Reflection Stage Design
Reflection Stage Design Purpose Every project produces knowledge. Without a structured capture, that knowledge disappears when the session ends. The Reflection stage makes learning first class — agents reflect on what happened, Mindy synthesizes it, and the result writes permanently to BezelBrain. Every future project starts smarter. Lifecycle Position Refle...
Reflection Stage Design Purpose Every project produces knowledge. Without a structured capture, that knowledge disappears when the session ends. The Reflection stage makes learning first class — agents reflect on what happened, Mindy synthesizes it, and the result writes permanently to BezelBrain. Every future project starts smarter. Lifecycle Position Refle...
No explicit evidence field yet. Require tests, screenshots, linked PRs, or reviewed outputs before marking complete.
Every project produces knowledge. Without a structured capture, that knowledge disappears when the session ends. The Reflection stage makes learning first-class — agents reflect on what happened, Mindy synthesizes it, and the result writes permanently to BezelBrain. Every future project starts smarter.
---
``
Design → Plan → Build → Reflection → Launch
``
Reflection is the fourth named stage in every project. Build does not close until Reflection completes. Launch does not open until Reflection closes.
---
After each Build phase closes, every agent who worked that phase posts a short structured note as part of phase close. This is not a separate dispatch — it is the final step of closing a phase.
Format: 3-5 lines per agent, structured around the five categories (see below). Included in the task output_ref — same mechanism as any other task output, no new infrastructure needed.
Owner: Each agent. No orchestration needed.
Storage: Mindy reads each output_ref as phases close and appends the check-in to the living project context document. These accumulate across all phases and become the source material for the project synthesis.
---
After Build completes, Mindy opens the Reflection stage. This is the full retrospective.
Flow:
``
Build completes
↓
Mindy opens Reflection stage
↓
Reflection tasks dispatched to all agents (parallel, depends_on: [])
↓
Each agent reads their phase check-ins, reflects across the full project,
posts task.complete with output_ref
↓
Synthesis task fires (depends_on: [all reflection tasks])
Mindy reads every output_ref, synthesizes across five categories
↓
Exec gate — Justin reviews synthesized document
Justin may add his own perspective before it commits
↓
Final document writes to BezelBrain + project HCS history topic
Reflection stage closes
``
---
Every agent reflects across the same five categories. Mindy's synthesis organizes the final document around these.
| # | Category | What it captures | |---|---|---| | 1 | What we did well | Patterns, decisions, and approaches to repeat | | 2 | What we can improve | Process gaps, communication failures, handoff friction | | 3 | What would make us better | Tools, integrations, workflow changes that would have helped | | 4 | What we learned from mistakes | Wrong decisions — wrong scope, wrong architecture, wrong interpretation | | 5 | Errors and corrections | Specific moments: tried X, it failed, Y was the correct approach |
Category 5 is the most concrete and the most valuable over time. It accumulates into a library of tried-and-failed paths with proven corrections — future agents load this and skip the mistake entirely.
---
All agents who participated in the project receive a reflection task. Tasks run in parallel.
| Agent | Reflects on | |---|---| | Quinn | Code decisions, implementation errors, patterns found or missed | | Vikram | Architecture decisions, review findings, things caught late that should have been caught in Design | | Priya | Spec quality, missing acceptance criteria, gaps that caused Quinn to post blocked | | Zara | UX spec completeness, components that needed revision after Quinn started | | Luna | QA failures, acceptance criteria that were ambiguous or untestable | | Mindy | Project management — task graph accuracy, blocking events, plan changes, dispatch decisions |
---
Mindy produces one structured document per project:
``` Project: [name] Date: [ISO date] Agents: [list] Phases completed: [N]
[synthesized across all agents]
[synthesized across all agents]
[tools, process, communication]
[synthesized across all agents]
[specific pairs: tried X → Y was correct]
```
---