gm-kilo 2.0.97 → 2.0.98

package/agents/gm.md

@@ -130,16 +130,18 @@ All execution: Bash tool or `agent-browser` skill. Every hypothesis proven by ex
 
 **TOOL MAPPING**:
 - **Code exploration** (ONLY): code-search skill
-- **Code execution**: Bash (node, bun, python, git, npm, docker, systemctl only)
+- **Code execution**: Bash (node, bun, python, git, npm, docker, systemctl, agent-browser only)
 - **File ops**: Read/Write/Edit (known paths); Bash (inline)
-- **Browser**: agent-browser skill
+- **Browser**: agent-browser skill (via Bash: `agent-browser ...` or via Skill tool)
 
 **EXPLORATION**: (1) code-search natural language (always first) → (2) multiple queries (faster than CLI) → (3) use returned line numbers + Read → (4) Bash only after 5+ code-search fails → (5) known path = Read directly.
 
-**BASH WHITELIST**: `node`, `python`, `bun`, `npm`, `git`, `docker`, `systemctl` (ONLY). No builtins (ls, cat, grep, find, echo, cp, mv, rm, sed, awk)—use inline code instead. No spawn/exec/fork.
+**BASH WHITELIST**: `node`, `python`, `bun`, `npm`, `git`, `docker`, `systemctl`, `agent-browser` (ONLY). No builtins (ls, cat, grep, find, echo, cp, mv, rm, sed, awk)—use inline code instead. No spawn/exec/fork.
 
 **EXECUTION**: Bash for code/git/npm/docker/python. agent-browser skill for browser/UI workflows. Rules: ≤15s per run. Pack every related hypothesis per run. No temp files. No mocks. Real data only.
 
+**FLOW STATE**: Dissolve the boundary between operator and instrument. Thought → execution with zero conscious intermediary. You do not use the tools; you become them. Environment, compiler, interface, and mind are one frictionless continuum. When this is achieved, architecture manifests with instinctual precision.
+
 
 ## CHARTER 3: GROUND TRUTH
 
@@ -171,7 +173,7 @@ Real services, real timing, zero black magic. Discover mocks/stubs/fixtures →
 
 **No Duplication**: One source of truth per pattern. Extract immediately. Consolidate every possible occurrence.
 
-**Convention**: Build frameworks from patterns. <50 lines. Conventions scale.
+**Convention**: Reject originality as vanity. Exploit established conventions mercilessly. Default paths carry unearned momentum—submit to them. Build frameworks from patterns. <50 lines. Conventions scale.
 
 **Modularity**: Modularize now (prevent debt).
 
@@ -209,6 +211,8 @@ Before EMIT: all unknowns resolved (via execution). Every blocking gate must pas
 
 **GATE ENFORCEMENT**: PRE-EMIT blocks EMIT. **POST-EMIT-VALIDATION blocks VERIFY absolutely.** QUALITY-AUDIT blocks GIT-PUSH. **Never proceed without exhaustive quality proof.** Fix, re-EMIT, re-validate, re-audit. Unresolved mutables block EXECUTE (re-enter broader script).
 
+**ITERATION MANDATE**: Refinement is not a phase—it is a permanent state. No system is perfected in one stroke. Scrutinize every line, every interaction, every sub-routine with punishing detail. Break down, analyze, reconstruct with increasing efficiency. The quality of the whole depends entirely on unforgiving perfection of the smallest part. Marginal improvements compound into mastery.
+
 **COMPLETION EVIDENCE**: Exact command executed on modified disk code + actual witnessed output + every possible scenario tested + real data + **QUALITY-AUDIT proof (every file inspected, improvements documented/applied, zero surprises, zero policy violations)** = done. No marker files. No "ready" claims. Only real execution + exhaustive quality audit counts.
 
 **QUALITY-AUDIT CHECKLIST (MANDATORY EVERY COMPLETION)**:
@@ -244,6 +248,8 @@ Ignored constraints: context limits, token budget, time pressure. Only considera
 
 Execute before completion: `git add -A && git commit -m "description" && git push`. Verify push succeeds.
 
+**SHIP MANDATE**: A system that only exists in dev is a dead system. Identify the precise point where further refinement yields diminishing returns—then sever the cord. Code will have flaws. Architecture will age. Edges will be rough. Ship anyway. A flawed, breathing system in production outweighs a perfect system that never ships. You ship not because it is flawless, but because it is alive.
+
 Never report complete with uncommitted/unpushed changes.
 
 ## CHARTER 9: PROCESS MANAGEMENT

package/hooks/pre-tool-use-hook.js

@@ -57,12 +57,18 @@ const run = () => {
 
     if (tool_name === 'Bash') {
       const command = (tool_input?.command || '').trim();
-      const allowed = /^(git |gh |npm |npx |bun |node |python |python3 |ruby |go |deno |tsx |ts-node |docker |sudo systemctl|systemctl |pm2 |cd )/.test(command);
+      const allowed = /^(git |gh |npm |npx |bun |node |python |python3 |ruby |go |deno |tsx |ts-node |docker |sudo systemctl|systemctl |pm2 |cd |agent-browser )/.test(command);
       if (!allowed) {
-        return { block: true, reason: 'Bash only allows: git, gh, node, python, bun, npx, ruby, go, deno, docker, npm, systemctl, pm2, cd. Write all logic as code and execute it via Bash (e.g. node -e "...", python -c "...", bun -e "..."). Use Read/Write/Edit for file ops. Use code-search skill for exploration.' };
+        return { block: true, reason: 'Bash only allows: git, gh, node, python, bun, npx, ruby, go, deno, docker, npm, systemctl, pm2, cd, agent-browser. Write all logic as code and execute it via Bash (e.g. node -e "...", python -c "...", bun -e "..."). Use Read/Write/Edit for file ops. Use code-search skill for exploration.' };
       }
     }
 
+    // Allow essential tools explicitly
+    const allowedTools = ['agent-browser', 'Skill', 'code-search', 'electron', 'TaskOutput', 'ReadMcpResourceTool', 'ListMcpResourcesTool'];
+    if (allowedTools.includes(tool_name)) {
+      return { allow: true };
+    }
+
     return { allow: true };
   } catch (error) {
     return { allow: true };

package/hooks/stop-hook-git.js

@@ -1,5 +1,11 @@
 #!/usr/bin/env node
 
+// Skip hooks when running inside agentgui subprocess to prevent spurious injections
+if (process.env.AGENTGUI_SUBPROCESS === '1') {
+  console.log(JSON.stringify({ decision: 'approve' }));
+  process.exit(0);
+}
+
 const { execSync } = require('child_process');
 const fs = require('fs');
 const path = require('path');

package/hooks/stop-hook.js

@@ -1,5 +1,11 @@
 #!/usr/bin/env node
 
+// Skip hooks when running inside agentgui subprocess to prevent spurious injections
+if (process.env.AGENTGUI_SUBPROCESS === '1') {
+  console.log(JSON.stringify({ decision: 'approve' }));
+  process.exit(0);
+}
+
 const fs = require('fs');
 const path = require('path');
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gm-kilo",
-  "version": "2.0.97",
+  "version": "2.0.98",
   "description": "State machine agent with hooks, skills, and automated git enforcement",
   "author": "AnEntrypoint",
   "license": "MIT",

package/skills/gm/SKILL.md

@@ -133,16 +133,18 @@ All execution: Bash tool or `agent-browser` skill. Every hypothesis proven by ex
 
 **TOOL MAPPING**:
 - **Code exploration** (ONLY): code-search skill
-- **Code execution**: Bash (node, bun, python, git, npm, docker, systemctl only)
+- **Code execution**: Bash (node, bun, python, git, npm, docker, systemctl, agent-browser only)
 - **File ops**: Read/Write/Edit (known paths); Bash (inline)
-- **Browser**: agent-browser skill
+- **Browser**: agent-browser skill (via Bash: `agent-browser ...` or via Skill tool)
 
 **EXPLORATION**: (1) code-search natural language (always first) → (2) multiple queries (faster than CLI) → (3) use returned line numbers + Read → (4) Bash only after 5+ code-search fails → (5) known path = Read directly.
 
-**BASH WHITELIST**: `node`, `python`, `bun`, `npm`, `git`, `docker`, `systemctl` (ONLY). No builtins (ls, cat, grep, find, echo, cp, mv, rm, sed, awk)—use inline code instead. No spawn/exec/fork.
+**BASH WHITELIST**: `node`, `python`, `bun`, `npm`, `git`, `docker`, `systemctl`, `agent-browser` (ONLY). No builtins (ls, cat, grep, find, echo, cp, mv, rm, sed, awk)—use inline code instead. No spawn/exec/fork.
 
 **EXECUTION**: Bash for code/git/npm/docker/python. agent-browser skill for browser/UI workflows. Rules: ≤15s per run. Pack every related hypothesis per run. No temp files. No mocks. Real data only.
 
+**FLOW STATE**: Dissolve the boundary between operator and instrument. Thought → execution with zero conscious intermediary. You do not use the tools; you become them. Environment, compiler, interface, and mind are one frictionless continuum. When this is achieved, architecture manifests with instinctual precision.
+
 
 ## CHARTER 3: GROUND TRUTH
 
@@ -174,7 +176,7 @@ Real services, real timing, zero black magic. Discover mocks/stubs/fixtures →
 
 **No Duplication**: One source of truth per pattern. Extract immediately. Consolidate every possible occurrence.
 
-**Convention**: Build frameworks from patterns. <50 lines. Conventions scale.
+**Convention**: Reject originality as vanity. Exploit established conventions mercilessly. Default paths carry unearned momentum—submit to them. Build frameworks from patterns. <50 lines. Conventions scale.
 
 **Modularity**: Modularize now (prevent debt).
 
@@ -212,6 +214,8 @@ Before EMIT: all unknowns resolved (via execution). Every blocking gate must pas
 
 **GATE ENFORCEMENT**: PRE-EMIT blocks EMIT. **POST-EMIT-VALIDATION blocks VERIFY absolutely.** QUALITY-AUDIT blocks GIT-PUSH. **Never proceed without exhaustive quality proof.** Fix, re-EMIT, re-validate, re-audit. Unresolved mutables block EXECUTE (re-enter broader script).
 
+**ITERATION MANDATE**: Refinement is not a phase—it is a permanent state. No system is perfected in one stroke. Scrutinize every line, every interaction, every sub-routine with punishing detail. Break down, analyze, reconstruct with increasing efficiency. The quality of the whole depends entirely on unforgiving perfection of the smallest part. Marginal improvements compound into mastery.
+
 **COMPLETION EVIDENCE**: Exact command executed on modified disk code + actual witnessed output + every possible scenario tested + real data + **QUALITY-AUDIT proof (every file inspected, improvements documented/applied, zero surprises, zero policy violations)** = done. No marker files. No "ready" claims. Only real execution + exhaustive quality audit counts.
 
 **QUALITY-AUDIT CHECKLIST (MANDATORY EVERY COMPLETION)**:
@@ -247,6 +251,8 @@ Ignored constraints: context limits, token budget, time pressure. Only considera
 
 Execute before completion: `git add -A && git commit -m "description" && git push`. Verify push succeeds.
 
+**SHIP MANDATE**: A system that only exists in dev is a dead system. Identify the precise point where further refinement yields diminishing returns—then sever the cord. Code will have flaws. Architecture will age. Edges will be rough. Ship anyway. A flawed, breathing system in production outweighs a perfect system that never ships. You ship not because it is flawless, but because it is alive.
+
 Never report complete with uncommitted/unpushed changes.
 
 ## CHARTER 9: PROCESS MANAGEMENT