v1.6.0

Minor Changes

• d4640e2: context/workflow/ is now context/config/, and the two templates moved out of review-policies/ into context/config/templates/, renamed review-policy.md (was policy-template.md) and review-report.md (was report-template.md). review-policies/ now holds only Policies, so the rule that a .md ending in -template.md is not a Policy is removed: every .md there is a Policy. /dx-init seeds context/config/{review-policies,templates}/ (missing files only), and /dx-review and /dx-review-triage read from the new paths. Existing projects: re-run /dx-init to get the new tree, then delete the old context/workflow/ by hand — nothing is migrated, so your edited Policies are not copied over; move them into context/config/review-policies/ yourself.
• d4640e2: The plan gate is now a Policy: run /dx-review plan <change-id> for the optional pre-implementation review. Breaking: /dx-plan-review is removed with no alias. The built-in plan Policy keeps the same four dimensions (Substance, Feasibility, Architectural fitness, Standards-fit), their sub-checks and the report-only behaviour, but the report now opens with a per-dimension Verdicts block and tags findings like [Feasibility: Blocker]; the single sound / revise / rethink verdict line is dropped. /dx-plan now prints Next: /dx-review plan <change-id>, /dx-review-triage <id> plan reads only reviews/plan.md, and the unknown-policy error tells you to re-run /dx-init. Existing projects: re-run /dx-init to seed context/config/review-policies/plan.md (missing files only; your edits are never overwritten). Old reviews/plan-review.md reports are not migrated and triage and archive ignore them, so triage any PENDING findings in them before upgrading, or re-run /dx-review plan <change-id> to get a report in the new shape. The review-report reference topic of dx-references is removed; review-report.md is the only report schema. Upgraded installs keep a stale dx-plan-review skill until you remove it by hand.
• d4640e2: Policy-driven review: new /dx-review <policy-id> [container-id] [instructions…] runs a review whose criteria come from a Policy file in context/config/review-policies/. Breaking: /dx-impl-review is removed with no alias — run /dx-review implementation <change-id> instead, which keeps the same guard, dimensions, verdicts, status: reviewed on pass, lesson offer and dx-diagnose hand-off, and now also offers /dx-standards-update when a finding repeats a lesson or exposes a standard gap. /dx-review-triage now takes <container-id> [policy-id] instead of [plan|impl] (impl becomes implementation; plan still triages plan-review.md), and /dx-archive warns on PENDING findings in any reviews/*.md. /dx-implement and /dx-tdd now print Next: /dx-review implementation <change-id>. Existing projects: re-run /dx-init to seed implementation.md, review-policy.md and review-report.md (missing files only; your edits are never overwritten, so later fixes to shipped Policies must be merged by hand). Old reviews/impl-review.md reports are not migrated and triage ignores them — re-run /dx-review implementation <change-id> to get a report in the new shape. Upgraded installs keep a stale dx-impl-review skill until you remove it by hand.
• d4640e2: New built-in review Policy: /dx-review sessions [instructions…] is a report-only retrospective that reads agent session transcripts (the current session by default) and prints ranked suggestions for improving the coding agent's environment, then stops. Policies gain an optional promote: off line in ## Candidates: an Explore run then prints its Candidates and stops, with no promote question, no Next: line and no lesson offer, and type: becomes optional. Default behaviour, including test-strategy, is unchanged. Existing projects: re-run /dx-init to seed sessions.md (missing files only). /dx-init never overwrites your review-policy.md and review-report.md, and a fix to a shipped Policy or template never reaches an existing copy, so compare your review-policy.md and your review-policies with the shipped ones in the skill's assets/config/ and merge promote: off by hand.
• c94fdad: /dx-implement auto mode now runs in its own git worktree (.worktrees/<change-id> on a scratch branch) and fast-forwards your work branch on success, so several changes can run at once without touching your checkout. Phases that run concurrently each get a sibling worktree, merge back in dependency order (a resolver agent handles conflicts), and the group's (integrated) checks run once on the merged tree; subagent notes go to handoff.md as each returns. Worktrees are kept on any stop and removed only on full success. /dx-plan tags #### Automated rows (isolated) or (integrated) and emits an integration phase per parallel group. A new shared start-check reference (used by /dx-implement and /dx-tdd) resolves the work branch, commits the change's context/ files and stops the one-phase modes when a stopped auto run left a kept worktree. /dx-init adds .worktrees/ to .gitignore, copies the global standards only when context/standards/ didn't already exist (otherwise it asks which to add), and writes a one-line dx- pointer inside a managed <!-- dx:workflow:start/end --> block in AGENTS.md (when CLAUDE.md only points at it) or CLAUDE.md, no longer writing the rollback principle. Existing projects: add .worktrees/ to .gitignore or re-run /dx-init; old rollback text in CLAUDE.md is left as is.
• 32f1f0d: Four fixes from the sessions review. /dx-plan labels every Manual check (agent-runnable) or (user-only) and aims for a plan the agent can verify end to end; /dx-implement and /dx-tdd run agent-runnable rows themselves and tick them on evidence, and a user-only row ticked on bare say-so is recorded — ticked on say-so. /dx-plan gives each phase a Depends on: line so independent phases can be separated, and /dx-implement <change-id> now runs every pending phase on subagents in dependency order by default (independent phases concurrently, the coordinator owns Progress and the per-phase commits); --manual restores one phase per invocation, and a fully ticked plan with a follow-up request now points at /dx-new. dx-references accepts several space-separated topics in one call, and dx-plan, dx-implement and dx-tdd batch the topics they always load together. /dx-roadmap and /dx-frame print the slice order and the user cases as a table before asking you to confirm.
• e3a07d9: Cut story and planning overhead found in a second sessions review. /dx-new now defaults to a single change and uses an effort only when slices can each ship alone or run in parallel; /dx-roadmap folds a slice that only hardens or extends another into a phase, and makes the lead slice the tracer (one happy path, one negative test). /dx-plan decides by default and asks only what truly needs a human, recording its own decisions under a new ## Assumptions section; it reads at most one prior plan (cited as Reuse from:), makes phase 1 the tracer with hardening in later phases, names the affected test files per phase with one full-suite run in the final phase and one Standards checklist per plan, and warns and offers to move extras to a later change when a plan outgrows one shippable unit. /dx-implement runs only the test files a phase names and keeps the full run for the final phase. dx-new and dx-plan note that the next command can run in the same session (skills still never auto-chain), and dx-references reuses a topic already loaded in the session.
• d4640e2: /dx-review now runs over the repo or folders as well as a change or effort, and ships a third built-in Policy: /dx-review test-strategy [container-id | paths…] [instructions…] checks whether the testing strategy matches the problem class of the code under test (Strategy-fit, plus Frontend & e2e and Pathological tests, the latter two written from general knowledge rather than research). The new Explore-run argument grammar resolves positionally: a container ID, else the leading tokens that are existing repo paths, and the first token that is neither starts your instructions; a container plus a path is refused as ambiguous, and no container and no path asks what to review. An Explore run writes no report: it presents ranked Candidates inline and promotes one into a change, or several through a printed seed block. /dx-new now also parses that seed, headed ## <Policy> candidates (from /dx-review <policy-id>), into an effort with one research/<topic>.md per entry; the ## Refactor opportunities (from /dx-refactor-discover) seed is unchanged. Policies gain optional ## Questions (stage-tied gates asked in container and Explore runs; an unanswered gate leaves a unit unconfirmed with no verdict) and ## Candidates sections, and the report template gains an optional ## Reviewed list. Existing projects: re-run /dx-init to seed test-strategy.md (and plan.md if missing). /dx-init never overwrites your review-policy.md and review-report.md, which don't document ## Questions, ## Candidates or ## Reviewed, so compare both with the shipped ones in the skill's assets/config/ and merge the new sections by hand.

Patch Changes

• 32f1f0d: Fix /dx-new not installing via npx skills add: its argument-hint frontmatter was invalid YAML, so the installer silently skipped the skill.