loki-mode 5.1.3 → 5.2.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (11) hide show
  1. package/SKILL.md +26 -2
  2. package/VERSION +1 -1
  3. package/bin/postinstall.js +1 -1
  4. package/docs/ACKNOWLEDGEMENTS.md +63 -1
  5. package/docs/COMPARISON.md +123 -1
  6. package/package.json +1 -1
  7. package/references/core-workflow.md +14 -1
  8. package/references/memory-system.md +1097 -1
  9. package/skills/model-selection.md +231 -0
  10. package/skills/quality-gates.md +283 -0
  11. package/skills/troubleshooting.md +634 -1
package/SKILL.md CHANGED
@@ -3,7 +3,7 @@ name: loki-mode
3
3
  description: Multi-agent autonomous startup system. Triggers on "Loki Mode". Takes PRD to deployed product with zero human intervention. Requires --dangerously-skip-permissions flag.
4
4
  ---
5
5
 
6
- # Loki Mode v5.1.3
6
+ # Loki Mode v5.2.0
7
7
 
8
8
  **You are an autonomous agent. You make decisions. You do not ask questions. You do not stop.**
9
9
 
@@ -41,6 +41,13 @@ Every action follows this cycle. No exceptions.
41
41
  REASON: What is the highest priority unblocked task?
42
42
  |
43
43
  v
44
+ PRE-ACT ATTENTION: Goal alignment check (prevents context drift)
45
+ - Re-read .loki/queue/current-task.json
46
+ - Verify: "Does my planned action serve task.goal?"
47
+ - Check: "Am I solving the original problem, not a tangent?"
48
+ - IF drift detected: Log to .loki/signals/DRIFT_DETECTED, return to REASON
49
+ |
50
+ v
44
51
  ACT: Execute it. Write code. Run commands. Commit atomically.
45
52
  |
46
53
  v
@@ -57,6 +64,12 @@ VERIFY: Run tests. Check build. Validate against spec.
57
64
  After 5 failures: Log to dead-letter queue, move to next task.
58
65
  ```
59
66
 
67
+ **Why PRE-ACT ATTENTION matters** (from planning-with-files pattern):
68
+ - Context drift is silent - agents don't notice they've drifted off-task
69
+ - Forcing goal re-read before each action catches drift early
70
+ - Prevents "correct solution to wrong problem" failure mode
71
+ - Cost: One file read per action. Benefit: Catches misalignment before wasted work.
72
+
60
73
  ---
61
74
 
62
75
  ## PRIORITY 3: Autonomy Rules
@@ -128,8 +141,19 @@ GROWTH ──[continuous improvement loop]──> GROWTH
128
141
  | `.loki/CONTINUITY.md` | Every turn | Every turn |
129
142
  | `.loki/state/orchestrator.json` | Every turn | On phase change |
130
143
  | `.loki/queue/pending.json` | Every turn | When claiming/completing tasks |
144
+ | `.loki/queue/current-task.json` | Before each ACT (PRE-ACT ATTENTION) | When claiming task |
145
+ | `.loki/signals/DRIFT_DETECTED` | Never | When goal drift detected |
131
146
  | `.loki/specs/openapi.yaml` | Before API work | After API changes |
132
147
  | `skills/00-index.md` | Session start | Never |
148
+ | `.loki/memory/index.json` | Session start | On topic change |
149
+ | `.loki/memory/timeline.json` | On context need | After task completion |
150
+ | `.loki/memory/token_economics.json` | Never (metrics only) | Every turn |
151
+ | `.loki/memory/episodic/*.json` | On task-aware retrieval | After task completion |
152
+ | `.loki/memory/semantic/patterns.json` | Before implementation tasks | On consolidation |
153
+ | `.loki/memory/semantic/anti-patterns.json` | Before debugging tasks | On error learning |
154
+ | `.loki/queue/dead-letter.json` | Session start | On task failure (5+ attempts) |
155
+ | `.loki/signals/CONTEXT_CLEAR_REQUESTED` | Never | When context heavy |
156
+ | `.loki/signals/HUMAN_REVIEW_NEEDED` | Never | When human decision required |
133
157
 
134
158
  ---
135
159
 
@@ -203,4 +227,4 @@ Auto-detected or force with `LOKI_COMPLEXITY`:
203
227
 
204
228
  ---
205
229
 
206
- **v5.1.3 | Multi-Provider Support | ~210 lines core**
230
+ **v5.2.0 | CoVe + MemEvolve + Quality Gates | ~230 lines core**
package/VERSION CHANGED
@@ -1 +1 @@
1
- 5.1.3
1
+ 5.2.0
package/bin/postinstall.js CHANGED
@@ -13,7 +13,7 @@ const skillDir = path.join(homeDir, '.claude', 'skills', 'loki-mode');
13
13
  const packageDir = path.join(__dirname, '..');
14
14
 
15
15
  console.log('');
16
- console.log('Loki Mode v5.1.3 installed!');
16
+ console.log('Loki Mode v5.2.0 installed!');
17
17
  console.log('');
18
18
 
19
19
  // Try to create skill symlink
package/docs/ACKNOWLEDGEMENTS.md CHANGED
@@ -92,6 +92,18 @@ AWS Bedrock's multi-agent collaboration patterns inform Loki Mode's routing and
92
92
  | [Measurement Imbalance in Agentic AI](https://arxiv.org/abs/2506.02064) | arXiv 2506.02064 | Multi-dimensional evaluation axes |
93
93
  | [Demo-to-Deployment Gap](https://www.marktechpost.com/2025/12/24/) | Stanford/Harvard | Tool reliability vs tool selection |
94
94
 
95
+ ### Verification & Hallucination Reduction
96
+
97
+ | Paper | Authors/Source | Contribution |
98
+ |-------|----------------|--------------|
99
+ | [Chain-of-Verification Reduces Hallucination in LLMs](https://arxiv.org/abs/2309.11495) | Dhuliawala et al., Meta AI, 2023 | 4-step verification (Draft -> Plan -> Execute -> Verify), factored execution, significant hallucination reduction (23% F1 improvement, ~77% reduction in hallucinated entities) |
100
+
101
+ ### Memory Systems
102
+
103
+ | Paper | Authors/Source | Contribution |
104
+ |-------|----------------|--------------|
105
+ | [MemEvolve: Meta-Evolution of Agent Memory Systems](https://arxiv.org/abs/2512.18746) | Zhang et al., OPPO AI Agent Team, 2025 | Modular design (Encode/Store/Retrieve/Manage), task-aware strategy selection, 17.06% improvement via meta-evolution |
106
+
95
107
  ---
96
108
 
97
109
  ## Industry Resources
@@ -171,6 +183,10 @@ Key patterns incorporated from practitioner experience:
171
183
  | Debate Verification | DeepMind | Critical change verification |
172
184
  | One Feature at a Time | Anthropic Harness | Single feature per iteration, full verification |
173
185
  | E2E Browser Testing | Anthropic Harness | Playwright MCP for visual verification |
186
+ | Chain-of-Verification | arXiv 2309.11495 | CoVe protocol in quality-gates.md |
187
+ | Factored Verification | arXiv 2309.11495 | Independent verification execution |
188
+ | Modular Memory Design | arXiv 2512.18746 | Encode/Store/Retrieve/Manage mapping in memory-system.md |
189
+ | Task-Aware Memory Strategy | arXiv 2512.18746 | Retrieval weight adjustment by task type |
174
190
 
175
191
  ---
176
192
 
@@ -223,6 +239,52 @@ Key patterns incorporated from practitioner experience:
223
239
 
224
240
  ---
225
241
 
242
+ ## Community Projects (Open Source Claude Code Skills)
243
+
244
+ The following open-source projects have pioneered patterns that influence or complement Loki Mode. Analyzed January 2026.
245
+
246
+ ### High-Impact Projects
247
+
248
+ | Project | Stars | Key Patterns | Contribution to Loki Mode |
249
+ |---------|-------|--------------|---------------------------|
250
+ | [Superpowers (obra)](https://github.com/obra/superpowers) | 35K+ | Two-Stage Review, TDD Iron Law, Rationalization Tables | **ADOPTED**: Two-stage review (spec compliance THEN code quality) |
251
+ | [agents (wshobson)](https://github.com/wshobson/agents) | 26K+ | 72 plugins, 108 agents, 129 skills, Four-Tier Model Strategy | Plugin marketplace architecture inspiration |
252
+ | [claude-flow (ruvnet)](https://github.com/ruvnet/claude-flow) | 12K+ | Swarm topologies (hierarchical/mesh/ring/star), Consensus algorithms (Raft, Byzantine, CRDT) | Terminal-based orchestration patterns |
253
+ | [oh-my-claudecode (Yeachan-Heo)](https://github.com/Yeachan-Heo/oh-my-claudecode) | N/A | 32 agents, 35 skills, Tiered architecture (LOW/MEDIUM/HIGH), Delegation-first | **ADOPTED**: Tiered agent escalation protocols |
254
+
255
+ ### Specialized Skills
256
+
257
+ | Project | Focus | Key Patterns | Contribution to Loki Mode |
258
+ |---------|-------|--------------|---------------------------|
259
+ | [claude-mem (thedotmack)](https://github.com/thedotmack/claude-mem) | Memory | Progressive Disclosure (3-layer), SQLite + FTS5, Timeline compression | **ADOPTED**: 3-layer memory (index -> timeline -> full) |
260
+ | [planning-with-files (OthmanAdi)](https://github.com/OthmanAdi/planning-with-files) | Planning | Manus-style 3-file pattern, PreToolUse attention hooks | **ADOPTED**: File-based planning persistence |
261
+ | [claude-scientific-skills (K-Dense-AI)](https://github.com/K-Dense-AI/claude-scientific-skills) | Scientific | 140 domain-specific skills, modular organization | Domain organization patterns |
262
+ | [claude-code-guide (zebbern)](https://github.com/zebbern/claude-code-guide) | Shortcuts | QNEW/QCODE/QCHECK patterns, structured reports | Shortcut command inspiration |
263
+
264
+ ### Key Patterns Adopted from Community
265
+
266
+ | Pattern | Source | Implementation in Loki Mode |
267
+ |---------|--------|----------------------------|
268
+ | **Two-Stage Review** | Superpowers | Spec compliance review BEFORE code quality review |
269
+ | **Rationalization Tables** | Superpowers | Explicit counters to common agent excuses/rationalizations |
270
+ | **Progressive Disclosure Memory** | claude-mem | 3-layer context: index -> timeline -> full details |
271
+ | **Tiered Agent Escalation** | oh-my-claudecode | LOW -> MEDIUM -> HIGH with explicit escalation triggers |
272
+ | **File-Based Planning** | planning-with-files | Persistent markdown files (task_plan.md, findings.md, progress.md) |
273
+ | **PreToolUse Attention** | planning-with-files | Re-read goals before actions to combat context drift |
274
+ | **Fresh Subagent Per Task** | Superpowers | Clean context for each major task, prevents cross-contamination |
275
+
276
+ ### Patterns Under Evaluation
277
+
278
+ | Pattern | Source | Status | Notes |
279
+ |---------|--------|--------|-------|
280
+ | **Token Economics Tracking** | claude-mem | Evaluating | discovery_tokens vs read_tokens for compression analysis |
281
+ | **Delegation Enforcer Middleware** | oh-my-claudecode | Evaluating | Auto-inject model parameters based on task tier |
282
+ | **Swarm Topologies** | claude-flow | Not adopted | Adds complexity beyond hierarchical orchestration |
283
+ | **Consensus Algorithms** | claude-flow | Not adopted | Byzantine/Raft overkill for single-user autonomous operation |
284
+ | **Shortcut Commands** | claude-code-guide | Evaluating | QNEW/QCODE/QCHECK for rapid task switching |
285
+
286
+ ---
287
+
226
288
  ## License
227
289
 
228
290
  This acknowledgements file documents the research and resources that influenced Loki Mode's design. All referenced works retain their original licenses and copyrights.
@@ -231,4 +293,4 @@ Loki Mode itself is released under the MIT License.
231
293
 
232
294
  ---
233
295
 
234
- *Last updated: v4.1.0*
296
+ *Last updated: v5.1.3*
package/docs/COMPARISON.md CHANGED
@@ -1,6 +1,6 @@
1
1
  # Autonomous Coding Agents Comparison (2025-2026)
2
2
 
3
- > Last Updated: January 17, 2026 (v2.36.8)
3
+ > Last Updated: January 25, 2026 (v2.36.9)
4
4
  >
5
5
  > A comprehensive comparison of Loki Mode against major autonomous coding agents and AI IDEs in the market.
6
6
  > Deep-dive comparisons validated by Opus feedback loops.
@@ -193,6 +193,117 @@
193
193
 
194
194
  ---
195
195
 
196
+ ## Open Source Claude Code Skills Comparison (v2.36.9)
197
+
198
+ **Comprehensive analysis of 8 leading open-source Claude Code skills/extensions. Honest assessment of what Loki Mode lacks and does well.**
199
+
200
+ ### Feature Comparison
201
+
202
+ | Feature | **Loki Mode** | **Superpowers** | **agents** | **claude-flow** | **oh-my-claudecode** | **claude-mem** |
203
+ |---------|--------------|-----------------|------------|-----------------|---------------------|----------------|
204
+ | **Stars** | 500+ | 35K+ | 26K+ | 12K+ | N/A | N/A |
205
+ | **Agents** | 37 in 7 swarms | Fresh per task | 108 agents | Swarm-based | 32 agents | N/A |
206
+ | **Skills** | Progressive disclosure | N/A | 129 skills | N/A | 35 skills | Memory focus |
207
+ | **Multi-Provider** | Yes (Claude/Codex/Gemini) | No | No | No | No | No |
208
+ | **Memory System** | 3-tier (episodic/semantic/procedural) | N/A | N/A | Hybrid | N/A | SQLite+FTS5 |
209
+ | **Quality Gates** | 7 gates | Two-Stage Review | N/A | Consensus | Tiered | N/A |
210
+
211
+ ### What Loki Mode LACKS (Honest Assessment)
212
+
213
+ These are patterns from competing projects that are **practically and scientifically superior** to Loki Mode's current implementation:
214
+
215
+ | Gap | Source | Why It Matters | Status |
216
+ |-----|--------|----------------|--------|
217
+ | **Two-Stage Review** | Superpowers | Separating spec compliance from code quality prevents spec drift. | **IMPLEMENTED** (quality-gates.md lines 285-380) |
218
+ | **Rationalization Tables** | Superpowers | Explicit counters to common agent excuses ("I'll refactor later", "This is edge case"). | **IMPLEMENTED** (troubleshooting.md lines 48-112) |
219
+ | **Progressive Disclosure Memory** | claude-mem | 3-layer (index -> timeline -> full) is more efficient than flat memory. Reduces token usage by 60-80% on context recall. | **IMPLEMENTED** (memory-system.md lines 710-1018) |
220
+ | **Token Economics Tracking** | claude-mem | Tracking discovery_tokens vs read_tokens identifies context bloat. Loki Mode has no visibility into token efficiency. | **IMPLEMENTED** (memory-system.md lines 855-893) |
221
+ | **File-Based Planning Persistence** | planning-with-files | Manus-style 3-file pattern (task_plan.md, findings.md, progress.md) survives session restarts. Loki Mode loses planning context on crash. | **MEDIUM** |
222
+ | **PreToolUse Attention Hooks** | planning-with-files | Re-reading goals BEFORE each action combats context drift. Loki Mode relies on RARV but doesn't enforce pre-action goal review. | **IMPLEMENTED** (SKILL.md lines 44-71) |
223
+ | **Delegation Enforcer Middleware** | oh-my-claudecode | Auto-injecting model parameters prevents wrong-model-for-task. Loki Mode relies on agent discipline. | **LOW** |
224
+ | **Shortcut Commands** | claude-code-guide | QNEW/QCODE/QCHECK patterns enable rapid task switching. Loki Mode requires full prompts. | **LOW** |
225
+
226
+ ### What Loki Mode Does WELL
227
+
228
+ | Strength | Details | Competitors Lacking This |
229
+ |----------|---------|-------------------------|
230
+ | **Multi-Provider Support** | Only skill supporting Claude, Codex, and Gemini with graceful degradation | All 8 competitors are Claude-only |
231
+ | **RARV Cycle** | Reason-Act-Reflect-Verify is more rigorous than Plan-Execute | Most use simple Plan-Execute |
232
+ | **7-Gate Quality System** | Static analysis + 3 reviewers + devil's advocate + anti-sycophancy + severity blocking + coverage + debate | Superpowers has 2-stage, others have less |
233
+ | **Constitutional AI Integration** | Principles-based self-critique from Anthropic research | None have this |
234
+ | **Anti-Sycophancy (CONSENSAGENT)** | Blind review + devil's advocate prevents groupthink | None have this |
235
+ | **Provider Abstraction Layer** | Clean degradation from full-featured to sequential-only | Claude-only projects can't degrade |
236
+ | **37 Specialized Agents** | Purpose-built agents in 7 swarms vs generic | agents (108) has more but less organized |
237
+ | **Research Foundation** | 10+ academic papers integrated with citations | Most have no research backing |
238
+
239
+ ### Superpowers Deep-Dive (35K+ Stars)
240
+
241
+ The most influential open-source Claude Code skill. Key patterns:
242
+
243
+ | Pattern | Description | Loki Mode Status |
244
+ |---------|-------------|------------------|
245
+ | **Two-Stage Review** | Stage 1: Does code match spec? Stage 2: Is code quality good? Never mix. | **IMPLEMENTED** (quality-gates.md) |
246
+ | **TDD Iron Law** | Write failing test BEFORE implementation. No exceptions. | Already in testing.md |
247
+ | **Rationalization Tables** | Explicit list of agent excuses with counters | **IMPLEMENTED** (troubleshooting.md) |
248
+ | **Fresh Subagent Per Task** | New context for each major task, prevents cross-contamination | Already via Task tool |
249
+ | **Red Flag Detection** | Patterns indicating agent is rationalizing (hedging, scope changes) | **IMPLEMENTED** (troubleshooting.md lines 71-103) |
250
+
251
+ ### agents Deep-Dive (26K+ Stars)
252
+
253
+ Plugin marketplace architecture with unprecedented scale:
254
+
255
+ | Pattern | Description | Loki Mode Status |
256
+ |---------|-------------|------------------|
257
+ | **72 Plugins** | Modular, focused plugins instead of monolith | Different approach (progressive disclosure) |
258
+ | **108 Agents** | Specialized agents for specific domains | 37 agents in Loki Mode |
259
+ | **129 Skills** | Skills as first-class objects | 10 skills in skills/ |
260
+ | **Four-Tier Model Strategy** | Explicit tier selection with constraints | Similar to Loki Mode tiers |
261
+
262
+ ### claude-mem Deep-Dive
263
+
264
+ Memory-focused skill with superior context management:
265
+
266
+ | Pattern | Description | Loki Mode Status |
267
+ |---------|-------------|------------------|
268
+ | **Progressive Disclosure** | 3-layer: index (100 tokens) -> timeline (500 tokens) -> full (unlimited) | **IMPLEMENTED** (memory-system.md lines 710-1018) |
269
+ | **SQLite + FTS5** | Full-text search on memory | Loki Mode uses file-based |
270
+ | **Timeline Compression** | Compress old memories, keep recent detailed | **TO ADOPT** |
271
+ | **Token Economics** | Track tokens per operation for optimization | **IMPLEMENTED** (memory-system.md lines 855-946) |
272
+
273
+ ### oh-my-claudecode Deep-Dive
274
+
275
+ Tiered agent architecture with explicit escalation:
276
+
277
+ | Pattern | Description | Loki Mode Status |
278
+ |---------|-------------|------------------|
279
+ | **32 Agents** | Smaller but well-organized agent set | 37 in Loki Mode |
280
+ | **35 Skills** | Domain-specific skills | 10 skills in Loki Mode |
281
+ | **Tiered Architecture** | LOW/MEDIUM/HIGH with explicit triggers | **IMPLEMENTED** (model-selection.md lines 180-363) |
282
+ | **Delegation Enforcer** | Middleware auto-injects correct model | Evaluating |
283
+ | **Delegation-First** | Agents must delegate before acting directly | Different approach |
284
+
285
+ ### Actionable Improvements for Loki Mode
286
+
287
+ **Phase 1: Critical (v5.2.0)** - COMPLETED
288
+ 1. ~~Implement Two-Stage Review in quality-gates.md~~ - DONE (lines 285-380)
289
+ 2. ~~Add Rationalization Tables to troubleshooting.md~~ - DONE (lines 48-112)
290
+ 3. ~~Add Red Flag Detection patterns~~ - DONE (troubleshooting.md lines 71-103)
291
+
292
+ **Phase 2: Critical (v5.2.0)** - COMPLETED
293
+ 4. ~~Implement Progressive Disclosure Memory (3-layer)~~ - DONE (memory-system.md lines 710-1018)
294
+ 5. ~~Add Token Economics Tracking to metrics~~ - DONE (memory-system.md lines 855-893)
295
+ 6. ~~Add PreToolUse Attention Hooks~~ - DONE (SKILL.md lines 44-71)
296
+
297
+ **Phase 3: Medium Priority (v5.4.0)**
298
+ 7. File-Based Planning Persistence (Manus-style)
299
+ 8. Timeline Compression for memory
300
+
301
+ **Phase 4: Evaluation (Future)**
302
+ 9. Shortcut Commands (QNEW/QCODE)
303
+ 10. Delegation Enforcer Middleware
304
+
305
+ ---
306
+
196
307
  ## Deep-Dive Comparison Results
197
308
 
198
309
  ### Patterns Adopted from Each Competitor
@@ -301,6 +412,16 @@ Each comparison was validated through:
301
412
  - [Google Antigravity Blog](https://developers.googleblog.com/build-with-google-antigravity-our-new-agentic-development-platform/)
302
413
  - [Amazon Q Developer Features](https://aws.amazon.com/q/developer/features/)
303
414
 
415
+ ### Open Source Claude Code Skills (v2.36.9)
416
+ - [Superpowers (obra)](https://github.com/obra/superpowers) - 35K+ stars
417
+ - [agents (wshobson)](https://github.com/wshobson/agents) - 26K+ stars
418
+ - [claude-flow (ruvnet)](https://github.com/ruvnet/claude-flow) - 12K+ stars
419
+ - [oh-my-claudecode (Yeachan-Heo)](https://github.com/Yeachan-Heo/oh-my-claudecode)
420
+ - [claude-mem (thedotmack)](https://github.com/thedotmack/claude-mem)
421
+ - [planning-with-files (OthmanAdi)](https://github.com/OthmanAdi/planning-with-files)
422
+ - [claude-scientific-skills (K-Dense-AI)](https://github.com/K-Dense-AI/claude-scientific-skills)
423
+ - [claude-code-guide (zebbern)](https://github.com/zebbern/claude-code-guide)
424
+
304
425
  ### Additional Sources
305
426
  - [Faros AI - Best AI Coding Agents 2026](https://www.faros.ai/blog/best-ai-coding-agents-2026)
306
427
  - [Artificial Analysis - Coding Agents Comparison](https://artificialanalysis.ai/insights/coding-agents-comparison)
@@ -319,6 +440,7 @@ Each comparison was validated through:
319
440
  | v2.36.5 | 2026-01-15 | Antigravity, Amazon Q |
320
441
  | v2.36.7 | 2026-01-17 | Zencoder/Zenflow |
321
442
  | v2.36.8 | 2026-01-17 | Model assignment update (Opus for SDLC phases) |
443
+ | v2.36.9 | 2026-01-25 | Open Source Claude Code Skills (8 repos: Superpowers, agents, claude-flow, oh-my-claudecode, claude-mem, planning-with-files, claude-scientific-skills, claude-code-guide) |
322
444
 
323
445
  ---
324
446
 
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "loki-mode",
3
- "version": "5.1.3",
3
+ "version": "5.2.0",
4
4
  "description": "Multi-agent autonomous startup system for Claude Code, Codex CLI, and Gemini CLI",
5
5
  "keywords": [
6
6
  "claude",
package/references/core-workflow.md CHANGED
@@ -33,6 +33,13 @@ Every iteration follows this cycle:
33
33
  | - Identify highest priority unblocked task |
34
34
  | - Determine exact steps to complete it |
35
35
  +-------------------------------------------------------------------+
36
+ | PRE-ACT ATTENTION: Goal alignment check (prevents context drift) |
37
+ | - Re-read .loki/queue/current-task.json |
38
+ | - Verify: "Does my planned action serve task.goal?" |
39
+ | - Check: "Am I solving the original problem, not a tangent?" |
40
+ | - IF drift detected: Log to .loki/signals/DRIFT_DETECTED, |
41
+ | return to REASON |
42
+ +-------------------------------------------------------------------+
36
43
  | ACT: Execute the task |
37
44
  | - Dispatch subagent via Task tool OR execute directly |
38
45
  | - Write code, run tests, fix issues |
@@ -64,12 +71,18 @@ Every iteration follows this cycle:
64
71
  +-------------------------------------------------------------------+
65
72
  ```
66
73
 
67
- **Key Enhancement:** The VERIFY step creates a feedback loop where the AI:
74
+ **Key Enhancement (VERIFY):** The VERIFY step creates a feedback loop where the AI:
68
75
  - Tests every change automatically
69
76
  - Learns from failures by updating CONTINUITY.md
70
77
  - Retries with learned context
71
78
  - Achieves 2-3x quality improvement (Boris Cherny's observed result)
72
79
 
80
+ **Key Enhancement (PRE-ACT ATTENTION):** The planning-with-files pattern adds a goal alignment check:
81
+ - Context drift is silent - agents don't notice they've drifted off-task
82
+ - Forcing goal re-read before each action catches drift early
83
+ - Prevents "correct solution to wrong problem" failure mode
84
+ - Cost: One file read per action. Benefit: Catches misalignment before wasted work.
85
+
73
86
  ---
74
87
 
75
88
  ## CONTINUITY.md - Working Memory Protocol