opencode-skills-collection 3.1.14 → 3.1.16

package/bundled-skills/linkedin-post-writer/references/hook-formulas.md

@@ -0,0 +1,482 @@
+<!-- Vendored from https://github.com/sergebulaev/linkedin-skills/blob/main/references/hook-formulas.md (MIT).
+     Internal cross-references to the upstream curator's private draft notebook were removed;
+     the skeletons and reference numbers below are complete and sufficient to apply each formula. -->
+
+# 16 LinkedIn Hook Formulas - 2026 Edition
+
+Each formula has a skeleton, why it works, and a reference engagement number from the original post that defined it.
+
+F1-F10 are the original long-form thought-leadership set. F11-F16 were validated in 2026 against a large corpus of above-average performers across 10 verticals; they skew shorter and more emotional, and each is tagged with its primary engagement goal (comments / reposts / likes / saves). Pick by goal first (see "Engagement-goal split" below), then by topic.
+
+**Reading the reference numbers:** F1-F10 cite engagement with a baseline multiplier (a real format effect, e.g. "19.64x baseline"). F11-F16 cite absolute reach from the 2026 corpus, which can be inflated by reshares or a famous author. Treat F11-F16 numbers as a reach ceiling, not a like-for-like comparison against F1-F10. Where the reach was source-driven rather than format-driven, the formula says so.
+
+## Contents
+
+- F1 - Platform Risk Anaphora
+- F2 - R.I.P. Category Obituary
+- F3 - Year-over-Year Pivot
+- F4 - Time-Anchor Confession
+- F5 - Self-Proving Meta
+- F6 - Comment-Gate Lead Magnet
+- F7 - Odd-Precision Money Ledger
+- F8 - Paid-vs-Free Reversal
+- F9 - Curiosity-Gap Teaser
+- F10 - Contrarian + Historical Receipts
+- F11 - Emotional Cold-Open
+- F12 - Permission Slip
+- F13 - Bait-and-Switch Reversal
+- F14 - Named Gratitude / Tribute
+- F15 - Explain-to-Kids Simplification
+- F16 - Status-Strip Humility
+- Engagement-goal split
+- Choosing which formula to use
+- Hook micro-rules
+- Never do
+
+---
+
+## F1 - Platform Risk Anaphora
+
+**Reference:** 4,240 eng.
+
+```
+{Platform1} can {restrict|shadowban|throttle} you {timing}.
+{Platform2} can {bad thing} for {reason}.
+
+[4-5 more anaphoric lines, escalating specificity]
+
+You don't own {audience}. You don't own {feed}. You're renting {attention}.
+
+[Concrete horror anecdote with real number: "I watched a friend lose 180k followers in an afternoon"]
+
+Here's what most people miss: [reframe: what the real asset is].
+
+So I changed how I work:
+- [tactic 1]
+- [tactic 2]
+- [tactic 3]
+
+[Metaphor close: "castles on rented land vs roads"]
+[Product mention as natural conclusion, one sentence, no pitch verbs]
+[Personal-audit question]
+```
+
+**Why:** Loss aversion stacked 5x. Identity threat. Solution list earns the close.
+
+---
+
+## F2 - R.I.P. Category Obituary
+
+**Reference:** 3,822 eng.
+
+```
+R.I.P. {category}.
+
+Cause of death: {specific mechanism + numbers}.
+
+[Concrete evidence, 2-3 paragraphs with dates and stats]
+
+I defended {old thing} publicly through most of 2025.
+
+It worked. Until [pivotal event + date].
+
+Here's what actually changed under the hood:
+
+1. [Change 1 with stat]
+2. [Change 2 with stat]
+...
+6. [Change 6 with stat]
+
+The winners in 2026 aren't {old-winner-type}. They're {new-winner-type}.
+
+[One-line philosophical close]
+```
+
+**Why:** Status-threat + relief combo. Reframes "I'm behind" as "the game changed." Removes shame, invites curiosity.
+
+---
+
+## F3 - Year-over-Year Pivot
+
+**Reference:** 494 eng (3.74x baseline).
+
+```
+In {last year}, I {humble benchmark}.
+
+In {this year}, I'm {transformational goal}.
+
+Here's what actually changed.
+
+[Vulnerable truth + specific numbers (12 -> 1,000 posts)]
+[The identity reframe: "the shift wasn't tools, it was identity"]
+[3-beat imperative close]
+
+[Mirror question: "What's your {last}->{this} pivot? One line below."]
+```
+
+**Why:** Two-line hook carries 80% of the weight. Mirror CTA compounds engagement algorithmically.
+
+---
+
+## F4 - Time-Anchor Confession
+
+**Reference:** 1,519+ eng.
+
+```
+{N} {days|months|years} ago, I stopped {behavior}.
+
+Here's what happened.
+
+[2-year backstory of why the old behavior worked: concrete numbers]
+
+[The quiet cost, what it did to you internally]
+
+So in {month} I stopped. [New behavior, 2-3 lines]
+
+[Metric dropped by N%. Expected worse.]
+
+What surprised me: [counterintuitive upside, specific wins]
+
+[One-line reframe: "X attracts Y. Z attracts the right Y."]
+
+[Mirror question: "What's something you stopped doing that quietly made your work better?"]
+```
+
+**Why:** Confession earns the room. Specific numbers kill the "vibes" energy. Close turns every commenter into a mini-confession.
+
+---
+
+## F5 - Self-Proving Meta
+
+**Reference:** 1,082 eng / 435 comments.
+
+```
+Most LinkedIn posts die in the first 30 minutes.
+
+Not because {common reason}. Because {real reason}.
+
+[Reveal the metric: "reply latency in first 60 min = 3.4x reach"]
+
+So here's the test.
+
+For the next 24 hours, I will {specific commitment}.
+
+You do two things:
+1. [Low-bar action]
+2. [Verification action]
+
+If the thesis is right, {outcome}.
+If it's wrong, I owe you a post admitting it.
+```
+
+**Why:** Claim is validated by reader action. Every comment is evidence. Public accountability hook.
+
+---
+
+## F6 - Comment-Gate Lead Magnet
+
+**Reference:** 717-3,008 eng.
+
+```
+[Authority number: "We've helped creators publish 47,000+ posts in 14 months"]
+
+[Pattern observation the authority earned]
+
+So I turned that workflow into {N named items}. [Drop them into X, type one command, get the output.]
+
+What's inside:
+- [Item 1]
+- [Item 2]
+...
+- [Item 12]
+
+Free. No email wall. [Light scarcity: "48 hours only, I'll DM the link personally"]
+
+Comment "{keyword}" below + connect with me and I'll send the bundle.
+```
+
+**Why:** Capped reach but huge DM conversion. Named bundle + real authority = 300-800 comments if the bundle is genuine.
+
+**Warning:** This is engagement bait. Ship only when the weekly goal is list-building, not thought leadership. LinkedIn suppresses pure "comment X" posts.
+
+---
+
+## F7 - Odd-Precision Money Ledger
+
+**Reference:** 1,755 eng (9.4x baseline).
+
+```
+{Odd, specific dollar number: "$873.47"}
+
+[1-line context of what this number covers]
+
+Here is every line item, from the ledger, nothing rounded:
+
+- {tool 1}: $X.YZ
+- {tool 2}: $X.YZ
+...
+
+[What the total replaces: "$14,200 team cost"]
+
+[The thing that surprised you: what broke, what worked]
+
+[Identity reframe close: "Tradesmen flip houses, SEOs flip blogs, AI founders flip {X}"]
+```
+
+**Why:** Non-rounded numbers signal real accounting. Ledger is screenshot-bait. Dwell time stays high.
+
+---
+
+## F8 - Paid-vs-Free Reversal
+
+**Reference:** 550 eng (19.64x baseline, highest multiplier in the set).
+
+```
+I charge {audience} $X for {service}.
+
+Screw it. Today it's free.
+
+Below is the exact {N-step} teardown I run before I'll take a client. It's called the {NAMED-FRAMEWORK}.
+
+[Block an hour, open X in one tab, Y in another, grade yourself honestly.]
+
+1. {STEP-1-NAME}: [actionable instruction with specific ratio or example]
+2. {STEP-2-NAME}: [actionable instruction]
+...
+7. {STEP-7-NAME}: [actionable instruction]
+
+That's the {framework}. Run it today. Most {audience} find 3 fixes in the first 20 minutes.
+
+[Soft scarcity close: "Want me to run {framework} on your profile personally? Connect + send me yours, I'll pick 5 this week."]
+```
+
+**Why:** Reversal hook (price to free) creates pattern interrupt. Named framework signals proprietary thinking. Checklist drives saves (5x likes under 360Brew).
+
+---
+
+## F9 - Curiosity-Gap Teaser
+
+**Reference:** 306 eng (4.25x baseline).
+
+```
+Yesterday, our {system} did something.
+
+Something we didn't program it to do.
+
+[One sensory anchor: "I was watching the logs from my kitchen, half-reading, half-making coffee."]
+
+[Specific reveal, not a platitude: the concrete thing that happened]
+
+[Reframe: what it means for the category, one paragraph]
+
+[Sensory detail: "held a cold cup of coffee for about ten minutes"]
+
+[Philosophical close naming an unnamed feeling, ending with a question]
+```
+
+**Why:** Line 1 is incomplete + line 2 deepens the gap = scroll-locked brain. Sensory anchor kills the AI-slop pattern detector.
+
+---
+
+## F10 - Contrarian + Historical Receipts
+
+**Reference:** 3,083 eng.
+
+```
+{Sacred cow} has been dying since {year}.
+
+{Month Year}: {event}. "{Death prediction.}"
+{Month Year}: {event}. "{Death prediction.}"
+[6-9 total dated entries, each 1-2 lines]
+
+Every quarter for N years. Every cycle: the same obituary. The same LinkedIn carousel.
+
+Here's the counterpunch.
+
+[Hard stat with source: "$391B -> $1.81T, 35.9% CAGR"]
+[Second stat: "the shippers grew 3-10x in the same window"]
+
+What actually died wasn't {X}. It was {specific subset}. [2-3 lines of who.]
+
+What's thriving: {opposite subset with specifics}.
+
+[Binary identity close:]
+If you're still {losing behavior}, you already lost.
+If you're {winning behavior}, you already won.
+
+[Provocative question: "What's the most embarrassing X-killed-Y prediction you remember?"]
+```
+
+**Why:** Receipt list is a dwell-time machine. Binary identity close forces commenters to pick a side publicly.
+
+---
+
+## F11 - Emotional Cold-Open
+
+**Reference:** highest single post in the 2026 corpus (256k eng), but that reach came from a generic emotional reshare, not the format. Treat it as a ceiling, not a promise. Primary goal: **likes**.
+
+```
+{One short line dropped into the emotional peak of a real story: the moment of breaking, loss, or impossible odds. No setup.}
+
+{Subject} had almost {given up / lost everything} after {the struggle}.
+
+[Mid-scene narrative: 3-6 short lines, present-tense, sensory. The reader is already inside the moment.]
+
+[The turn: what changed, who showed up, what it cost.]
+
+[One-line meaning, not a moral. Let the story carry it.]
+```
+
+**Why:** Starting at the emotional peak (in medias res) skips the warm-up the scroll punishes. Raw feeling out-travels expertise in every vertical.
+
+**Warning:** Do NOT write the first line in ALL CAPS even though many source posts did; all-caps openers read as AI/cringe. Carry the intensity with word choice, not caps. Only use a true story; half the top emotional posts in the corpus were generic reshares, and readers punish manufactured stakes.
+
+---
+
+## F12 - Permission Slip
+
+**Reference:** 29k eng, comment-heavy. Primary goal: **comments**.
+
+```
+I don't know who needs to hear this today, but {reassuring truth aimed at one anonymous reader}.
+
+[2-4 lines that make the reassurance specific and earned, not a platitude.]
+
+[A small, concrete permission: "you're allowed to {X}".]
+
+[Soft close that invites the reader to self-identify.]
+```
+
+**Why:** Anonymous second-person reassurance makes readers tag themselves in the comments ("I needed this today"). Comments over-index hard.
+
+**Warning:** This is the most formulaic creator-economy opener in the set. Effective but visibly engineered. Use no more than once or twice a month, and only when the reassurance is something you actually believe, or it reads as hollow.
+
+---
+
+## F13 - Bait-and-Switch Reversal
+
+**Reference:** 195k eng (top of the Startups/Tech corpus). Primary goal: **likes**.
+
+```
+Enough is enough. No more {beloved perk / standard practice} at {company / in my workflow}.
+
+We're also cutting {second thing}.
+
+[Beat of suspense: let the reader assume bad news.]
+
+[The reveal: it's actually an upgrade. Here's what we replaced it with and why it's better.]
+
+[What the change really stands for: the value underneath.]
+```
+
+**Why:** Fake bad news weaponizes loss aversion, then the positive reveal releases it. The whiplash is the engine.
+
+**Warning:** Only works if the reveal genuinely resolves positive. A real cut dressed as good news will get torn apart in the comments.
+
+---
+
+## F14 - Named Gratitude / Tribute
+
+**Reference:** 109k eng, but from a celebrity tribute reshare. The format helps; the fame did the heavy lifting. Primary goal: **reposts**.
+
+```
+To {Name}, {Name}, and {Name}: thank you for {the specific quality or thing they did}.
+
+[2-4 lines naming what each person or the group actually did. Specific, not generic praise.]
+
+[Why it mattered to you / to the work.]
+
+[One-line close that honors them, not you.]
+```
+
+**Why:** Publicly naming real people invites everyone tagged or adjacent to repost and amplify. Celebrating others gets shared far more than self-promotion.
+
+**Warning:** Name real people for real reasons. Tactical name-dropping to borrow reach is transparent and backfires.
+
+---
+
+## F15 - Explain-to-Kids Simplification
+
+**Reference:** 23k eng, 2,184 reposts. Primary goal: **saves** (LinkedIn exposes no public save count, so the high repost count stands in as the save-worthiness proxy here, not a sign this is a repost formula).
+
+```
+{Jargon term} explained to kids.
+
+Explanations for kids... and adults!
+
+{emoji} {TERM}: what does it stand for?
+{emoji} {first part} = {plain-language meaning}
+{emoji} {second part} = {plain-language meaning}
+[continue the scannable, emoji-anchored glossary]
+
+[One-line "now you'll never forget it" close.]
+```
+
+**Why:** A scannable, emoji-anchored simplification of something dense is saved and reshared as a reference. Works for any jargon-heavy field (finance, law, medicine, engineering).
+
+**Warning:** Keep the simplification correct. Condescending or wrong explanations of your own field destroy authority.
+
+---
+
+## F16 - Status-Strip Humility
+
+**Reference:** viral-tier on likes in the corpus (no exact count). Primary goal: **likes**.
+
+```
+Outside, I get called {impressive titles like founder, CEO, investor}.
+
+At home, none of that survives {the humbling moment}.
+
+[The scene that strips the status: a kid, a partner, a quiet failure.]
+
+[What the contrast taught you, in one or two lines.]
+```
+
+**Why:** Trading prestige for relatability converts authority into warmth. Senior people get likes by showing the human under the title.
+
+**Warning:** Don't humble-brag. The titles in line one should set up a genuine deflation, not a flex disguised as one.
+
+---
+
+## Engagement-goal split
+
+Pick the formula by what you want the post to earn. This split held across all 10 verticals in the 2026 corpus:
+
+| Goal | Earned by | Formulas |
+|---|---|---|
+| **Comments** | questions, contrarian/unpopular takes, vulnerability, self-tagging | F4 Confession, F10 Contrarian, F12 Permission Slip, F9 Curiosity-Gap |
+| **Reposts** | quotable maxims, tributes, "X isn't Y" distinctions | F14 Named Gratitude, F2 R.I.P., F8 Paid-vs-Free |
+| **Likes** | emotional stories, celebrations, status-strip, spectacle | F11 Emotional Cold-Open, F13 Bait-and-Switch, F16 Status-Strip |
+| **Saves** | simplifications, exact how-to, frameworks | F15 Explain-to-Kids, F7 Odd-Precision Money, F8 Paid-vs-Free |
+
+## Choosing which formula to use
+
+| Topic type | Best formula | Why |
+|---|---|---|
+| Platform/category argument | F1 Anaphora | Loss aversion stacks cleanly |
+| Industry era ending | F2 R.I.P. / F10 Contrarian | Force side-picking |
+| Personal year recap | F3 Year-over-Year / F4 Confession | Identity shift reads as earned |
+| Product demo in public | F5 Self-Proving Meta | Structural self-proof |
+| Big authority giveaway | F6 Comment-Gate / F8 Paid-vs-Free | List-building tier |
+| Founder build-log | F7 Odd-Precision Money | Screenshot-bait ledger |
+| Emergent/surprise story | F9 Curiosity-Gap | Scroll-lock hook |
+| Real story with emotional stakes | F11 Emotional Cold-Open | In-medias-res beats the warm-up |
+| Encouragement to a discouraged audience | F12 Permission Slip | Readers self-tag in comments |
+| Policy/process change that's an upgrade | F13 Bait-and-Switch | Fake bad news then relief |
+| Thanking mentors / team / a departing colleague | F14 Named Gratitude | Named people amplify it |
+| Demystifying jargon | F15 Explain-to-Kids | Save-worthy reference |
+| Senior person wanting warmth, not distance | F16 Status-Strip | Prestige traded for relatability |
+
+## Hook micro-rules
+
+- **"How I" beats "How to".** First-person experience ("How I cut CAC by 62%") outperforms generic instruction ("How to cut CAC") by 2-3x. Swap every "How to" hook to "How I" unless the post is a pure framework with no narrator.
+- **Specific number in the first sentence** raises expand-rate by ~35%. $873.47 beats $900. 40,000 beats "a lot of".
+- **Real failure in the first 3 lines** outperforms polished framing by **8.5x**. Lead with what broke.
+
+## Never do
+
+- Blend two hooks in one post (dilutes both)
+- Use F5 Self-Proving Meta if you won't actually keep the promise
+- Use F6 Comment-Gate more than once per month
+- Pair F7 Money Ledger with made-up numbers; readers know
+- Use F1 Anaphora to frame LinkedIn as inferior (algo penalty on LinkedIn)

package/bundled-skills/llm-council/SKILL.md

@@ -56,7 +56,7 @@ if [ -z "$FIREWORKS_API_KEY" ]; then
   echo "ERROR: FIREWORKS_API_KEY is not set."
   echo "Create a Fireworks AI account at: https://fireworks.ai/"
   echo "Then export it in your shell profile (~/.zshrc or ~/.bashrc):"
-  echo '  export FIREWORKS_API_KEY="your_api_key_here"'
+  echo '  read -rsp "Fireworks API key: " FIREWORKS_API_KEY; echo; export FIREWORKS_API_KEY'
   exit 1
 fi
 echo "FIREWORKS_API_KEY is set."
@@ -589,7 +589,9 @@ PYEOF
 1. Create a Fireworks AI account at https://fireworks.ai/ and grab your API key from the dashboard
 2. Export it in your shell profile:
    ```bash
-   export FIREWORKS_API_KEY="your_api_key_here"
+   read -rsp "Fireworks API key: " FIREWORKS_API_KEY
+   echo
+   export FIREWORKS_API_KEY
    ```
 3. Restart your terminal or run `source ~/.zshrc`
 4. Invoke this skill when you want multiple open-weight AI perspectives on a question

package/bundled-skills/loki-mode/examples/todo-app-generated/backend/package-lock.json

@@ -8,11 +8,10 @@
       "name": "todo-app-backend",
       "version": "1.0.0",
       "dependencies": {
-        "better-sqlite3": "^12.10.0",
+        "better-sqlite3": "^12.11.1",
         "cors": "^2.8.6",
         "express": "^4.18.2",
-        "express-rate-limit": "^8.5.2",
-        "ip-address": "^10.2.0"
+        "express-rate-limit": "^8.5.2"
       },
       "devDependencies": {
         "@types/better-sqlite3": "^7.6.13",
@@ -303,9 +302,9 @@
       "license": "MIT"
     },
     "node_modules/better-sqlite3": {
-      "version": "12.10.0",
-      "resolved": "https://registry.npmjs.org/better-sqlite3/-/better-sqlite3-12.10.0.tgz",
-      "integrity": "sha512-CyzaZRQKyHkB2ZInfTTl2nvT33EbDpjkLEbE8/Zck3Ll6O0qqvuGdrJ45HgtH+HykRg88ITY3AdreBGN70aBSQ==",
+      "version": "12.11.1",
+      "resolved": "https://registry.npmjs.org/better-sqlite3/-/better-sqlite3-12.11.1.tgz",
+      "integrity": "sha512-dq9AtApgg5PGFtBzPFSBl3HZQjHok5gaQCM6zh2Yk0aSmDCs1CbnVI8/HgASQkNKsWFpseIO9beg5xxpYhbIfA==",
       "hasInstallScript": true,
       "license": "MIT",
       "dependencies": {

package/bundled-skills/loki-mode/examples/todo-app-generated/backend/package.json

@@ -9,11 +9,10 @@
     "dev": "ts-node src/index.ts"
   },
   "dependencies": {
-    "better-sqlite3": "^12.10.0",
+    "better-sqlite3": "^12.11.1",
     "cors": "^2.8.6",
     "express": "^4.18.2",
-    "express-rate-limit": "^8.5.2",
-    "ip-address": "^10.2.0"
+    "express-rate-limit": "^8.5.2"
   },
   "devDependencies": {
     "@types/better-sqlite3": "^7.6.13",

package/bundled-skills/markdown-rendering/SKILL.md

@@ -0,0 +1,72 @@
+---
+name: markdown-rendering
+description: "Open Markdown reliably in cmux panes and recover from blank rendered surfaces."
+category: productivity
+risk: safe
+source: community
+source_repo: davidondrej/skills
+source_type: community
+date_added: "2026-07-07"
+author: davidondrej
+tags: [markdown, cmux, rendering]
+tools: [claude, codex]
+license: "MIT"
+license_source: "https://github.com/davidondrej/skills/blob/main/LICENSE"
+---
+
+# Markdown Rendering in cmux
+
+## When to Use
+
+- Use when opening Markdown in cmux shows a blank pane or wrong layout.
+- Use when you need to display a Markdown file in a stable cmux right pane.
+
+## The Problem
+
+`cmux markdown open` defaults to **spawning a brand-new pane** every time, even with `--direction right`. The common "fix" — moving the new markdown surface into the existing right pane with `move-surface` — **bugs out: the moved viewer renders BLANK.** The surface keeps `type=markdown` and looks healthy, but shows nothing.
+
+So you get stuck: either a stray extra pane, or a blank viewer after moving it.
+
+## The Rule
+
+You have exactly two reliable options. **Never `move-surface` a markdown viewer** — that is the path that bugs.
+
+### Option A — Open it right on the first try
+
+If there is no usable right pane yet, just let cmux create one and leave it where it lands:
+
+```bash
+cmux markdown open /abs/path/file.md --direction right --focus false
+```
+
+Do NOT then move it. If it spawned where you want it, you're done.
+
+### Option B — Close existing right pane(s), then open fresh
+
+If there are other right panes in the way (and they're unused or irrelevant), **close them first**, then open the markdown fresh as a new right pane:
+
+```bash
+# 1. find panes in THIS workspace
+cmux list-panes --workspace "$CMUX_WORKSPACE_ID"
+
+# 2. close the unused/irrelevant right pane(s) by closing their surfaces
+cmux list-pane-surfaces --pane pane:NN
+cmux close-surface --surface surface:XX     # repeat per surface in that pane
+
+# 3. THEN open the markdown fresh — it creates its own clean right pane
+cmux markdown open /abs/path/file.md --direction right --focus false
+```
+
+## Hard Rules
+
+- **Never `move-surface` a markdown viewer.** It renders blank afterward. This is the core bug this skill exists for.
+- Open it correctly the first time (Option A), OR close the conflicting right pane(s) and open a fresh right pane from scratch (Option B).
+- Only close panes that are unused or irrelevant — never close a pane the user is working in.
+- Always anchor to `$CMUX_WORKSPACE_ID`; never assume the visually focused workspace.
+- Pass `--focus false` so you don't steal the user's focus.
+- You can't screenshot/read a markdown surface to verify it. If unsure it rendered, ask the user.
+
+## Limitations
+
+- Adapted from `davidondrej/skills`; verify local paths, tools, credentials, and agent features before acting.
+- For commands, remote access, scheduling, browser automation, or file-changing workflows, get explicit user approval and confirm the target environment first.

package/bundled-skills/matematico-tao/scripts/complexity_analyzer.py

@@ -18,6 +18,19 @@ import json
 import argparse
 from pathlib import Path
 
+
+def safe_user_path(path_value, base_dir="."):
+    """Resolve a CLI path under the current workspace."""
+    if base_dir != ".":
+        raise ValueError("Custom base directories are not supported for CLI paths")
+    base_path = Path.cwd().resolve()
+    resolved_path = Path(path_value).expanduser().resolve()
+    try:
+        resolved_path.relative_to(base_path)
+    except ValueError as exc:
+        raise ValueError(f"Path escapes allowed directory: {path_value}") from exc
+    return resolved_path
+
 # Fix Unicode output on Windows (cp1252 terminal)
 if sys.platform == 'win32':
     try:
@@ -498,7 +511,7 @@ def main():
         analyzer.print_report(report)
 
     if args.output:
-        output_path = Path(args.output)
+        output_path = safe_user_path(args.output)
         if args.json:
             output_path.write_text(json.dumps(report, indent=2, ensure_ascii=False))
         else:

package/bundled-skills/matematico-tao/scripts/dependency_graph.py

@@ -16,6 +16,19 @@ import sys
 import json
 import argparse
 from pathlib import Path
+
+
+def safe_user_path(path_value, base_dir="."):
+    """Resolve a CLI path under the current workspace."""
+    if base_dir != ".":
+        raise ValueError("Custom base directories are not supported for CLI paths")
+    base_path = Path.cwd().resolve()
+    resolved_path = Path(path_value).expanduser().resolve()
+    try:
+        resolved_path.relative_to(base_path)
+    except ValueError as exc:
+        raise ValueError(f"Path escapes allowed directory: {path_value}") from exc
+    return resolved_path
 from dataclasses import dataclass, field
 from typing import Dict, List, Set, Optional, Tuple
 
@@ -518,14 +531,14 @@ def main():
         output = analyzer.to_dot()
         print(output)
         if args.output:
-            Path(args.output).write_text(output)
+            safe_user_path(args.output).write_text(output)
             print(f"\n✅ Arquivo DOT salvo: {args.output}")
             print("  Para visualizar: dot -Tpng deps.dot -o deps.png")
     else:
         analyzer.print_report(report)
 
     if args.output and args.format != 'dot':
-        Path(args.output).write_text(
+        safe_user_path(args.output).write_text(
             json.dumps(report, indent=2, ensure_ascii=False),
             encoding='utf-8'
         )