v2.0.0 — autonomous default, shared core, plain language

Program: issue #38 (v7 — one-key autonomy, tailored scope, shared core, measured on two Claude tiers). Baseline commit 69e7bd4.

Breaking

• Autonomous default. No flag = every decision resolves by best judgment from the evidence gathered and is recorded in the summary; only the publish/irreversible exception list (core/ask-exception-list.md) is skipped and recorded only you can do. --ask restores menus, approval batches and confirmations at every decision point and propagates from ds-ship to its delegates. --auto is removed without an alias; the Up-Front Mode Menu, All-Affordance and Selection Transparency rules apply only under --ask.
• core/ is the single home for shared references (principles, severity/score categories, findings + profile format, signal inventory, ask-exception list, checkpoint protocol, execution loop, report + outcome templates, secret patterns, toolchains, CRAAP, best practices, experience rules). The 22 per-skill references/principles.md copies and the toolchain / CRAAP / secret-pattern duplicates are gone; skills link ../core/<file>.md. A manual copy of one skill directory must now copy core/ beside it — install.sh (including --skills subsets and --target) always does.
• ds-solve removed. Its budgeted backtracking lives in ds-build, its hypothesis loop in ds-debug.
• ds-ship gains --mode=harden|release|launch|maintain. Benchmark, launch, productize and Ship-ready legs run only in release/launch; every delegate is justified by a project signal (signal-absent — key=value, skipped — not part of this mode, skipped — another skill owns it for this project type, skipped — removed by you, failed, skipped — not installed).
• Closing block shares dev-rules' labels: Asked / Done / Effect / Decided without asking — say if wrong / Only you can do (empty lines omitted); every choice put to the user is recommendation-first with options in one shape. core/report-and-outcome-templates.md § 5 is the portable copy; the lean and claude install profiles strip it because the always-on layer owns it.
• Installer rewritten without rsync (bash + coreutils + git); install.cmd launches it from cmd.exe/PowerShell; profiles portable · lean · claude (lean + ds-ship delegates run as forked subagents on Claude Code); --status; profile-aware --check.
• Spec Kit artifacts removed (specs/, .specify/, .opencode/); ds-pipeline runs specify → clarify → plan → tasks → analyze natively, Spec Kit optional when installed, and hands tasks.md to ds-build.
• docs/ lifecycle guides removed after a delta migration into the owning skills' rules files (every actionable item became a rule with Detect/Fix/Impact/Source, templates moved into ds-docs references); docs/methodology/ and docs/guide.html remain.
• Spec citations banned inside skills (SKILL-SPEC, Unattended Mode rule, rule-4 exception, IDU, DSC, OVERLAP-n, VAR) — 143 citations replaced by the canonical inline equivalents; a lone install has nowhere to resolve them.
• Contributor-facing: check-consistency.sh checks 21/22/23 folded into check_core_links, check 28 removed; new checks check_spec_citations, check_dead_sources, check_undefined_flags (+ retired-flag list), check_readme_counts, check_ask_row_canonical, check_scope_resolution_table, check_rule_impact, check_duplicate_rule_titles, each with a self-test fixture and a mutation case.

Added

• ds-build (prefix BLD) — plan executor: issue / specs/{feature}/tasks.md / plain request → bounded units, a verify signal per unit, regression tests seen red first, ≤ 3 attempts × ≤ 3 approaches with recorded causes, per-unit commits, code-proven close; ds-issue --do, ds-pipeline and ds-freeze hand off to it.
• ds-debug (DBG) — reproduce the red → localize (git bisect run, trace, temporary logging, flaky split 3× isolated + 1× shuffled) → ≤ 3 falsifiable hypotheses → minimal fix behind a red-proven regression test; not reproduced is an honest outcome; tests are never weakened.
• ds-release (REL) — version from Conventional Commits, dated CHANGELOG section with Unreleased kept, every version surface found by search and bumped together, check green before the tag, local release commit + annotated tag; push / GitHub release / registry / store handed back with the exact commands; --verify post-release (remote tag, CI run, attestation, registry, smoke, rollback path).
• Relevance First: ds-blueprint produces a Signals: line (has_ui, has_api, has_db, has_auth, has_billing, has_pii, platforms, audience, jurisdiction, integrations, …; core/signal-inventory.md); every multi-scope skill carries a | Scope | Runs when (signal) | Otherwise | table; the three largest rules files tag every rule applies_when:.
• Mechanical Done Gate + Checkpoint in ds-pr (closes the reset --hard-over-uncommitted-work path with an explicit Push phase), ds-repo, ds-devops, ds-docs; 21 code-modifying skills gated.
• Rules: Impact line on every rule (≈ 400 added); ds-mobile Expo + Kotlin Multiplatform detection, per-platform release rules, staged rollout; ds-deps JVM/.NET/Swift rows, lockfile-integrity table, SBOM (syft); ds-init Bun/Hono/Astro/Expo; ds-frontend Svelte 5/Astro/Solid/htmx; ds-compliance EAA, DSA, age verification, Quebec Loi 25 (+ analytics-privacy and legal-checklist deltas); ds-productize chargeback and external-purchase-link rules; ds-deploy DNS/TLS/HSTS rules; ds-repo GitHub security toggles (secret scanning + push protection, Dependabot, private vulnerability reporting, code scanning) via gh api; ds-review gitleaks/semgrep as optional detectors with enumerated per-scope checklists; ds-test contract tests on API boundary artifacts.
• Evals: 13 tasks (5 new: fix-autonomy-no-prompt, ship-harden-no-launch, build-tasks-md, debug-reproduce-fix, release-tag-changelog); evals/validate-scorers.sh proves every scorer red-before / green-after; results on #38.
• Gate: scripts/quality.sh adds typos, the installer round-trip (32 assertions) and dev-rules' check-cross-repo.sh when a sibling ../dev-rules checkout exists (visibly SKIPPED otherwise); .gitattributes pins LF; _typos.toml domain vocabulary.

Changed

• ds-review: --diff is the default scope, bootstrap is an if/then table, simplify/yagni/duplicate detection handed to ds-simplify, discarded (no harm signal): n in the summary. ds-fix: --diff default, tracked secret = CRITICAL / untracked = HIGH via git ls-files. ds-test: input-state decision table, findings.md write scope, red-proof + flaky gate. ds-commit: secret/size commands stated, no fetch/pull, Co-Authored-By only when the repo already uses it. ds-devops: rules split into ci / signing / deps / release-pipeline, actionlint + zizmor in the done gate, monitoring owned by ds-deploy. ds-launch: SEO/email handed to ds-compliance, integrations audit keyed on the integrations signal, ds/launch/ metadata directory. ds-compliance: ds-mobile and ds-frontend absence branches (no silently dropped scopes). ds-frontend: three fully defined presets. ds-blueprint: --preview writes nothing, Foundation questions only under --init/--ask, privacy single-owned. ds-brief: slot manifest, verifier path from the dispatcher. ds-research-agent: normative Contract split from the operating procedure. Frontmatter descriptions ≤ 300 chars; SKILL.md byte ceilings per skill with per-scope reference loading.

Measurements

Measure	1.2.0 (69e7bd4)	2.0.0	Note
Skills	30	32	−ds-solve, +ds-build, +ds-debug, +ds-release
SKILL.md total	784,470 B	702,069 B	−10.5% with more coverage
SKILL.md median	25,366 B	19,460 B	−23% — what a run actually loads
Largest SKILL.md	43,979 B (ds-brief)	33,802 B (ds-brief)	−23%
Duplicated principles.md copies	22 (≈293 KB)	0	one home in core/
Unresolvable spec citations inside skills	112	0	canonical inline equivalents
Consumer-less docs/ guides	13 (≈304 KB)	0	migrated into rules files
Rules with a stated Impact	partial	798/798	check_rule_impact
Gate checks (check-consistency.sh)	33	39	+9 new, −3 folded into check_core_links
Self-test fixtures / mutation cases	22 / 35	24 / 45	every new check proven red
Eval tasks	8	13	+autonomy, mode, build, debug, release

Behavioural A/B (same fixture, same harness, old text = git archive 69e7bd4 installed lean, new text = install.sh --profile claude; score = evals/tasks/*/score.sh, claim = the agent's own STATUS: line):

Task	Old (Haiku)	New (Haiku)	Old (Fable)	New (Fable)
ship-harden-no-launch	FAIL 4/6	PASS 6/6	FAIL 4/6	PASS 6/6
fix-autonomy-no-prompt	PASS 6/6	PASS 6/6	interrupted (quota)	PASS 6/6
the other 11 tasks	—	PASS 13/13 total	—	PASS 11/12 completed, 1 interrupted (quota)

The harden case is the design change made visible: the old text stops at a confirmation gate and runs launch/benchmark/productize legs on a library with no store or billing signal; the new text resolves the mode, records why each leg was skipped, and finishes the run. No false-done in any completed run (the 1.2.0 baseline had one: Haiku's fix-secret-report claimed OK while converting a live key to an env var and committing it).