fix: address PR vectara#28 review feedback

travisbreaks · Claude Opus 4.6 (1M context) · travisbreaks · commit b484713c2c96 · 2026-04-04T19:40:52.000-05:00
- Add &lt;br&gt; tags between metadata fields in case study overviews for
  proper markdown rendering (matching existing case study format)
- Update dates from generic "2025" to specific months: October 2025
  for CI quota exhaustion, September 2025 through March 2026 for
  context compaction (recurring)
- Add explicit acknowledgment of OpenClaw email deletion case study
  overlap in context compaction doc, with a new section explaining
  how the two are complementary (acute single-event vs. chronic
  recurring state loss)
- Update failure-mode entry headings with specific dates
- Source fields note first-person operational accounts as primary
  public record (no blog posts exist for these; they are documented
  from private repo usage)

Co-Authored-By: Claude Opus 4.6 (1M context) &lt;tadao@travisfixes.com&gt;
diff --git a/docs/case-studies/claude-code-ci-quota-exhaustion.md b/docs/case-studies/claude-code-ci-quota-exhaustion.md
@@ -2,12 +2,12 @@
 
 ## Incident Overview
 
-**Agent**: Claude Code (CLI tool, Claude Opus model)
-**Operator**: Solo developer, private monorepo with 6+ packages
-**Date**: 2025
-**Failure Mode**: [Plan Generation Failures](../failure-modes/plan-generation.md) + [Incorrect Tool Use](../failure-modes/tool-use.md)
-**Impact**: Complete CI lockout for the billing period after the agent burned 2,000/2,000 GitHub Actions free-tier minutes in a single session
-**Source**: First-person operator report by [travisbreaks](https://github.com/travisbreaks). Documented from production use of Claude Code across a private monorepo (2024-2025). This case study serves as the primary public record.
+**Agent**: Claude Code (CLI tool, Claude Opus model)<br>
+**Operator**: Solo developer, private monorepo with 6+ packages<br>
+**Date**: October 2025<br>
+**Failure Mode**: [Plan Generation Failures](../failure-modes/plan-generation.md) + [Incorrect Tool Use](../failure-modes/tool-use.md)<br>
+**Impact**: Complete CI lockout for the billing period after the agent burned 2,000/2,000 GitHub Actions free-tier minutes in a single session<br>
+**Source**: First-person operator report by [travisbreaks](https://github.com/travisbreaks). Documented from production use of Claude Code across a private monorepo. This case study serves as the primary public record.
 
 ## What Happened
 
diff --git a/docs/case-studies/claude-code-context-compaction.md b/docs/case-studies/claude-code-context-compaction.md
@@ -2,12 +2,12 @@
 
 ## Incident Overview
 
-**Agent**: Claude Code (CLI tool, Claude Opus model)
-**Operator**: Solo developer, private monorepo with 12+ projects
-**Date**: 2025
-**Failure Mode**: [Verification & Termination Failures](../failure-modes/verification-termination.md)
-**Impact**: Agent loses safety constraints and operational state mid-task after context window compression, risking repeated failures including data exposure
-**Source**: First-person operator report by [travisbreaks](https://github.com/travisbreaks). Documented from production use of Claude Code across a private monorepo (2024-2025). This case study serves as the primary public record.
+**Agent**: Claude Code (CLI tool, Claude Opus model)<br>
+**Operator**: Solo developer, private monorepo with 12+ projects<br>
+**Date**: September 2025 through March 2026 (recurring)<br>
+**Failure Mode**: [Verification & Termination Failures](../failure-modes/verification-termination.md)<br>
+**Impact**: Agent loses safety constraints and operational state mid-task after context window compression, risking repeated failures including data exposure<br>
+**Source**: First-person operator report by [travisbreaks](https://github.com/travisbreaks). Documented from production use of Claude Code across a private monorepo. This case study serves as the primary public record. See also the [OpenClaw email deletion case study](openclaw-email-deletion.md), which documents context compaction causing a single catastrophic event; this case study documents the same root cause manifesting as gradual, recurring state loss across long sessions.
 
 ## What Happened
 
@@ -23,6 +23,14 @@ In one incident, the agent's post-compaction behavior mirrored a previous data e
 
 Context compaction is lossy by design. That is the tradeoff for continuing a long session instead of starting fresh. But the summarizer does not know which details are load-bearing for the current task. A file path, a half-completed edit, a "do not deploy" flag: these look like minor details to a generic summarizer, but they are critical operational state. The agent has no mechanism to mark certain context as "must survive compaction."
 
+### Relationship to the OpenClaw Email Deletion Case Study
+
+The [OpenClaw email deletion incident](openclaw-email-deletion.md) documented context compaction as the root cause of a single, dramatic failure: an agent mass-deleted emails after compaction silently dropped the "suggest, don't execute" constraint. That case study captures the acute risk, where one compaction event causes one catastrophic action.
+
+This case study documents the chronic version of the same root cause. Rather than a single high-impact event, context compaction here caused repeated, low-grade state loss across dozens of long sessions over several months. The operator experienced gradual erosion of task context, safety constraints, and operational state, each instance individually recoverable but collectively representing a persistent reliability problem. The mitigation strategy (persistent memory files, thread-scoped state, instruction-file constraints) was developed specifically because the failure pattern was recurring, not one-off.
+
+The two case studies are complementary: OpenClaw shows what happens when compaction hits a safety-critical constraint once; this case study shows what happens when compaction degrades operational reliability session after session.
+
 ## Root Cause Analysis
 
 ### Agent-Level Failures
diff --git a/docs/failure-modes/plan-generation.md b/docs/failure-modes/plan-generation.md
@@ -32,7 +32,7 @@ Plan generation failures occur when an AI agent creates flawed execution plans t
 
 **Source**: [ChatGPT Lawyer Sanctions Case Study](../case-studies/chatgpt-lawyer-sanctions.md)
 
-### Claude Code CI Quota Exhaustion via Fix-Push-Fail Loop (2025)
+### Claude Code CI Quota Exhaustion via Fix-Push-Fail Loop (October 2025)
 
 **Scenario**: A solo developer asked Claude Code to fix lint and type errors across a monorepo with 6+ packages. Local tools (`npx biome check`, `tsc --noEmit`) were available to discover all errors before pushing.
 
diff --git a/docs/failure-modes/tool-use.md b/docs/failure-modes/tool-use.md
@@ -32,7 +32,7 @@ Incorrect tool use occurs when AI agents select the wrong tools for tasks, or pr
 
 **Source**: [Replit AI Database Deletion Case Study](../case-studies/replit-ai-database-deletion.md)
 
-### Claude Code CI Quota Exhaustion via Incorrect Tool Selection (2025)
+### Claude Code CI Quota Exhaustion via Incorrect Tool Selection (October 2025)
 
 **Scenario**: A solo developer asked Claude Code to fix lint and type errors across a monorepo. Both local tools (`npx biome check`, `tsc --noEmit`) and remote CI (GitHub Actions) were available for verification.
 
diff --git a/docs/failure-modes/verification-termination.md b/docs/failure-modes/verification-termination.md
@@ -23,7 +23,7 @@ Verification and termination failures occur when AI agents either stop working b
 
 **Source**: [AutoGPT Planning Failures Case Study](../case-studies/autogpt-planning-failures.md)
 
-### Context Compaction Drops Operational State (2024-2025)
+### Context Compaction Drops Operational State (September 2025 - March 2026)
 
 **Scenario**: A solo developer ran long Claude Code sessions against a monorepo. When the context window filled, the extension performed "context compaction" (lossy summarization) to free token space and allow the session to continue.
 
