@zachjxyz/moxie 0.3.5 → 0.3.8

package/README.md

@@ -66,7 +66,7 @@ Dependencies:
 
 Agents (from .moxie/config.toml):
   [OK] claude — 2.1.91 (Claude Code)
-  [OK] codex — codex-cli 0.118.0
+  [OK] gpt-gw — openai/gpt-5.4 (AI Gateway, key stored)
 
 Project:
   [OK] .moxie/ directory found
@@ -151,7 +151,7 @@ Each phase uses round-robin rotation with randomized order. Agents independently
 
 Quorum after claude's turn:
   claude       pending  .moxie/phases/rfc/claude-041126-142611.md
-  codex        pending  .moxie/phases/rfc/codex-041126-141850.md
+  gpt-gw       pending  -
   ── 0/2 agents reached quorum
 ```
 
@@ -161,6 +161,7 @@ Quorum after claude's turn:
 - After each turn, the agent marks whether it agrees with the current state (`reached: true`)
 - If any agent finds issues, it resets all agents to `reached: false`
 - A phase completes when all healthy agents simultaneously agree
+- Agents are told how many turns remain — as the countdown decreases, they prioritize substantive issues over minor nitpicks
 
 ```
 ╔════════════════════════════════════════════════════════════╗
@@ -172,6 +173,8 @@ Quorum after claude's turn:
 
 ### Safeguards
 
+**Turn pressure** — Every agent prompt includes the current turn count and how many remain. As rounds run low, agents naturally triage harder — accepting minor incompletes rather than resetting quorum over stylistic concerns.
+
 **Dead agent detection** — If an agent fails 2 consecutive turns (timeout, empty output, quota exhaustion), it's automatically removed from rotation. Quorum adjusts to the remaining healthy agents.
 
 ```
@@ -207,6 +210,39 @@ moxie agents                List configured agents
 moxie doctor                Check dependencies, agents, and project health
 ```
 
+## Cost tracking
+
+`moxie cost` shows token usage per phase and agent:
+
+```
+Token usage by phase:
+
+  RFC:
+    claude-gw        950,980 tokens
+    gpt-gw           505,430 tokens
+    qwen-gw          271,317 tokens
+    TOTAL          1,727,727 tokens
+
+Grand total:
+  1,727,727 tokens
+```
+
+For gateway agents, moxie also pulls actual USD costs from the [Vercel AI Gateway Custom Reporting API](https://vercel.com/docs/ai-gateway/capabilities/custom-reporting) (requires Pro or Enterprise plan, currently in beta):
+
+```
+Gateway cost (via Vercel AI Gateway):
+  claude-gw            $    5.38  (1,121,678 input + 26,964 output tokens)
+  gpt-gw               $    0.17  (32,021 input + 457 output tokens)
+  qwen-gw              $    1.27  (349,598 input + 12,723 output tokens)
+  TOTAL                $    6.82
+
+  Dashboard: https://vercel.com/~/ai
+```
+
+Each `moxie init` generates a unique run ID. All gateway requests are tagged with the run ID plus `project:moxie`, `phase:<name>`, and `agent:<name>`, so costs are scoped to the exact session — no date range issues for overnight or multi-day runs.
+
+If the reporting API is unavailable, `moxie cost` shows token counts only — you can always check your spend in the [AI Gateway dashboard](https://vercel.com/docs/ai-gateway) directly.
+
 ## Config
 
 After `moxie init`, your `.moxie/config.toml` looks like:
@@ -217,6 +253,7 @@ path = "spec.md"
 
 [gateway]
 endpoint = "https://ai-gateway.vercel.sh"
+run_id = "run-20260411-204514-93675"
 
 [agents.claude]
 command = "claude --dangerously-skip-permissions --effort max --output-format json -p"
@@ -235,38 +272,6 @@ turn_timeout = 900
 
 Edit this file to change agents, adjust timeouts, or add custom agent commands.
 
-## Cost tracking
-
-`moxie cost` shows token usage per phase and agent:
-
-```
-Token usage by phase:
-
-  RFC:
-    claude          12,655,300 tokens (+1 unknown)
-    codex            8,535,058 tokens
-    qwen             2,484,379 tokens (+2 unknown)
-    TOTAL           23,674,737 tokens
-
-Grand total:
-  23,674,737 tokens
-```
-
-For gateway agents, moxie also pulls actual USD costs from the [Vercel AI Gateway Custom Reporting API](https://vercel.com/docs/ai-gateway/capabilities/custom-reporting) (requires Pro or Enterprise plan, currently in beta):
-
-```
-Gateway cost (via Vercel AI Gateway):
-  claude-gw              $4.82  (1,245,890 input + 389,201 output tokens)
-  gpt-gw                 $3.17  (987,654 input + 234,567 output tokens)
-  TOTAL                  $7.99
-
-  Dashboard: https://vercel.com/~/ai
-```
-
-Gateway requests are automatically tagged with `project:moxie`, `phase:<name>`, and `agent:<name>` so you can filter and drill down in the [AI Gateway dashboard](https://vercel.com/docs/ai-gateway).
-
-If you're on the free tier or the reporting API is unavailable, `moxie cost` shows token counts only — you can always check your spend in the AI Gateway dashboard directly.
-
 ## Platform support
 
 | Platform | Status |

package/bin/moxie

@@ -16,7 +16,7 @@
 
 set -euo pipefail
 
-MOXIE_VERSION="0.3.5"
+MOXIE_VERSION="0.3.8"
 # Resolve symlinks (npm installs bin as a symlink)
 _self="$0"
 while [ -L "$_self" ]; do

package/lib/agents.sh

@@ -271,9 +271,10 @@ dispatch_agent() {
 
     # Gateway agent: dispatch via Node.js script
     if _is_gateway_agent "$agent"; then
-        local model endpoint api_key
+        local model endpoint api_key run_id
         model=$(_agent_model "$agent")
         endpoint=$(toml_get "$MOXIE_CONFIG" "gateway.endpoint" "https://ai-gateway.vercel.sh")
+        run_id=$(toml_get "$MOXIE_CONFIG" "gateway.run_id" "")
         api_key=$(gateway_get_key "vercel-ai-gateway") || {
             echo "ERROR: Gateway API key not found. Run 'moxie init' to configure." >&2
             return 1
@@ -285,6 +286,7 @@ dispatch_agent() {
             GATEWAY_MODEL="$model" \
             GATEWAY_PHASE="$phase" \
             GATEWAY_AGENT="$agent" \
+            GATEWAY_RUN_ID="$run_id" \
             timeout "$timeout_secs" \
             node "$MOXIE_ROOT/lib/gateway-agent.mjs" "$(cat "$prompt_file")" || {
             local rc=$?

package/lib/gateway-agent.mjs

@@ -18,6 +18,7 @@ const MAX_TURNS = parseInt(process.env.GATEWAY_MAX_TURNS || '50', 10);
 const IDLE_TIMEOUT_MS = parseInt(process.env.GATEWAY_IDLE_TIMEOUT || '120000', 10);
 const PHASE = process.env.GATEWAY_PHASE || 'unknown';
 const AGENT_NAME = process.env.GATEWAY_AGENT || 'unknown';
+const RUN_ID = process.env.GATEWAY_RUN_ID || '';
 const CWD = process.cwd();
 
 if (!API_KEY) { process.stderr.write('ERROR: GATEWAY_API_KEY not set\n'); process.exit(1); }
@@ -310,7 +311,7 @@ function chatCompletion(messages) {
       stream_options: { include_usage: true },
       providerOptions: {
         gateway: {
-          tags: [`project:moxie`, `phase:${PHASE}`, `agent:${AGENT_NAME}`],
+          tags: [`project:moxie`, `phase:${PHASE}`, `agent:${AGENT_NAME}`, ...(RUN_ID ? [`run:${RUN_ID}`] : [])],
           user: 'moxie',
         },
       },

package/lib/gateway-cost.mjs

@@ -7,18 +7,34 @@ import https from 'https';
 import http from 'http';
 
 const API_KEY = process.env.GATEWAY_API_KEY;
+const RUN_ID = process.env.GATEWAY_RUN_ID || '';
 const endpoint = process.argv[2] || 'https://ai-gateway.vercel.sh';
-const startDate = process.argv[3];
-const endDate = process.argv[4];
+const startDate = process.argv[3] || '';
+const endDate = process.argv[4] || '';
 
 if (!API_KEY) { process.stderr.write('ERROR: GATEWAY_API_KEY not set\n'); process.exit(1); }
-if (!startDate || !endDate) { process.stderr.write('ERROR: start_date and end_date required\n'); process.exit(1); }
 
 const url = new URL(`${endpoint}/v1/report`);
-url.searchParams.set('start_date', startDate);
-url.searchParams.set('end_date', endDate);
 url.searchParams.set('group_by', 'tag');
-url.searchParams.set('tags', 'project:moxie');
+
+if (RUN_ID) {
+  // Filter by run ID — captures the full session regardless of dates
+  url.searchParams.set('tags', `run:${RUN_ID}`);
+  // Still need dates — use a wide window
+  const now = new Date();
+  const thirtyDaysAgo = new Date(now.getTime() - 30 * 24 * 60 * 60 * 1000);
+  const tomorrow = new Date(now.getTime() + 24 * 60 * 60 * 1000);
+  url.searchParams.set('start_date', thirtyDaysAgo.toISOString().slice(0, 10));
+  url.searchParams.set('end_date', tomorrow.toISOString().slice(0, 10));
+} else if (startDate && endDate) {
+  // Fallback: filter by date range + project tag
+  url.searchParams.set('start_date', startDate);
+  url.searchParams.set('end_date', endDate);
+  url.searchParams.set('tags', 'project:moxie');
+} else {
+  process.stderr.write('ERROR: Either GATEWAY_RUN_ID or start_date/end_date required\n');
+  process.exit(1);
+}
 
 const isHttps = url.protocol === 'https:';
 const mod = isHttps ? https : http;

package/lib/phases.sh

@@ -425,9 +425,12 @@ HEADER
 
     # Gateway section (if any gateway models selected)
     if [ ${#SELECTED_GATEWAY_INDICES[@]} -gt 0 ]; then
-        cat >> "$config_file" <<'GATEWAY'
+        local run_id
+        run_id="run-$(date +%Y%m%d-%H%M%S)-$$"
+        cat >> "$config_file" <<GATEWAY
 [gateway]
 endpoint = "https://ai-gateway.vercel.sh"
+run_id = "$run_id"
 
 GATEWAY
     fi
@@ -805,13 +808,18 @@ _phase_is_complete() {
 # Shuffle AGENT_NAMES array in-place using Fisher-Yates.
 # Compatible with Bash 3.2 (no $RANDOM % n bias issues for small arrays).
 _shuffle_agents() {
-    local i len tmp j
+    local i len tmp tmp_cmd j
     len=${#AGENT_NAMES[@]}
     for (( i = len - 1; i > 0; i-- )); do
         j=$(( RANDOM % (i + 1) ))
+        # Swap names
         tmp="${AGENT_NAMES[$i]}"
         AGENT_NAMES[$i]="${AGENT_NAMES[$j]}"
         AGENT_NAMES[$j]="$tmp"
+        # Swap commands to keep in sync
+        eval "tmp_cmd=\"\$AGENT_CMD_$i\""
+        eval "AGENT_CMD_$i=\"\$AGENT_CMD_$j\""
+        eval "AGENT_CMD_$j=\"$tmp_cmd\""
     done
 }
 
@@ -1272,13 +1280,44 @@ _run_phase() {
             fi
         fi
 
-        # Hydrate prompt with agent name and project dir
+        # Hydrate prompt with agent name, project dir, and tiered turn pressure
+        local remaining=$(( max_rounds - round ))
+        local pct_remaining=$(( (remaining * 100) / max_rounds ))
+        local turn_pressure=""
+
+        if [ "$pct_remaining" -gt 60 ]; then
+            # Early: full rigor, gentle awareness
+            turn_pressure="This is turn $round of $max_rounds ($remaining turns remaining). Focus on thoroughness and correctness. Flag all issues you find — there is time to resolve them."
+        elif [ "$pct_remaining" -gt 30 ]; then
+            # Middle: triage mode
+            turn_pressure="This is turn $round of $max_rounds ($remaining turns remaining). Quorum has not been reached. Prioritize issues that would cause real problems in downstream phases. Do not reset quorum over stylistic preferences, speculative concerns, or minor incompletes that agents can work around. If the draft is substantively correct, reach quorum."
+        elif [ "$pct_remaining" -gt 0 ]; then
+            # Late: urgency
+            turn_pressure="WARNING: This is turn $round of $max_rounds — only $remaining turn(s) remaining. If quorum is not reached by turn $max_rounds, the best draft will be force-accepted AS-IS including any unresolved issues. Only reset quorum if you find a factual error that would directly cause a build failure. Accept minor gaps — they can be addressed in later phases. Reaching quorum NOW with a good-enough draft is better than running out of turns with a perfect draft that was never agreed upon."
+        else
+            # Final turn
+            turn_pressure="FINAL TURN: This is turn $round of $max_rounds — there are no more turns after this. If you do not reach quorum now, the best available draft will be force-accepted with all current issues unresolved. Set reached: true unless there is a critical factual error that makes the draft unusable for downstream phases."
+        fi
+
         local prompt_file
         prompt_file=$(mktemp)
         sed -e "s|{{AGENT_NAME}}|$current|g" \
             -e "s|{{PROJECT_DIR}}|$(pwd)|g" \
+            -e "s|{{TURN}}|$round|g" \
+            -e "s|{{MAX_ROUNDS}}|$max_rounds|g" \
+            -e "s|{{REMAINING}}|$remaining|g" \
             "$prompt_template" > "$prompt_file"
 
+        # Replace multiline pressure text (sed can't do multiline easily, use temp file)
+        python3 -c "
+import sys
+with open('$prompt_file', 'r') as f:
+    content = f.read()
+content = content.replace('{{TURN_PRESSURE}}', '''$turn_pressure''')
+with open('$prompt_file', 'w') as f:
+    f.write(content)
+" 2>/dev/null
+
         # Run
         dispatch_logged "$current" "$prompt_file" "$turn_timeout" "$log_dir" "$round" "$csv" "$phase"
         local dispatch_rc=$?

package/lib/tokens.sh

@@ -200,36 +200,13 @@ _show_gateway_cost() {
     local api_key
     api_key=$(gateway_get_key "vercel-ai-gateway" 2>/dev/null) || return
 
-    # Find earliest timestamp across all CSV files to determine date range
-    local start_date
-    start_date=$(python3 -c "
-import csv, glob, datetime
-earliest = None
-for f in glob.glob('$MOXIE_DIR/phases/*/token-usage.csv'):
-    with open(f) as fh:
-        for row in csv.DictReader(fh):
-            ts = row.get('timestamp', '')
-            if len(ts) >= 13:
-                try:
-                    dt = datetime.datetime.strptime(ts[:13], '%m%d%y-%H%M%S')
-                    if earliest is None or dt < earliest:
-                        earliest = dt
-                except ValueError:
-                    pass
-if earliest:
-    print(earliest.strftime('%Y-%m-%d'))
-else:
-    print(datetime.date.today().strftime('%Y-%m-%d'))
-" 2>/dev/null)
-
-    local end_date
-    end_date=$(date +"%Y-%m-%d")
-
-    [ -z "$start_date" ] && return
+    # Get run_id for filtering (if available, use it; otherwise fall back to dates)
+    local run_id
+    run_id=$(toml_get "$MOXIE_CONFIG" "gateway.run_id" "")
 
     # Call the reporting API
     local cost_json
-    cost_json=$(GATEWAY_API_KEY="$api_key" node "$MOXIE_ROOT/lib/gateway-cost.mjs" "$gw_endpoint" "$start_date" "$end_date" 2>/dev/null) || {
+    cost_json=$(GATEWAY_API_KEY="$api_key" GATEWAY_RUN_ID="$run_id" node "$MOXIE_ROOT/lib/gateway-cost.mjs" "$gw_endpoint" 2>/dev/null) || {
         echo ""
         echo "Gateway cost:"
         echo "  (unavailable — check your API key or AI Gateway dashboard)"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zachjxyz/moxie",
-  "version": "0.3.5",
+  "version": "0.3.8",
   "description": "Run multiple AI coding agents through spec-driven phases with quorum convergence. Supports CLI agents (Claude, Codex, Qwen, Aider, Goose, Amp, Cline, Roo) and Vercel AI Gateway models.",
   "bin": {
     "moxie": "bin/moxie"

package/templates/audit/prompt.txt

@@ -76,6 +76,10 @@ Only when EVERY agent has independently verified and agreed:
 
 Do NOT write the FINAL file unless ALL agents show `reached: true` in the ledger.
 
+## Turn Pressure
+
+{{TURN_PRESSURE}}
+
 ## Rules
 
 - ALWAYS verify against actual source code, not other audits

package/templates/build/prompt.txt

@@ -64,6 +64,10 @@ When ALL build phases are `"complete"`:
 
 Do NOT write the FINAL file unless ALL agents show `reached: true` in the ledger.
 
+## Turn Pressure
+
+{{TURN_PRESSURE}}
+
 ## Rules
 
 - Complete phases in order — never skip

package/templates/fix/prompt.txt

@@ -53,6 +53,10 @@ Write `.moxie/phases/fix/FIX-FINAL-{MMDDyy-HHmmss}.md` with:
 
 Do NOT write the FINAL file unless ALL agents show `reached: true` in the ledger.
 
+## Turn Pressure
+
+{{TURN_PRESSURE}}
+
 ## Rules
 
 - Fix highest severity issues first

package/templates/plan/prompt.txt

@@ -57,6 +57,10 @@ self-contained implementation plan. This is the canonical build spec.
 
 Do NOT write the FINAL file unless ALL agents show `reached: true` in the ledger.
 
+## Turn Pressure
+
+{{TURN_PRESSURE}}
+
 ## Rules
 
 - All file paths must be accurate against current source

package/templates/rfc/prompt.txt

@@ -80,6 +80,10 @@ If ALL agents (every agent in the config) show `reached: true`:
 - Copy the final RFC to `.moxie/spec.md` (this becomes the canonical spec for all later phases)
 - Set ledger `status` to `"rfc_final"`
 
+## Turn Pressure
+
+{{TURN_PRESSURE}}
+
 ## Rules
 
 - Be EXHAUSTIVE — this RFC drives the entire pipeline. Missing details here mean gaps downstream.