as an imported helper without registering it as a duplicate lifecycle hook.

0.9.0 — 2026-07-17

Added

• Added Kimi Code 0.26+ as a fourth native plugin host with namespaced commands, twenty allowlisted skills including its session bootstrap, lifecycle hooks, bundled review-worker prompt templates, native plugin registration, and read-only install diagnostics.
• Added packed-package and live MCP conformance coverage for the Kimi-managed plugin layout while preserving unrelated registry entries and config.toml.

Changed

• Advanced .ultra/state.db to schema 9.1 so Kimi is a valid runtime for events and sessions. Existing databases rebuild only the constrained tables transactionally while preserving rows, foreign keys, telemetry, incidents, and migration history.
• Extended CLI, handbook, orchestrator runtime selection, package metadata, and runtime compatibility documentation to include Kimi without reintroducing Ultra-owned memory, graph, browser, deployment, or framework capabilities.

Fixed

• Made the Kimi MCP launcher recover the active project from the inherited working-directory contract even though Kimi starts plugin MCP processes from the managed plugin root. It fails closed instead of creating .ultra inside the installed plugin.
• Avoided Kimi's embedded Node native-module ABI mismatch by launching the bundled MCP with the installer-compatible Node executable from PATH.

0.8.1 — 2026-07-17

Fixed

• Removed a retired-surface name accidentally reintroduced in the D53 decision note. The unchanged retirement contract caught it before v0.8.0 reached npm; this immutable patch release carries the complete harness closure forward.

0.8.0 — 2026-07-17

Added

• Added a first-class incident debug lane: incident changes now create a registered diagnosis.md artifact whose reproduction, hypotheses, root cause, regression-test, and recovery sections are required by convergence.
• Added normalized install-provenance manifests for Claude Code, OpenCode, and Codex plus read-only ubp --doctor [--json] checks for asset hashes and host-specific plugin, hook, MCP, launcher, and runtime entry points.

Fixed

• Turned workflow compaction checkpoints into a real recovery consumer: resume now rejects malformed or terminal checkpoints, selects the newest valid live/checkpoint state, and atomically restores missing or older live workflow state before re-injecting context.

0.7.0 — 2026-07-17

Added

• Added ultra-change and ultra-doctor as native workflows on Claude Code, OpenCode, and Codex. Daily fixes, features, redesigns, and incidents now stay attached to an explicit change packet after the initial baseline delivery.
• Added seven change.* MCP tools and system.doctor, advancing the live MCP contract to 29 tools and .ultra/state.db to schema 9.0 with 15 tables for changes, artifacts, context snapshots, trace links, incidents, projection jobs, and durable event-consumer cursors.
• Added deterministic convergence and archive gates for task completion, spec deltas, tests, documentation impact, review evidence, context freshness, and baseline reconciliation.

Changed

• task.init_project now initializes state.db immediately and creates active and archived change roots. tasks.json remains a generated projection.
• MCP mutations now commit authority first, enqueue a durable projection job, process it, and expose projection status in MCP metadata instead of silently swallowing projection failures.
• Session-start and edit hooks now understand initialized baselines and active changes. Health checks surface incidents, projection lag, orphan sessions, and missing change artifacts; direct writes to the task projection remain blocked even outside a command-scoped workflow.
• Memory and code-graph content are exclusively owned by separately installed providers. Ultra accepts only provider metadata references in compiled change contexts and never stores recalled content, graph payloads, or embeddings.

Fixed

• Wired spec-staleness consumption, crash recovery, projection processing, and durable cursors into the real orchestrator loop instead of leaving production helpers unconsumed.
• Kept intent.md synchronized with authoritative change updates, linked task status events to their newly assigned change, and included plan.export events in projection accounting.
• Added atomic projection claims and backup-first recovery for failed or stale interrupted jobs so a crashed projector cannot remain invisibly running.
• Kept all three native plugin bundles, MCP launchers, agents, hooks, commands, and user-handbook renderings on one explicit Ultra-owned asset allowlist.

Removed

• Removed the internal filesystem graph watcher and its dependency; external graph providers own indexing and refresh.
• Removed the retired command-proxy detector, installer option, managed handbook block, cache artifacts, prompts, configuration, and documentation surface.
• Stopped publishing the archived pre-CLI manual in the npm package; it remains repository history and cannot be mistaken for an active plugin contract.

0.6.0 — 2026-07-17

Changed

• Reduced the supported host matrix to Claude Code, OpenCode, and Codex.
• Reworked ultra-verify into independent two-model verification: the current host remains primary, one read-only external advisor supplies a second view, and evidence rather than model voting determines confidence.
• Restricted orchestrator routing, MCP session/event inputs, telemetry writes, state schema enums, installer flags, package metadata, and handbooks to the three supported runtimes.

Removed

• Removed the retired fourth-runtime adapter, installer surface, extension and conformance packaging, pricing entries, collaboration companion, and all active Ultra prompts and documentation that offered it as an advisor.

Fixed

• Added a retirement regression contract so unsupported runtimes cannot re-enter active source, package, schema, prompt, or generated plugin surfaces.
• Rejects unsupported runtime labels before state, event, or telemetry writes, including old databases whose historical SQLite constraints were broader.

0.5.3 — 2026-07-17

Fixed

• Removed the final bundled agent instruction that asked workers to update and consult project memory. Test workers now return reusable findings to the parent agent, while persistent memory remains owned by a separately installed host provider.
• Added a release regression guard that rejects private memory ownership, retired internal memory paths, and obsolete recall instructions from every bundled agent.

0.5.2 — 2026-07-16

Fixed

• Preserved Codex hook entry points for already-running tasks across plugin refreshes, including cache versions whose adapter file or entire cache directory had already disappeared. Historical cache versions are now kept in the managed runtime manifest and restored as validated forwarders to the current adapter.

0.5.1 — 2026-07-16

Changed

• Made task lifecycle workflows fail closed when MCP state is unavailable or reports an authority conflict. Removed the nonexistent task create/update/list CLI fallbacks from Claude Code, OpenCode, and Codex workflow assets.
• Advanced the state schema to 8A.2 so estimated_days is authoritative and survives MCP create/update, projection, and legacy migration.

Fixed

• Fixed the v4.4→v4.5 migration for the real Ultra task shape: top-level version, dependencies, relative contexts/... paths, date-only project timestamps, and task rows without individual timestamps.
• Added a state-authority gate that returns LEGACY_STATE_MIGRATION_REQUIRED when a non-empty v4.4 task projection meets an empty state database, instead of returning a misleading empty task list.
• Made migration task/event inserts atomic, refused merges into a non-empty task table, projected immediately after success, and preserved context bodies while removing obsolete duplicate status banners and retired memory-hook references.
• Added deterministic projection-write guards for Claude Code Edit|Write, OpenCode tool.execute.before, and Codex Edit|Write|apply_patch during active Ultra workflows.

0.5.0 — 2026-07-16

Added

• Added first-class native Claude Code and OpenCode plugin builders alongside the Codex personal plugin. All hosts build from one explicit runtime asset allowlist; the then-supported compatibility extension used the same boundary.
• Added ubp-handbook for previewed, backed-up managed-block integration with user-level Claude CLAUDE.md, Codex AGENTS.md, and OpenCode AGENTS.md.
• Added ultra-tools legacy-memory inspect|archive|prune; prune requires the exact DELETE_ULTRA_LEGACY_MEMORY confirmation token and never runs during install or update.
• Added release verification for a freshly packed npm consumer, including executable CLI links and live Claude Code, OpenCode, and Codex MCP round trips.

Changed

• Reduced the package to ten Ultra-owned public workflows, four internal agent-rule skills, and host-specific collaboration companions. The internal code-review-expert skill is no longer implicitly user-invocable.
• Replaced the imported hook suite with seven workflow-only hooks that are no-ops outside an active .ultra/workflow-state.json. OpenCode uses native JavaScript lifecycle hooks; Claude Code and Codex use native plugin manifests.
• Upgraded @anthropic-ai/sdk to the patched ^0.111.0 line.
• Reduced the published MCP contract to the 21 tools the server actually exposes; review, impact, skill discovery, and user interaction use native Host surfaces.

Fixed

• Made MCP database initialization lazy, synchronized the handshake version with the npm package, and fixed session.subscribe_events sid filtering.
• Aligned session/plan input schemas with runtime behavior, including circuit breaker output, session spawn intent semantics, and removal of ignored inputs.
• Bundled the compatibility MCP runtime durably and restored project-local .ultra/state.db ownership across all adapters.
• Exported the documented ultra-tools executable and included every README documentation target in the npm tarball.

Removed

• Removed Ultra's memory MCP tools, store, wrapper, recall skill, memory-capture and summary hooks, and private agent-memory declarations. Persistent memory is now solely the responsibility of a separate cloud-mem/claude-mem plugin.
• Removed bundled external browser, discovery, deployment, framework, and inherited design-skill assets from Ultra output.
• Removed generic dangerous-command and post-edit governance hooks from the Ultra plugin; unrelated user/repository hooks remain outside this package.

0.4.0 — 2026-07-15

Changed

• Replaced the deprecated Codex prompts/ and config.toml marker projection with a complete personal plugin: 25 adapted skills, nine legacy-command maps, plugin MCP registration, and current native hook events.
• Converted all nine Ultra agents to native .codex/agents/*.toml definitions.
• Renamed the cross-model skill from codex-collab to cc-collab; Codex remains the primary agent and Claude Code is an explicitly requested, read-only advisor.
• Made hook memory, compaction, health, subagent, transcript, and summary behavior runtime-aware. Codex no longer writes fallback state under ~/.claude or launches a nested model CLI for session summaries.
• Bundled the Codex MCP and status CLI with @vercel/ncc, preserving the active task cwd for project-local state. Generated Codex plugins now distinguish the 24 live MCP tools from nine upstream scheduled contracts and document each Codex-native replacement.

Fixed

• Preserved live Codex task hooks across plugin cache refreshes by restoring retired adapter paths as forwarding shims to the current cached adapter. Failed refreshes recover through the managed plugin source, and corrupt runtime manifests now stop uninstall before managed assets are removed.

0.3.0 — 2026-04-18

Added

• Phase 8B — executor line (D48, 8224159): parallel session orchestrator that consumes .ultra/execution-plan.json waves, spawns sessions per dispatch-rules.cjs (declarative priority-sorted rule table), manages N concurrent git worktrees via worktree-manager.cjs, and auto-merges session branches back to main with conflict detection via auto-merge.cjs. Events: wave_started / wave_completed / plan_completed / merged_back / merge_conflict. Opt-in autoMerge on runPlan and closeSession.
• Phase 8A — planner line (D47, a932cb8): task.parse_prd with dual-provider LLM client (official @anthropic-ai/sdk + openai); lib/topo.cjs Kahn + Tarjan SCC for dependency waves; task.expand atomic subtask creation; lib/plan-store.cjs for atomic .ultra/execution-plan.json write / section projection; skills/ultra-plan PRD-Direct workflow with human-gate via dry_run parameter. Schema bump 7.1 → 8A.1 for tasks.parent_id.
• Phase 7 — intelligence layer (D46): memory_entries FTS5 store with auto-recall on session spawn and auto-retain on close (event-type heuristic, zero LLM cost); tagged task lists per git branch (deriveBranchTag); skill mining to skills/learned/*_unverified.md on task completion / breaker trip / session crash.

Changed

• STATUS_TRANSITIONS contract now strictly requires pending → in_progress → completed — parallel-orchestrator transitions tasks to in_progress before the child's exit code may flip them to completed/pending.
• daemon.routeTask is now a thin wrapper over dispatch-rules.evaluate (6 default rules: breaker-blocked / deps-not-ready / no-runtimes / wave-conflict / by-preference / fallback-first-available).
• closeSession gains autoMerge / mergeBaseBranch opt-in params.

Fixed

• D49 tech-debt sweep: stale Phase comments ("scheduled for Phase 1", "not implemented in Phase 3.1") replaced with accurate unknown verb messages; docs/ROADMAP.md status table synchronized.

Tests

• 397 tests total across state (182) + orch (103) + spec (6) + rest (106). Zero regressions from v0.2.

0.2.0 — 2026-04-17

Added

• Phase 6 — monitoring + live code graph (D44): token / cost telemetry with per-runtime pricing table, telemetry table + daily .ultra/telemetry/YYYY-MM-DD.jsonl; ultra-tools status --cost panel with by-runtime / by-task / by-session aggregation; chokidar-based then-experimental incremental graph watcher with debounce and batch thresholds.
• Phase 5 — execution resilience (D43): recovery.cjs boot-time orphan scan using live PID probe; circuit_breaker table with recordTaskFailure counter and resetCircuitBreaker escape hatch; spec-change staleness propagation via cursor-consumed spec_changed events; orchestrator/daemon.cjs polling loop with routeTask by complexity hint; bin/orchestrator.js run/start/stop/status subcommands, gated by settings.json#orchestrator.auto_dispatch.
• Phase 4.6b — full conformance suite (D45): 4 runtime × 5 capability (command / skills / hooks / MCP + no env._source leak / install idempotency) = 20 additional conformance tests; table-driven resolve-target.test.cjs (21 tests) covering configDir > env > home/cwd precedence and extension root resolution; schema bump 5.2 → 7.1 alongside memory store.

Changed

• _source leak fix (D45 P2 #9): four adapters now emit a sibling _ubp: {source} envelope instead of polluting env._source; Codex uses a marker-fence instead of nested TOML object.

0.1.0 — 2026-04-17

Added

• Phase 4.5 — execution-lite (D42, 0d3e5ed): 7 session.* MCP tools (spawn / close / get / list / admission_check / heartbeat / subscribe_events); orchestrator/session-runner.cjs creates real git worktree + child process; sessions table holds lease / heartbeat (no lease.json file — D32); admission control with takeover / resume / abandon strategies (D33).
• Phase 4 — cross-runtime distribution (D41, 5aa1fd0): shared adapter toolkit (adapters/_shared/ — file-ops / frontmatter / settings-merge / path-rewrite / md-to-toml); the initial runtime adapters (Claude sentinel-block settings.json merge; OpenCode opencode.json.mcp; Codex config.toml marker-block + prompts/; plus the then-supported compatibility package); docs/RUNTIME-COMPAT-MATRIX.md with 10-section capability matrix; install / uninstall round-trip tests + idempotency.
• Phase 3 — thin-shell commands (D40, b3d1797): 9 commands migrated to skill + MCP + CLI three-layer (36-54 LOC per command); 7 new skills under skills/ultra-*/; task.init_project MCP tool with .ultra-template/; docs/AGENT-CONTEXT.md canonical runtime contract.
• Phase 2 — authoritative state (D39, e286e41): .ultra/state.db SQLite
  • WAL with 7 tables; mcp-server/lib/{state-db, state-ops, projector}.cjs write API and state machine; mcp-server/server.cjs stdio MCP server with 7 task.* tools; ultra-tools db (init/checkpoint/vacuum/integrity/ backup) and migrate (v4.4→v4.5 with dry-run + rollback).
• Phase 1 — spec contracts (D38): spec/ locks the three-layer contract — 30 MCP tools in mcp-tools.yaml, 7-table SQLite schema, skill manifest, CLI protocol + mapping table; 5 spec validators.
• Phase 0 — skeleton: multi-runtime installer scaffolding.