Version History

Version History

Side-by-side comparison. Detailed per-version notes live on each release page.

Version	Date	Theme	Headline change
v0.10.0	2026-04-26	Honesty Cleanup	SKILL-full split reverted (Mode-gated single file), audit-specialist merges 3 audit agents, runtime backend hooks (route/verify/record/update/drift), install.sh, 12 default checklists, incremental backend save (route-memory.jsonl)
v0.9.2	2026-04-23	Gravity Pattern	Two-tier knowledge (Centerlite hub + Dcenterlite project), prompt log hook, /kasi-init, /kasi-promote, /kasi-pull, /kasi-sync, /kasi-wiki-sync
v0.9.1	2026-04-22	Master Orchestrator	Master delegates all strong work to 10 specialist agents; dispatch brief format
v0.9.0	2026-04-20	Claude Design	Design/Visual Mode, .kasidit/prototypes/, DESIGN_SYSTEM.md, mockup-to-code parity
v0.8.0	2026-04-15	Tier Cascade	Opus plans → Sonnet works → Haiku greps. Local embedding layer for .kasidit/knowledge/
v0.7.4	2026-04-10	SWE-bench + rules	60.7% PASS on 56 SWE-bench Lite tasks. Rules 2.3 (no fake metrics), 2.4 (numbered options), 2.5 (native language), 2.6 (mandatory git-log/git-S on bug fix)
v0.3.0	2026-03	Tier adaptation	Opus/Sonnet/Haiku rules diverge. Confidence labels [high/medium/low/unsure]. CHECKLISTS/. PATTERNS.md. Multi-agent orchestration. Verifier pass
v0.2.1	2026-02	Docs protocol	Trust hierarchy (source → official docs → release notes → blogs), version matching, .kasidit/knowledge/ cache
v0.2.0	2026-02	UI Override	Scoped class + !important, cache-aware, one-change-per-round, override > audit for legacy CSS
v0.1.0	2026-01	Core	Core principles, mission counter, สารบัญ (INDEX/RELATIONS/MEMORY)

What changed between each pair

v0.9.2 → v0.10.0 (this release)

• Honesty Cleanup: the prior SKILL-full.md split is reverted. Full Framework merges back into SKILL.md behind a prompt-level /kasi off|router|lite|full|ultra Mode gate. One file, mode-gated load depth.
• Backend hooks (runtime-enforced for the first time): kasidit-route.py (UserPromptSubmit classifier + memory query), kasidit-verify.py (PostToolUse + Stop confidence + orchestrator check), kasidit-record.py (Stop emit-line parser → JSONL stores), kasidit-update-check.sh (1×/day release tag check), kasidit-drift-check.sh (SessionStart Centerlite-sync reminder).
• audit-specialist consolidates code-reviewer + security-auditor + perf-profiler via --focus=quality|security|perf|all. Old agents kept as name-resolution stubs (no automatic mapping; users must invoke audit-specialist --focus= explicitly). Stubs disappear in v0.11. Active registry now 8 + 3 stubs.
• install.sh canonical installer — copies hooks, merges ~/.claude/settings.json (jq primary / python3 fallback, idempotent), seeds Gravity hub with 5 JSONL stores + 12 default checklists + 2 knowledge templates, writes config.json, manages .last_sync / .last_update_check stamps.
• 12 default checklists seeded: PHP / Node / Python / Go × code-review / security / perf.
• Incremental backend save ("ออม") — AI emits [kasidit-log|pattern|memory|rule] lines at mission end; kasidit-record.py appends to matching JSONL. Router learns shortest successful route per mission kind over time.
• sudo <mission> clarified — parallel fan-out, min 2 agents, assumption-narrated pacing. Not a permission escalation. /kasi-multi --fast is an equivalent flag form.
• Docs honesty: /kasi precedence chain marked spec, not runtime. No code currently merges the three configs into a resolved value — applied by user + AI reading the files.

v0.9.1 → v0.9.2

• New pattern: Gravity — a formalized two-tier knowledge layout. ~/.claude/skills/kasidit/center/ (Centerlite) + <project>/.kasidit/ (Dcenterlite). Before this, the split existed implicitly (per-project .kasidit/ plus a vague "user scope") but had no name, no sync commands, and no discipline about what belonged where.
• New commands: /kasi-init, /kasi-promote, /kasi-pull, /kasi-sync, /kasi-wiki-sync.
• New hook: UserPromptSubmit → ~/.claude/hooks/kasidit-log.sh → writes every user prompt to ~/.claude/skills/kasidit/center/logs/YYYY-MM-DD.jsonl. 200-line trim with head/tail preservation.
• Permissions: default allow-list expanded for common Kasidit bash / git / read patterns.
• Docs: a docs/wiki/ source tree lands for the GitHub wiki (this page is part of it).

v0.9.0 → v0.9.1

• Added Master Orchestrator Rule — the main agent is forbidden from executing strong work (multi-file changes, migrations, perf hunts, security audits, new features, deep research). It only narrows, dispatches, synthesizes.
• Added 7 specialist agents: bug-hunter, architect-planner, perf-profiler, test-writer, refactor-surgeon, deep-researcher, migration-specialist. Plus the pre-existing code-reviewer, security-auditor, legacy-specialist = 10 total. (In v0.10 this collapses to 8 + 3 stubs as the 3 audit agents merge into audit-specialist.)
• Dispatch brief format standardized: MISSION / INPUTS / CONSTRAINTS / EXPECTED OUTPUT / PRIOR CONTEXT.
• The older Multi-Agent Orchestration section became an implementation detail of the master rule.

v0.8.0 → v0.9.0

• Introduced Design / Visual Mode. Triggers: mockup, wireframe, deck, slide, one-pager, landing page, pre-code sketch, ยังไม่ code.
• Rule 3 (Design before code) now offers Claude Design prototypes for UI-touching missions instead of text diagrams.
• UI Override Rule 1 adds "Claude Design mockup" as a third valid visual target (besides screenshot and raw CSS values).
• New storage: .kasidit/prototypes/<mission-id>-{before,after}.png|url|pptx.
• DESIGN_SYSTEM.md — visual counterpart to PATTERNS.md; records colors, spacing, typography, components extracted from the codebase.
• Mockup → code handoff with per-round token-level parity check (UI Override Rule 8).
• Haiku rule: never hand-code mockups. Always route to Claude Design.
• New commands: design <what>, mockup <screen>, extract-system, parity <mockup-id>, report visual.

v0.7.4 → v0.8.0

• Tier Cascade orchestration — Opus plans, Sonnet works, Haiku greps. Before this, tier awareness existed (v0.3.0) but was per-session; v0.8.0 routed work across tiers within one mission.
• Local embedding layer — sentence-transformers over .kasidit/knowledge/; enables /kasi-search semantic queries without phoning home.
• /kasi-cascade command added.

v0.3.0 → v0.7.4

• Rule 2.3 — ban on analytical / theoretical / expected / projected metrics. Numbers must be measured before claimed. Prior versions tolerated "expected 10x speedup" language.
• Rule 2.4 — every list of options must be numbered (1/2/3) so the user can reply by number instead of paraphrasing.
• Rule 2.5 — reply in the user's native language. Thai user → Thai response (code/identifiers stay English).
• Rule 2.6 — mandatory git log --grep=<term> and git log -S <symbol> before proposing a bug fix. Prior versions allowed skipping when the cause "seemed obvious".
• Validation: SWE-bench Lite sequential sample — 56/300 tasks, 60.7% strict PASS, 87.5% valid rate.

v0.2.1 → v0.3.0

• Model tier adaptation. Before this, the skill was tier-blind — the same rules for Opus and Haiku. v0.3.0 tightens Haiku (no architecture decisions, checklist-driven only, confidence labels mandatory, verifier pass required) and loosens Opus (cross-file reasoning OK, architecture suggestions OK).
• Confidence labels — [high / medium / low / unsure]. [unsure] items are always listed separately for user decision, never silently guessed.
• CHECKLISTS/ directory — pre-built mechanical audit lists per stack. This is "checklist does the thinking" — Haiku's biggest enabler.
• PATTERNS.md — project-specific patterns that AI copies, not reinvents.
• Multi-agent orchestration — split heavy missions (code review, refactor) across subagents for context isolation + parallelism.
• Verifier pass — mandatory on Haiku reviews. A second agent reads the first's findings and removes false positives.
• Vague mission detection — specific trigger words (check / review / audit / "ดูดี") force narrowing before work begins.
• CSS width audit protocol — grep-before-fix for UI work; no assuming selector scope.

v0.2.0 → v0.2.1

• Documentation retrieval protocol — explicit trust hierarchy (project source > official docs at exact version > docs latest > framework source at release tag > release notes > blogs). No Stack Overflow / Medium / AI memory for version-specific syntax.
• Version detection first — composer.json, package.json, go.mod, requirements.txt read before coding.
• Knowledge caching at .kasidit/knowledge/<stack>-<version>-<topic>.md. Never cache whole pages — only the snippet that answered the question.

v0.1.0 → v0.2.0

• UI Override Mode — for legacy CSS: scoped class + !important, cache-aware, one change per round. Does not refactor cascade.
• Cache protocol — every CSS/JS change bumps ?v=X, user hard-refreshes, confirms new version loaded before evaluating the fix.
• Domain detection — splits backend / UI / review missions; each gets its own flow.
• Override-first strategy — when you can see the bug, override beats cascade audit.

Upgrade path

Each version is backward-compatible with prior .kasidit/ directories. Upgrading = re-enabling the plugin from the marketplace; old .kasidit/ content remains valid. v0.9.2 adds center/ at user scope — no project edits required.

The only behavior that changes silently on upgrade is the Centerlite prompt-log hook: if you installed the hook per the v0.9.2 instructions, every prompt from that point on is logged. Remove the UserPromptSubmit block from ~/.claude/settings.json to opt out.