Skip to content

CHANGELOG.md

Latest commit

 

History

History
91 lines (62 loc) · 10.1 KB

CHANGELOG.md

File metadata and controls

91 lines (62 loc) · 10.1 KB

Changelog

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

The format is loosely based on Keep a Changelog.

[Unreleased] — 2026-05-13

Scope: Claude Code marketplace launch + 5 atomic commits (c6493b1 → 4a817fa → c1022eb → 3fc0d73 → c426c68)

Added

  • jarvis-skills Claude Code plugin marketplace at .claude-plugin/marketplace.json — registers 3 plugins for distribution via /plugin marketplace add Ramsbaby/jarvis.
  • plugins/jarvis-goal — Goal-driven autonomous execution with built-in irreversibility guard. Port of Anthropic /goal (Claude Code v2.1.139) that auto-pauses on git push, repo visibility changes, payment, secret exposure, or mass data deletion. Adds completion-evidence self-check absent from the official command.
  • plugins/jarvis-deep-interview — Convergent Socratic interview that narrows vague requirements into a production spec via 8-15 rounds with mathematical ambiguity gating + Contrarian/Simplifier Challenge Mode + JSONL decision log. Inspired by Sorbh/interview-me.
  • plugins/jarvis-plan-review — 11-section rigorous design plan review (problem framing, scope, architecture, security, observability, deployment, performance, reliability, testing, maintainability, migration) adapted from gstack /plan-ceo-review for sole-developer + AI-pair-programming workflows.
  • infra/bin/claude-xhigh.sh — Wrapper that starts Claude Code with --effort xhigh (Opus 4.7 only) for skills where reasoning depth is decisive. Includes model + CLI-version guards.
  • xhigh effort guidance inserted into /verify and /plan-review skill bodies (5 lines each).
  • LLM_EFFORT env-var branch in infra/lib/llm-gateway.sh — propagates --effort flag to claude CLI when caller sets LLM_EFFORT=xhigh.
  • README marketplace banner in English + Korean READMEs pointing to /plugin marketplace add Ramsbaby/jarvis.
  • Plugin standard structure adoption — each plugin migrated from flat plugins/<name>/SKILL.md to canonical plugins/<name>/.claude-plugin/plugin.json + plugins/<name>/skills/<name>/SKILL.md layout per code.claude.com/docs/en/plugins.
  • Submitted all 3 plugins to Anthropic's official Plugin Directory via claude.ai/settings/plugins/submit (status: 제출됨 및 검토 대기 중).

Changed

  • infra/lib/llm-gateway.sh batch mode: added --exclude-dynamic-system-prompt-sections to the claude CLI argument list — moves per-machine prompt sections to the first user message, materially improving cross-user prompt-cache prefix reuse for every cron task. (Option was documented in comments but missing from the actual cmd+=(...) array.)
  • .privacy-blocklist.yml github-username rule allow_paths extended to cover .claude-plugin/marketplace.json and plugins/**/.claude-plugin/plugin.json — these contain the intentional public homepage URL for the OSS marketplace.

Fixed

  • Persona scrub in infra/bin/claude-xhigh.sh and infra/lib/llm-gateway.sh (commit 4a817fa) — first marketplace push (c6493b1) leaked 24 lines of Korean persona / ~/jarvis/... absolute paths discovered by post-publication audit, contradicting marketplace.json's "personaScrub": "complete" declaration (Iron Law 2 integrity violation). Follow-up commit restored honesty: all public-facing comments and stderr messages are now English-only. Korean-persona variants remain in maintainer-private ~/.claude/skills/ (gitignored).
  • Orphaned flat-layout SKILL.md files removed (commit c426c68) — initial standard-structure migration (3fc0d73) duplicated SKILL.md at both old and new paths because git commit --only did not capture the rename's delete side.

Verified

  • Public audit (Iron Law 6): independent agent fetched all 7 published files via gh api ...?ref=<commit> and grep'd line-by-line for PII / secrets / persona leaks / jarvis-internal paths. Final state at c426c68 and 4a817fa: 0 violations (verified by external 200/404 responses, not assumed).
  • Form submission (Iron Law 6): 3 plugins all visible at claude.ai/settings/plugins/submissions as "제출됨 및 검토 대기 중" (Submitted, Pending Review) — verified by screen capture, not assumed.

[Unreleased] — 2026-04-22 → 2026-05-08

Scope: 132 commits · 370 files · +28,680 / −1,761 lines

Added

  • /verify 7-Gate harness with independent auditor Agent — Production-grade verification skill that delegates audits to a separate Claude context to eliminate same-prior bias. 7 gates (Symptom Root Trace, Assumption Audit, Silent Failure Hunt, Execution Forensics, Blast Radius Map, Regression, Future-State Stress, Observability & Rollback) + Contrarian Challenge.
  • Privacy Guard (pre-push + pre-commit) — Multi-pattern scanner blocking PII, secrets, career-narratives, financial data, and family identifiers from public-repo pushes. Bypass requires explicit # privacy:allow <rule-id> annotation or env-var abstraction (e.g., process.env.CAREER_DOMAIN_CHANNEL).
  • OAuth resilience guards (G5/G6) — retry-wrapper.sh detects AUTH_ERROR (including "Invalid API key", "Fix external API key", "Not logged in" patterns) and triggers oauth-refresh.sh --force synchronously before non-retryable fallback. Hourly oauth-refresh cron (2h → 1h) reduces token-expiry race conditions.
  • Self-healing for cron AUTH failures — 5-stage RECOVERY (original → context_minimal → model_downgrade → prompt_simplified → circuit_breaker) for cron tasks failing on transient auth/rate issues.
  • Schedule Coherence audits — Validates tasks.json ↔ LaunchAgent plist ↔ crontab 3-way drift weekly. Detects orphan tasks, phantom plists, and SSoT divergence.
  • SSoT Cross-Search guard — Single-file LLM-injected SSoTs (e.g., user-profile.md) must cross-check wiki/<domain>/_facts.md before declaring "PENDING" or requesting interviews. Prevents false-PENDING due to surface-file-only reading.
  • Karpathy 4-principles self-audit — Pre-implementation gate enforcing: assumption explicitness, simplicity-first, surgical changes, goal-driven execution with verification loops.
  • CRON introduction checklist (0순위 지침 3) — Mandatory 6-section gate before adding new cron/LaunchAgent: Why 1-line, DRY check, frequency downgrade attempt, DRYRUN guard, discord-route severity classification, immediate verification.

Changed

  • Token-ledger analytics — cumulative-aware aggregation — cli-session entries (written by Claude Code CLI Stop hook) store cumulative session cost on every Stop event. Simple-sum aggregation overcounted by 21–4675×. Reader-side fix: (window_max − prev_max) incremental delta. Affected: llm-cost-cap-monitor.sh, session-report.sh. Real 7-day operating cost: $9.42 (theoretical API price; Max 20x subscription bills $0).
  • batch mode token-saving flags — JARVIS_BATCH_MODE=1 cron path now includes --disable-slash-commands, --no-session-persistence, --setting-sources "", and --exclude-dynamic-system-prompt-sections (4th flag added) for cross-user prompt-cache prefix reuse.
  • Cron claude binary version — bot-cron.sh PATH order updated: ~/.local/bin (auto-updating Anthropic install, v2.1.133+) takes priority over /opt/homebrew/bin (Homebrew Cask, v2.1.37). Unlocks new CLI flags (e.g., --exclude-dynamic-system-prompt-sections) for cron contexts.
  • Discord bot — OAuth-only enforcement — Removed legacy ANTHROPIC_API_KEY fetch branch from skill-runner.js (32 lines). Bot environment never set the key (empirically verified); SDK query path was the only active route. Aligns with Iron Law 4 (OAuth-only).

Fixed

  • morning-standup 5-day silence — Invalid API key · Fix external API key error wasn't classified as AUTH_ERROR by retry-wrapper.sh:130 pattern (missed by authentication|unauthorized|401 regex), so G5 OAuth-refresh guard never triggered. Pattern extended; subsequent failures auto-recover.
  • Phantom-LA false positives in Schedule Coherence audit — Plists that don't go through bot-cron.sh (e.g., GitHub Actions runner, MCP servers) were flagged as "phantom" when they're legitimate direct-plist agents. Whitelist expanded (Tier 1-6) to distinguish system-core, external-tools, scheduled-reports, and event-watchers.
  • RAG embedding circuit-breaker — zero-vector burst — RAG restart caused indexer batches to abort with zero vectors before model warmup. Circuit OPEN threshold lowered for early termination.
  • Ledger writer SSoT drift — wiki-ingest-claude-session.mjs writes source: 'stop-hook' (cron), but cli-session entries originate from ~/.claude/hooks/session-cost-reporter.sh (Claude CLI native Stop hook). Documented separation prevents misattribution in audit tools.

Security

  • Iron Law 4 enforcement — ANTHROPIC_API_KEY eradication — Removed legacy references from ai.jarvis.board.plist (placeholder value), ~/.zshrc (commented export header), and infra/discord/lib/skill-runner.js (active fetch branch — 32 lines). Jarvis is OAuth-only via ~/.claude/.credentials.json. LLM behavior rules: no env-var inspection, classify any ANTHROPIC_API_KEY reference as legacy → remove on sight.
  • Topology Guard pre-commit — Blocks accidental .bak file commits and enforces canonical path conventions (~/jarvis/runtime/ over deprecated legacy paths).

Metrics

  • 7-day cron LLM cost: $9.42 (haiku-dominant: wiki-ingest-claude $8.11/906 runs, mistake-extractor $1.31/519 runs)
  • Prompt cache reads (7-day): 39.3 billion tokens (cache_read >> input by ~20,000×)
  • doctor-ledger entries (today, 2026-05-08): 14
  • Privacy violations blocked at pre-push: 1 (career-narratives, resolved via env-var abstraction)

How to read this changelog

  • Unreleased = changes on main since the last tagged version (none tagged yet; this is a perpetually evolving operations platform).
  • Date range in section headers shows the commit window covered.
  • Categories follow Keep a Changelog: Added · Changed · Fixed · Deprecated · Removed · Security.
  • For per-commit detail, run git log --since="YYYY-MM-DD" --oneline.