loki-mode 6.80.1 → 6.81.1

package/SKILL.md

@@ -3,7 +3,7 @@ name: loki-mode
 description: Multi-agent autonomous startup system. Triggers on "Loki Mode". Takes PRD to deployed product with minimal human intervention. Requires --dangerously-skip-permissions flag.
 ---
 
-# Loki Mode v6.80.1
+# Loki Mode v6.81.1
 
 **You are an autonomous agent. You make decisions. You do not ask questions. You do not stop.**
 
@@ -131,7 +131,6 @@ GROWTH ──[continuous improvement loop]──> GROWTH
 
 - Load only 1-2 skill modules at a time (from skills/00-index.md)
 - Use Task tool with subagents for exploration (isolates context)
-- IF context feels heavy: Create `.loki/signals/CONTEXT_CLEAR_REQUESTED`
 - **Context Window Tracking (v5.40.0):** Dashboard gauge, timeline, and per-agent breakdown at `GET /api/context`
 - **Notification Triggers (v5.40.0):** Configurable alerts when context exceeds thresholds, tasks fail, or budget limits hit. Manage via `GET/PUT /api/notifications/triggers`
 
@@ -154,7 +153,6 @@ GROWTH ──[continuous improvement loop]──> GROWTH
 | `.loki/memory/semantic/patterns.json` | Before implementation tasks | On consolidation |
 | `.loki/memory/semantic/anti-patterns.json` | Before debugging tasks | On error learning |
 | `.loki/queue/dead-letter.json` | Session start | On task failure (5+ attempts) |
-| `.loki/signals/CONTEXT_CLEAR_REQUESTED` | Never | When context heavy |
 | `.loki/signals/HUMAN_REVIEW_NEEDED` | Never | When human decision required |
 | `.loki/state/checkpoints/` | After task completion | Automatic + manual via `loki checkpoint` |
 
@@ -266,10 +264,9 @@ The following features are documented in skill modules but not yet fully automat
 
 | Feature | Status | Notes |
 |---------|--------|-------|
-| PRE-ACT goal drift detection | Planned | Agent-level attention check before each action; no automated enforcement yet |
 | CONTINUITY.md working memory | Implemented (v5.35.0) | Auto-managed by run.sh, updated each iteration |
 | GitHub integration | Implemented (v5.42.2) | Import, sync-back, PR creation, export. CLI: `loki github`, API: `/api/github/*` |
 | Quality gates 3-reviewer system | Implemented (v5.35.0) | 5 specialist reviewers in `skills/quality-gates.md`; execution in run.sh |
 | Benchmarks (HumanEval, SWE-bench) | Infrastructure only | Runner scripts and datasets exist in `benchmarks/`; no published results |
 
-**v6.80.1 | [Autonomi](https://www.autonomi.dev/) flagship product | ~260 lines core**
+**v6.81.1 | [Autonomi](https://www.autonomi.dev/) flagship product | ~260 lines core**

package/VERSION

@@ -1 +1 @@
-6.80.1
+6.81.1

package/autonomy/CONSTITUTION.md

@@ -433,8 +433,8 @@ git reset --hard ${checkpoint_hash}
 
 - Load only 1-2 skill modules at a time
 - Use Task tool with subagents for exploration (isolates context)
-- After 25 iterations: Consolidate learnings to CONTINUITY.md
-- If context feels heavy: Create `.loki/signals/CONTEXT_CLEAR_REQUESTED`
+- Periodically consolidate learnings to CONTINUITY.md
+- Rely on provider-native context management when context feels heavy (no explicit clear signal)
 
 ---
 

package/autonomy/loki

@@ -522,6 +522,7 @@ cmd_start() {
     local mirofish_docker_image=""
     local mirofish_bg=false
     local mirofish_disabled=false
+    local no_plan=false  # v6.81.1: --no-plan opts out of auto-plan display
 
     while [[ $# -gt 0 ]]; do
         case "$1" in
@@ -554,6 +555,7 @@ cmd_start() {
                 echo "  --mirofish-timeout S  Max pipeline wait in seconds (default: 3600)"
                 echo "  --mirofish-bg         Run MiroFish pipeline in background"
                 echo "  --no-mirofish         Disable MiroFish even if env var is set"
+                echo "  --no-plan             Skip auto-shown PRD analysis at startup"
                 echo "  --yes, -y             Skip confirmation prompts (auto-confirm)"
                 echo ""
                 echo "Environment Variables:"
@@ -776,6 +778,10 @@ cmd_start() {
                 mirofish_disabled=true
                 shift
                 ;;
+            --no-plan)
+                no_plan=true
+                shift
+                ;;
             --budget)
                 if [[ -n "${2:-}" ]]; then
                     if ! echo "$2" | grep -qE '^[0-9]+(\.[0-9]+)?$'; then
@@ -1032,6 +1038,13 @@ cmd_start() {
         fi
     fi
 
+    # v6.81.1: auto-show PRD analysis so users see complexity / cost / time
+    # up front without needing a separate `loki plan` invocation. Skipped via
+    # --no-plan or in non-TTY contexts (CI/pipe). No-op when PRD is absent.
+    if [ -n "$prd_file" ]; then
+        maybe_show_auto_plan "$prd_file" "$no_plan" || true
+    fi
+
     # Load saved provider if not specified on command line
     if [ -z "$provider" ]; then
         if [ -f "$LOKI_DIR/state/provider" ]; then
@@ -4083,6 +4096,7 @@ cmd_run() {
                 echo "  --complex          Force complex complexity tier"
                 echo "  --no-dashboard     Disable web dashboard"
                 echo "  --sandbox          Run in Docker sandbox"
+                echo "  --no-plan          Skip auto-shown PRD analysis at startup"
                 echo "  --budget USD       Set cost budget limit"
                 echo ""
                 echo "Progressive Isolation:"
@@ -4155,7 +4169,7 @@ cmd_run() {
                 start_args+=("--provider" "$provider_override")
                 shift
                 ;;
-            --parallel|--bg|--background|--simple|--complex|--no-dashboard|--sandbox)
+            --parallel|--bg|--background|--simple|--complex|--no-dashboard|--sandbox|--no-plan)
                 start_args+=("$1")
                 shift
                 ;;
@@ -4394,6 +4408,29 @@ INNER_SCRIPT_EOF
     echo "$prd_content" > "$output_file"
     echo -e "${GREEN}PRD generated:${NC} $output_file"
 
+    # v6.81.1: auto-show PRD analysis once the issue-derived PRD is on disk,
+    # so users see cost / complexity / time before execution begins. When the
+    # user will continue into `cmd_start` below, cmd_start also has its own
+    # auto-plan call -- but that second call respects --no-plan / TTY checks
+    # and would typically re-emit the same analysis. We therefore pass
+    # --no-plan along in start_args to prevent duplicate output. Passing
+    # --no-plan here is safe because `cmd_run` itself already displayed the
+    # plan.
+    local _run_no_plan=false
+    for _arg in "${start_args[@]+"${start_args[@]}"}"; do
+        if [[ "$_arg" == "--no-plan" ]]; then
+            _run_no_plan=true
+            break
+        fi
+    done
+    if [ -f "$output_file" ]; then
+        maybe_show_auto_plan "$output_file" "$_run_no_plan" || true
+    fi
+    # Ensure cmd_start does not re-show the plan now that we just did.
+    if [[ "$_run_no_plan" != "true" ]]; then
+        start_args+=("--no-plan")
+    fi
+
     # Start Loki Mode unless --no-start
     if [[ "$no_start" == "true" ]]; then
         echo ""
@@ -10524,58 +10561,25 @@ with open('$failover_file', 'w') as f: json.dump(d, f, indent=2)
     done
 }
 
-# Dry-run PRD analysis and cost estimation (v6.18.0)
-cmd_plan() {
-    local prd_file=""
-    local show_json=false
-    local show_verbose=false
-
-    while [[ $# -gt 0 ]]; do
-        case "$1" in
-            --help|-h)
-                echo -e "${BOLD}loki plan${NC} - Dry-run PRD analysis and cost estimation"
-                echo ""
-                echo "Usage: loki plan <PRD> [options]"
-                echo ""
-                echo "Analyzes a PRD file without executing anything. Outputs complexity,"
-                echo "estimated iterations, token usage, cost, and execution plan."
-                echo ""
-                echo "Options:"
-                echo "  --json        Machine-readable JSON output"
-                echo "  --verbose     Show detailed per-iteration breakdown"
-                echo "  --help, -h    Show this help"
-                echo ""
-                echo "Examples:"
-                echo "  loki plan ./prd.md"
-                echo "  loki plan ./prd.md --json"
-                echo "  loki plan ./prd.md --verbose"
-                return 0
-                ;;
-            --json) show_json=true; shift ;;
-            --verbose) show_verbose=true; shift ;;
-            *)
-                if [ -z "$prd_file" ]; then
-                    prd_file="$1"
-                fi
-                shift
-                ;;
-        esac
-    done
-
-    if [ -z "$prd_file" ]; then
-        echo -e "${RED}Usage: loki plan <PRD file>${NC}"
-        echo "Run 'loki plan --help' for usage."
+# Reusable PRD plan analysis (v6.81.1)
+# Shared by `loki plan`, `loki start` (auto-plan), and `loki run` (auto-plan).
+# Arguments:
+#   $1 - absolute PRD path
+#   $2 - show_json (true|false)
+#   $3 - show_verbose (true|false)
+# Respects:
+#   LOKI_SESSION_MODEL          - Pins ALL iterations to one tier (default: sonnet)
+#   LOKI_LEGACY_TIER_SWITCHING  - When 'true', restores per-iteration rotation
+show_prd_plan() {
+    local prd_path="$1"
+    local show_json="${2:-false}"
+    local show_verbose="${3:-false}"
+
+    if [ -z "$prd_path" ] || [ ! -f "$prd_path" ]; then
+        echo -e "${RED}show_prd_plan: PRD file not found: $prd_path${NC}" >&2
         return 1
     fi
 
-    if [ ! -f "$prd_file" ]; then
-        echo -e "${RED}PRD file not found: $prd_file${NC}"
-        return 1
-    fi
-
-    local prd_path
-    prd_path="$(cd "$(dirname "$prd_file")" && pwd)/$(basename "$prd_file")"
-
     python3 -c "
 import json, sys, os, re, math
 
@@ -10583,6 +10587,22 @@ prd_path = sys.argv[1]
 show_json = sys.argv[2] == 'true'
 show_verbose = sys.argv[3] == 'true'
 
+# v6.81.1: Respect LOKI_SESSION_MODEL (pinned by v6.81.0) so the estimate
+# matches what the main loop actually invokes. LOKI_LEGACY_TIER_SWITCHING=true
+# restores the legacy per-iteration Opus/Sonnet/Haiku rotation for users who
+# explicitly opt out.
+session_model_env = (os.environ.get('LOKI_SESSION_MODEL', 'sonnet') or 'sonnet').strip().lower()
+legacy_tier_switching = (os.environ.get('LOKI_LEGACY_TIER_SWITCHING', 'false') or 'false').strip().lower() == 'true'
+
+# Map session model name to tier key used in tokens_per_tier below.
+# Unknown models fall through to 'development' (Sonnet) as a safe default.
+_session_tier_map = {
+    'opus': 'planning',
+    'sonnet': 'development',
+    'haiku': 'fast',
+}
+session_tier = _session_tier_map.get(session_model_env, 'development')
+
 # Colors (disabled for JSON mode)
 if show_json:
     RED = GREEN = YELLOW = BLUE = CYAN = BOLD = DIM = NC = ''
@@ -10770,18 +10790,28 @@ tier_iterations = {'Opus': 0, 'Sonnet': 0, 'Haiku': 0}
 for i in range(estimated_iterations):
     rarv_step = i % 4
     if rarv_step == 0:
-        tier = 'planning'
         phase_label = 'Reason (Planning)'
     elif rarv_step == 1:
-        tier = 'development'
         phase_label = 'Act (Implementation)'
     elif rarv_step == 2:
-        tier = 'development'
         phase_label = 'Reflect (Review)'
     else:
-        tier = 'fast'
         phase_label = 'Verify (Testing)'
 
+    if legacy_tier_switching:
+        # Legacy pre-v6.81.0 behavior: rotate tier per RARV step.
+        if rarv_step == 0:
+            tier = 'planning'
+        elif rarv_step in (1, 2):
+            tier = 'development'
+        else:
+            tier = 'fast'
+    else:
+        # v6.81.0+: main loop is pinned to LOKI_SESSION_MODEL. Estimate
+        # MUST use the same tier for all iterations to avoid quoting a
+        # cost higher than the user will actually pay.
+        tier = session_tier
+
     info = tokens_per_tier[tier]
     model = info['model']
     inp = info['input']
@@ -10957,7 +10987,20 @@ print(f'  RARV cycles: {cycle_count} (4 iterations per cycle)')
 opus_n = tier_iterations.get('Opus', 0)
 sonnet_n = tier_iterations.get('Sonnet', 0)
 haiku_n = tier_iterations.get('Haiku', 0)
-print(f'  Model distribution: Opus x{opus_n} | Sonnet x{sonnet_n} | Haiku x{haiku_n}')
+if legacy_tier_switching:
+    print(f'  Model distribution: Opus x{opus_n} | Sonnet x{sonnet_n} | Haiku x{haiku_n}')
+    print(f'  {DIM}(legacy rotation: LOKI_LEGACY_TIER_SWITCHING=true){NC}')
+else:
+    # Session-pinned: exactly one tier has all iterations.
+    pinned_parts = []
+    if opus_n > 0:
+        pinned_parts.append(f'Opus x{opus_n}')
+    if sonnet_n > 0:
+        pinned_parts.append(f'Sonnet x{sonnet_n}')
+    if haiku_n > 0:
+        pinned_parts.append(f'Haiku x{haiku_n}')
+    print('  Model distribution: ' + (' | '.join(pinned_parts) if pinned_parts else '(none)'))
+    print(f'  {DIM}(pinned via LOKI_SESSION_MODEL={session_model_env}; set LOKI_LEGACY_TIER_SWITCHING=true to rotate){NC}')
 
 # Tokens
 total_tok = total_input_tokens + total_output_tokens
@@ -10971,6 +11014,11 @@ ds = chr(36)  # dollar sign
 print(f'\n{CYAN}Cost Estimate{NC}')
 print(f'  {BOLD}Total: {ds}{total_cost:.2f}{NC}')
 for model in ['Opus', 'Sonnet', 'Haiku']:
+    # v6.81.1: suppress zero-cost tiers (e.g. session-pinned runs). Keeps
+    # the breakdown focused on models actually used. Legacy mode still
+    # shows all three because each tier has non-zero iterations.
+    if tier_iterations.get(model, 0) == 0 and tier_totals.get(model, 0) == 0:
+        continue
     pct = (tier_totals[model] / total_cost * 100) if total_cost > 0 else 0
     bar_len = int(pct / 5)
     bar = '#' * bar_len + '.' * (20 - bar_len)
@@ -11025,6 +11073,99 @@ print()
 " "$prd_path" "$show_json" "$show_verbose"
 }
 
+# Dry-run PRD analysis and cost estimation (v6.18.0; refactored v6.81.1)
+# Thin wrapper around show_prd_plan() so the same analysis is reusable from
+# `loki start`, `loki run`, and other commands without duplicated code.
+cmd_plan() {
+    local prd_file=""
+    local show_json=false
+    local show_verbose=false
+
+    while [[ $# -gt 0 ]]; do
+        case "$1" in
+            --help|-h)
+                echo -e "${BOLD}loki plan${NC} - Dry-run PRD analysis and cost estimation"
+                echo ""
+                echo "Usage: loki plan <PRD> [options]"
+                echo ""
+                echo "Analyzes a PRD file without executing anything. Outputs complexity,"
+                echo "estimated iterations, token usage, cost, and execution plan."
+                echo ""
+                echo "Options:"
+                echo "  --json        Machine-readable JSON output"
+                echo "  --verbose     Show detailed per-iteration breakdown"
+                echo "  --help, -h    Show this help"
+                echo ""
+                echo "Environment Variables:"
+                echo "  LOKI_SESSION_MODEL            Pin cost estimate to a single tier"
+                echo "                                (opus|sonnet|haiku, default: sonnet)"
+                echo "  LOKI_LEGACY_TIER_SWITCHING    Set to 'true' to restore the legacy"
+                echo "                                Opus/Sonnet/Haiku per-iteration rotation"
+                echo ""
+                echo "Examples:"
+                echo "  loki plan ./prd.md"
+                echo "  loki plan ./prd.md --json"
+                echo "  loki plan ./prd.md --verbose"
+                return 0
+                ;;
+            --json) show_json=true; shift ;;
+            --verbose) show_verbose=true; shift ;;
+            *)
+                if [ -z "$prd_file" ]; then
+                    prd_file="$1"
+                fi
+                shift
+                ;;
+        esac
+    done
+
+    if [ -z "$prd_file" ]; then
+        echo -e "${RED}Usage: loki plan <PRD file>${NC}"
+        echo "Run 'loki plan --help' for usage."
+        return 1
+    fi
+
+    if [ ! -f "$prd_file" ]; then
+        echo -e "${RED}PRD file not found: $prd_file${NC}"
+        return 1
+    fi
+
+    local prd_path
+    prd_path="$(cd "$(dirname "$prd_file")" && pwd)/$(basename "$prd_file")"
+
+    show_prd_plan "$prd_path" "$show_json" "$show_verbose"
+}
+
+# Helper: decide whether to auto-show the plan from a major command.
+# v6.81.1: default ON, opt-out via --no-plan. Also auto-skipped when stdout is
+# not a TTY and --no-plan was not explicitly set (CI/automation contexts),
+# logging a one-line "plan skipped: non-interactive" to stderr for visibility.
+maybe_show_auto_plan() {
+    local prd_path="$1"
+    local no_plan_flag="${2:-false}"  # user passed --no-plan explicitly
+
+    # Explicit opt-out wins.
+    if [[ "$no_plan_flag" == "true" ]]; then
+        return 0
+    fi
+
+    # PRD missing: silently skip (other commands will emit their own errors).
+    if [ -z "$prd_path" ] || [ ! -f "$prd_path" ]; then
+        return 0
+    fi
+
+    # Non-TTY heuristic: keep CI / pipe output clean. Log to stderr so it is
+    # still surfaced in logs without corrupting stdout parsers.
+    if [ ! -t 1 ]; then
+        echo "plan skipped: non-interactive (pass --no-plan to silence, or run 'loki plan <PRD>' explicitly)" >&2
+        return 0
+    fi
+
+    local abs_prd
+    abs_prd="$(cd "$(dirname "$prd_path")" && pwd)/$(basename "$prd_path")"
+    show_prd_plan "$abs_prd" "false" "false"
+}
+
 # Main command dispatcher
 main() {
     if [ $# -eq 0 ]; then

package/autonomy/run.sh

@@ -294,7 +294,6 @@ parse_simple_yaml() {
     set_from_yaml "$file" "model.planning" "LOKI_MODEL_PLANNING"
     set_from_yaml "$file" "model.development" "LOKI_MODEL_DEVELOPMENT"
     set_from_yaml "$file" "model.fast" "LOKI_MODEL_FAST"
-    set_from_yaml "$file" "model.compaction_interval" "LOKI_COMPACTION_INTERVAL"
 
     # Parallel
     set_from_yaml "$file" "parallel.enabled" "LOKI_PARALLEL_MODE"
@@ -432,7 +431,6 @@ parse_yaml_with_yq() {
         "model.prompt_repetition:LOKI_PROMPT_REPETITION"
         "model.confidence_routing:LOKI_CONFIDENCE_ROUTING"
         "model.autonomy_mode:LOKI_AUTONOMY_MODE"
-        "model.compaction_interval:LOKI_COMPACTION_INTERVAL"
         "parallel.enabled:LOKI_PARALLEL_MODE"
         "parallel.max_worktrees:LOKI_MAX_WORKTREES"
         "parallel.max_sessions:LOKI_MAX_PARALLEL_SESSIONS"
@@ -637,8 +635,14 @@ PROMPT_REPETITION=${LOKI_PROMPT_REPETITION:-true}
 CONFIDENCE_ROUTING=${LOKI_CONFIDENCE_ROUTING:-true}
 AUTONOMY_MODE=${LOKI_AUTONOMY_MODE:-perpetual}  # perpetual|checkpoint|supervised
 
-# Proactive Context Management (OpenCode/Sisyphus pattern, validated by Opus)
-COMPACTION_INTERVAL=${LOKI_COMPACTION_INTERVAL:-25}  # Suggest compaction every N iterations
+# Session-pinned model (S0.1): pin a single tier for the whole main loop.
+# Default is "sonnet" -> resolved via the "development" tier in the abstract
+# tier map. Set LOKI_LEGACY_TIER_SWITCHING=true to restore the old
+# per-iteration RARV-driven tier rotation in the main loop. The
+# get_rarv_tier function is preserved for subagent dispatch regardless.
+LOKI_SESSION_MODEL="${LOKI_SESSION_MODEL:-sonnet}"
+LOKI_LEGACY_TIER_SWITCHING="${LOKI_LEGACY_TIER_SWITCHING:-false}"
+export LOKI_SESSION_MODEL LOKI_LEGACY_TIER_SWITCHING
 
 # Parallel Workflows (Git Worktrees)
 PARALLEL_MODE=${LOKI_PARALLEL_MODE:-false}
@@ -7618,16 +7622,6 @@ check_max_iterations() {
     return 1
 }
 
-# Check if context clear was requested by agent
-check_context_clear_signal() {
-    if [ -f ".loki/signals/CONTEXT_CLEAR_REQUESTED" ]; then
-        log_info "Context clear signal detected from agent"
-        rm -f ".loki/signals/CONTEXT_CLEAR_REQUESTED"
-        return 0
-    fi
-    return 1
-}
-
 # Load latest ledger content for context injection
 load_ledger_context() {
     local ledger_content=""
@@ -8389,13 +8383,7 @@ build_prompt() {
     local analysis_instruction="CODEBASE_ANALYSIS_MODE: No PRD. FIRST: Analyze codebase - scan structure, read package.json/requirements.txt, examine README. THEN: Generate PRD at .loki/generated-prd.md. FINALLY: Execute SDLC phases."
 
     # Context Memory Instructions (integrated with new memory system)
-    local memory_instruction="MEMORY SYSTEM: Relevant context from past sessions is provided below (if any). Your actions will be automatically recorded for future reference. For complex handoffs: create .loki/memory/handoffs/{timestamp}.md. For important decisions: they will be captured in the timeline. Check .loki/CONTINUITY.md for session-level working memory. If context feels heavy, create .loki/signals/CONTEXT_CLEAR_REQUESTED and the wrapper will reset context with your ledger preserved."
-
-    # Proactive Compaction Reminder (every N iterations)
-    local compaction_reminder=""
-    if [ $((iteration % COMPACTION_INTERVAL)) -eq 0 ] && [ $iteration -gt 0 ]; then
-        compaction_reminder="PROACTIVE_CONTEXT_CHECK: You are at iteration $iteration. Review context size - if conversation history is long, consolidate to CONTINUITY.md and consider creating .loki/signals/CONTEXT_CLEAR_REQUESTED to reset context while preserving state."
-    fi
+    local memory_instruction="MEMORY SYSTEM: Relevant context from past sessions is provided below (if any). Your actions will be automatically recorded for future reference. For complex handoffs: create .loki/memory/handoffs/{timestamp}.md. For important decisions: they will be captured in the timeline. Check .loki/CONTINUITY.md for session-level working memory."
 
     # Load existing context if resuming
     local context_injection=""
@@ -8691,15 +8679,15 @@ except Exception:
     else
         if [ $retry -eq 0 ]; then
             if [ -n "$prd" ]; then
-                echo "Loki Mode with PRD at $prd. $human_directive $gate_failure_context $queue_tasks $bmad_context $openspec_context $mirofish_context $magic_context $checklist_status $app_runner_info $playwright_info $memory_context_section $rarv_instruction $memory_instruction $compaction_reminder $completion_instruction $sdlc_instruction $autonomous_suffix"
+                echo "Loki Mode with PRD at $prd. $human_directive $gate_failure_context $queue_tasks $bmad_context $openspec_context $mirofish_context $magic_context $checklist_status $app_runner_info $playwright_info $memory_context_section $rarv_instruction $memory_instruction $completion_instruction $sdlc_instruction $autonomous_suffix"
             else
-                echo "Loki Mode. $human_directive $gate_failure_context $queue_tasks $bmad_context $openspec_context $mirofish_context $magic_context $checklist_status $app_runner_info $playwright_info $memory_context_section $analysis_instruction $rarv_instruction $memory_instruction $compaction_reminder $completion_instruction $sdlc_instruction $autonomous_suffix"
+                echo "Loki Mode. $human_directive $gate_failure_context $queue_tasks $bmad_context $openspec_context $mirofish_context $magic_context $checklist_status $app_runner_info $playwright_info $memory_context_section $analysis_instruction $rarv_instruction $memory_instruction $completion_instruction $sdlc_instruction $autonomous_suffix"
             fi
         else
             if [ -n "$prd" ]; then
-                echo "Loki Mode - Resume iteration #$iteration (retry #$retry). PRD: $prd. $human_directive $gate_failure_context $queue_tasks $bmad_context $openspec_context $mirofish_context $magic_context $checklist_status $app_runner_info $playwright_info $memory_context_section $rarv_instruction $memory_instruction $compaction_reminder $completion_instruction $sdlc_instruction $autonomous_suffix"
+                echo "Loki Mode - Resume iteration #$iteration (retry #$retry). PRD: $prd. $human_directive $gate_failure_context $queue_tasks $bmad_context $openspec_context $mirofish_context $magic_context $checklist_status $app_runner_info $playwright_info $memory_context_section $rarv_instruction $memory_instruction $completion_instruction $sdlc_instruction $autonomous_suffix"
             else
-                echo "Loki Mode - Resume iteration #$iteration (retry #$retry). $human_directive $gate_failure_context $queue_tasks $bmad_context $openspec_context $mirofish_context $magic_context $checklist_status $app_runner_info $playwright_info $memory_context_section Use .loki/generated-prd.md if exists. $rarv_instruction $memory_instruction $compaction_reminder $completion_instruction $sdlc_instruction $autonomous_suffix"
+                echo "Loki Mode - Resume iteration #$iteration (retry #$retry). $human_directive $gate_failure_context $queue_tasks $bmad_context $openspec_context $mirofish_context $magic_context $checklist_status $app_runner_info $playwright_info $memory_context_section Use .loki/generated-prd.md if exists. $rarv_instruction $memory_instruction $completion_instruction $sdlc_instruction $autonomous_suffix"
             fi
         fi
     fi
@@ -9728,9 +9716,25 @@ except Exception as exc:
         echo "=== Provider: ${PROVIDER_NAME:-claude} ===" | tee -a "$log_file" "$agent_log"
         echo "=== Prompt (truncated): ${prompt:0:200}... ===" | tee -a "$log_file" "$agent_log"
 
-        # Dynamic tier selection based on RARV cycle phase
-        CURRENT_TIER=$(get_rarv_tier "$ITERATION_COUNT")
-        # NEW BUG FIX: Export LOKI_CURRENT_TIER so provider helper functions
+        # Tier selection (S0.1):
+        # Default: pin a single tier for the whole session via LOKI_SESSION_MODEL.
+        # Legacy: LOKI_LEGACY_TIER_SWITCHING=true restores RARV-driven rotation.
+        if [ "${LOKI_LEGACY_TIER_SWITCHING:-false}" = "true" ]; then
+            CURRENT_TIER=$(get_rarv_tier "$ITERATION_COUNT")
+        else
+            # Map session-pinned model name to an abstract tier so provider
+            # helpers (which expect tier names) resolve correctly. Unknown
+            # model strings are passed through as-is; provider loaders fall
+            # back to a sane default.
+            case "${LOKI_SESSION_MODEL:-sonnet}" in
+                opus)   CURRENT_TIER="planning" ;;
+                sonnet) CURRENT_TIER="development" ;;
+                haiku)  CURRENT_TIER="fast" ;;
+                planning|development|fast) CURRENT_TIER="${LOKI_SESSION_MODEL}" ;;
+                *)      CURRENT_TIER="${LOKI_SESSION_MODEL}" ;;
+            esac
+        fi
+        # Export LOKI_CURRENT_TIER so provider helper functions
         # (e.g., gemini.sh:provider_get_current_model) can resolve the correct model.
         # Without this, LOKI_CURRENT_TIER is always empty and defaults to "planning".
         LOKI_CURRENT_TIER="$CURRENT_TIER"

package/dashboard/__init__.py

@@ -7,7 +7,7 @@ Modules:
     control: Session control API (start/stop/pause/resume)
 """
 
-__version__ = "6.80.1"
+__version__ = "6.81.1"
 
 # Expose the control app for easy import
 try:

package/docs/INSTALLATION.md

@@ -2,7 +2,7 @@
 
 The flagship product of [Autonomi](https://www.autonomi.dev/). Complete installation instructions for all platforms and use cases.
 
-**Version:** v6.80.1
+**Version:** v6.81.1
 
 ---
 

package/mcp/__init__.py

@@ -57,4 +57,4 @@ try:
 except ImportError:
     __all__ = ['mcp']
 
-__version__ = '6.80.1'
+__version__ = '6.81.1'

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "loki-mode",
-  "version": "6.80.1",
+  "version": "6.81.1",
   "description": "Loki Mode by Autonomi - Multi-agent autonomous startup system for Claude Code, Codex CLI, and Gemini CLI",
   "keywords": [
     "agent",

package/references/core-workflow.md

@@ -33,13 +33,6 @@ Every iteration follows this cycle:
 | - Identify highest priority unblocked task                        |
 | - Determine exact steps to complete it                            |
 +-------------------------------------------------------------------+
-| PRE-ACT ATTENTION: Goal alignment check (prevents context drift)  |
-| - Re-read .loki/queue/current-task.json                           |
-| - Verify: "Does my planned action serve task.goal?"               |
-| - Check: "Am I solving the original problem, not a tangent?"      |
-| - IF drift detected: Log to .loki/signals/DRIFT_DETECTED,         |
-|   return to REASON                                                |
-+-------------------------------------------------------------------+
 | ACT: Execute the task                                             |
 | - Dispatch subagent via Task tool OR execute directly             |
 | - Write code, run tests, fix issues                               |
@@ -77,12 +70,6 @@ Every iteration follows this cycle:
 - Retries with learned context
 - Achieves 2-3x quality improvement (Boris Cherny's observed result)
 
-**Key Enhancement (PRE-ACT ATTENTION):** The planning-with-files pattern adds a goal alignment check:
-- Context drift is silent - agents don't notice they've drifted off-task
-- Forcing goal re-read before each action catches drift early
-- Prevents "correct solution to wrong problem" failure mode
-- Cost: One file read per action. Benefit: Catches misalignment before wasted work.
-
 ---
 
 ## CONTINUITY.md - Working Memory Protocol

package/skills/model-selection.md

@@ -1,5 +1,23 @@
 # Model Selection & Task Tool
 
+## Session-Pinned Model (default)
+
+Since v6.81.0, the main RARV loop uses **one pinned model per session** instead of rotating models per RARV phase. This eliminates per-iteration tier churn and keeps the main loop behavior predictable.
+
+- Controlled by `LOKI_SESSION_MODEL` (default: `sonnet`).
+- Accepted values: `opus`, `sonnet`, `haiku`, or a raw tier name (`planning`, `development`, `fast`). Model names are mapped to abstract tiers (opus -> planning, sonnet -> development, haiku -> fast) so provider helpers resolve correctly.
+- `get_rarv_tier()` is retained and still used for subagent Task-tool dispatches (planning vs implementation vs fast work), not for the main loop.
+
+**Rollback (legacy RARV tier rotation in the main loop):**
+
+```bash
+LOKI_LEGACY_TIER_SWITCHING=true ./autonomy/run.sh ./prd.md
+```
+
+Set `LOKI_LEGACY_TIER_SWITCHING=true` to restore the previous per-iteration `get_rarv_tier "$ITERATION_COUNT"` behavior in the main loop.
+
+---
+
 ## Multi-Provider Support (v5.0.0)
 
 Loki Mode supports five AI providers. Claude has full features; all others run in **degraded mode** (sequential execution only, no Task tool, no parallel agents).

package/skills/troubleshooting.md

@@ -577,7 +577,7 @@ Signals are inter-process communication files in `.loki/signals/` that trigger a
 | Any drift detected | Return to REASON phase immediately |
 | 1 drift in task | Log warning, continue with corrected action |
 | 2 drifts in same task | Escalate to orchestrator for task review |
-| 3+ drifts accumulated | Trigger context clear + full state reload |
+| 3+ drifts accumulated | Escalate to orchestrator, consolidate state to CONTINUITY.md |
 | High/Critical severity | Pause task, dispatch opus reviewer |
 
 **Automated Responses:**
@@ -590,9 +590,9 @@ immediate:
 
 accumulated_threshold: 3
 accumulated_action:
-  - Create CONTEXT_CLEAR_REQUESTED
-  - Wait for wrapper to reset context
-  - Resume with fresh context + ledger state
+  - Consolidate working memory to CONTINUITY.md
+  - Escalate to orchestrator for task review
+  - Rely on provider-native context management (no explicit clear signal)
 ```
 
 **How to Write:**
@@ -605,42 +605,6 @@ EOF
 
 ---
 
-### CONTEXT_CLEAR_REQUESTED Signal
-
-**Purpose:** Agent requests context window reset while preserving state.
-
-**Schema:** File presence only (empty file is sufficient).
-
-**When to Create:**
-- Context feels heavy/slow
-- After 25+ iterations
-- Conversation history exceeds 50KB
-- Agent detects own confusion or loops
-- 3+ accumulated DRIFT_DETECTED events
-
-**Processing (by run.sh wrapper):**
-1. Detect signal file
-2. Load ledger context from `.loki/memory/ledgers/`
-3. Load recent handoffs from `.loki/memory/handoffs/`
-4. Delete signal file
-5. Start new Claude session with injected context
-6. Continue from last CONTINUITY.md state
-
-**How to Create:**
-```bash
-touch .loki/signals/CONTEXT_CLEAR_REQUESTED
-```
-
-**Thresholds:**
-| Trigger | Threshold |
-|---------|-----------|
-| Iteration count | Every 25 iterations (compaction reminder) |
-| Drift accumulation | 3+ DRIFT_DETECTED events |
-| Agent self-assessment | "Context feels heavy" |
-| Error loop detection | Same error 3+ times |
-
----
-
 ### HUMAN_REVIEW_NEEDED Signal
 
 **Purpose:** Escalate to human when autonomous action is inappropriate.
@@ -762,7 +726,6 @@ These signals coordinate parallel worktrees (see `skills/parallel-workflows.md`)
 ```
 .loki/signals/
   DRIFT_DETECTED              # Append-only log (JSON lines)
-  CONTEXT_CLEAR_REQUESTED     # Presence flag (empty file OK)
   HUMAN_REVIEW_NEEDED         # JSON with review context
   HUMAN_APPROVED              # JSON with decision
   FEATURE_READY_auth          # Worktree signal