mdcontext 0.0.1 → 0.2.0

package/research/dogfood/strategy-c/c2-diver-memory.md

@@ -0,0 +1,145 @@
+# Report: C2 - Diver: Memory Architecture
+
+## Mission
+
+Deep-dive into: Three-Layer Memory Architecture
+
+## Command Log
+
+| #   | Command                                                                                            | Purpose                            | Result                                     | Useful? |
+| --- | -------------------------------------------------------------------------------------------------- | ---------------------------------- | ------------------------------------------ | ------- |
+| 1   | `mdcontext search "memory"`                                                                        | Find memory content                | 10 results, found key files and references | Yes     |
+| 2   | `mdcontext search "event memory"`                                                                  | Find event memory specifics        | 10 results with detailed sections          | Yes     |
+| 3   | `mdcontext search "semantic memory"`                                                               | Find semantic memory specifics     | 10 results covering definition and usage   | Yes     |
+| 4   | `mdcontext search "status memory"`                                                                 | Find status memory specifics       | 10 results with operational details        | Yes     |
+| 5   | `mdcontext context docs/05-MEMORY_MODEL.md "## 2."`                                                | Try positional context (failed)    | ENOENT error - wrong syntax                | No      |
+| 6   | `mdcontext context docs/05-MEMORY_MODEL.md --section "The Three-Layer Architecture"`               | Get architecture overview          | Clean section output                       | Yes     |
+| 7   | `mdcontext context docs/05-MEMORY_MODEL.md --section "Layer 1: Event Memory"`                      | Get event memory details           | No match found                             | No      |
+| 8   | `mdcontext context --sections docs/05-MEMORY_MODEL.md`                                             | List available sections            | Full section tree (17 top-level)           | Yes     |
+| 9   | `mdcontext context docs/05-MEMORY_MODEL.md --section "3. Layer 1: Event / Fact Memory"`            | Get event memory with correct name | Full section with all subsections          | Yes     |
+| 10  | `mdcontext context docs/05-MEMORY_MODEL.md --section "4. Layer 2: Status Memory"`                  | Get status memory details          | Complete operational truth section         | Yes     |
+| 11  | `mdcontext context docs/05-MEMORY_MODEL.md --section "5. Layer 3: Semantic Memory"`                | Get semantic memory details        | Full knowledge graph layer info            | Yes     |
+| 12  | `mdcontext context docs/05-MEMORY_MODEL.md --section "8. Time Travel and Branching"`               | Get time travel features           | Branching and replay mechanics             | Yes     |
+| 13  | `mdcontext context docs/05-MEMORY_MODEL.md --section "7. Memory and Execution Context Resumption"` | Get resumption pattern             | Critical 5-step resumption flow            | Yes     |
+| 14  | `mdcontext context docs/05-MEMORY_MODEL.md --section "1. Memory Philosophy"`                       | Get design rationale               | Core philosophy defined                    | Yes     |
+| 15  | `mdcontext search "immutable"`                                                                     | Find immutability references       | 10 cross-doc references                    | Yes     |
+| 16  | `mdcontext context docs/05-MEMORY_MODEL.md --section "13. Memory Guarantees"`                      | Get guarantees                     | 5 guarantee categories                     | Yes     |
+| 17  | `mdcontext context docs/05-MEMORY_MODEL.md --section "14. Anti-Patterns (Explicitly Forbidden)"`   | Get anti-patterns                  | 7 forbidden patterns                       | Yes     |
+| 18  | `mdcontext context docs/01-ARCHITECTURE.md --section "Workspace Memory"`                           | Get architecture view              | Concise 3-layer summary                    | Yes     |
+| 19  | `mdcontext context docs/05-MEMORY_MODEL.md --section "17. Summary"`                                | Get summary                        | Concluding philosophy                      | Yes     |
+| 20  | `mdcontext search "replay"`                                                                        | Find replay capability             | 7 replay references                        | Yes     |
+| 21  | `mdcontext search "derived"`                                                                       | Find derived state refs            | 10 references showing derivation chain     | Yes     |
+| 22  | `mdcontext context docs/05-MEMORY_MODEL.md --section "6. Memory Composition"`                      | Get composition details            | Layer relationships and consumers          | Yes     |
+| 23  | `mdcontext context docs/05-MEMORY_MODEL.md --section "9. Org-Level Memory"`                        | Get org memory                     | Workspace-to-org flow                      | Yes     |
+| 24  | `mdcontext search "knowledge graph"`                                                               | Find knowledge graph refs          | 7 results covering semantic layer          | Yes     |
+| 25  | `mdcontext context docs/05-MEMORY_MODEL.md --section "16. Why This Model Works"`                   | Get rationale                      | Benefits and scaling                       | Yes     |
+| 26  | `mdcontext context docs/05-MEMORY_MODEL.md --section "15. Implementation Guidance"`                | Get implementation tech            | Tech stack recommendations                 | Yes     |
+| 27  | `mdcontext context docs/04-EXECUTION_MODEL.md --section "What Agents Receive"`                     | Get agent context details          | 4 memory sources for agents                | Yes     |
+
+## Findings
+
+### Key Discoveries
+
+1. **Event Memory is the Single Source of Truth**
+   - Immutable, append-only log of all Workspace events
+   - Contains Records: execution_context.started, artifact.produced, step.completed, etc.
+   - All other memory layers can be rebuilt from Event Memory
+   - Optimized for correctness/durability, not query performance
+
+2. **Status Memory is Derived Operational Truth**
+   - Computed from Event Memory, never authoritative
+   - Stores: active jobs, pending signals, resource usage, attention state
+   - Can be deleted and rebuilt at any time
+   - Hard rule: "Status Memory must be fully derivable from Event Memory"
+
+3. **Semantic Memory is Understanding, Not Facts**
+   - Knowledge graph of relationships, concepts, patterns
+   - Probabilistic and explainable (every edge has provenance)
+   - Advises but never controls execution
+   - Uses LLMs, heuristics, and annotations for extraction
+
+4. **Time Travel and Branching are Natural Consequences**
+   - Rewind to any point by replaying Events up to timestamp T
+   - Branches share history up to fork point, then diverge
+   - Each branch has independent Status and Semantic Memory
+   - "Time travel is a query, not a feature"
+
+5. **5-Step Context Resumption Pattern**
+   - Query Event Memory (what happened)
+   - Query Artifact Store (what exists)
+   - Query Status Memory (what's pending)
+   - Query Semantic Memory (what it means)
+   - Assemble context for Agent
+
+### Relevant Quotes/Sections Found
+
+> "Event Memory is the immutable, append-only log of everything that occurred in the Workspace. This is the source of truth."
+> Source: docs/05-MEMORY_MODEL.md, Section 3 Definition
+
+> "Status Memory must be fully derivable from Event Memory. You must be able to delete Status Memory entirely, rebuild it from Events, get identical results."
+> Source: docs/05-MEMORY_MODEL.md, Section 4.6 The Hard Rule
+
+> "Semantic Memory advises but never controls. It may suggest, recommend, highlight. It may never block execution, override decisions, mutate artifacts, drive workflows."
+> Source: docs/05-MEMORY_MODEL.md, Section 5.6 The Non-Authoritative Rule
+
+> "Memory is the connective tissue that enables long-running, interruptible human work with perfect continuity and trust."
+> Source: docs/05-MEMORY_MODEL.md, Section 17 Summary
+
+> "No circular dependencies. Clean derivation hierarchy."
+> Source: docs/05-MEMORY_MODEL.md, Section 6 Memory Composition
+
+### Summary of Theme
+
+The Three-Layer Memory Architecture is the foundational innovation of HumanWork that enables its core promises: interruptibility, auditability, and time travel.
+
+**Layer 1: Event Memory (Facts)** - The immutable append-only log that records every meaningful event. This is the only source of truth. Technologies: EventStoreDB, append-only DBs, WAL.
+
+**Layer 2: Status Memory (Operational Truth)** - Derived projections optimized for fast queries about current state. Ephemeral and recomputable. Technologies: Redis, PostgreSQL projections.
+
+**Layer 3: Semantic Memory (Understanding)** - Derived knowledge graph capturing relationships, concepts, and patterns. Advisory only, never authoritative. Technologies: Neo4j, Graphiti, vector DBs.
+
+The key insight is the strict derivation hierarchy: Event -> Status -> Semantic. No circular dependencies. This enables:
+
+- **Time travel**: Replay events to any timestamp
+- **Branching**: Fork at any event boundary, independent evolution
+- **Resumption**: Any agent can continue any work with full context
+- **Auditability**: Complete traceable history
+- **Agent independence**: No hidden agent state required
+
+The model explicitly forbids anti-patterns like treating Status Memory as authoritative, letting Semantic Memory drive execution, or bypassing Event Memory for performance.
+
+## Proposed Spec Changes
+
+- [ ] Clarify the distinction between Event Memory and Artifact Store (both immutable, but different purposes)
+- [ ] Add concrete examples of Record types and their schema
+- [ ] Define retention policy defaults for each layer
+- [ ] Specify maximum acceptable latency for Status Memory queries
+- [ ] Document conflict resolution when Semantic Memory inferences contradict
+
+## Tool Evaluation
+
+### What Worked Well
+
+- `mdcontext search` quickly found all relevant documents and sections
+- `mdcontext context --section` returned clean, well-formatted output with full section content
+- `mdcontext context --sections` listing was invaluable for finding exact section names
+- Keyword search with compound terms ("event memory", "status memory") worked well
+- Token counts in section listings help estimate content size
+
+### What Was Frustrating
+
+- Initial attempt at context extraction failed due to positional argument confusion (thought section name was second positional, not a flag)
+- Section name matching requires exact match including numbering prefix (e.g., "3. Layer 1: Event / Fact Memory" not just "Layer 1")
+- No fuzzy/partial matching for section names
+- Had to list sections first to get exact names, adding an extra step
+- Search results capped at 10, potentially missing relevant content
+
+### Confidence Level
+
+[X] High / Medium / Low
+
+The three-layer memory architecture is extremely well documented with clear definitions, properties, guarantees, anti-patterns, and implementation guidance. The documentation is comprehensive and internally consistent.
+
+### Would Use Again? (1-5)
+
+4 - The tool is effective for systematic exploration once you learn the patterns. The `--sections` listing feature is essential. The combination of broad search followed by targeted context extraction works well. Would rate 5 if section matching was more forgiving and search results weren't capped.

package/research/dogfood/strategy-c/c3-diver-control.md

@@ -0,0 +1,148 @@
+# Report: C3 - Diver: Human-Agent Control
+
+## Mission
+
+Deep-dive into: Human-Agent Control Model
+
+## Command Log
+
+| #   | Command                                                                             | Purpose                               | Result                                                            | Useful?   |
+| --- | ----------------------------------------------------------------------------------- | ------------------------------------- | ----------------------------------------------------------------- | --------- |
+| 1   | `mdcontext --help`                                                                  | Understand available commands         | Full command list shown                                           | Yes       |
+| 2   | `mdcontext tree`                                                                    | List all markdown files               | 23 files found                                                    | Yes       |
+| 3   | `mdcontext search "human control"`                                                  | Find human control content            | 10 results across multiple files                                  | Yes       |
+| 4   | `mdcontext search "control plane"`                                                  | Find control plane architecture       | 10 results showing core authority layer                           | Yes       |
+| 5   | `mdcontext search "preemptive authority"`                                           | Find preemptive authority details     | 5 results with key definitions                                    | Yes       |
+| 6   | `mdcontext search "interrupt"`                                                      | Find interruptibility patterns        | 10 results showing guaranteed interrupts                          | Yes       |
+| 7   | `mdcontext tree docs/01-ARCHITECTURE.md`                                            | Show document outline                 | Full section hierarchy (5037 tokens)                              | Yes       |
+| 8   | `mdcontext context docs/01-ARCHITECTURE.md --section "3. Control Plane"`            | Get Control Plane section             | Full section with responsibilities and authority model            | Excellent |
+| 9   | `mdcontext tree docs.amorphic/04-THE_HUMAN-AGENT_COLLABORATION_MODEL.md`            | Show collaboration model outline      | Two-part structure revealed                                       | Yes       |
+| 10  | `mdcontext context ... --section "The Authority Gradient"`                          | Get authority gradient details        | Complete section on dynamic control spectrum                      | Excellent |
+| 11  | `mdcontext context ... --section "Intervention Points"`                             | Get intervention details              | Redirect, Override, Inject, Escalate actions                      | Excellent |
+| 12  | `mdcontext search "checkpoint"`                                                     | Find checkpoint patterns              | 10 results showing deliberate pause points                        | Yes       |
+| 13  | `mdcontext context docs/06-WORKFLOWS.md --section "Checkpoints"`                    | Get checkpoint details                | Comprehensive checkpoint types and why they matter                | Excellent |
+| 14  | `mdcontext search "pause"`                                                          | Find pause control operations         | 10 results showing control actions                                | Yes       |
+| 15  | `mdcontext context docs/02-PRIMITIVES.md --section "Control Actions"`               | Get control action primitives         | YAML schema for pause, resume, cancel, etc.                       | Excellent |
+| 16  | `mdcontext search "human decision"`                                                 | Find human decision points            | 8 results on explicit decision points                             | Yes       |
+| 17  | `mdcontext context docs/00-README.md --section "Core Principles"`                   | Get core principles                   | Human Control as explicit commitment                              | Excellent |
+| 18  | `mdcontext context ... --section "Spectrum of Collaboration Modes"`                 | Get collaboration modes               | Four modes: Instructional, Consultative, Supervisory, Exploratory | Excellent |
+| 19  | `mdcontext search "governance"`                                                     | Find governance patterns              | 10 results showing accountability structures                      | Yes       |
+| 20  | `mdcontext search "trust"`                                                          | Find trust-related concepts           | 10 results showing designed trust model                           | Yes       |
+| 21  | `mdcontext search "autonomous"`                                                     | Find autonomous vs controlled context | 10 results contrasting with human control                         | Yes       |
+| 22  | `mdcontext context docs/01-ARCHITECTURE.md --section "Control Flow"`                | Get control flow details              | Three flows: Initiate, Interrupt, Resume                          | Excellent |
+| 23  | `mdcontext search "override"`                                                       | Find override capabilities            | 10 results showing override principle                             | Yes       |
+| 24  | `mdcontext context docs/06-WORKFLOWS.md --section "The Override Principle"`         | Get override details                  | Every suggestion overridable, logged, not penalized               | Excellent |
+| 25  | `mdcontext search "directive"`                                                      | Find directive patterns               | 4 results on control plane directives                             | Yes       |
+| 26  | `mdcontext context docs/01-ARCHITECTURE.md --section "Architectural Invariants"`    | Get invariants                        | "No loss of human authority" is an invariant                      | Excellent |
+| 27  | `mdcontext search "escalate OR escalation"`                                         | Find escalation patterns              | 10 results on escalation mechanisms                               | Yes       |
+| 28  | `mdcontext context ... --section "Event-Driven Transparency"`                       | Get transparency details              | Every action becomes observable event                             | Excellent |
+| 29  | `mdcontext context docs/01-ARCHITECTURE.md --section "Why This Architecture Works"` | Get architecture rationale            | "Deep autonomy without loss of control"                           | Excellent |
+| 30  | `mdcontext context docs/00-README.md --section "Key Differentiators"`               | Get differentiators table             | Control: Explicit & global vs Implicit                            | Excellent |
+
+## Findings
+
+### Key Discoveries
+
+1. **Control Plane as Authority Layer**: The Control Plane is the system's central authority mechanism. It has "preemptive authority" over all execution - no agent, workflow, or job may refuse a control plane directive. This is the architectural guarantee of human control.
+
+2. **Authority Gradient Concept**: HumanWork implements a dynamic "authority gradient" - humans can move anywhere along a spectrum from instructional (step-by-step) to supervisory (broad autonomy with monitoring) at any moment, even mid-execution.
+
+3. **Checkpoints as Core Control Mechanism**: Checkpoints are explicitly called "the most important concept for human-first workflows." They provide deliberate pauses where humans are re-engaged, direction is confirmed, and escalation is considered.
+
+4. **Four Collaboration Modes**: The system defines Instructional, Consultative, Supervisory, and Exploratory modes - each with different balances of human involvement and agent autonomy.
+
+5. **Guaranteed Interruptibility**: Unlike "best-effort" interrupts in typical agent frameworks, HumanWork guarantees interruptibility as a core differentiator. This is achieved through immutable event records and control plane architecture.
+
+6. **Override Principle**: Every workflow suggestion can be overridden by humans - including skipping phases, ignoring activities, bypassing checkpoints (with acknowledgment), or detaching workflows entirely. Overrides are logged but not penalized.
+
+7. **Architectural Invariant**: "No loss of human authority" is explicitly listed as a system-wide invariant that guides all implementation decisions.
+
+8. **Control Actions Primitive**: The system defines a formal schema for control actions: pause, resume, cancel, reassign, modify_metadata, inject_step, fork, terminate.
+
+### Relevant Quotes/Sections Found
+
+> "The Control Plane has **preemptive authority** over all execution. No agent, workflow, or job may refuse a control plane directive. This guarantees human control at all times."
+> Source: docs/01-ARCHITECTURE.md, Authority Model section
+
+> "Humans retain authority over execution at all times. Any step can be paused, redirected, modified, or terminated immediately."
+> Source: docs/00-README.md, Core Principles
+
+> "Critically, humans can move anywhere along this gradient at any moment, even mid-execution."
+> Source: docs.amorphic/04-THE_HUMAN-AGENT_COLLABORATION_MODEL.md, The Authority Gradient
+
+> "Checkpoints are where: Human control is enforced, Cost is consciously acknowledged, Autonomy is bounded, Trust is built. Most agent systems lack this concept entirely."
+> Source: docs/06-WORKFLOWS.md, Checkpoints (Guardrails)
+
+> "Every workflow suggestion can be overridden by humans... Overrides are: Logged, Explained, Not penalized"
+> Source: docs/06-WORKFLOWS.md, The Override Principle
+
+> "No loss of human authority" [listed as architectural invariant]
+> Source: docs/01-ARCHITECTURE.md, Architectural Invariants
+
+> "The system doesn't break - it adapts. The workflow doesn't restart - it continues from the intervention point with new constraints."
+> Source: docs.amorphic/04-THE_HUMAN-AGENT_COLLABORATION_MODEL.md, Intervention Points
+
+### Summary of Theme
+
+The Human-Agent Control Model in HumanWork is fundamentally different from typical autonomous agent frameworks. Rather than treating human control as a constraint or afterthought, it is architected as the primary design principle.
+
+**Core Philosophy**: "The system favors clarity, safety, and trust over raw autonomy."
+
+**Key Architectural Elements**:
+
+1. **Control Plane** - An always-active authority layer that governs (but doesn't participate in) execution
+2. **Checkpoints** - Deliberate pause points for human re-engagement built into workflow design
+3. **Authority Gradient** - Dynamic spectrum allowing humans to adjust autonomy level at any moment
+4. **Intervention Points** - Every event boundary is an opportunity for human redirect, override, inject, or escalate
+
+**Guarantees**:
+
+- Preemptive authority (agents cannot refuse directives)
+- Guaranteed interruptibility (not best-effort)
+- All overrides logged but not penalized
+- No loss of human authority (architectural invariant)
+
+**Differentiation from Typical Frameworks**:
+
+| Aspect           | Typical     | HumanWork         |
+| ---------------- | ----------- | ----------------- |
+| Control          | Implicit    | Explicit & global |
+| Interruptibility | Best-effort | Guaranteed        |
+| Trust            | Assumed     | Designed          |
+
+The model enables "deep autonomy without loss of control" - agents can operate with significant independence while humans retain the ability to intervene, redirect, or terminate at any moment. This is achieved through immutable event recording, the Control Plane architecture, and systematic checkpoint placement.
+
+## Proposed Spec Changes
+
+- [ ] Add explicit section on "Human Authority Guarantees" listing all control mechanisms
+- [ ] Document the Control Plane API/interface more concretely (how do humans issue directives?)
+- [ ] Add examples of authority gradient transitions (moving from Supervisory to Instructional mid-task)
+- [ ] Clarify checkpoint enforcement mechanisms (how is "required" vs "optional" enforced technically?)
+- [ ] Add comparison table showing HumanWork control model vs other frameworks (LangGraph, AutoGPT, etc.)
+
+## Tool Evaluation
+
+### What Worked Well
+
+- `mdcontext search` with keyword queries was highly effective for finding relevant content
+- `mdcontext context --section` was excellent for extracting focused, complete sections
+- `mdcontext tree <file>` provided perfect document outlines for navigation
+- Boolean operators (`OR`) worked as expected for combining search terms
+- Search results showed line numbers and context, making it easy to understand matches
+- The workflow of search -> tree -> context was very natural for deep-diving
+
+### What Was Frustrating
+
+- No way to search within a specific file (had to search globally then filter mentally)
+- Could not request multiple sections in one `context` command
+- No embeddings indexed, so semantic search unavailable (would have helped with concept exploration)
+- Search result limit of 10 sometimes truncated useful results (no way to paginate or increase)
+- Phrase search with multiple words ("pause resume terminate") returned 0 results - had to search individually
+
+### Confidence Level
+
+[X] High - I found comprehensive documentation on the human-agent control model across multiple files and was able to extract detailed information on the Control Plane, checkpoints, authority gradient, and intervention mechanisms.
+
+### Would Use Again? (1-5)
+
+**4** - Effective for focused documentation research. The search-tree-context workflow is well-designed for finding and extracting specific information. Would be a 5 if embeddings were available for semantic search and if there were better multi-section extraction options.

package/research/dogfood/strategy-c/c4-diver-failure.md

@@ -0,0 +1,151 @@
+# Report: C4 - Diver: Failure Analysis
+
+## Mission
+
+Deep-dive into: Failure of Pure Automation - Understanding why current agent frameworks fail
+
+## Command Log
+
+| #   | Command                                                                                       | Purpose                          | Result                               | Useful? |
+| --- | --------------------------------------------------------------------------------------------- | -------------------------------- | ------------------------------------ | ------- |
+| 1   | `mdcontext search "failure automation"`                                                       | Find failure content             | 0 results (multi-word keyword issue) | No      |
+| 2   | `mdcontext search "brittleness"`                                                              | Find brittleness content         | 6 results across docs                | Yes     |
+| 3   | `mdcontext search "coordination trap"`                                                        | Find coordination trap           | 2 results                            | Yes     |
+| 4   | `mdcontext context --help`                                                                    | Learn context command syntax     | Showed section filtering options     | Yes     |
+| 5   | `mdcontext context --sections docs.amorphic/02-THE_FAILURE_OF_PURE_AUTOMATION.md`             | List all sections                | Found 6 failure modes                | Yes     |
+| 6   | `mdcontext context --full --section "1.1" docs.amorphic/02-THE_FAILURE_OF_PURE_AUTOMATION.md` | Get Brittleness section          | Full content retrieved               | Yes     |
+| 7   | `mdcontext context --full --section "1.2" docs.amorphic/02-THE_FAILURE_OF_PURE_AUTOMATION.md` | Get Coordination Trap            | Full content retrieved               | Yes     |
+| 8   | `mdcontext context --full --section "1.3" docs.amorphic/02-THE_FAILURE_OF_PURE_AUTOMATION.md` | Get Innovation Strangulation     | Full content retrieved               | Yes     |
+| 9   | `mdcontext context --full --section "1.4" docs.amorphic/02-THE_FAILURE_OF_PURE_AUTOMATION.md` | Get Human Bottleneck Paradox     | Full content retrieved               | Yes     |
+| 10  | `mdcontext context --full --section "1.5" docs.amorphic/02-THE_FAILURE_OF_PURE_AUTOMATION.md` | Get Context Collapse             | Full content retrieved               | Yes     |
+| 11  | `mdcontext context --full --section "1.6" docs.amorphic/02-THE_FAILURE_OF_PURE_AUTOMATION.md` | Get Judgment Gap                 | Full content retrieved               | Yes     |
+| 12  | `mdcontext search "autonomous agent"`                                                         | Find autonomous agent content    | 3 results                            | Yes     |
+| 13  | `mdcontext search "judgment"`                                                                 | Find judgment content            | 10 results showing solution approach | Yes     |
+| 14  | `mdcontext context --full --section "1" docs.amorphic/02-THE_FAILURE_OF_PURE_AUTOMATION.md`   | Get full document                | Complete failure analysis            | Yes     |
+| 15  | `mdcontext search "human-agent"`                                                              | Find collaboration model         | 10 results                           | Yes     |
+| 16  | `mdcontext search "symbiosis"`                                                                | Find symbiosis concept           | 2 results                            | Yes     |
+| 17  | `mdcontext context --sections docs.amorphic/01-EXECUTIVE_SUMMARY.md`                          | List exec summary sections       | 4 sections found                     | Yes     |
+| 18  | `mdcontext context --full --section "1.1" docs.amorphic/01-EXECUTIVE_SUMMARY.md`              | Get Core Innovation              | Solution overview                    | Yes     |
+| 19  | `mdcontext search "enterprise adoption"`                                                      | Find enterprise context          | 1 result in feedback.md              | Yes     |
+| 20  | `mdcontext context --full --section "The Core Thesis" docs.llm/feedback.md`                   | Get enterprise adoption analysis | Core thesis on opacity/risk          | Yes     |
+
+## Findings
+
+### Key Discoveries
+
+- The docs.amorphic/ folder contains a rigorous philosophical analysis of WHY pure automation fails in knowledge work
+- Six distinct failure modes are identified, each with concrete examples
+- The analysis establishes the intellectual foundation for why HumanWork's approach is necessary
+- Enterprise adoption has stalled due to two core issues: Opacity and Risk
+
+### Failure Modes Identified
+
+#### 1. The Brittleness of Complete Systems
+
+- Pure automation assumes complete knowledge of problem space
+- Works for manufacturing/transactions, fails for knowledge work
+- Combinatorial explosion of rules creates rigid systems
+- **Key insight**: "The system becomes brittle not because any individual rule is wrong, but because the combinatorial explosion of rules creates a rigid lattice that cannot bend without breaking"
+
+#### 2. The Coordination Trap
+
+- Pure automation promises to eliminate coordination overhead
+- Actually MULTIPLIES coordination by forcing human work into machine-readable formats
+- Rich contextual communication gets decomposed into discrete units
+- **Key insight**: "The tools become the work, rather than enablers of work"
+
+#### 3. The Innovation Strangulation
+
+- Pure automation actively inhibits exploratory/creative work
+- Cannot accommodate "work that hasn't been work before"
+- Creates perverse incentive: teams avoid innovative approaches because they're automation-incompatible
+- **Key insight**: "The system optimizes for processable work rather than valuable work"
+
+#### 4. The Human Bottleneck Paradox
+
+- Pure automation tries to route around humans as bottlenecks
+- In knowledge work, humans are often the HIGHEST-bandwidth component
+- Creates new bottlenecks: system configuration, exception handling, cross-system integration
+- **Key insight**: "The very attempts to systematize human judgment create new categories of problems that require... human judgment to resolve"
+
+#### 5. The Context Collapse
+
+- Pure automation treats context as configuration rather than conversation
+- Context is fluid - shifts with circumstance, evolves with understanding
+- Automation has no mechanism for negotiation, no capacity for doubt
+- **Key insight**: "When automation encounters the unexpected, it has nowhere to turn. The rigid pathways that make it fast and reliable become brittle barriers."
+
+#### 6. The Judgment Gap
+
+- Judgment is the ability to recognize when existing rules don't apply
+- Work increasingly happens in spaces between the known - edge cases, exceptions
+- Even sophisticated AI cannot make situated judgments humans make instinctively
+- **Key insight**: "Systems that handle 80% of cases flawlessly but create chaos in the remaining 20%. And often, it's precisely those exceptional 20% of cases that matter most."
+
+### Relevant Quotes/Sections Found
+
+> "The allure of pure automation is irresistible to the technical mind. The promise is elegant: replace human unpredictability with algorithmic precision, eliminate coordination overhead through perfect orchestration, achieve scale through mechanical repetition. Yet in the domain of knowledge work... pure automation consistently fails not because it's implemented poorly, but because it's conceptually mismatched to the substrate it attempts to govern."
+> Source: docs.amorphic/02-THE_FAILURE_OF_PURE_AUTOMATION.md, Introduction
+
+> "Enterprise adoption of autonomous agents has stalled due to **Opacity** (we don't know how the agent works) and **Risk** (we can't trust it to run unsupervised)."
+> Source: docs.llm/feedback.md, The Core Thesis
+
+> "Human-AI Symbiosis: This isn't about making AI more human or humans more machine-like. Instead, Amorphic creates genuine cognitive partnerships where each intelligence type contributes its strengths. Humans bring intuition, creativity, and contextual judgment. AI provides systematic analysis, pattern recognition, and execution consistency. The magic happens in their interaction."
+> Source: docs.amorphic/01-EXECUTIVE_SUMMARY.md, The Core Innovation
+
+> "The fundamental question in human-agent collaboration isn't whether humans or agents are better at specific tasks—it's how to design interfaces that amplify human judgment while leveraging computational power."
+> Source: docs.amorphic/04-THE_HUMAN-AGENT_COLLABORATION_MODEL.md
+
+### Summary of Theme
+
+The Failure of Pure Automation analysis provides the philosophical and practical foundation for why HumanWork's approach is necessary. The key critique is that pure automation is **conceptually mismatched** to knowledge work for six fundamental reasons:
+
+1. Knowledge work resists enumeration - too many states/edge cases
+2. Coordination overhead transforms rather than disappears
+3. Innovation requires spaces automation cannot accommodate
+4. Humans are bandwidth, not bottlenecks in knowledge work
+5. Context is conversational, not configurable
+6. Judgment emerges from interaction, not rules
+
+The solution is not better automation, but **Human-AI Symbiosis** - cognitive partnerships where human judgment guides computational execution while maintaining granular control. This directly addresses:
+
+- **Brittleness** -> Fluid human intervention points
+- **Coordination Trap** -> Rich contextual communication preserved
+- **Innovation** -> Emergent patterns supported, not suppressed
+- **Human Bottleneck** -> Humans as high-bandwidth collaborators
+- **Context Collapse** -> Context as conversation, evolving
+- **Judgment Gap** -> Human judgment at decision points
+
+## Proposed Spec Changes
+
+- [ ] Add a "Failure Modes Addressed" section in the spec that maps HumanWork features to specific failure modes
+- [ ] Include the 80/20 insight: emphasize that HumanWork shines on the exceptional 20% of cases
+- [ ] Document the "tools become the work" anti-pattern as a key problem statement
+- [ ] Add "Opacity and Risk" as top-level enterprise adoption barriers to address
+- [ ] Include concrete examples of each failure mode with corresponding HumanWork solutions
+
+## Tool Evaluation
+
+### What Worked Well
+
+- `mdcontext context --sections` is excellent for discovering document structure
+- `mdcontext context --full --section` retrieved complete sections cleanly
+- Section numbering system (1.1, 1.2, etc.) is intuitive
+- Search results show context lines around matches
+- Token counts on sections help understand document scope
+
+### What Was Frustrating
+
+- First `context` attempt failed due to wrong syntax (tried passing section name as second argument)
+- Multi-word searches like "failure automation" returned 0 results even though both words appear
+- No embeddings available, so semantic search not possible
+- Help text was needed to understand context command properly
+- Some duplicate content in docs.llm/ mirrors docs.amorphic/ which can be confusing
+
+### Confidence Level
+
+[X] High - The failure analysis content was comprehensive and well-structured. All six failure modes were retrieved in full with excellent context.
+
+### Would Use Again? (1-5)
+
+**4** - Very effective for structured markdown exploration. The section filtering via `--section` with number syntax is powerful once you know it exists. Lost one point for initial learning curve with context command syntax and multi-word search limitations. Would be 5/5 with embeddings enabled.