RELEASES.md

Releases

natbag v1.2.3 (2026-08-21)

Bugfix: SSL fallback for Python installs without a CA bundle.

• python.org macOS Python builds ship with an empty certificate store until Install Certificates.command is run. When such a build owns python3 on PATH, every data.gov.il request failed with CERTIFICATE_VERIFY_FAILED, and the daily snapshot hook swallowed the error — history silently stopped accumulating.
• New shared scripts/ssl_fallback.py used by snapshot.py and query_flights.py: when the interpreter's default trust store is usable (including capath-based stores that report zero cached CAs), requests keep stdlib default behavior untouched; otherwise the first working CA bundle is used — certifi if importable, then OS bundles for macOS/BSD, Debian/Ubuntu, RHEL/Fedora, and openSUSE. Missing or corrupt bundles fall through to the next candidate, the resolved context is cached per process, and if nothing is usable a one-line actionable hint goes to stderr. No new hard dependencies.
• The SessionStart snapshot hook now logs to ~/.natbag/snapshot.log (size-capped) instead of /dev/null, so the next failure class leaves a trace instead of silently stopping history accumulation.
• Native manifests (Codex/Grok/Cursor) regenerated via scripts/sync-native-manifests.py.

Native multi-CLI marketplace support (2026-08-14)

The marketplace now ships first-class manifests for Claude Code, Codex, Grok, and Cursor:

• Claude Code: .claude-plugin/marketplace.json and per-plugin .claude-plugin/plugin.json.
• Codex: .agents/plugins/marketplace.json and per-plugin .codex-plugin/plugin.json.
• Grok: .grok-plugin/marketplace.json and per-plugin .grok-plugin/plugin.json.
• Cursor: .cursor-plugin/marketplace.json and per-plugin .cursor-plugin/plugin.json.

scripts/sync-native-manifests.py discovers new Claude marketplace entries automatically, infers presentation metadata, generates the three additional catalogs, wraps command-only plugins as Codex skills, and offers --check for CI-style drift detection. copy v2.4.0 adds generated Agent Skill forms of its eight commands, coacher v0.3.0 adds a portable skill plus Cursor always-on rule, natbag v1.2.2 refreshes its daily snapshot from the invocation workflow when hooks are unavailable, and find-session v1.2.1 fixes strict YAML parsing for its argument hint.

Marketplace rename (2026-08-14)

The repository and root marketplace were renamed from omri-cc-stuff to omri-marketplace to reflect support for multiple AI coding agents. Use omriariav/omri-marketplace for new installs; existing GitHub links redirect to the new repository. Individual plugin names and versions were unchanged by the rename.

For Claude Code and Codex, this is an identity migration: a plugin installed as x@omri-cc-stuff is distinct from x@omri-marketplace. Existing users should migrate in this order:

1. Register omriariav/omri-marketplace with the relevant plugin manager.
2. Reinstall each existing plugin under its new plugin@omri-marketplace ID.
3. Verify the new installation and preserve the previous enabled/disabled state.
4. Remove the corresponding plugin@omri-cc-stuff registrations, keeping plugin data where the manager supports it.
5. Remove the obsolete omri-cc-stuff marketplace source.

See the README for Claude Code, Codex, and Grok registration commands.

find-session v1.2.0 (2026-06-24)

New feature: /find-session now searches Codex transcripts alongside Claude Code sessions.

• Codex sessions (~/.codex/sessions/<Y>/<M>/<D>/rollout-<ts>-<uuid>.jsonl) are scanned in the same pass as Claude Code sessions. Results from both sources are interleaved by recency, each tagged with its source and the correct resume command (codex resume <uuid> vs claude --resume <uuid>).
• New flags: --source claude|codex|both (default both, with shorthand --claude / --codex / --both) to scope by source, and --include-subagents to include Codex subagent sessions (guardian auto-review judges and spawned children) that are otherwise excluded as non-resumable noise.
• Codex previews show the first genuine user message — injected role: user boilerplate (# AGENTS.md, <environment_context>, <skill>, <turn_aborted>, <<ccr:) is filtered from both the preview and keyword matching. Codex has no /rename, so Codex results are always preview-based.
• Current-project scope resolves Codex sessions by the cwd recorded in each transcript's session_meta, skipping non-matching files after a single line read (the default path stays ~1s). Cross-project --all keyword search uses a ripgrep/grep pre-filter over Codex transcripts (graceful fallback to a full scan when neither tool is present).
• --json output gains a source field and a per-result resume_cmd.
• Claude Code session parsing is unchanged (verified against the prior release).

natbag v1.2.1 (2026-06-03)

New plugin — added to the marketplace by migrating omriariav/natbag-skill in via git subtree (full commit history preserved).

• Ben Gurion Airport (TLV) live flight data, destination weather, and historical delay analysis. /natbag shows departures, arrivals, specific flights, status, destination weather, and accumulated on-time/delay history.
• New entry in the root marketplace.json (source: ./plugins/natbag). The marketplace's first plugin with hooks (SessionStart snapshot + PreToolUse freshness guard); both use ${CLAUDE_PLUGIN_ROOT} and resolve correctly under the monorepo layout.
• Runtime history accrues in ~/.natbag/flights.db (install-independent), so existing data carries over untouched.
• Dropped the imported repo's redundant nested .claude-plugin/marketplace.json (the monorepo reads only the root manifest).
• The standalone omriariav/natbag-skill repo is now deprecated and points here. If you had it installed, uninstall it to avoid double-firing the SessionStart hook.

coacher v0.2.0 (2026-06-03)

Moved into the marketplace as a local plugin — was previously listed in marketplace.json as an external url source pointing at omriariav/claude-coacher; now vendored locally via git subtree (history preserved).

• marketplace.json entry flipped from the git URL to source: ./plugins/coacher. No behavior change: still injects the SessionStart collaborator frame and provides /coacher:status, /coacher:audit, /coacher:reset, /coacher:rant.
• Dropped the imported repo's redundant nested .claude-plugin/marketplace.json.
• The standalone omriariav/claude-coacher repo is now deprecated and points here. If you had it installed from there, reinstall as coacher@omri-marketplace to avoid double-firing the SessionStart hook.

Marketplace housekeeping: the root plugin.json no longer carries a version field — versioning is strictly per-plugin (Claude Code keys installs on each plugin's own version, namespaced by marketplace; the root version drove nothing).

claude-reviewer v1.2.0 (2026-05-28)

New feature: fleet-level skill-cleaner signals on every config audit.

• Companion to skill-reviewer v1.1.0. Where /skill-reviewer surfaces skill-cleaner signals per-skill, /claude-reviewer now surfaces them fleet-wide across the audited scope. Every /claude-reviewer [project] or /claude-reviewer --global invocation now includes a ## Skill Fleet Audit section in its report with:
  • Fleet budget — total always-loaded description tokens across all skills in scope vs cap (default 2% of 200,000-tok context = 4,000 tok). Pressure: OK / WARN / OVER.
  • Description trim candidates — every skill with description ≥ heavy threshold (default 120 tok), sorted by token cost.
  • Duplicates — every skill name that appears in multiple roots, with keep-priority plugin > personal > repo.
  • Unused candidates — opt-in via --with-logs. Conservative path-only scan of ~/.claude/projects/*.jsonl for any skill in scope. Candidate signal only.
• New script: scripts/fleet_audit.py. Flags: --global, --root (repeatable), --with-logs, --months, --context-tokens, --budget-percent, --heavy-threshold, --json.
• Read-only — the script informs; per-skill trims flow through /skill-reviewer --fix <skill>.
• argument-hint exposes --with-logs; SKILL.md adds Phase 4 Step 1 documenting the integration; references/report-template.md adds the ## Skill Fleet Audit section between Inventory and Scores (in both project-mode and global-mode templates).
• Same fixes applied as cleaner_checks.py from skill-reviewer v1.1.0: tolerant YAML block-scalar parsing, os.walk directory pruning, realpath-deduped scanning.

skill-reviewer v1.1.0 (2026-05-28)

New feature: every per-skill review now surfaces skill-cleaner signals.

• Adapts the skill-cleaner methodology by @steipete for Claude Code as per-skill checks, not a separate audit mode. Every /skill-reviewer <skill> invocation now includes a "Skill Cleanliness Signals" section in its report with:
  • Description budget — token cost for the target skill (ceil(utf8_bytes/4)) vs cap (default 2% of a 200,000-token context). Flags CANDIDATE for trim above the heavy threshold (default 120 tok).
  • Duplicates — scans all skill roots for other copies of the same skill name; suggests which to keep via plugin > personal > repo priority.
  • Usage evidence — opt-in via --with-logs. Conservative path-only scan of ~/.claude/projects/*.jsonl for transcript paths like skills/<name>/SKILL.md. Candidate signal only.
• New script: scripts/cleaner_checks.py. Flags: --root (repeatable), --with-logs, --months, --context-tokens, --budget-percent, --heavy-threshold, --json.
• Read-only — the script informs; fixes still flow through /skill-reviewer --fix <skill>.
• argument-hint exposes --with-logs; SKILL.md adds the integration step as Phase 5, Step 1; references/report-template.md adds the section between Inventory and Scores.
• Fleet-level audits (totals across all skills, all-duplicates-at-once, all-unused) will land separately as an addition to /claude-reviewer, whose scope already covers the whole .claude/ setup.

x v2.1.2 (2026-05-27)

Maintenance — no behavior change.

• /x:read and /x:tweet command files collapsed into thin pointers to their respective skills/<name>/SKILL.md. The skills are now the single source of truth for both commands; the previously duplicated workflow prose (which had to be hand-synced on every change, e.g. the v2.1.1 URL-counting fix) is gone. Commands instruct the model to resolve scripts/... paths relative to the skill directory.

gdoc-math v1.0.2 (2026-05-27)

Maintenance — no behavior change.

• Trimmed the SKILL.md frontmatter description (tightened the opening line and compressed the NOT-for clauses) to cut always-loaded prompt-budget cost. All trigger phrases preserved.

skill-reviewer v1.0.1 (2026-05-27)

Maintenance.

• Trimmed the SKILL.md frontmatter description (dropped the inline dimension list — documentation, not a trigger) to reduce always-loaded prompt-budget cost.
• Fixed broken YAML in the SKILL.md frontmatter: argument-hint was unquoted and began with [, which a strict YAML parser reads as a flow sequence and rejects. Now single-quoted (matching the sibling reviewer skills). Surfaced by Codex review of this PR.

claude-reviewer v1.1.1 (2026-05-27)

Maintenance — no behavior change.

• Trimmed the SKILL.md frontmatter description (dropped the rubric-stats clause) to reduce always-loaded prompt-budget cost. Trigger phrases unchanged.

copy v2.3.1 (2026-05-27)

Maintenance — no behavior change.

• Removed an orphaned legacy unified-router command (.claude/commands/copy.md) that sat in a non-loading location and duplicated the eight split /copy:* commands. No effect on the active command surface.

gdoc-math v1.0.1 (2026-05-25)

Docs/example only — no pipeline changes.

• Bundled example swapped to a general math showcase: examples/math-demo.md (quadratic formula, Euler's identity, finite + geometric series, the derivative limit, the Gaussian integral, Pythagoras). Replaces the ML-specific laplace-smoothing.md sample so the demo reads as broadly recognizable math.
• SKILL.md and README.md updated to point at the new sample.

gdoc-math v1.0.0 (2026-05-25)

New plugin. /gdoc-math converts a Markdown file containing LaTeX math into a native Google Doc whose equations are real, editable equation objects — solving the gap that /copy:gdocs (Paste from Markdown) and Google's own "Export to Docs" can't: both leave $...$ / $$...$$ as literal text.

How it works

markdown+LaTeX  →[pandoc]→  .docx (OMML)  →[gws drive upload + convert --to docs]→  Google Doc
                                                                └─ intermediate .docx auto-trashed

pandoc renders LaTeX into Office MathML; Google Drive imports OMML as native, editable equations. Verified by round-tripping the converted Doc back to .docx — equations return as m:oMath objects with zero embedded images.

Features

• Inline ($…$), display ($$…$$), and \(…\) / \[…\] math.
• Input modes: a .md file path, inline content, or the last math-bearing content in the conversation.
• --name (Doc title), --folder (target Drive folder), --keep-docx (debug).
• Optional default_folder_id in config.json; empty by default (My Drive root), never required.
• Ships a runnable sample at examples/laplace-smoothing.md.

Requirements (verified by verify-setup.sh at Step 0)

• pandoc — brew install pandoc
• python3 — parses gws output (default on macOS)
• gws (Google Workspace CLI), authenticated for Drive — go install github.com/omriariav/workspace-cli/cmd/gws@latest, then gws auth login

Notes

• Simple fractions/subscripts/Greek round-trip cleanly; complex LaTeX (stacked fractions, matrices, large operators) can lose fidelity on Google's import and may need a touch-up.
• Kept standalone (not folded into copy) on purpose: it writes to Google Drive over the network with external dependencies — a different contract from copy's instant, offline, clipboard-only commands.

copy v2.3.0 (2026-05-04)

/copy:slack now produces a native Slack table on paste when the resolved content is just a table.

New

• Native Slack table paste (table-only path) — When the content is a bare table, /copy:slack puts a Google-Sheets-flavored payload on the macOS pasteboard: public.utf8-plain-text carries TSV, and public.html carries a <table> wrapped with <google-sheets-html-origin> and data-sheets-root="1". Slack's composer recognizes that exact shape and converts the paste into a real native table block — same render as a bot posting via the Block Kit table API.
• New ## Native Table Paste section in commands/slack.md with the full TSV + HTML + JXA NSPasteboard recipe (plain pbcopy is not enough for this path).

Changed

• Router (/copy slack): step 4 of the workflow now branches Slack into table-only (JXA path) vs mixed/prose (pbcopy); the destination rules and the example explicitly call out the same branch so a model following either prompt can't fall back to plain text and silently break native rendering.
• Mixed/prose Slack guidance (bold, links, QuickChart) is now scoped to the mixed path. Adding any prose or chart URL to the table-only path would convert the paste into mixed content and Slack would drop the native render — the docs spell this out.

Why "table-only"?

Slack's composer recognizes the Sheets-flavored clipboard only when the paste is a table. Add prose around the table and Slack falls back to plain text. For prose + native table together, the realistic options are:

• Two pastes (prose first, table-only second), or
• Use a Slack MCP / webhook that posts via chat.postMessage with a Block Kit table block.

End users still cannot construct a native table block in the composer manually — that's a Slack platform constraint, not a plugin gap.

x v2.0.0 (2026-04-29)

Breaking change: plugin renamed tweet → x. Command /tweet is now /x:tweet. Update any scripts or muscle memory.

New

• /x:read — Fetch a tweet by URL or ID via X API v2 (GET /2/tweets/:id) and load it into the conversation as markdown (author, timestamp, public metrics, replied-to / quoted tweets, media). Reuses /x:tweet Keychain credentials. Metered by X at ~$0.005 / post read.
• Plugin description updated to cover both posting and reading.

Changed

• Plugin directory plugins/tweet/ → plugins/x/.
• plugins/x/.claude-plugin/plugin.json name field is now x.
• marketplace.json plugin entry renamed and now lists both ./skills/tweet and ./skills/read.

Migration

If you had the old tweet plugin installed, Claude Code's plugin cache still points at plugins/tweet/ and /plugin update will fail with Plugin source not found at .../plugins/tweet. The directory and entry name both moved, so this is a fresh install, not an upgrade.

To migrate:

1. /plugin → select tweet → Uninstall
2. /plugin → install x from this marketplace
3. Restart Claude Code so the new commands register

Keychain credentials live under service x-api and are untouched — /x:tweet will pick them up immediately. Just update muscle memory: /tweet → /x:tweet, and you get /x:read for free.

tweet v1.3.0 (2026-03-30)

Voice learning and skill structure improvements.

New

• Voice learning: Reads LEARNINGS.json before drafting to match your established voice, appends observations after each posted tweet
• config.json for operational defaults (char target, thread style)
• references/ directory with extracted char-counting rules and drafting guidelines

Changed

• Scoped allowed-tools to Bash(python3*), Bash(bash scripts/*) for security
• Expanded description with negative trigger cases
• Trimmed SKILL.md from 134 to 92 lines via progressive disclosure
• Removed duplicate credential setup and verbose script output sections

copy v2.2.2 (2026-04-29)

Bugfix: /copy:slack now emits Markdown [text](url) links instead of Slack's <url|text> Block Kit syntax.

Fixed

• Slack: Links use Markdown [text](url) — renders as a hyperlink in the composer when "Format messages with markup" is enabled in Slack preferences. The previous <url|text> form is API/mrkdwn syntax and renders literally in the composer. Raw URLs remain a safe fallback (Slack auto-unfurls them). (#6)
• Router (/copy slack) updated to match.

copy v2.2.1 (2026-04-05)

Bugfix: resolve format inconsistencies between router and subcommands.

Fixed

• Jira: Router had stale wiki markup (||header||, *bold*); now matches jira.md Markdown syntax
• Gmail: gmail.md incorrectly used pbcopy; now uses textutil → RTF → osascript for rich text paste
• Router: Added richformat and cleartext to argument list, examples, and clipboard rules
• Router: Renamed /copy email → /copy gmail for consistency with actual command name

copy v2.2.0 (2026-03-30)

New

• /copy:md - Copy as standard Markdown (tables, headers, bold, links, code blocks)

---

v2.0.0 (2025-12-29)

Breaking Change: Restructured /copy command into separate format-specific commands.

New Commands

• /copy:slack - Format for Slack
• /copy:gchat - Format for Google Chat
• /copy:gmail - Format for Gmail/email (HTML)
• /copy:gdocs - Format for Google Docs (Markdown)
• /copy:jira - Format for Jira Cloud (Markdown)
• /copy:cleartext - Plain text
• /copy:richformat - Rich Text Format (RTF)

Changes

• Plugin renamed from copy-command to copy for cleaner namespacing
• Each format is now a separate command (/copy:slack instead of /copy slack)
• Jira format updated to use Markdown (Jira Cloud) instead of wiki markup
• All commands accept optional [what to copy] argument

Migration

Old: /copy slack the table New: /copy:slack the table

---

v1.x (2025-12-29)

Initial releases with single /copy command and various structural experiments.

• v1.3.0 - Skills-based structure attempt
• v1.2.0 - GitHub sync fixes
• v1.0.0 - Initial plugin release