Godmode

Activate When

• /godmode without subcommand
• /godmode:<skill> — read skills/<skill>/SKILL.md
• Natural language request → match to skill

Step 0a: Check Terse Mode

IF $GODMODE_TERSE is set OR /godmode:terse on was invoked this session: compress all emitted output per skills/terse/SKILL.md. TSV rows, code blocks, error messages, commit messages, and the final user-facing summary stay verbose. Terse is an emit contract only — loops, decisions, commits, and TSV writes are identical in both modes. See skills/terse/SKILL.md for the full contract and the before/after examples.

Also check .godmode/session-state.json for auto-activation. If the file exists and consecutive_rounds_without_human >= 2 AND terse_user_opted_out != true: auto-enable terse for this round and emit the activation message defined in skills/terse/SKILL.md § Auto-Activation. The threshold was lowered from 5 to 2 in Phase E so that normal /godmode commands get terse compression by round 2 instead of waiting 5 rounds — first round stays verbose for human context, all subsequent rounds are terse. This check runs at the START of every round, before the loop body. The counter is incremented at the END of each round (immediately after DECIDE and LOG) and reset to 0 at the START of any round that was triggered by a new user prompt. A user-issued /godmode:terse off sets terse_user_opted_out to true for the remainder of the session and disables auto-activation.

Step 0: Check for Resumable Session

Read .godmode/session-state.json if it exists. IF stop_reason is null: resume the interrupted skill at the saved round. IF stop_reason is set OR file missing: proceed with fresh detection. Print resume status at start of every session.

Step 0b: Load Lessons

Read .godmode/lessons.md if it exists. Surface relevant lessons for the detected skill. After session: append 1-3 new lessons.

IF lessons.md exceeds 100 lines: compress before loading. Keep entries from the last 30 days AND entries referenced (cited) more than twice. Mark stale entries [OBSOLETE] rather than deleting — the file is append-only history. After compression, re-read and continue. Target post-compression length: ≤60 lines. This keeps the lesson-loading context lean on long-running sessions where the file would otherwise grow unbounded.

Step 0c: Enable Token Logging (default on)

Unless GODMODE_TOKENS=0 is set in the environment, token logging per skills/tokens/SKILL.md is ON by default for this session. At the end of every round (after DECIDE and LOG), append one row to .godmode/token-log.tsv with the 10-column schema defined in the tokens skill. The logger uses the chars / 4 heuristic documented there. Opt out with GODMODE_TOKENS=0 only for privacy-sensitive workflows. This is part of Phase E Default Activations (SKILL.md §14) — no explicit /godmode:tokens invocation needed.

Step 1: Detect Stack (once, cache)

# Detect project stack from root files
ls package.json pyproject.toml Cargo.toml go.mod \
  Gemfile pom.xml 2>/dev/null

# Detect lockfile for package manager
ls yarn.lock pnpm-lock.yaml uv.lock \
  package-lock.json 2>/dev/null

# Verify commands work
$test_cmd --version 2>/dev/null
$lint_cmd --version 2>/dev/null

STACK DETECTION:
| Files                         | Stack      | test_cmd      | lint_cmd      | build_cmd      |
|-------------------------------|------------|---------------|---------------|----------------|
| package.json + next.config.*  | Next.js    | npm test      | eslint --fix  | npm run build  |
| package.json + tsconfig.json  | TypeScript | npx vitest    | eslint --fix  | tsc --noEmit   |
| pyproject.toml                | Python     | pytest        | ruff check .  | —              |
| Cargo.toml                    | Rust       | cargo test    | cargo clippy  | cargo build    |
| go.mod                        | Go         | go test ./... | golangci-lint | go build ./... |

IF no match: ask user for test/lint/build commands
IF lockfile found: use matching package manager

Step 2: Match Skill (Tier 1 routing)

Read ONLY Tier 1 of each skill file to match the user's request. Tier 1 = frontmatter + the ## Activate When block. The Tier 1 boundary is the first ## header that follows ## Activate When. Everything after that boundary is Tier 2 (loaded only once a skill is matched) or Tier 3 (loaded only on edge cases, marked by a literal tier-3 HTML comment).

# Tier 1 extractor — POSIX awk. Reads frontmatter + the Activate When block,
# stopping at the next `##` header. Skills without `## Activate When`
# (currently only `principles`) emit only their frontmatter.
for f in skills/*/SKILL.md; do
  awk '
    NR == 1 && $0 == "---" { in_fm = 1; print; next }
    in_fm && $0 == "---" { in_fm = 0; print; next }
    in_fm { print; next }
    /^## Activate When/ { in_aw = 1; print; next }
    in_aw && /^## / { exit }
    in_aw { print }
  ' "$f"
done

Measured Tier 1 cost (via tests/token-bench.sh): about 4,000 tokens across 135 skills, vs. about 54,000 tokens for full reads — roughly 92% reduction. Stacks with skills/terse/ (output-side compression), skills/stdio/ (input-side canonical commands), and rtk (if installed) for compound context savings.

Match process:

1. Scan all Tier 1 blocks for keyword hits in frontmatter description: plus ## Activate When bullets.
2. Pick the skill with the most trigger hits. Tie-break: shorter ## Activate When list wins (more specific trigger).
3. Once matched, read the FULL matched skill file up to the tier-3 HTML comment marker if present (Tier 2 auto-included).
4. If the skill has a Tier 3 section AND the loop hits a failure class from SKILL.md §8 (error recovery, quality target verification, platform fallback), read past the marker.

Canonical trigger shortcuts (fastest path — if a request matches one of these exactly, skip the Tier 1 scan):

| Trigger                                                      | Skill       |
|--------------------------------------------------------------|-------------|
| "integration test", "integration testing"                    | integration |
| "end to end", "end-to-end", "e2e test"                       | e2e         |
| "load test", "stress test"                                   | loadtest    |
| "docker image", "dockerfile", "container image"              | docker      |
| "rate limiting", "rate limit", "throttling"                  | ratelimit   |
| "react component", "react hook", "jsx"                       | react       |
| "vue page", "vue component", "composition api"               | vue         |
| "nextjs", "next.js", "app router"                            | nextjs      |
| "django view", "django model", "django orm"                  | django      |
| "fastapi route", "fastapi endpoint"                          | fastapi     |
| "rails controller", "rails model", "active record"           | rails       |
| "openapi spec", "openapi schema", "rest api spec"            | api         |
| "design the architecture", "system architecture"             | architect   |
| "event sourcing", "event-driven architecture"                | event       |
| "make faster", "optimize", "slow", "response time", "p99", "latency" | optimize |
| "debug", "why is this", "leaking", "segfault", "trace this"  | debug       |
| "is red", "failing", "errored", "fix", "broken", "error"     | fix         |
| "secure", "vulnerabilities"                                  | secure      |
| "review", "check my code", "look over this pr", "pull request" | review    |
| "research", "prior art"                                      | research    |
| "plan", "break down"                                         | plan        |
| "ship", "deploy"                                             | ship        |
| "finish", "done", "clean up", "wrap up"                      | finish      |
| "terse", "compress output"                                   | terse       |
| "tokens", "token budget"                                     | tokens      |
| "stdio", "command patterns"                                  | stdio       |
| "team", "bundle"                                             | team        |
| "tutorial", "onboarding", "get started", "first run"         | tutorial    |
| "bench", "benchmark"                                         | bench       |
| "test", "coverage"                                           | test        |
| "build", "implement", "create"                               | build       |

IF no match: fall through to phase detection (Step 3).

Skills whose Tier 1 block fails to parse (missing ## Activate When, malformed frontmatter) are routing-invisible. The only skill without ## Activate When is skills/principles/SKILL.md, which is imported directly as a prelude via @./skills/principles/SKILL.md and not routable — this is intentional.

Step 3: Detect Phase (State Machine)

PHASE DETECTION:
  non-trivial feature, no research  → RESEARCH  (see auto-dispatch below)
  no spec, no plan                  → THINK
  spec exists, no plan              → PLAN
  plan exists, tasks incomplete     → BUILD
  code exists, tests failing        → FIX
  tests passing, unreviewed         → REVIEW
  reviewed, metrics unoptimized     → OPTIMIZE
  all green                         → SHIP

THRESHOLDS:
  Stuck recovery: > 5 consecutive discards
    → try opposite approach
    → if still stuck: escalate to previous phase
    → if still stuck: log reason, move to next task

Research auto-dispatch rule (Phase E Default Activations — SKILL.md §14). Before routing to THINK, check whether the task is "non-trivial":

• Task mentions an external library, framework, or standard by name
• Task scope is >5 files (estimated from the user's phrasing)
• No .godmode/research.md exists for the feature

If ANY of these is true AND the user did NOT pass --no-research: dispatch skills/research/SKILL.md first. Research writes .godmode/research.md; THINK then reads it at its own Step 2 (Scan Codebase) instead of rescanning. Research auto-dispatch turns the "prior art gap" into a no-op for every normal command. Skip for trivial tasks (one-line fixes, typos, renames).

Step 3b: Failure-Aware Routing

Before routing to any skill, check .godmode/<skill>-failures.tsv if it exists. If the target skill has >10 consecutive failures: suggest an alternative skill or approach. Surface failure patterns in the session summary: "optimize had 5 noise failures — metric may be non-deterministic."

Step 3c: Resolve Model Per Role (dispatch time)

Resolve each dispatched role's model at dispatch time. Zero-config = every role inherits the session model; per-role model tiering is opt-in only. Resolution order: GODMODE_MODEL_<ROLE> env -> godmode.models.json (per-role merge: current project root file wins per-key over ~/.config/godmode/models.json) -> session model. Missing config = VALID zero-config default; every role inherits the session model.

A model value is valid iff it fully matches ^[A-Za-z0-9._-]+/[A-Za-z0-9._-]+$ OR is the literal string session. An INVALID value at ANY layer emits one stderr warning — [models] warning: ignoring invalid model value for role <role> — and is treated as absent (fall through to the next layer). Values are pasted into agent dispatch text, never shell-eval'd; still validate.

Inline resolver — portable subset, NO adapters/ dependency (skills-only installs never ship adapters/). Mirrors adapters/pi/models.sh, the full reference implementation (doctor/selftest live there):

gm_route() {  # usage: gm_route <role> -> "<model>\t<source>" on stdout
  local role name val f
  role="${1:?usage: gm_route <role>}"
  name="GODMODE_MODEL_$(printf '%s' "$role" | tr '[:lower:]._-' '[:upper:]___')"
  val="$(printenv "$name" 2>/dev/null || true)"  # set-but-empty = fall through
  case "$val" in *$'\n'*|*$'\r'*) val="" ;; esac  # newline would spoof grep anchors
  [ "${#val}" -le 128 ] || val=""
  if [ "$val" = session ] || \
     printf '%s' "$val" | grep -Eq '^[A-Za-z0-9._-]+/[A-Za-z0-9._-]+$'; then
    printf '%s\tenv\n' "$val"; return 0
  elif [ -n "$val" ]; then
    printf '[models] warning: ignoring invalid model value for role %s\n' "$role" >&2
  fi
  for f in "$(git rev-parse --show-toplevel 2>/dev/null || pwd)/godmode.models.json" \
           "$HOME/.config/godmode/models.json"; do
    [ -f "$f" ] || continue
    command -v python3 >/dev/null 2>&1 || continue  # no python3: skip file layer
    if val="$(python3 - "$f" "$role" <<'EOF'
import json, re, sys
try:
    roles = json.load(open(sys.argv[1], encoding="utf-8-sig")).get("roles", {})
except Exception:
    sys.exit(3)  # unreadable/invalid file: treat as absent
val = roles.get(sys.argv[2])
if not isinstance(val, str) or val == "":
    sys.exit(3)  # absent, non-string, or empty: silently try the next layer
if len(val) <= 128 and (val == "session" or
        re.fullmatch(r"^[A-Za-z0-9._-]+/[A-Za-z0-9._-]+$", val)):
    print(val)
else:
    print("[models] warning: ignoring invalid model value for role %s"
          % sys.argv[2], file=sys.stderr)
    sys.exit(4)  # present but invalid: warn once, fall through
EOF
)"; then
      printf '%s\tfile\n' "$val"; return 0
    fi  # helper exit 3/4 consumed by the if: fall through to next layer
  done
  printf 'session\tsession\n'
}

Dispatch table (godmode role -> pi-subagents child kind; the model param is resolved at dispatch time):

These are the canonical dispatch roles (AGENTS.md); verb aliases (plan, build, review, optimize, explore, test, docs — the doctor's default sweep) resolve identically because role keys are open-ended. Stamp the resolved model on the dispatch via the optional DispatchContext model field (model_profile for named profiles); unknown dispatch fields are logged, never a hard fail.

Zero-config rule: when gm_route <role> resolves to session, OMIT the model param entirely — the child inherits the session model. Example: GODMODE_MODEL_REVIEW=vendor/strong stamps model: vendor/strong on review dispatches only; every other role stays on the session model.

Roles are open-ended, not a fixed enum: unknown or custom roles resolve through gm_route identically (unknown -> session). STRONG/FAST-style names are user-defined aliases living in the user's own config — nothing built-in.

Env normalization: role -> env name by uppercasing and turning every hyphen, dot, and underscore into an underscore — exactly tr '[:lower:]._-' '[:upper:]___' — then prefixing GODMODE_MODEL_ (code-review -> GODMODE_MODEL_CODE_REVIEW). Case-folded roles (Plan and plan) collapse to one env var — document, do not special-case. A value that is set but empty is NOT an override; fall through.

Step 4: Execute

Read skills/{skill}/SKILL.md. Follow it literally. Pass: stack, test_cmd, lint_cmd, build_cmd.

Output Format

Print: Godmode: stack={stack}, skill={skill}, phase={phase}. Dispatching. After: Godmode: {skill} complete. Next: {next}.

Quality Targets

• Skill routing: <2s to match and dispatch
• Stack detection: <5s for full project analysis
• Target: >95% correct skill match on natural language input

Hard Rules

Never ask to continue. Loop autonomously until done.

1. Detect stack FIRST — cache result. Never guess.
2. One skill at a time — read SKILL.md, follow it.
3. Commit BEFORE verify — revert on failure.
4. Log every invocation to .godmode/session-log.tsv.
5. Stuck recovery: > 5 discards triggers escalation.

Rules

1. Iterative skills use WHILE loops with counter.
2. Iterations: N = run exactly N times then stop.
3. Commit before verify. Revert: git reset --hard HEAD~1.
4. Log: .godmode/<skill>-results.tsv (append only).
5. Session: .godmode/session-log.tsv with stop_reason: target_reached | budget_exhausted | diminishing_returns | stuck | user_interrupt
6. KEEP/DISCARD: atomic — commit before verify, revert if verify fails.
7. Multi-agent: <= 5 agents/round, worktree isolation.
8. Chain: think → plan → [predict] → build → test → fix → review → optimize → secure → ship.
9. Model routing: resolve each role's model at dispatch time (Step 3c); stamp a model param only when it resolves to something other than session.

Keep/Discard Discipline

KEEP if: metric improved AND guard passed
  (build_cmd && lint_cmd && test_cmd)
DISCARD if: metric worsened OR guard failed
On discard: git reset --hard HEAD~1. Log reason.

Meta-Loop (Outer Loop)

After each skill completes, analyze results and decide the next skill:

WHILE project_goal_not_met:
  result = run_current_skill()
  IF result.status == "DONE" AND result.findings > 0:
    next_skill = route_findings(result)  # e.g., optimize found security issue → secure
  ELIF result.status == "DONE" AND result.findings == 0:
    next_skill = advance_phase()          # move to next phase in chain
  ELIF result.status == "STUCK":
    next_skill = escalate_or_skip()       # try alternative skill
  LOG to session-log.tsv

The meta-loop enables: optimize → review finds issue → fix → re-optimize → secure → ship. Without it, each skill runs in isolation. With it, skills chain automatically.

Persistence

If session ends mid-loop, .godmode/session-state.json preserves state. The stop hook notifies the user to run /godmode to resume. For fully autonomous overnight runs: use Ralph Loop or /loop with godmode.

Stop Conditions

STOP when FIRST of:
  - target_reached: spec/goal fully achieved
  - budget_exhausted: max iterations hit
  - diminishing_returns: last 3 iters each < 1%
  - stuck: > 5 consecutive discards

TSV Logging

timestamp	skill	iterations	kept	discarded	stop_reason	outcome

Error Recovery

# Detect project stack from root files
ls package.json pyproject.toml Cargo.toml go.mod 2>/dev/null
git log --oneline -5

Model Routing Capability Matrix