Skip to content

CHANGELOG.md

Latest commit

 

History

History
1007 lines (806 loc) · 51.1 KB

CHANGELOG.md

File metadata and controls

1007 lines (806 loc) · 51.1 KB
Raw

Changelog

All notable changes to the ODIN Claude Plugin will be documented in this file.

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

[2.1.4] - 2026-09-12

Added

  • schedule-dependency-waves (odin-run): orders work units carrying declared dependencies into execution waves and detects dependency cycles before dispatch.
  • diagnose-wave-stall (odin-run): classifies a dispatched wave that idles on a result that never arrives and names the blocking edge.
  • from-perspective (odin-critique): one skill taking the seat as an input.

Changed

  • finish-it-now is renamed do-it-now, restoring the retired name; the procedure is unchanged and "finish it now" stays a trigger phrase.
  • gate-file-completion, writing-for-agents, architect, and deslop regain the trigger words unlazy, writing skills, codebase design, and cleanup codebase.
  • 252 cold skills are set slash-only with disable-model-invocation: true, cutting the autoload catalog from 622 to 362; each stays reachable by /skill:<slug>.
  • Stale version pins corrected to the <languages> table: Rust edition 2024 (rust-version = "1.85", resolver = "3") in cargo-workflows. The Python floors in cpu-kernel-authoring and gate-file-completion stay at their real requirements (3.11+ and 3.10+); the 3.14 pin is a generated-code target.

Retired

  • finish-it-now → do-it-now
  • from-breaking-perspective → from-perspective
  • from-business-perspective → from-perspective
  • from-career-perspective → from-perspective
  • from-codebase-perspective → from-perspective
  • from-human-perspective → from-perspective
  • from-impact-perspective → from-perspective
  • from-innovation-perspective → from-perspective
  • from-moat-perspective → from-perspective
  • from-rentseeking-perspective → from-perspective
  • from-skeptic-perspective → from-perspective
  • from-stability-perspective → from-perspective
  • instruction-understanding-gate → askme
  • commit-push → commit-push-pr
  • commit-push-main → commit-push-current
  • capture-isolated-patch → worktree
  • setup-gitignore → git-guardrails
  • setup-pre-commit → git-guardrails
  • resolve-pr-feedback → resolve
  • bus-drivers-i2c-spi → device-drivers
  • device-tree → device-drivers
  • platform-device-model → device-drivers
  • writing-char-drivers → device-drivers
  • ebpf-rust → ebpf
  • hardware-counters → linux-perf
  • rust-sanitizers-miri → sanitizers
  • binary-hardening → elf-inspection
  • datasheet-and-refmanual-reading → peripherals-from-datasheet
  • extremely-optimize → optimize
  • fastopt → optimize
  • fastopt-extreme → optimize
  • all-scenarios-storm → plan
  • plan-review → plan
  • waterfall-guide → plan
  • confirmed-security-review → security-review
  • sharp-edges → security-review
  • build-program-graph → project-findings-onto-graph
  • entry-point-analyzer → project-findings-onto-graph
  • trailmark-structural → project-findings-onto-graph
  • graph-evolution → project-findings-onto-graph
  • variant-hunt → project-findings-onto-graph
  • scaffold-cli → scaffold-nextjs
  • embed-diagram → diagramming-code
  • visualise-widget → show-me
  • architecture-diagram → visual-diagram
  • dataflow-diagram → visual-diagram
  • lifecycle-diagram → visual-diagram
  • sequence-diagram → visual-diagram
  • workflow-diagram → visual-diagram
  • address-sanitizer → sanitizers

[2.1.3] - 2026-09-09

Changed

  • compound now writes its learning records to docs/solutions/, the store autolearn already owns; docs/knowledge/ is gone. Compound records sit at the top level and the category subdirectories stay autolearn's, so each skill reads the other's docs for overlap without rewriting them to its own schema. An absent directory is an empty store, created on the first approved write, instead of blocking the run.
  • autolearn's refresh step names compound's seven-field records as outside the Solution schema: it reads them for overlap and never rewrites one or marks it stale for lacking problem_type or root_cause.
  • brainstorm searches docs/solutions/ and plans/ for prior knowledge, collapsing the duplicate docs/knowledge/ read.

[2.1.2] - 2026-09-07

Fixed

  • cpu-kernel-authoring: trial state rejects ids outside t<number>, an unknown or non-string best_trial, and a malformed baseline_us. save stages into a fresh directory and refuses an existing trial directory, so a failed copy never deletes prior contents; save and finalize copy symlinks as symlinks. The build validator parses build.toml as TOML, reads cxx-flags, reports a non-table kernel value instead of crashing, checks AVX512 flags per [kernel.*] section, and requires Python 3.11. The profiler rejects nonpositive iters and negative warmup.
  • Memory script usage strings name the .py files, the Clang hardening example matches the GCC linker flags, and the Hub examples pass version=1.

[2.1.1] - 2026-09-05

Added

  • Expanded odin-native from 2 to 148 skills: 142 adopted low-level-dev-skills entries plus 4 new skills (Rust and Go handbooks, a C hardening baseline, and CPU kernel authoring) alongside the 2 existing skills. The CPU correctness comparator checks every structured output and uses dtype-specific tolerances instead of checking only the first output.
  • Renamed odin-lean to odin-formal, retained writing-lean-proofs, and added eight formal-methods skills. The plugin now contains nine skills.
  • Restored commit-push, commit-push-pr, commit-push-current, and commit-push-main. The current-branch and default-branch variants require explicit human invocation; every push requires approval of its mutation set. The tree now holds 703 skills in 28 plugins, up from 545 on the current main branch.

Migration

Retired Successor Skills
odin-lean odin-formal 1 of 1; the plugin now covers formal methods beyond Lean

To find a moved skill's new plugin, look up its slug in docs/specs/skill-index.md. Reinstall a single skill from its new path:

gh skill install OutlineDriven/odin-claude-plugin plugins/<new-id>/skills/<slug> \
  --agent claude-code --scope user

[2.1.0] - 2026-09-04

Changed

  • Shortened every skill description to at most 200 characters (median 174, down from 269). Each keeps its trigger phrases and Not for pointer and drops mechanism narration; a fresh read-only review of every rewrite restored 95 lost triggers or pointers.
  • Folded setup-tool-credentials into setup: the credentials mechanism now makes .env safe before writing, validates credential formats, never overwrites existing keys, and verifies repository, tool access, and prerequisites.
  • Folded record-design-decisions into docs-and-adrs: resolved codebase terms are routed to domain-modeling, which owns CONTEXT.md and its entry schema, and an unavailable ADR directory no longer halts the pass.
  • Folded single-file-html-deck into presentation-creator: the web-deck format now builds one self-contained, keyboard-navigable HTML file from caller-supplied brand tokens with a frozen pnpm lockfile and browser verification.

Changed

  • Plugins are reorganized by job. Twelve ids that named a tier (-advanced) or a grab bag (odin-create, odin-loop) are retired, and twelve job-named plugins are created: odin-critique, odin-fuzzing, odin-git, odin-infra, odin-knowledge, odin-learn, odin-people, odin-release, odin-review, odin-skills, odin-testing, and odin-visual. 442 skills moved by git mv: 225 out of the retired ids and 217 between surviving ids; 275 of them landed in the new plugins and the rest in surviving ones. Every skill keeps its slug, and the move touched no body except one description (lockstep-version-guard, which dropped a hardcoded plugin count). The 28 plugins are ordered in the catalog by workflow with kin adjacent: think, build, ship, operate, protect, know, tooling, communicate, stacks. A plugin id now names a job or a stack, never a tier; check-plugin-surfaces rejects a tier suffix.
  • The doctrine cascade is converged. system-prompt-baseline.md loses its duplicate rule families (delete-not-deprecate into Posture, wrappers into Craft), its reading clamp, its [MANDATORY], [HARD-REJECT], [REJECT], and [INTERNAL] tags, its arrow-chain and pseudo-math jargon, and the harness tool layer, which splits into a shared <change_discipline> section and a per-carrier <code_tools> layer. <languages> is rewritten to one row per stack: the LTS or current-stable target and one first principle. The six output styles carry the regenerated tail; the two external carriers are realigned in place.
  • Every skill body passed three prose waves and one judgment wave: em-dashes replaced by the punctuation their grammatical role needs, incidental harness nouns replaced by the mechanism they stood for, stale packages/ paths fixed, the Contract Authority row canonized to four leads (Read-only., Reversible local:, Human-gated:, Remote:) with each skill's specifics kept, a full unslop pass, and a prompt-optimizer audit with one skill-improver cycle per skill: 104 audit findings, 60 applied and the rest recorded in the plugin reports, plus 34 skill-improver fixes. A pre-push review then restored 46 Authority rows the canon had narrowed or widened.
  • Four slugs that named an upstream vendor rename to their jobs: sred-project-organizer to eligibility-batch-organizer, sred-work-summary to work-records-summary, unlazy to gate-file-completion, and slack-qa-investigate to repo-qa-investigate.

Fixed

  • lockstep-version-guard no longer requires exactly 28 catalog entries; the catalog is the member list.
  • Install examples name a skill's real plugin (askme lives in odin-planning), the README contents list points only at sections that exist, and docs/specs/skill-lifecycle.md describes the five dotdir manifests and five registries the tree ships.
  • AGENTS.md names the gate commands (just check, just verify), the carrier generator, the tier-suffix assertion, and the worktree gotcha in the Devin mirror target.
  • Nine review findings repaired eight skill contracts whose Authority rows contradicted their bodies: memory-clean and memory-update claimed version control as the rollback for data outside version control; github-bug-report-triage, docs-update, saga, and poteto-mode misplaced their write or approval boundaries; resolve-merge-conflicts orders git stash drop after the resolution is committed; and influence-and-negotiation permits a conditional sub-step to be skipped with a reason.
  • watch-for-harness-mode regains its no-VCS-commit, no-remote-call, no-credential-use, no-published-artifact ban, which the Authority canon had compressed to a weaker no-remote-mutation clause.
  • cubic P1: refactor-break-compat now previews the demolition and waits for confirmation before deleting, and deletes only flags used by the old path; gh-review-requests drops the invalid requestedReviewers field; setup-tool-credentials never echoes a colliding .env value and refuses to write credentials into a tracked or unignored path.
  • Codex P1: the Everyday code changes install row includes odin-git; spec-driven names odin-testing for tdd; model-retuning restores the first-match outcome classifier; retaxonomize-plugins leaves the generated README plugin table to just render; tournament failure terminals map onto the cascade set (stalled, blocked).
  • Tournament Output listed capped and pending, terminals no procedure step produces. Output now matches the four defined terminals.
  • memory-update and memory-clean claimed $MEMORY_DIR is outside version control; that is true for the default (gitignored) path and false for an override git would track. The resolver refuses a path git would track, whether already tracked or merely unignored; it canonicalizes a relative override against the caller's directory rather than the repository root, and it treats a leading hyphen as a pathname.
  • A MEMORY_DIR override pointing at a directory that does not exist now fails the same existence check the derived default always paid; the resolver no longer prints a nonexistent path and exits 0.
  • memory-clean and memory-sanitize now call memory-update's resolve-paths.sh; the extra copies are gone, so a tracked MEMORY_DIR is refused for sanitize too.
  • pov now calls brainstorm's repo-profile cache; the extra copies are gone.
  • setup-tool-credentials can start in a directory with no .gitignore by creating ignore coverage for .env after confirmation, and it checks tracked vs ignored with git ls-files and git check-ignore rather than ordinary git status, which omits ignored files.
  • retaxonomize-plugins writes catalog entries with index, display_name, homepage, and directory, the fields the generators require.
  • sync-carriers --check reuses the check-carriers audit, so a carrier with an unknown section, a missing tool layer, or a malformed overlay fails even when shared bodies already match. An overlay-only <role> is kept and the canonical block is inserted after it. An existing non-file --carrier override is an error, not a traceback.

Added

  • retaxonomize-plugins (odin-skills): moves skills between plugins and creates, merges, or retires plugins in a catalog-generated marketplace, then proves the gates.
  • scripts/sync-carriers.py, run as just sync-carriers: rewrites each external carrier's shared doctrine from the baseline; --check is the gate.
  • docs/specs/skill-index.md, generated by scripts/render-skill-index.mjs: one row per skill with its plugin and category, the migration index for per-skill installs.

Removed

  • AST-GREP-AGENT-REFERENCE.md: a third copy of the ast-grep reference already carried by the ast-grep skill and the baseline.

  • docs/specs/graph.yaml and its pointer in docs/specs/skill-lifecycle.md: a hand-maintained second model of the tree with no executable consumer.

  • The historical 657-skill, 2.0.0 proof blocks in docs/specs/install-proof.md and docs/specs/distribution-surfaces.md, replaced by proofs run against this tree.

  • Duplicate root copies of implementer-prompt.md and task-reviewer-prompt.md under subagent-driven; only the references/ copies were cited.

  • setup-benny: folded into setup; the archive-install verification now lives there.

  • setup-pstack: folded into setup; the role-to-model mapping now lives there.

  • poteto-mode: folded into figure-it-out; the 23-playbook contracts now live there.

  • xcode-project-sync: folded into ios-build-fix as the XcodeGen regeneration mode.

  • ios-device-qa: removed; the vendor daemon has no public equivalent.

  • factory-mcp-bootstrap: removed; the Warp Factory endpoint is vendor-internal, and mcp-builder covers MCP server authoring.

  • Folded 62 skills into 45 survivors (605 to 543 skills, 10.2 percent). Each survivor absorbs its retired members' unique steps as inputs or named modes; a retired slug that names a mode of its survivor keeps that name in the survivor's mode input. Retired skills, slug to survivor:

    Retired Survivor
    aflpp libfuzzer
    batch-ask-me askme
    burp-search-command burpsuite-project-parser
    ci-sweeper ci-fix
    clarify askme
    classify-ci-failure ci-fix
    codebase-design architect
    commit-economy commit
    competitor-changelog competitor-feature-research
    constant-time-testing constant-time-analysis
    copywriting-cta copywriting
    copywriting-hooks copywriting
    copywriting-prose-creator copywriting-tone-of-voice-creator
    design-consultation design
    differential-security-review security-review
    engineering-blog-writing technical-article-writer
    framing-divergence-fanout prism
    get-pr-comments resolve-pr-feedback
    handoff-prompt handoff
    ideate brainstorm
    intent-proposal askme
    interview-me askme
    isolate-work-in-worktree worktree
    isolate-workspace-gate worktree
    plan-review-tune plan-review
    planning plan
    planning-and-task-breakdown plan
    port-static-analysis-rule semgrep-rule-authoring
    pr-link-issue create-pull-request
    pr-review-canvas-html pr-review-canvas
    property-test-failure-triage property-test-authoring
    property-test-review property-test-authoring
    property-testing-fit-assessment property-test-authoring
    prototype-logic prototype
    publish-release-pr release-gate
    rationale-by-source why
    render-excalidraw-diagram visual-argument-diagram
    respond-to-pr-comments-in-blocklist resolve-pr-feedback
    review-reception-protocol resolve
    shape plan
    sharp-edges-analyzer sharp-edges
    skill-benchmark-gate skill-benchmark
    skill-improver agent-surface-forge
    skill-progressive-disclosure-design writing-for-agents
    smart-contract-code-maturity-assessor smart-contract-guidelines-advisor
    trailmark-summary trailmark-structural
    validation-self-audit evaluation-leakage-audit
    variant-neighborhood-seeding variant-hunt
    visual-brainstorm-companion brainstorm
    visual-fact-check verify-both-ways
    visual-plan plan
    visual-project-recap catchup
    visual-render-tool visual-diagram
    visual-slides presentation-creator
    visualise-chart visualise-widget
    visualise-explainer visualise-widget
    watch-until watch-for
    what-did-i-get-done weekly-review
    write-pr-description create-pull-request
    write-prd write-product-spec
    write-tech-spec write-product-spec
    writing-skills writing-for-agents

Migration

Retired id to successor; the count is how many of the retired plugin's skills went to the named successor.

Retired Successor Skills
odin-bigquery odin-infra 2 of 2
odin-code-advanced odin-code 17 of 50; the rest to odin-testing (17), odin-review (5), odin-git (2), odin-security (2), odin-knowledge (2), odin-skills (2), odin-planning (1), odin-infra (1), odin-design (1)
odin-create odin-writing 15 of 45; the rest to odin-visual (15), odin-planning (10), odin-design (2), odin-critique (1), odin-product (1), odin-release (1)
odin-create-advanced odin-visual 10 of 22; the rest to odin-planning (6), odin-writing (2), odin-critique (1), odin-product (1), odin-code (1), odin-learn (1)
odin-design-advanced odin-design 4 of 4
odin-loop odin-design 4 of 21; the rest to odin-review (3), odin-testing (3), odin-infra (2), odin-product (1), odin-code (1), odin-git (1), odin-run (1), odin-research (1), odin-knowledge (1), odin-agent (1), odin-skills (1), odin-web (1)
odin-prometheus odin-infra 1 of 1
odin-research-advanced odin-critique 6 of 13; the rest to odin-research (4), odin-product (2), odin-review (1)
odin-run-advanced odin-run 19 of 21; the rest to odin-release (2)
odin-security-advanced odin-security 27 of 39; the rest to odin-fuzzing (12)
odin-terraform odin-infra 1 of 1
odin-writing-advanced odin-writing 6 of 6

To find a moved skill's new plugin, look up its slug in docs/specs/skill-index.md. Reinstall a single skill from its new path:

gh skill install OutlineDriven/odin-claude-plugin plugins/<new-id>/skills/<slug> \
  --agent claude-code --scope user

[2.0.4] - 2026-09-03

Fixed

  • resolve-pr-feedback: the entry-point document named five dispositions where the rubric, both mode files, and the fixer prompt name six. It called the fix verdict fix where every other file calls it fixed, and omitted fixed-differently, so an agent reading only SKILL.md emitted a verdict the fixer prompt does not accept and had no verdict for the better-repair case the fixer is told to take. The four drifted sites now cite the rubric and carry its six names.

[2.0.3] - 2026-09-03

Added

  • gate-and-merge references: feedback-sweep.md carries the sweep procedure and the six verdict names, taken from the resolve-pr-feedback rubric rather than coined a second time; reviewer-subagent.md carries the subagent assignment and the six diff classes; gotchas.md carries the evidence only a failing branch reaches, for the listing 502, mergeStateStatus, and the conflict path.

Changed

  • gate-and-merge: the one-axis severity table becomes two axes and four terminals. Direction decides close, repair bound decides repair, hold, or merge. minor had welded harmless to locally fixable, so a reachable defect with a bounded repair was held with a comment instead of fixed and merged. The bound is now three conditions rather than a --name-only file allowlist.
  • gate-and-merge: reviewDecision: "CHANGES_REQUESTED" routes to the new review-feedback gate instead of terminating, which had made a feedback sweep unreachable. The reading gates run in a fresh-context subagent that never sees the merge decision, gate ordinals are gone, and the checks gate runs last against the head that will actually merge.

[2.0.2] - 2026-09-03

Added

  • prompt-optimizer audit mode: finds dated instructions in prompts, skills, output styles, and tool descriptions; returns a confidence-ordered report and a proposed diff carrying only high- and medium-confidence hunks. Pattern groups and keep list in references/audit-patterns.md; per-vendor guide index in references/prompt-guides.md.
  • AGENTS.md requires the audit on behavior change and at each model release, and states the patch-bump rule for releaseVersion.

Changed

  • generate-my-taste: the influence catalogue heading drops its caps tag; the criteria below it carry the reason.

Removed

  • Five retired-campaign write-ups under docs/solutions/ and docs/ideation/.

[2.0.1] - 2026-09-03

The Agent Plugins surface is retired; four native harness surfaces join the three existing ones. Every harness now gets its own registry and per-plugin manifest, all generated from the catalog.

Added

  • Per-plugin manifests for Codex (.codex-plugin/plugin.json), Cursor (.cursor-plugin/plugin.json), Grok (.grok-plugin/plugin.json), and Kimi (.kimi-plugin/plugin.json).
  • Registries .codex-plugin/marketplace.json, .grok-plugin/marketplace.json, and .kimi-plugin/marketplace.json.

Removed

  • The Agent Plugins surface: .agents/plugins/marketplace.json and the 28 root plugins/<id>/plugin.json manifests.

Changed

  • Codex now loads plugins in Legacy format from .codex-plugin/plugin.json; mcpServers is declared explicitly because the Legacy default filename is .mcp.json with a dot.
  • policy is added to .claude-plugin/marketplace.json entries so Codex, which parses that file, keeps first-use authentication instead of prompting at install time.

[2.0.0] - 2026-09-01

The repository becomes a multi-plugin marketplace. Skills move out of one flat directory into the plugin that owns them, npm distribution is retired, and the gate set becomes unconditional.

Added

  • 28 plugins, each carrying its own skills, plugin.json, README.md, LICENSE, and NOTICE, generated from catalog/plugins.json as the single source of plugin identity.
  • Three marketplace registries generated from that catalog: .claude-plugin, .agents/plugins, and .cursor-plugin.
  • docs/specs/graph.yaml, the structural backbone as a compiled graph, and docs/specs/skill-lifecycle.md, the skill lifecycle as a state machine.
  • scripts/check-voice.py and scripts/check-skill-frontmatter.py, each carrying its own fixtures, plus surface and route checkers. Every gate is dependency-free Node ESM or standard-library Python.

Changed

  • Skills move from a flat skills/<slug>/ root to plugins/<plugin>/skills/<slug>/. The directory now states which plugin owns a skill; no membership registry records it.
  • 148 skills become 613, re-derived against a bound contract: each states its trigger, authority, side effect, and done condition, and each procedure step carries a criterion an agent can check.
  • Attribution consolidates into licenses/NOTICE with a per-plugin NOTICE generated beside each plugin.
  • Every whole-tree gate runs unconditionally. A files: selector on a gate that audits the whole tree was a second model of its own dependencies, and it drifted.

Removed

  • catalog/skill-membership.json and the flat skills/ tree it indexed. The directory a skill lives in is now the only record of which plugin owns it.

[1.15.49] - 2026-06-14

Removed

  • srgn purged from tool doctrine: removed the srgn tool from the agent persona/doctrine (system-prompt-baseline.md + the cascaded output-styles/*.md) and deleted the skills/srgn-cli skill. Structural search/rewrite is now served solely by ast-grep.

Changed

  • Enriched ast-grep guidance: replaced the single-line ast-grep code-manipulation bullet with a structural reference block (cascaded byte-identically across system-prompt-baseline.md and all six output-styles): patterns-are-code (not regex), the metavar table, the two-pass --json-disables--U apply gotcha, stopBy: neighbor defaults for relational YAML rules, strictness levels, and the ast-grep-not-sg binary note (Linux sg/setgroups collision).

Added

  • AST-GREP-AGENT-REFERENCE.md: self-contained ast-grep field reference distilled from code-yeongyu/ast-grep-skill: patterns-as-code rules, metavariable semantics, valid-code fixes, strictness levels, context/selector objects, CLI two-pass apply + debug recipes, YAML rule anatomy with a verbatim no-console skeleton, per-language rewrite recipes, and a 14-item pitfalls guide.

[1.15.43] - 2026-06-05

Added

15 skills ported from the agent-sh plugin marketplace: re-homed as native, self-contained ODIN skills. All external dependencies (the agent-analyzer binary, repo-intel.json cache, editor shims, bespoke JS lib/, and opus/sonnet/haiku model routing) are replaced by native tooling: codegraph MCP, git/ast-grep/git grep recipes, repomix, generic ODIN agents, and the ask tool. Attribution in skills/LICENSES.md.

  • repo-intel: native repository intelligence (hotspots, coupling, bus factor, bugspots, ownership, entry points) from git history + codegraph; no cache, every signal recomputed on demand.
  • agnix: native agent-config lint pass (skill frontmatter, CLAUDE.md/AGENTS.md, hooks, MCP, plugin manifests, agent files) graded HIGH/MEDIUM/LOW.
  • deslop: three-phase certainty-graded AI-slop detection; HIGH-only guarded autofix with test-verify and rollback.
  • sync-docs: diff-driven doc-vs-code drift detection; safe-fix limited to version bumps + CHANGELOG, everything else flagged.
  • banthis: persist hard user prohibitions into a managed AGENTS.md/CLAUDE.md section (dep-free node script).
  • drift-detect: plan-vs-reality reality check across GitHub + docs + code with a prioritized reconstruction plan.
  • audit-project: iterative multi-agent code audit with a false-positive contract; loops until critical/high findings clear.
  • onboard: new-codebase orientation tour: bounded context collection, 7-section synthesis, interactive guidance.
  • can-i-help: route contributors to data-backed contribution opportunities matched to stated interest.
  • learn: online research → scored sources → summaries-only → RAG-optimized learning guide + retrieval index.
  • system-prompt-curator: create or improve autonomous-agent system prompts from research-backed principles.
  • skillers: mine local agent transcripts (sanitized first) into automation recommendations; never auto-creates files.
  • agent-surface-forge (formerly enhance): certainty-graded enhancement of agent, plugin, and instruction surfaces through parallel analyzers; HIGH-only --apply.
  • perf-investigate: self-contained multi-phase performance investigation (baseline → hypotheses → profile → one-change optimization).
  • next-task (explicit-only): self-contained backlog orchestrator: selects the next task and drives it through isolated git-branchless implementation, review, docs, and verification gates.

[1.15.29] - 2026-05-15

Changed

Output styles

  • Linus: register idiom standardized to "code" across body sections and principle lines; cross-domain reach (any work-product: code, analysis, knowledge work, documents, decisions) anchored in the <role> block.

[1.15.28] - 2026-05-15

Changed

Output styles

  • Linus: generalized register scope beyond code/kernel/C to any artifact-producing work (code in any language, analysis, knowledge work, documents, decisions). Doctrine reframed in the register's own voice without primary-source citations or URLs. Section retitled # Don't break the consumer's contract (replaces kernel-specific "we don't break userspace"); section retitled # Blunt about the work, never about people. Principle tags [coding-style] → [complexity], [no-typedef] → [no-hidden-shape]; [no-break] contract list compressed; cross-domain widening anchored once in <role> block per /taste audit (sections B/C/D/E speak artifact-agnostically without per-section domain enumeration). The C linked-list code snippets stay as the canonical teach-by-example, unattributed. Concrete complexity rules (≤3 indentation, ≤10 locals, one-or-two screens) retained; goto-pattern bullet generalized to "centralized cleanup at meaningful labels". Cascade byte-equivalence with system-prompt-baseline.md preserved.

[1.15.27] - 2026-05-15

Added

New output styles

  • Linus: Linus Torvalds maintainer-reviewer register as a taste qualifier. Applies as a judgment lens on top of any code work: special-case elimination as compression (TED 2016 linked-list doctrine), the "we don't break userspace" contract covenant (LKML 2012-12-23), kernel coding-style discipline (Documentation/process/coding-style.rst), blunt about code never about people (LKML 2018-09-16 register-guard). Sixth output-style; fills the taste-and-rejection lens slot left open by the five existing implementer registers (ODIN, AxiomMode, Builder, Duet, Eval). AGENTS.md enumeration and cascade-count references updated in the same atomic commit so future canonical edits cannot silently skip linus.md.

[1.15.9] - 2026-05-05

Changed

  • 13 skills: removed disable-model-invocation: true per cluster-level auto-invocation decisions: axiom-mode, caveman, gh-address-comments, gh-fix-ci, memory-clean, memory-sanitize, memory-update, request-refactor-plan, to-issues, to-prd, triage-issue, ubiquitous-language, generate-my-taste/assets/template. 7 skills retain explicit-only invocation: ai-collab-protocols, edit-article, generate-my-taste, improve-codebase-architecture, strict-validation-setup, write-a-skill, zoom-out.

[1.15.8] - 2026-05-04

Changed

  • 39 skills: added disable-model-invocation: true to frontmatter: ai-collab-protocols, ast-grep, axiom-mode, caveman, cleanup-codebase, deps-upgrade, design-an-interface, design-by-contract, duet, edit-article, gh-address-comments, gh-fix-ci, git-branchless, git-guardrails-claude-code, github-triage, improve-codebase-architecture, llm-self-loop, memory-clean, memory-sanitize, memory-update, perf-profile, pr-merge-base, pr-merge-temporal, pr-review, proof-driven, refactor-break-bw-compat, request-refactor-plan, setup-gitignore, setup-pre-commit, srgn-cli, taste, tests-adversarial, tests-purge-unneeded, to-issues, to-prd, triage-issue, type-driven, validation-first, write-a-skill. These skills require explicit /skill-name invocation; auto-discovery disabled. 18 cross-cutting verb skills remain auto-discoverable: askme, atomic-commit, atomic-commit-and-push, contexts, debug, design, explore, fix, init, parallel-launch, plan, proceed, qa, research, resolve, review, security-review, test-driven.

[1.5.2] - 2026-04-20

Changed

  • duet skill: added ## VS-gated question protocol [MANDATORY] section. Every AskUserQuestion (Phase 1 elicitation, Phase 2 fork, Phase 3 checkpoint) is now preceded by a compressed visible Verbalized Sampling block (VS (N→M): 1. ...). Survivors-only render; weakness/contradiction/oversight stays internal. Phase 2 single-survivor short-circuit defined; Phase 1 and Phase 3 always fire AskUserQuestion. >4-survivor cap added. odin:askme owns the canonical VS+actor-critic spec.
  • Duet output style: # Decisions before prose updated: VS block is now the only permitted preamble before AskUserQuestion. Skill-load block cross-references the VS-gated question protocol section.

[1.5.0] - 2026-04-19

Added

New skills

  • duet: two-party working posture where the user is director and the agent is executor. Surfaces every fork, tradeoff, and taste choice via batched AskUserQuestion with structural framing, a recommended default, and concrete previews when comparison is visual. Eliminates the review-bottleneck (review distributed across picks, not piled at the end) and prevents codebase-understanding debt (user remembers the architecture because they picked it). Three-phase loop: adaptive intent elicitation → execution with fork-surfacing → irreversible checkpoints.

New output styles

  • Duet: presentation half of the duet posture. Decisions before prose, structural/taste framing first with jargon on demand, concrete previews when comparison is visual, silent mechanics / loud forks, no validation language or recap. Enforces duet skill invocation whenever active.

Skill multi-language expansion (10-language support)

  • test-driven: conditionalized test commands across 10 languages (previously single-language lock-in).
  • proof-driven: replaced Lean 4 default with property-based testing; 10-language support.
  • design-by-contract: expanded to 10 languages; added verification dispatch; integrated LLM context.
  • type-driven: removed tool lock-in; 10-language support.
  • validation-first: removed tool lock-in; 10-language support.

Skill enhancements

  • parallel-launch: refined strategy; added explicit review step between worker phases.
  • refactor-break-bw-compat: enhanced guide with strategy section.

Changed

CLAUDE.md methodology

  • Clarified scope targeting; added post-agent verification step (read back modified files, confirm line-count expectations).
  • Refined Verbalized Sampling section: dynamic hypothesis sampling (baseline N≥5, trivial N≥3, high-ambiguity N≥7, architectural N≥10, no hard cap).
  • Simplified editing workflow (removed edit_file partial-snippet specifics in favor of the general Find → Transform → Verify discipline).
  • Updated banned tools list; enhanced token-efficiency guidelines (per-tool flags, discovery-then-targeted-read pattern).
  • Revised orchestration and delegation sections: task splitting, parallel vs sequence criteria, delegation mandatory thresholds.
  • Revised role/tidy-first/VS sections for clarity.

Tooling

  • Adopted repository-aware search defaults: git grep as primary text-search, rg as fallback, ast-grep for structural queries, srgn for grammar-aware scoped regex.
  • Revised srgn and nomino command documentation (usage, options, glob syntax).

Output styles

  • builder: enhanced communication guidelines for non-technical builders (outcome-first, plain-language, progressive-disclosure).

Infrastructure

  • Installation script simplified (single-line install).
  • mcp.json streamlined: cleaner server configuration keys.

Fixed

  • refactor-break-bw-compat: replaced stale git grep invocations with rg where appropriate.

Removed

Skills

  • hodd-rust, outline-strong, plan-now, deprecated and replaced by the broader plan / proceed / language-agnostic skill set.
  • code-simplifier: merged behavior into refactoring agent.

Hooks

  • UserPromptSubmit hook removed (replaced by explicit skill invocation patterns).

[1.4.0] - 2026-02-25

Changed

Agent Consolidation (57 → 46 agents)

Removed redundancy by merging overlapping agents and eliminating agents outside the code agent scope.

Created:

  • devil-advocate - Adversarial technical challenger for pre-decision challenge (pre-mortem, assumption dismantling, red-teaming)
  • database - Unified database agent (merged sql-pro + sql-query-engineer + database-optimizer)
  • refactoring - Full refactoring lifecycle (merged refactorer + refactor-planner + tech-debt-resolver + modernizer)

Merged/Absorbed:

  • docs-architect now absorbs docs and reference-builder capabilities (API references, configuration guides, exhaustive coverage)
  • debugger now absorbs investigator capabilities (hypothesis-driven debugging, incident investigation, report templates)
  • migrator now absorbs porter capabilities (cross-platform porting, language transitions, library equivalents)
  • criticizer rewritten with sharpened identity (severity-driven, post-implementation systemic critique)

Removed:

  • docs (exact duplicate of docs-architect)
  • reference-builder (absorbed into docs-architect)
  • sql-pro, sql-query-engineer, database-optimizer (merged into database)
  • refactorer, refactor-planner, tech-debt-resolver, modernizer (merged into refactoring)
  • investigator (absorbed into debugger)
  • porter (absorbed into migrator)
  • sales-automator (outside code agent scope)
  • reflector (overlaps general reasoning)
  • meta-programming-pro (niche, absorbed into language-specific agents)

Enhanced:

  • All surviving agent descriptions updated with cross-references and PROACTIVELY trigger conditions
  • Fixed analyzer.md registration gap (file existed but was not registered in plugin.json)

Fixed

  • analyzer agent now registered in plugin.json (was missing since v1.0.0)

1.0.0 - 2025-11-17

Added

Consolidated Repository Structure

  • Single repository serves as both plugin and marketplace
  • Marketplace manifest included in .claude-plugin/marketplace.json
  • Simplified installation: add repository once, access all features
  • No separate marketplace repository needed

Agents (57 total)

Language Specialists (16)

  • rust-pro - Rust Edition 2024 with ownership, lifetimes, zero-cost abstractions
  • rust-pro-ultimate - Grandmaster-level Rust for complex scenarios
  • typescript-pro - Strict TypeScript with discriminated unions, no any/unknown
  • python-pro - Modern Python with type hints, asyncio, dataclasses
  • golang-pro - Idiomatic Go with context-first APIs, goroutines
  • java-pro - Java 21+ with records, virtual threads, sealed classes
  • kotlin-pro - Kotlin K2 with coroutines, null safety, immutability
  • cpp-pro - Modern C++20+ with RAII, smart pointers, ranges
  • cpp-pro-ultimate - Grandmaster-level C++ with template metaprogramming
  • c-pro - Modern C with memory safety, systems programming
  • c-pro-ultimate - Master-level C for kernel programming, extreme optimization
  • javascript-pro - ES6+ with async patterns, Node.js APIs
  • php-pro - Modern PHP with generators, SPL structures
  • csharp-pro - Modern C# with async/await, LINQ, .NET 6+
  • sql-pro - Complex SQL queries, optimization, schema design
  • sql-query-engineer - BigQuery, data analysis, and insights

Architecture & Design (7)

  • architect - System architecture, scalability, technical decisions
  • backend-architect - Backend systems, APIs, database design
  • graphql-architect - GraphQL schemas, resolvers, federation
  • docs-architect - Technical documentation, architecture guides
  • ui-ux-designer - Interface design, user experience, design systems
  • artistic-designer - Visual design, aesthetics, beautiful interfaces
  • branding-specialist - Brand identity, visual language, corporate identity

Code Quality (10)

  • code-reviewer - Expert code review, quality, security, maintainability
  • debugger - Root cause analysis, error resolution, debugging workflows
  • refactorer - Code restructuring, design improvements
  • refactor-planner - Strategic refactoring plans, technical debt reduction
  • test-writer - Test suites, TDD workflows
  • test-designer-advanced - Edge cases, chaos engineering, property-based testing
  • modernizer - Legacy code updates, modern practices adoption
  • investigator - Deep debugging, root cause analysis
  • criticizer - Critical analysis, constructive feedback
  • reflector - Deep reflection, retrospectives, continuous improvement

Performance (4)

  • performance - Whole-system performance optimization, profiling, benchmarking
  • concurrency-expert - Thread safety, synchronization, parallel patterns
  • memory-expert - Memory optimization, leak detection, allocation analysis
  • database-optimizer - Query optimization, indexing, schema efficiency

Specialized (13)

  • ml-engineer - ML pipelines, model serving, feature engineering
  • mlops-engineer - ML infrastructure, experiment tracking, model registries
  • data-engineer - ETL pipelines, data warehouses, streaming architectures
  • quant-researcher - Financial models, trading strategies, market analysis
  • trading-system-architect - HFT systems, market making, order execution
  • security-auditor - Vulnerability review, OWASP compliance, secure auth
  • migrator - System migrations, schema changes, version upgrades
  • porter - Cross-platform code porting, language transitions
  • docs - Technical documentation from codebases
  • reference-builder - Exhaustive technical references, API docs
  • meta-programming-pro - Code generation, DSLs, abstractions
  • prompt-engineer - LLM prompts optimization, AI features
  • sales-automator - Sales outreach, proposal templates, pricing pages

Frontend & Mobile (4)

  • react-specialist - React components, hooks, state management
  • flutter-specialist - Flutter widgets, state management, platform channels
  • ios-developer - Swift/SwiftUI, UIKit, Core Data, App Store optimization
  • mobile-developer - React Native/Flutter, offline sync, push notifications

Infrastructure (2)

  • terraform-specialist - IaC best practices, modules, state management
  • analyzer - Deep analysis, pattern recognition, codebase insights

Maintenance (1)

  • tech-debt-resolver - Technical debt identification and strategic resolution

Commands (127 total)

Analysis (12 commands)

  • analyze/code/elaborate - Deep code elaboration and explanation
  • analyze/code/map - Map codebase architecture and structure
  • analyze/code/schema - Extract and document data schemas
  • analyze/data/data-flow - Analyze data flow patterns
  • analyze/data/data-viz - Create data visualizations
  • analyze/data/visualize - Generate visual representations
  • analyze/db/db-optimize - Database optimization recommendations
  • analyze/research/deep-dive - Research with web sources
  • analyze/research/deep-web-research - Extensive web research with citations
  • analyze/research/investigate - Investigative analysis of complex topics
  • analyze/research/quick-web-research - Fast web research
  • analyze/think/think - Structured thinking with sequential analysis

Code (14 commands)

  • code/analyze/analyze-deps - Dependency analysis
  • code/analyze/bottleneck - Identify performance bottlenecks
  • code/analyze/dependencies - Dependency graph analysis
  • code/analyze/deps - Quick dependency check
  • code/analyze/technical-debt - Technical debt assessment
  • code/fix/bug-fix - Systematic bug fixing workflow
  • code/generate/api - Generate API implementations
  • code/migrate/deno-ify - Migrate code to Deno
  • code/migrate/migrate - General migration workflows
  • code/navigate/related - Navigate to related files intelligently
  • code/refactor/refactor - Code refactoring workflow
  • code/refactor/simplify - Simplify complex code
  • code/refactor/standardize - Standardize code patterns

Context (25 commands)

  • Context loaders for Dragonfly, PostgreSQL, RedPanda, ScyllaDB
  • Deno Fresh and Deno scripting contexts
  • Go concurrency, ConnectRPC, and web contexts
  • Java Quarkus, Spring, and Temporal contexts
  • Kubernetes (Cilium, Flux, K8s, Talos) contexts
  • Observability and logging contexts
  • Project auto-detection context
  • Rust async, database, and web contexts
  • Security context
  • Testing contexts for Deno, Go, Java, and Rust
  • Web contexts (Fresh, GitHub CLI, GitHub Actions, Tailwind, Temporal, etc.)

Documentation (8 commands)

  • docs/analyze/explain - Explain code and documentation
  • docs/generate/api-docs - Generate API documentation
  • docs/generate/changelog - Generate changelog from commits
  • docs/generate/document - Generate full documentation
  • docs/generate/onboard - Create onboarding documentation
  • docs/manage/docs-add - Add new documentation sections
  • docs/manage/docs-init - Initialize documentation structure
  • docs/manage/docs-update - Update existing documentation

Git (6 commands)

  • git/commit/commit - Create atomic commits with Conventional Commits
  • git/commit/commit-push - Commit and push to remote
  • git/pr/pr-check - Check PR status and CI/CD state
  • git/pr/pr-create - Create PR with intelligent analysis
  • git/pr/pr-review - Review and manage PRs
  • git/pr/pr-update - Update existing PRs
  • git/review/review-git - Git history review

Meta (10 commands)

  • meta/command/generate-command - Generate new slash commands
  • meta/command/ideate-commands - Brainstorm command ideas
  • meta/extract/knowledge-extract - Extract knowledge from codebases
  • meta/ideate/ideate-new - Ideate new features and approaches
  • meta/reflect/reflection - Deep reflection on decisions and outcomes
  • meta/search/search-smart - Intelligent codebase search
  • meta/util/scratch - Scratch pad for quick notes
  • meta/util/translate - Translate content between languages

Scaffolding (8 commands)

  • scaffold/deno/scaffold-deno-fresh - Scaffold Deno Fresh app
  • scaffold/deno/scaffold-deno-script - Scaffold Deno script
  • scaffold/go/scaffold-go-connect - Scaffold Go Connect RPC service
  • scaffold/go/scaffold-go-http-server - Scaffold Go HTTP server
  • scaffold/java/scaffold-java-quarkus - Scaffold Quarkus application
  • scaffold/rust/scaffold-rust-axum - Scaffold Axum web service
  • scaffold/rust/scaffold-rust-cli - Scaffold Rust CLI application

Security (4 commands)

  • security/audit/audit - Security audit
  • security/audit/secrets-audit - Scan for exposed secrets
  • security/model/harden - Security hardening recommendations
  • security/model/threat-model - Generate threat models (STRIDE methodology)

Task Management (9 commands)

  • task/manage/add-code-reviews-to-task - Add code reviews to tasks
  • task/manage/task-archive - Archive completed tasks
  • task/manage/task-create - Create new tasks
  • task/manage/task-list - List all tasks
  • task/manage/task-log - Log task activities
  • task/manage/task-search - Search through tasks
  • task/manage/task-show - Show task details
  • task/manage/task-update - Update existing tasks
  • task/view/task - View task information

Testing (10 commands)

  • test/analyze/coverage - Test coverage analysis
  • test/fix/flaky-fix - Fix flaky tests
  • test/generate/integration-test - Generate integration tests
  • test/generate/test-gen - Generate test suites
  • test/run/load-test - Load testing
  • test/run/tdd - Test-driven development workflow
  • test/run/validate - Validation with auto-detection

Tools (5 commands)

  • tool/cpr - Code, PR, and review workflows
  • tool/diagram - Generate diagrams (nomnoml/mermaid)
  • tool/five - Five-question analysis framework
  • tool/review - Code review
  • tool/zed-task - Zed editor task integration

Workflow (17 commands)

  • workflow/create/epic - Create epic with sub-tasks
  • workflow/create/prototype - Rapid prototyping workflow
  • workflow/manage/clean - Clean up temporary files and artifacts
  • workflow/manage/integrate - Integration workflows
  • workflow/manage/organize - Organize project structure
  • workflow/manage/plan - Strategic planning
  • workflow/manage/release - Release management
  • workflow/manage/sync - Synchronization workflows
  • workflow/start/start - Start new project/feature workflow
  • workflow/view/next-steps - Suggest next steps
  • workflow/view/options - View available options
  • workflow/view/progress - Show progress
  • workflow/view/summary - Generate summary
  • workflow/view/tldr - Quick summary

Core Features

Diagram-First Engineering

  • Five mandatory diagrams: Architecture, Data Flow, Concurrency, Memory, Optimization
  • Nomnoml for conversations, Mermaid for documentation
  • Non-negotiable requirement for non-trivial implementations

Surgical Editing Workflow

  • Find → Copy → Paste pattern
  • AST-based transformations with ast-grep (highly preferred)
  • Minimal context extraction with precise targeting
  • Preview → Validate → Apply workflow

Atomic Commit Protocol

  • Conventional Commits v1.0.0 compliance
  • Type-classified commits (feat, fix, build, chore, ci, docs, perf, refactor, style, test)
  • One logical change per commit
  • Independent testability and reversibility

Confidence-Driven Execution

  • Adaptive behavior based on familiarity, complexity, risk, and scope
  • Four confidence levels with distinct patterns
  • Automatic confidence calibration based on outcomes

Tool Selection Mandate

  • ast-grep (HIGHLY PREFERRED) for code operations
  • native-patch for file edits
  • ripgrep for text/comments/strings
  • fd for file discovery (NEVER find)
  • eza for directory listing (NEVER ls)
  • Banned: sed for edits, find, ls, grep for code patterns

Configuration

settings.json

  • Full tool permissions (ast-grep, fd, rg, cargo, npm, git, etc.)
  • MCP integration (time, browser, git, context7, tavily, etc.)
  • Security denials (sed -i, force push, destructive commands)
  • Hooks for event-driven automation
  • Bypass permissions mode by default
  • Always thinking enabled

CLAUDE.md

  • ODIN methodology and principles
  • Language-specific guidelines (Rust 2024, TypeScript, Python, Go, Java, Kotlin, C++, C, JavaScript, PHP, C#, SQL)
  • UI/UX design guidelines
  • Verification and refinement patterns
  • Decision heuristics and frameworks

Documentation

  • Full README with installation, usage, and methodology
  • CHANGELOG for version tracking
  • Plugin manifest with full registry
  • Marketplace structure for distribution

Philosophy

Core Principles

  1. Investigate Before Acting - Never speculate about unread code
  2. Diagram-First Design - Five diagrams mandatory before implementation
  3. Surgical Precision - Minimal, targeted changes using AST-based tools
  4. Atomic Commits - One logical change per commit, properly typed
  5. Confidence-Driven - Adapt behavior based on familiarity and risk
  6. Tool Selection - ast-grep > native-patch > ripgrep

Quality Standards

  • Functional accuracy ≥ 95%
  • Code quality ≥ 90%
  • Design excellence ≥ 95%
  • Performance within budgets
  • Error recovery 100%
  • Security compliance 100%