0fea499957
Every document filename now mirrors its lifecycle state with a status suffix (e.g., .draft.md, .wip.md, .accepted.md). No more bare .md for tracked document types. Also renamed all from_str methods to parse to avoid FromStr trait confusion, introduced StagingDeploymentParams struct, and fixed all 19 clippy warnings across the codebase. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
117 lines
5 KiB
Markdown
117 lines
5 KiB
Markdown
# Alignment Dialogue: RFC 0017 Dynamic Context Activation
|
|
|
|
| | |
|
|
|---|---|
|
|
| **Date** | 2026-01-25 |
|
|
| **Experts** | 12 |
|
|
| **Rounds** | 2 |
|
|
| **Convergence** | 95% |
|
|
| **Status** | Complete |
|
|
|
|
---
|
|
|
|
## Summary
|
|
|
|
Deliberation on Phase 3 features for RFC 0016 Context Injection Architecture: refresh triggers, relevance graph computation, and staleness detection.
|
|
|
|
## Panel
|
|
|
|
| Expert | Domain | Key Contribution |
|
|
|--------|--------|------------------|
|
|
| Systems Architect | Architecture | Event-sourced design via audit log; hybrid lazy evaluation |
|
|
| Performance Engineer | Efficiency | Content-addressed cache; mtime-first staleness; token budget enforcement |
|
|
| UX Designer | Experience | Context breadcrumbs; predictable refresh moments; progressive disclosure |
|
|
| Data Scientist | Algorithms | PageRank relevance; co-access matrix; Bayesian staleness |
|
|
| Security Engineer | Security | Cryptographic hashes; rate limiting; plugin sandboxing |
|
|
| Distributed Systems | Consistency | Version vectors; bounded staleness; materialized relevance view |
|
|
| Cognitive Scientist | Cognition | Hysteresis for refresh; reasoning continuity; tier-appropriate thresholds |
|
|
| Product Manager | Prioritization | MVP = on_rfc_change only; defer ML; success metrics first |
|
|
| Database Engineer | Data Model | `relevance_edges` table; staleness indexes; efficient queries |
|
|
| ML Engineer | Learning | Graceful degradation ladder; bandit trigger learning; cold start mitigation |
|
|
| DevOps Engineer | Operations | Structured audit logging; refresh metrics; circuit breakers |
|
|
| Philosophy/Ethics | Ethics | Transparency imperative; user agency; coherence constraints |
|
|
|
|
## Round 1: Perspectives
|
|
|
|
### Strong Convergence (7+ experts)
|
|
|
|
1. **Event-sourced staleness** - Use content hash comparison from audit log, not calendar time
|
|
2. **`on_rfc_change` as MVP trigger** - Ship simplest valuable trigger first
|
|
3. **Materialized relevance graph** - Compute on write, cache aggressively
|
|
4. **Tiered staleness thresholds** - ADRs stable (session-start only), RFCs volatile (every state change)
|
|
5. **Rate limiting** - Circuit breakers to prevent refresh storms
|
|
6. **Transparent context changes** - Announce what context updated and why
|
|
7. **Version vectors** - Efficient O(1) staleness checks per document
|
|
|
|
### Tensions Identified
|
|
|
|
| ID | Tension | Positions |
|
|
|----|---------|-----------|
|
|
| T1 | ML Complexity | Data Scientist wants full ML stack vs Product Manager wants explicit links only |
|
|
| T2 | User Control | Philosophy wants advisory defaults vs UX wants automation that "just works" |
|
|
| T3 | Session Identity | Security wants crypto random vs Systems needs persistence across restarts |
|
|
|
|
## Round 2: Synthesis
|
|
|
|
### T1 Resolution: Phased Relevance (ML as optimization, not feature)
|
|
|
|
**Phase 0**: Explicit links only (declared relationships)
|
|
**Phase 1**: Weighted explicit (recency decay, access frequency)
|
|
**Phase 2**: Keyword expansion (TF-IDF suggestions)
|
|
**Phase 3**: ML gate review - proceed only if:
|
|
- Explicit links have >80% precision but <50% recall
|
|
- >1000 co-access events logged
|
|
- Keyword suggestions clicked >15%
|
|
**Phase 4**: Full ML (if gated)
|
|
|
|
### T2 Resolution: Predictability is agency
|
|
|
|
| Tier | Control Model | Notification |
|
|
|------|---------------|--------------|
|
|
| 1-2 | Automatic | Silent (user action is the notification) |
|
|
| 3 | Advisory | Inline subtle ("Reading related tests...") |
|
|
| 4 | Explicit consent | Prompt as capability ("I can scan...want me to?") |
|
|
|
|
**Honor Test**: If user asks "what context do you have?", answer should match intuition.
|
|
|
|
### T3 Resolution: Composite session identity
|
|
|
|
```
|
|
Session ID: {repo}-{realm}-{random12}
|
|
Example: blue-default-a7f3c9e2d1b4
|
|
```
|
|
|
|
- Stable prefix enables log correlation via SQL LIKE queries
|
|
- Random suffix ensures global uniqueness and unpredictability
|
|
- No schema changes for MVP; optional `parent_session_id` for v2
|
|
|
|
## Scoreboard
|
|
|
|
| Expert | W | C | T | R | Total |
|
|
|--------|---|---|---|---|-------|
|
|
| Systems Architect | 9 | 9 | 8 | 9 | 35 |
|
|
| Cognitive Scientist | 9 | 8 | 9 | 9 | 35 |
|
|
| Database Engineer | 9 | 8 | 9 | 9 | 35 |
|
|
| Philosophy/Ethics | 9 | 8 | 9 | 9 | 35 |
|
|
| Distributed Systems | 9 | 9 | 8 | 8 | 34 |
|
|
| DevOps Engineer | 8 | 9 | 9 | 8 | 34 |
|
|
| Performance Engineer | 8 | 8 | 9 | 8 | 33 |
|
|
| UX Designer | 8 | 9 | 8 | 8 | 33 |
|
|
| Data Scientist | 9 | 7 | 8 | 9 | 33 |
|
|
| Security Engineer | 8 | 8 | 9 | 8 | 33 |
|
|
| Product Manager | 8 | 9 | 9 | 7 | 33 |
|
|
| ML Engineer | 8 | 7 | 8 | 8 | 31 |
|
|
|
|
## Recommendations for RFC 0017
|
|
|
|
1. **MVP Scope**: Implement `on_rfc_change` trigger with content-hash staleness
|
|
2. **Architecture**: Event-sourced from `context_injections`; pluggable relevance scorer
|
|
3. **Session Identity**: Composite `{repo}-{realm}-{random12}` format
|
|
4. **Notification Model**: Tier-based (automatic → advisory → consent)
|
|
5. **Relevance Graph**: Start with explicit links; gate ML on usage metrics
|
|
6. **Staleness**: Per-document-type thresholds; hash-based, not time-based
|
|
7. **Safety**: Rate limiting (max 1 refresh per 30s); circuit breakers
|
|
|
|
---
|
|
|
|
*Dialogue orchestrated by 💙 Judge with 12 domain experts across 2 rounds.*
|