skills/forge-compression-help/SKILL.md
Quick reference card for all Forge compression features -- output modes, inter-agent compression, input compression, and configuration. Use when you need a reminder of compression options or want to understand how compression works.
npx skillsauth add quantumbitcz/dev-pipeline forge-compression-helpInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
None. This skill is a reference guide.
Display the reference card below. If the user asks about a specific section, navigate directly to it.
None. This skill displays static content.
Controls how Forge talks to YOU. Does not affect inter-agent communication.
| Mode | Command | Effect | Estimated Savings |
|------|---------|--------|-------------------|
| Off | /forge-caveman off | Standard verbose output | 0% |
| Lite | /forge-caveman lite | Drop filler/hedging, keep grammar | ~20% |
| Full | /forge-caveman | Fragments OK, short synonyms, drop articles | ~45% |
| Ultra | /forge-caveman ultra | Max abbreviation, arrows, no conjunctions | ~65% |
Auto-clarity exceptions: Security warnings, irreversible confirmations, AskUserQuestion, escalations, and PR descriptions always use full verbosity regardless of mode.
Persistence: Stored in .forge/caveman-mode. Survives /forge-reset.
Controls how agents talk to EACH OTHER inside the pipeline. Configured in forge-config.md.
| Level | Name | Token Range | Savings | Default Stages |
|-------|------|-------------|---------|----------------|
| 0 | verbose | 800-2000 | 0% | User reports, escalations |
| 1 | standard | 800-2000 | ~20% | Planning, docs, retrospective |
| 2 | terse | 400-1200 | ~45% | Implementation, verification, review |
| 3 | minimal | 100-600 | ~65% | Inner-loop lint/test, mutation, scaffolding |
Configuration:
output_compression:
enabled: true
default_level: terse
auto_clarity: true
per_stage:
PREFLIGHT: standard
EXPLORING: standard
PLANNING: standard
VALIDATING: terse
IMPLEMENTING: terse
VERIFYING: terse
REVIEWING: terse
DOCUMENTING: standard
SHIPPING: verbose
LEARNING: standard
Reduces token cost of agent .md system prompts. Applied offline, not at runtime.
| Command | What It Does |
|---------|-------------|
| /forge-compress | Compress agent prompts (default: agents/ scope) |
| /forge-compress --dry-run | Estimate savings without modifying files |
| /forge-compress --restore | Restore original files from .original.md backups |
| /forge-compress --scope all | Compress agents + modules + shared + config |
| /forge-compress --level 1 | Conservative (~20% reduction) |
| /forge-compress --level 2 | Aggressive (~45% reduction) |
| /forge-compress --level 3 | Ultra (~65% reduction) |
Rules: Preserves frontmatter, code blocks, tables, technical terms, category codes, severity levels, and all numeric thresholds. Compresses prose only.
| Skill | Purpose |
|-------|---------|
| /forge-caveman | Toggle user-facing output compression |
| /forge-compress | Compress input files (agent prompts) |
| /forge-commit | Terse conventional commit messages |
| /forge-review | Review with optional terse output format |
| /forge-help | Find the right skill |
All compression settings in forge-config.md:
# User-facing output (caveman mode)
caveman:
enabled: true # Auto-activate via SessionStart hook
default_mode: ultra # lite | full | ultra
# Inter-agent output compression
output_compression:
enabled: true
default_level: terse # verbose | standard | terse | minimal
auto_clarity: true # Suspend compression for security/escalation
per_stage: { ... } # Per-stage overrides (see section 2)
# Input compression (agent prompts)
# No config -- controlled via /forge-compress flags
output-compression.md in sharedinput-compression.md in sharedbenchmarks/ -- compression measurements (if available)development
[writes] Build, fix, deploy, review, or modify code in this project. Universal entry for the forge pipeline. Auto-bootstraps on first run; brainstorms before planning when given a feature description. Use when you want to take any productive action: implementing features, fixing bugs, reviewing branches, deploying, committing, running migrations.
tools
[writes] Manage forge state and configuration: recovery, abort, config edits, session handoff, automations, playbooks, output compression, knowledge graph maintenance. Use when you need to recover from broken pipeline state, edit settings, or manage long-lived state.
development
[writes] Create, list, show, resume, or search forge session handoffs. Use when context is getting heavy and you want to transfer a forge run or conversation into a fresh Claude Code session, or to resume from a prior handoff artefact. Subcommands - no args (write), list, show, resume, search.
development
[writes] Manage the Neo4j knowledge graph. Subcommands: init, rebuild (writes); status, query <cypher>, debug (read-only). Requires Docker. No default — an explicit subcommand is required. Use when setting up the graph for the first time, rebuilding after major refactors, checking graph health, or running ad-hoc Cypher diagnostics.