centsible

Manage repository-specific micro-facts, inspect token optimization status, and record reusable discoveries to minimize context bloat and token consumption.

Centsible Skill 🪙

This skill allows the agent and user to manage experience-driven micro-facts and token optimization.

When to Use

• User asks about token optimization status or active facts (centsible status, centsible facts list).
• You resolved a non-trivial repo quirk (e.g. non-standard test runner, mandatory environment flag, path alias) that future turns will need.
• Testing pattern matching against prompts.

Commands

1. Check Status & Token ROI Stats

bash

centsible status
centsible stats
centsible stats --transcript <path_to_transcript.jsonl> --json

Loop Guard Thresholds & Reflection

Pick a mode preset, or fine-tune thresholds and reflection modes.

• Modes (errors / identical calls): ultra 2/3, normal 3/4 (default), weak 4/5.
• Reflection modes:
  • inline (default): in-turn reflection, agent saves facts before final response; zero extra turns, answers never collapsed in UI.
  • hook: Stop hook reflection; blocks session termination to force a continuation turn.
  • off: disables reflection prompts.

bash

centsible config mode ultra
centsible config reflect inline # In-turn reflection without extra turns (default)
centsible config reflect hook   # Force extra continuation turn on wrap-up
centsible config reflect off    # Disable reflection nudges
centsible config list
centsible config set max_consecutive_errors 4
centsible config unset max_consecutive_errors

Keys: mode, reflect, max_consecutive_errors, max_identical_calls. Env vars CENTSIBLE_LOOP_MODE, CENTSIBLE_REFLECT, CENTSIBLE_MAX_* override the file.

2. List Micro-Facts

bash

centsible facts list

3. Add a Micro-Fact

Always keep the fact to a single, direct sentence without conversational padding:

bash

centsible facts add --pattern "<regex_pattern>" --fact "<one_line_fact>"

Examples:

bash

centsible facts add --pattern "pytest.{0,40}(fail|error|slow)" --fact "Run unit tests with 'pytest tests/ -q --tb=short'."
centsible facts add --pattern "bruno.{0,30}(run|uat|stg)" --fact "Append --condense to Bruno commands to save 80% JSON tokens."
centsible facts add --pattern "auth/login\.py" --fact "Never mutate token keys directly."

The pattern decides when the fact is injected, so it must describe the situation, not a keyword. Saving is rejected (exit 2) for a lone short word (pytest, bruno, pip), for generic words only (run.*test), and for facts over 240 characters. Anchor on, in order of strength:

1. an identifier: file, script, table, API path or flag (query_argus, fast-e2e, a/b.py);
2. a tool plus a symptom (pytest.{0,40}(fail|error));
3. an action plus object plus context ((create|book).{0,12}order.{0,40}(uat|stg)).

Before saving, check a couple of prompts that must not match with centsible facts test "<prompt>". --force overrides the check; use it only when the pattern is deliberately broad.

4. Test Fact Matching

Verify what facts trigger for a specific user prompt:

bash

centsible facts test "How do I run the bruno tests?"

5. Remove Stale Fact

bash

centsible facts remove <id_or_pattern>

6. Harvest Golden Facts (Agentic Multi-Turn Discovery)

When working in a new repository or after a fresh install, discover 5–8 golden architectural facts:

bash

centsible harvest
centsible harvest --prompt

• In Claude Code: type /centsible-harvest
• In Antigravity: ask your agent "Run centsible harvest to discover golden facts for this repo"