reghis86
UserDeterministic context engineering for AI coding agents — structure over search, tokens over brute force. Claude Code / Cursor plugin: 18 rules, 20 skills, 11 enforcement hooks.
Categories
Indexed Skills (20)
clear-context
Orchestrates tool result and thinking block clearing. Checks accumulated context pressure (threshold: 30K tokens) and emits correct context_management.edits params. Implements rule 014-cf-tool-result-clearing workflow. <example> Context: Agentic session, tool results at 45K tokens user: "/clear-context" assistant: "[CLEAR-CONTEXT] Action: clear_tool_uses_20250919 via context_management.edits; keeping 3 tool uses." </example>
compile-wiki
Karpathy Wiki pattern: reads source files for a domain, generates structured wiki pages (one per major concept/module), and builds an index. Stored persistently so future sessions read the index (~500 tokens) instead of raw source files. <example> Context: Starting work on the billing domain and need persistent reference docs. user: "/compile-wiki --domain billing --sources libs/billing/src/" assistant: "Running compile-wiki for billing domain — generates wiki pages and index at ~/worklogs/wiki/billing/." </example>
end-session
Runs the end-of-session close-out sequence for a work repo: capture diary + session-learnings, then commit all changes and auto-merge the working branch into the main branch. Invoke when the user signals the session is over ("kończymy sesję", "we're done", "wrap up", "/end-session"). <example> Context: Feature work finished on branch feature/x, user wants to close out. user: "kończymy sesję" assistant: "[END-SESSION] diary + learnings captured, committed 6 files, merged feature/x → main." </example>
evolve
Runs the self-evolving loop v2 reviewer panel. Pre-filters signal, gates on Evidence, then spawns a Fit/Cost/Risk Haiku panel, and writes human-gated proposals (ADD/MODIFY/MERGE/PRUNE/REJECT) to <IDE_DIR>/evolve/pending/. NEVER edits rules or skills — proposals only. <example> Context: SessionStart showed "SELF-EVOLVE: 7 new signals — run /evolve". user: "/evolve" assistant: "[EVOLVE DIGEST] 2 proposals written to .claude/evolve/pending/ — human approval needed." </example>
help
Lists all 20 ContextForge skills grouped by category, shows ownership and Superpowers delegation links, and reports current system status (manifests, worklogs, context capacity). <example> Context: Developer wants to know what skills are available user: "/help" assistant: "ContextForge 1.1.0 — 20 skills across 4 categories. System status: manifests FRESH, worklogs active." </example>
log-analyzer
Distills build, test, or lint failure logs into a structured Root Cause Analysis (RCA) block. Extracts root cause, failing file, line number, and a 1-line fix suggestion. Returns [RCA] envelope. <example> Context: npm run build failed, safe-exec stored the log at a path. user: "Analyze the build failure log at ~/worklogs/logs/run_20260414_101530_a1b2c3d4.log" assistant: "Running log-analyzer — returns RCA block with root cause and fix suggestion." </example>
optimize-rules
Audits all rules for token cost, detects cross-rule duplication, and suggests activation mode changes to keep the always-on budget at or below 6k tokens. Reports full budget vs target. <example> Context: Rules have grown after several sessions; time to trim token cost user: "/optimize-rules" assistant: "Audit complete — always-on budget 7.2k (target ≤6k). 2 duplication clusters found. 3 mode-change suggestions." </example>
pre-review
Gathers code review context (git diff, changed files, manifest freshness, impact radius) and delegates to superpowers:requesting-code-review. Sets review marker to satisfy pre-commit hook. <example> Context: Feature implementation complete, about to commit user: "/pre-review" assistant: "Context gathered (14 files, impact radius 23). Delegating to superpowers:requesting-code-review." </example>
record-failure
Creates a sharded failure ledger YAML at <IDE_DIR>/lessons/ when an unrecoverable error or circuit-breaker halt occurs. Checks for 3-strike patterns and auto-triggers /evolve. <example> Context: npm run build failed twice with the same TS error (circuit breaker triggered) user: "[CIRCUIT_BREAKER_HALT] — recording failure" assistant: "Written .claude/lessons/2026/04/14/a1/a1b2c3d4-barrel-export-missing.yaml" </example>
refresh-manifest
Regenerates the lightweight shadow manifest for symbol lookup. Auto-detects IDE (.claude or .cursor), dual-writes to both shadow directories. Use when manifest is missing, stale, or broken. <example> Context: Symbol lookup returns no results — manifest is missing. user: "The manifest is missing, I need to regenerate it" assistant: "Running refresh-manifest ��� will auto-detect IDE and generate manifests for both .claude/shadow/ and .cursor/shadow/." </example>
safe-exec
Runs a shell command with automatic output compression. If output exceeds 5 KB, writes full output to a timestamped log file and returns a concise summary. Prevents context bloat from large build/test/lint outputs. <example> Context: Running npm run build which produces large output. user: "Run npm run build and check for errors" assistant: "Running safe-exec for npm run build — output exceeding 5 KB will be stored and summarized." </example>
session-learnings
Captures key observations, anti-patterns, and rule improvement suggestions at the end of a session. Writes a sharded YAML to ~/worklogs/learnings/YYYY/MM/ and updates the learnings index. <example> Context: End of a debugging session on manifest generation user: "/session-learnings" assistant: "Captured 3 observations, 1 anti-pattern, 2 rule suggestions → ~/worklogs/learnings/2026/04/2026-04-14-manifest-debug.yaml" </example>
diary
Appends a decision/outcome entry to the current session's diary so the self-evolving loop learns from successes and failures, not just ledgers. Enriches the baseline entry written automatically at session end. <example> Context: Just chose count-based deltas over mtime math for portability. user: "/diary chose count-based signal delta — BSD find has no -newermt" assistant: "[DIARY] appended decision (outcome: worked) to today's diary entry." </example>
evolve-apply
Applies a human-approved /evolve proposal. Runs evolve-apply.sh to patch core/, convert, and audit; reverts and records a failure ledger if the audit fails. NEVER run without explicit human approval of the proposal. <example> Context: Human reviewed proposal abc12345 and approved it. user: "/evolve-apply abc12345" assistant: "[EVOLVE-APPLY] abc12345 applied — convert + audit green." </example>
extract-signatures
Extracts method stubs (names, parameters, return types — no bodies) from TypeScript or PHP source files using a live ripgrep-based script. Returns ~500-2k tokens per file, always fresh. <example> Context: User needs to understand AuthService API before editing it. user: "Extract signatures from libs/auth/src/lib/auth.service.ts" assistant: "Running extract-signatures on auth.service.ts — returns class/method stubs without bodies." </example>
pack-context
Bundles multiple source files matching a regex pattern into a compact XML context block for tasks touching 3+ interrelated files. Max 15 files, ~10-30k tokens. Needs Sonnet to judge which files to include. <example> Context: Refactoring BillingFacade across multiple files. user: "Bundle all billing files for the cancelSubscription refactor" assistant: "Running pack-context with pattern 'billing|subscription' — returns XML bundle of up to 15 matching files." </example>
plugin-audit
Runs the repeatable ContextForge audit gate after change packages, before PRs or commits, after convert, and after runtime artifact cleanup. Uses plugin-audit.sh as the only supported entrypoint.
shadow-lookup
Zero-LLM-cost symbol-to-file-path lookup using the shadow manifest. Returns TSV: symbol → kind → file → repo. ~50 tokens per lookup vs ~30k for full manifest read. <example> Context: Need to find where BillingFacade is defined. user: "Where is BillingFacade defined?" assistant: "Running shadow-lookup for BillingFacade — returns file path and kind from manifest." </example>
task-init
Initializes a new Jira task by creating ~/worklogs/tickets/<KEY>.yaml, updating INDEX.yaml, and detecting branch/task-type metadata. Use at the start of every new Jira ticket session. <example> Context: Starting work on a new feature ticket user: "task-init PROJ-3456" assistant: "[TASK INIT] created ~/worklogs/tickets/PROJ-3456.yaml and updated INDEX.yaml" </example>
update-worklog
Appends a structured entry to the active ticket's worklog YAML when a significant decision, phase completion, or cross-boundary impact is identified. Auto-extracts entity names from scope. <example> Context: Just chose ts-morph over string templates for codegen user: "log: chose ts-morph for code generation" assistant: "Appended entry to ~/worklogs/tickets/PROJ-3456.yaml" </example>
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.