pakka
Claude Code harness — fewer tokens, fewer bugs, audit-ready. Apache-2.0.
Install
/plugin marketplace add amargautam/pakka-marketplace
/plugin install pakka@pakka-marketplace
Zero-config. Uses your existing Claude Code auth. No API key required.
Upgrade
/plugin marketplace update
/plugin install pakka@pakka-marketplace
/reload-plugins
/plugin marketplace update must run first — it pulls the latest catalog ref. Without it, install resolves to a stale cached version.
What it does
8 commands — context-inferred, discipline-driven
Pakka injects discipline into every session. Commands infer what you need from context — no mode flags, no guessing.
| Command | Infers from | What it does |
|---|
/pakka:plan | "build X", "design", "challenge this", "probe me", "break into tickets" | Design hub. Writes spec to docs/specs/. Routes to spec · challenge · probe · slice based on context. Never auto-chains to build. |
/pakka:build | "implement", "write tests", "broken", "how does X work", "hard to test" | Implementation hub. Checks for spec approval first. Routes to TDD · debug · map · audit based on context. Blocks completion claims without exit-code evidence. |
/pakka:review | "done?", "ship?", "they said...", "merge?" | Quality hub. Verifies first (exit codes), then runs reviewer + security + architect + performance agents in parallel. Handles incoming feedback and branch landing. |
/pakka:triage | "triage", "look at issue #N", "what needs attention" | Issue queue. Routes bugs and features through classification state machine. Produces agent-ready briefs. |
/pakka:setup | one-time setup | Detects stack, writes permissions overlay. setup guard installs git guard hook. |
/pakka:compress | — | Compression control. [lite|strict|ultra|super-ultra|status]. Default: super-ultra. Hook-handled — instant, no LLM round-trip. |
/pakka:recall [query] | — | search audit trail across sessions via FTS5 index |
/pakka:help | — | Show pakka status — active level, gate config, hooks. |
Ambient disciplines (always active, no invocation needed)
Verification: before any "done", "working", "passing" claim — pakka requires actual exit-code evidence. Injected at session start.
Skill-check: before each response — pakka checks whether the message calls for /pakka:plan, /pakka:build, or /pakka:review. Catches cases that explicit invocation misses.
4-vector compression: output tokens · input context · tool results · subagent returns — all compressed independently.
Review gate: reviewer + security + architect + performance subagents run in parallel on every Claude-authored commit. Confidence threshold ≥ 80. Blocks on severity=error findings.
Deny-by-default permissions: secrets, destructive git, shell-fetched-then-executed commands blocked at the permission layer. Guard covers secret reads and writes — the hook matches Read|Write|Edit|MultiEdit|Bash (was Read|Bash before v0.11.0). Repeated overrides teach a per-repo allowlist (.pakka/guard-allowlist.json) with override-count decay; secret categories are never allowlistable.
Audit trail: every tool call appended to ~/.pakka/audit/<session>.jsonl. No dial-home.
recall: /pakka:recall searches your audit trail. cross-session memory backed by local FTS5 index (SQLite). no remote storage.
skill-check: UserPromptSubmit hook keyword-scans every message. if a build/plan/review signal matches, targeted alert fires before the model responds. no more relying on model memory.
Status line: pakka [super-ultra] · ~$72.47 saved · 21 bugs caught — compression level, token savings, and bugs caught, always visible. Savings priced per model — Fable 5, Mythos 5, and Opus 4.x covered; dated model IDs resolve by prefix.
Kill-switch: PAKKA_DISABLED=1 turns off every pakka hook for the session. Used by make bench to isolate the raw arm; works anywhere.
Benchmarks: make bench A/Bs pakka vs raw Claude Code via claude -p on your existing OAuth session. No API key.
Results (v0.8.0)
Three absolute numbers, each verifiable from artifacts in this repo.
-
Bug catch rate: 9/10. Combined reviewer + security + architect agents caught 9 of 10 seeded bugs on the Pass 5b in-session corpus.
-
Bytes saved: 307,451 cumulative since 2026-04-24. Estimated tokens: 88,073 (bytes ÷ 3.5). Total estimated savings: ~$72.47. Source: RECEIPTS.md, regenerated via make self-report.
-
Gate enforcement: every Claude-authored commit path. Architectural claim — gate runs and blocks on findings. Verify: git log --format='%H' | while read sha; do git show -s --format='%(trailers:key=Reviewed-by-pakka,valueonly=true)' "$sha" | grep -q . && echo "$sha"; done | wc -l.
Attribution
Every commit pakka reviews carries:
