🚀 Grab a bottle

Install for Codex in the current project:

npx --yes @dotsisdev/condiments@latest --host codex-cli --target .

Restart Codex, then choose a strength:

/cond full

That is it. Condiments checks the model and task, adds only proven controls, and keeps mystery combinations on baseline.

| Where | Command | | --- | --- | | Claude Code | npx --yes @dotsisdev/condiments@latest --host claude-code --target . | | Cursor | npx --yes @dotsisdev/condiments@latest --host cursor --target . | | OpenClaw | npx --yes @dotsisdev/condiments@latest --host openclaw --target . | | Codex everywhere (PowerShell) | npx --yes @dotsisdev/condiments@latest --host codex-cli --target "$env:USERPROFILE" | | Codex everywhere (Bash) | npx --yes @dotsisdev/condiments@latest --host codex-cli --target "$HOME" |

Add --force to update or repair an installation.

🍟 Meet the bottles

| Bottle | Nickname | Job | Full squeeze | | :---: | --- | --- | --- | | 🥚 | Mayo | Trims replies | Smallest complete answer | | 🟡 | Mustard | Picks context | Loads exact files, symbols, and evidence | | 🍅 | Ketchup | Remembers | Retrieves exact local memory before summarizing | | 🌿 | Ranch | Handles tools | Batches, deduplicates, and bounds noisy output | | 🌶️ | Hot Sauce | Controls thinking | Starts cheap and turns up effort only when needed |

Use none to put every bottle away, some for a light squeeze, or full for the strongest validated recipe.

📊 The scoreboard

Official measured claim: Condiments produced 5.1–32.2% fewer total tokens in its validated Luna workloads with no quality loss. This is a measured range, not a promise for every task.

| What happened | Decision | | --- | --- | | Zero-token memory: 32.2% less, 8/8 exact recalls | ✅ Ship for validated Luna recall | | Progressive skill disclosure: 5.1% less, 8/8 checks | ✅ Use on active routes | | Prompt-only tool state: model reran tools in 6/8 pairs | 🛑 Require native blocking | | Unknown model or workload | 🧊 Stay on baseline |

The complete numbers live in the 48-run paired evaluation.

• Luna debugging and noisy-command workloads saved 20.4–23.2% logical tokens, but cache-cost savings remain unproven.
• General Luna tasks, Sol exact edits, and Astra exact edits showed regressions, so the router keeps them on baseline.
• Context and tool-schema experiments showed large byte reductions, but Condiments does not label those as provider-token savings.
• Reasoning savings are still estimates until authenticated provider A/B runs finish.

Open every evaluation and methodology note.

🧭 How the sauce gets picked

Condiments treats /cond full as a ceiling, not an order to optimize everything. If a combination has not passed paired testing, it receives no extra policy.

known winner  → tiny active recipe
unknown combo → baseline
quality drop  → reject and recover

🧪 From research to recipe

Four ideas became working controls

| Recipe | What Condiments built | Current label | | --- | --- | :---: | | 🧩 Progressive disclosure | Stable core + current host + active bottles only | ✅ Validated | | 🧠 Zero-token memory | Local exact-evidence index with zero model calls | ✅ Validated | | 🧰 Compact tool state | Hash-addressed results, facts, gaps, and retry blocking | 🛡️ Native guard only | | ✂️ Workflow pruning | 8-pair quality, confidence, and break-even gate | ✅ Active gate |

Read the source reviews:

• Chinese net-token savings update
• Chinese token-efficiency research
• Latest algorithmic token-reduction research
• DeepSeek-V4.1-Flash report

🎮 Command menu

| Command | Effect | | --- | --- | | /cond | Show the current sauce | | /cond full | Strongest validated recipe | | /cond some | Lighter quality-first recipe | | /cond none | Plain baseline | | /cond ranch full | Change one bottle | | /cond status | Show levels and host abilities | | /cond v | Show the installed release |

/condiments accepts the same commands. Codex also accepts $cond and $condiments.

🤖 Where it works

| Agent or API | Skill commands | Native extras | | --- | :---: | --- | | Codex | ✅ | Compaction, large-read guard, routing, cache/reasoning telemetry | | Claude Code | ✅ | Tool-result replacement, checkpoints, cache telemetry | | Cursor | ✅ | MCP result replacement, compaction, cache fields | | OpenClaw | ✅ | Tokenjuice, pruning, compaction, cache retention | | OpenAI / Anthropic APIs | Helpers | Reasoning, output caps, cache controls, state reuse | | Qwen / DeepSeek | Helpers | Capability-gated thinking and reasoning controls |

Run /cond status to see what the current installation can control. Missing telemetry is shown as unavailable—never as zero.

🔬 Proof locker

• Net-savings controls: 48 paired live runs
• Savings-controls replication
• Paired live evaluation
• Patch-first coding
• Mayo output
• Tool context
• Lossless logs and learned caps
• Three-repository context evaluation

1. Count input, output, reasoning, cache traffic, retries, helpers, and failures.
2. Compare the same model, effort, task, and starting state.
3. Keep exact files, tests, facts, errors, and citations correct.
4. Require at least eight paired samples before workflow pruning.
5. Publish failures too. A shorter answer is not a saving if the task needs another turn.

• BINGO — Not All Tokens Matter
• IterCOMP
• REFRAIN — Stop When Enough
• Lossless Prompt Compression via Dictionary-Encoding
• TALE — Token-Budget-Aware LLM Reasoning
• LLMLingua-2
• Qwen3 Technical Report
• OpenAI agent-harness efficiency
• Anthropic cost and intelligence optimization
• DeepSeek context cache

🛠️ Builder corner

git clone https://github.com/DotsIsDev/condiments.git
cd condiments
npm install
npm test
node scripts/install-adapter.mjs --host codex-cli --target /path/to/project

| Pantry shelf | What is inside | | --- | --- | | SKILL.md | Portable skill contract | | src/ | Policies, routing, memory, telemetry, compression | | scripts/ | Installer, hooks, helpers, evaluations | | references/ | Runtime and platform notes | | research/ | Full paper reviews | | evals/ | Workloads, fixtures, and results | | test/ | Deterministic test suite |

Contributions are welcome. Bring reproducible workloads, sanitized traces, provider adapters, or a clever new sauce. Start with CONTRIBUTING.md.