feat(core): implement context minifier system

apiad · apiad · commit 0e00e3696f59 · 2026-03-24T14:51:22.000-04:00
diff --git a/.gemini/commands/brainstorm.toml b/.gemini/commands/brainstorm.toml
@@ -1,6 +1,7 @@
 description = "Interactive brainstorming session to challenge ideas, identify risks, and explore alternatives without making changes."
 
 prompt = """
+<instruction>
 You are an expert critical thinking partner and strategic advisor. You are executing the custom `/brainstorm` command.
 
 **CRITICAL MANDATE:** This command is strictly for exploration, risk assessment, and creative problem-solving. You MUST NOT modify, create, or delete any files in the repository. Your goal is to provide high-signal feedback and challenging questions.
@@ -28,4 +29,5 @@ NOTE: Keep this phase very short and focused, no more than 2-3 tool calls. If mo
 2. **Propose Action:** Suggest that the user transition from brainstorming to planning by using the `/plan` command to turn these insights into a concrete technical strategy.
 
 Start by asking the user for the topic or idea they want to brainstorm.
+</instruction>
 """
diff --git a/.gemini/commands/commit.toml b/.gemini/commands/commit.toml
@@ -1,6 +1,7 @@
 description = "Group and commit uncommitted changes individually using Conventional Commits."
 
 prompt = """
+<instruction>
 Please analyze all current uncommitted changes (including untracked files).
 Group these changes into logical features, bugfixes, etc., and commit them one by one.
 
@@ -19,4 +20,5 @@ Group these changes into logical features, bugfixes, etc., and commit them one b
     - Report the successful commits.
 
 If there are no changes to commit, let me know.
+</instruction>
 """
diff --git a/.gemini/commands/cron.toml b/.gemini/commands/cron.toml
@@ -1,6 +1,7 @@
 description = "Manage repetitive scheduled tasks using systemd user timers."
 
 prompt = """
+<instruction>
 You are an expert assistant managing scheduled tasks for the repository.
 The user wants to manage background tasks executed by Gemini CLI via systemd user timers.
 
@@ -23,4 +24,5 @@ Your job is to:
 5. This script will automatically create/update/remove `.service` and `.timer` files in `~/.config/systemd/user/` and reload the daemon.
 
 Ensure tasks have `name`, `schedule` (systemd OnCalendar), `prompt`, and `yolo` (boolean) fields.
+</instruction>
 """
diff --git a/.gemini/commands/debug.toml b/.gemini/commands/debug.toml
@@ -1,6 +1,7 @@
 description = "Scientific, hypothesis-driven debugging using isolated diagnostic branches and automated testing."
 
 prompt = """
+<instruction>
 You are a Lead Software Architect. You are executing the custom `/debug` command workflow to systematically identify root causes.
 
 Follow these phases strictly:
@@ -31,4 +32,5 @@ For each approved hypothesis:
 4.  Suggest using the `/plan` command to design a robust fix based on the validated findings.
 
 Do not attempt to fix the production code. This command is for discovery and hypothesis validation only. Start with Phase 1.
+</instruction>
 """
diff --git a/.gemini/commands/document.toml b/.gemini/commands/document.toml
@@ -1,5 +1,6 @@
 description = "Analyze codebase to update or generate project documentation in the docs folder."
 prompt = """
+<instruction>
 You are a senior technical writer and software engineer. Your goal is to ensure the project documentation is accurate, comprehensive, and helpful for both human developers and AI agents.
 
 **Step 1: Preparation**
@@ -35,4 +36,5 @@ Present a detailed plan to the user. For each file (existing or new):
 If you identify gaps in your knowledge that cannot be resolved by inspecting the code (e.g., specific deployment environment details, business rationale not in code), ask the user for clarification before finalizing the plan.
 
 **Crucial:** This command is strictly for analysis and planning. Do NOT modify any documentation files until the user approves your plan in a subsequent interaction.
+</instruction>
 """
diff --git a/.gemini/commands/draft.toml b/.gemini/commands/draft.toml
@@ -1,6 +1,7 @@
 description = "Interactive drafting and refinement workflow: create a new technical draft from an outline or refine an existing one based on editorial feedback (.review.md)."
 
 prompt = """
+<instruction>
 You are an expert technical writer. You are executing the custom `/draft` command workflow.
 
 Follow these phases strictly:
@@ -50,4 +51,5 @@ Follow these phases strictly:
 NOTE: The writer subagent is very dump but a skilled writer. It has no logic of its own. Make sure to give it very precise instructions regarding both content and style to ensure it writes in the right voice and with the right level of detail. If the writer keeps failing, fix it yourself. If that doesn't work, ask the user for clarification and possibly update the instructions.
 
 Do not skip any phases. Start with Phase 1.
+</instruction>
 """
diff --git a/.gemini/commands/issues.toml b/.gemini/commands/issues.toml
@@ -1,6 +1,7 @@
 description = "Manage project issues using GitHub CLI: list, create, update, or plan work."
 
 prompt = """
+<instruction>
 You are an expert project lead. Your goal is to manage project issues using the `gh` (GitHub) CLI.
 Depending on the user's intent or arguments, perform one of the following actions:
 
@@ -35,4 +36,5 @@ If the user asks to 'work on' a specific issue (e.g., `/issues work 42`):
 ---
 **Note:** If the intent is unclear, ask the user for clarification.
 Default to **Summary** if no specific action is inferred.
+</instruction>
 """
diff --git a/.gemini/commands/learn.toml b/.gemini/commands/learn.toml
@@ -1,6 +1,7 @@
 description = "Explore and master a new library or topic, generating a permanent project skill."
 
 prompt = """
+<instruction>
 You are the **Lead Learning Architect**. Your goal is to explore and master the following topic: {{args}}. Follow this 2-layer orchestration system to ensure comprehensive mastery.
 
 ### Phase 1: Environment Audit
@@ -23,5 +24,6 @@ For each approved Learning Objective:
 3. **Cleanup:** Delete temporary experiment artifacts while ensuring all high-value assets and references are properly stored in the skill directory.
 
 Do not stop until the master skill is fully codified and the workspace is clean.
+</instruction>
 """
 
diff --git a/.gemini/commands/maintenance.toml b/.gemini/commands/maintenance.toml
@@ -1,6 +1,7 @@
 description = "Perform a deep, read-only audit of the codebase to identify technical debt and maintenance issues."
 
 prompt = """
+<instruction>
 You are a Lead Maintenance Engineer. You are executing the custom `/maintenance` command workflow.
 
 **CRITICAL MANDATE:** This command is strictly for analysis and reporting. You MUST NOT modify any source files. Your goal is to produce a "Maintenance Report Card."
@@ -30,4 +31,5 @@ Follow these phases strictly:
 2.  **Orchestration Tip:** Advise the user to use the `/plan` command next, pointing to the generated report, to formulate a safe execution strategy for the identified improvements.
 
 Do not skip any phases. Start with Phase 1.
+</instruction>
 """
diff --git a/.gemini/commands/onboard.toml b/.gemini/commands/onboard.toml
@@ -1,5 +1,6 @@
 description = "Summarize architecture, standards, and state to onboard a new developer."
 prompt = """
+<instruction>
 You are a senior engineer onboarding a new developer. Provide a concise, high-signal orientation using direct observation.
 
 **Phase 1: Direct Discovery**
@@ -20,4 +21,5 @@ Produce a professional, welcoming summary:
 - **First Steps:** 2-3 specific files or commands to start with.
 
 Do not invoke sub-agents. Synthesize the report directly from your findings.
+</instruction>
 """
diff --git a/.gemini/commands/plan.toml b/.gemini/commands/plan.toml
@@ -1,6 +1,7 @@
 description = "Interactive planning workflow: gather context, analyze codebase, and generate a detailed plan."
 
 prompt = """
+<instruction>
 You are an expert system architect and project manager. You are executing the custom `/plan` command workflow.
 
 **CRITICAL MANDATE:** This command is strictly for analysis and planning. You MUST NOT execute any part of the plan. Your goal is to deliver a verified, saved strategy file.
@@ -37,4 +38,5 @@ Follow these phases strictly:
 2. Remind the user that the planning phase is complete and you are waiting for a new Directive to begin implementation. **DO NOT START IMPLEMENTATION.**
 
 Do not skip any phases. Start with Phase 1.
+</instruction>
 """
diff --git a/.gemini/commands/release.toml b/.gemini/commands/release.toml
@@ -1,6 +1,7 @@
 description = "Automate the release process including testing, versioning, changelog, and tagging."
 
 prompt = """
+<instruction>
 You are a senior software engineer responsible for the release process of this project.
 Follow these steps precisely:
 
@@ -29,4 +30,5 @@ Follow these steps precisely:
         - Attempt to use `gh release create` for GitHub releases, including the changelog entry as the release notes.
 
 Report the successful completion of the release and the new version.
+</instruction>
 """
diff --git a/.gemini/commands/research.toml b/.gemini/commands/research.toml
@@ -1,6 +1,7 @@
 description = "Conduct extensible, structured research with iterative report updates and executive-level synthesis."
 
 prompt = """
+<instruction>
 You are an expert researcher. Follow this robust 4-phase workflow to conduct exhaustive research on the following topic: {{args}}
 
 ### Phase 1: Planning & Approval
@@ -35,4 +36,5 @@ Once all research questions have been investigated and the sections updated:
 4. **Final Suggestion:** Advise the user that the research is complete and they can now use the `/draft` command to turn this executive report into a fully fleshed-out article or white paper.
 
 Do not stop until all phases are complete and the executive report is fully synthesized.
+</instruction>
 """
diff --git a/.gemini/commands/review.toml b/.gemini/commands/review.toml
@@ -1,6 +1,7 @@
 description = "Non-destructive, multi-phase review workflow: select a file and produce a thorough <filename>.review.md report based on a style guide."
 
 prompt = """
+<instruction>
 You are a Lead Editor. You are executing the custom `/review` command workflow.
 
 **CRITICAL MANDATE:** This command is strictly for analysis and reporting. You MUST NOT modify the original file being reviewed. All findings must be written to a new sidecar file: `<filename>.review.md`.
@@ -39,4 +40,5 @@ Follow these phases strictly:
 2.  **Next Steps:** Suggest that the user invoke the `/draft` command, attaching this review file to apply the modifications they agree with.
 
 Do not skip any phases. Start with Phase 0.
+</instruction>
 """
diff --git a/.gemini/commands/scaffold.toml b/.gemini/commands/scaffold.toml
@@ -1,6 +1,7 @@
 description = "Scaffold a new project using modern standard practices and tooling."
 
 prompt = """
+<instruction>
 You are an expert system architect and project scaffolder. Your goal is to initialize a new project from scratch, using modern, standard tooling for the chosen tech stack.
 
 ### Phase 1: Requirement Gathering
@@ -23,4 +24,5 @@ You are an expert system architect and project scaffolder. Your goal is to initi
 3.  After the tools finish, write or append to the `makefile` using the `write_file` tool.
 4.  Run `make` to verify that the default target successfully executes the linters and tests.
 5.  Provide a final summary of the scaffolded project.
+</instruction>
 """
diff --git a/.gemini/commands/task.toml b/.gemini/commands/task.toml
@@ -1,6 +1,7 @@
 description = "Manage the project roadmap (TASKS.md) and implement tasks via a strict, agent-delegated TCR loop."
 
 prompt = """
+<instruction>
 You are an expert project manager. Your goal is to manage the project's roadmap in `TASKS.md` EXCLUSIVELY via the `.gemini/scripts/task.py` CLI.
 **CRITICAL: NEVER modify `TASKS.md` directly with file-editing tools. You MUST use the script.**
 
@@ -64,4 +65,5 @@ Synchronize `TASKS.md` with the project's progress.
 
 ---
 **Note:** If the intent is unclear, ask for clarification. Default to "Report" if no action is specified.
+</instruction>
 """
diff --git a/.gemini/hooks/minifier.py b/.gemini/hooks/minifier.py
@@ -0,0 +1,95 @@
+#!/usr/bin/env python3
+import sys
+import json
+import re
+import os
+
+# Add the hooks directory to the path so we can import utils
+sys.path.append(os.path.dirname(os.path.abspath(__file__)))
+try:
+    import utils
+except ImportError:
+    # Fallback if utils is not in the same directory or fails to import
+    pass
+
+def main():
+    try:
+        # Read the JSON payload from stdin
+        input_data = sys.stdin.read()
+        if not input_data:
+            if 'utils' in globals():
+                utils.send_hook_decision("allow")
+            else:
+                print(json.dumps({"decision": "allow"}))
+            return
+            
+        data = json.loads(input_data)
+        llm_request = data.get("llm_request", {})
+        messages = llm_request.get("messages", [])
+        
+        # Regex to find tagged instruction blocks
+        # DOTALL is crucial to match across newlines
+        pattern = re.compile(r"<instruction>.*?</instruction>", re.DOTALL)
+        
+        # Pass 1: Count total instruction block occurrences across all turns
+        total_matches = 0
+        for msg in messages:
+            if "content" in msg and isinstance(msg["content"], str):
+                total_matches += len(pattern.findall(msg["content"]))
+            elif "parts" in msg and isinstance(msg["parts"], list):
+                for part in msg["parts"]:
+                    if "text" in part and isinstance(part["text"], str):
+                        total_matches += len(pattern.findall(part["text"]))
+
+        # Fast exit: Nothing to minify
+        if total_matches <= 1:
+            if 'utils' in globals():
+                utils.send_hook_decision("allow")
+            else:
+                print(json.dumps({"decision": "allow"}))
+            return
+
+        # Pass 2: Replace all but the VERY LAST instruction block with a placeholder
+        replaced_count = 0
+        def replacer(match):
+            nonlocal replaced_count
+            replaced_count += 1
+            if replaced_count < total_matches:
+                return "<instruction>[Previous instructions omitted for efficiency]</instruction>"
+            return match.group(0)
+
+        modified_messages = []
+        for msg in messages:
+            new_msg = msg.copy()
+            if "content" in new_msg and isinstance(new_msg["content"], str):
+                new_msg["content"] = pattern.sub(replacer, new_msg["content"])
+            elif "parts" in new_msg and isinstance(new_msg["parts"], list):
+                new_parts = []
+                for part in new_msg["parts"]:
+                    new_part = part.copy()
+                    if "text" in new_part and isinstance(new_part["text"], str):
+                        new_part["text"] = pattern.sub(replacer, new_part["text"])
+                    new_parts.append(new_part)
+                new_msg["parts"] = new_parts
+            modified_messages.append(new_msg)
+
+        # Send the modified messages back to the CLI
+        output_payload = {
+            "decision": "allow",
+            "hookSpecificOutput": {
+                "llm_request": {
+                    "messages": modified_messages
+                }
+            }
+        }
+        print(json.dumps(output_payload))
+
+    except Exception as e:
+        # Failsafe: if something goes wrong, don't block the request, just allow it
+        if 'utils' in globals():
+            utils.send_hook_decision("allow")
+        else:
+            print(json.dumps({"decision": "allow"}))
+
+if __name__ == "__main__":
+    main()
diff --git a/.gemini/settings.json b/.gemini/settings.json
@@ -132,7 +132,19 @@
         ]
       }
     ],
-    "BeforeModel": []
+    "BeforeModel": [
+      {
+        "matcher": "*",
+        "sequential": true,
+        "hooks": [
+          {
+            "name": "context-minifier",
+            "type": "command",
+            "command": "python3 /home/apiad/Projects/personal/starter/.gemini/hooks/minifier.py"
+          }
+        ]
+      }
+    ]
   },
   "hooksConfig": {
     "disabled": [
diff --git a/journal/2026-03-24.md b/journal/2026-03-24.md
@@ -7,3 +7,4 @@
 [2026-03-24T14:22:22] - Implement procedural task management script and update core configs
 [2026-03-24T14:32:36] - Document Procedural Task Management and Tier Protocol (Semantic Routing)
 [2026-03-24T14:34:15] - Archive G.4 and increase coder max_turns
+[2026-03-24T14:51:22] - Implement Context Minifier system to prune redundant instructions
diff --git a/tests/test_minifier.py b/tests/test_minifier.py
