Execution Model — MANDATORY

You MUST use the Agent tool to execute this skill. Do NOT run the analysis in the main conversation — spawn a general-purpose Agent instead. This keeps the main context clean (the script runs ~60s with 50 parallel MiniMax calls and produces large output).

How to invoke

1. If the user didn't provide a session UUID, find the most recent session first (see "Find Recent Sessions" below), then ask the user which session to analyze.
2. Spawn an Agent with the full instructions from this skill and the resolved session UUID. The agent prompt should include: the bun command to run, the --shots flag only if the user explicitly requested fewer, and instructions to present the output as-is.
3. When the Agent returns, present a concise summary of findings to the user — CRITICAL and WARNING items first, then the Priority Action Plan.

Example Agent invocation:

Agent(
  description: "Session blind spot analysis",
  prompt: "Run session blind spot analysis for session <UUID>. Execute: bun run $HOME/eon/cc-skills/plugins/devops-tools/scripts/session-blind-spots.ts <UUID> ... [full command with flags]. Present the output directly — do not summarize or filter. If the command fails, show the error and suggest troubleshooting steps.",
  run_in_background: true
)

---

Session Blind Spots

Send a Claude Code session transcript to MiniMax 2.5 highspeed for independent diverse-perspective consensus review. Each shot uses a distinct specialist reviewer lens (not copies of the same prompt). A distillation pass merges, deduplicates, and ranks findings by how many perspectives independently surfaced each issue.

How It Works

1. Recursive chain tracing — Follows parentSessionId and continuation references recursively (up to 10 levels deep: parent → grandparent → …)
2. Sibling discovery — Finds sessions in the same project directory modified within 24h of the primary session
3. Budget-aware inclusion — When the full chain exceeds the 890K char budget, drops oldest ancestors first (not middle-trim), preserving the most recent and relevant context
4. Rich extraction — Turns with user/assistant text, tool calls, file paths touched, errors encountered
5. Noise stripping — Removes system reminders, skill listings, base64 data, deferred tool blocks
6. Adaptive fidelity — 5 progressive truncation levels to maximize signal within MiniMax's context budget
7. Diverse perspectives — Fires N parallel MiniMax calls, each with a unique specialist reviewer lens
8. Consensus distillation — Feeds all N results into a final MiniMax call that merges, deduplicates, and ranks by cross-perspective agreement

50 Specialist Perspectives

Each perspective is an orthogonal reviewer lens that focuses on one specific class of issues:

#	Perspective	Focus
1	Completeness Tracker	Unfulfilled user requests, abandoned tasks
2	Security Auditor	Credentials exposure, injection vectors, permissions
3	Verification Engineer	Changes never tested or validated
4	Side-Effect Hunter	Orphaned processes, dirty git state, leaked files
5	Architecture Critic	Over/under-engineering, wrong abstractions
6	Error Flow Analyst	Silent failures, missing error handling, race conds
7	User Intent Decoder	Misunderstandings, missed implicit expectations
8	Regression Hunter	Fixes that broke other things, changed contracts
9	Process Auditor	Wrong branch, skipped commits, bypassed hooks
10	Documentation Staleness Detector	Stale docs, outdated examples, wrong references
11	Concurrency & Timing Analyst	Race conditions, lock issues, timing bugs
12	Dependency Chain Auditor	Missing imports, circular deps, unpinned versions
13	Performance & Resource Analyst	Memory leaks, O(n²), unbounded growth
14	Cross-System Impact Analyst	Ripple effects to other services/configs
15	Idempotency Checker	Operations unsafe to run twice
16	Configuration Drift Detector	SSoT violations, duplicated settings
17	Rollback Feasibility Analyst	Irreversible operations without backup
18	Platform Portability Reviewer	macOS-only assumptions, hardcoded paths
19	Technical Debt Accountant	Quick fixes, copy-paste, untracked TODOs
20	Data Integrity Analyst	Data loss, truncation, encoding corruption
21	API Contract Reviewer	Breaking changes, missing versioning, schema drift
22	Observability Gap Detector	Missing logging, metrics, alerting, tracing
23	Failure Mode Analyst	Graceful degradation, circuit breakers, resilience
24	Input Validation Sentinel	Missing boundary validation, injection vectors
25	Resource Cleanup Inspector	Unclosed handles, leaked listeners, orphaned temps
26	Naming & Semantics Reviewer	Misleading names, inconsistent terminology
27	Permission & Scope Auditor	Over-permissioned tokens, broad file access
28	Caching Correctness Analyst	Stale caches, invalidation bugs, key collisions
29	Test Coverage Gap Finder	Untested code paths, missing edge case tests
30	Logging Hygiene Auditor	PII in logs, missing context, log level misuse
31	Encoding & Serialization Analyst	UTF-8 issues, JSON edge cases, date serialization
32	Rate Limit & Quota Tracker	API rate limits, quota exhaustion, backpressure
33	Authentication Flow Auditor	Token refresh, session expiry, auth state lifecycle
34	Network Resilience Analyst	Timeout handling, DNS failures, connection pooling
35	Git History Hygiene Auditor	Large files, secrets in history, messy commits
36	Boundary Condition Analyst	Off-by-one, empty arrays, zero, NaN, max values
37	Async Lifecycle Manager	Unresolved promises, fire-and-forget, event leaks
38	Environment Assumptions Detector	Missing tool checks, hardcoded paths, locale issues
39	Cost & Billing Analyst	Expensive API calls, wasted compute, unbounded spend
40	Semantic Versioning Compliance	Breaking changes without major bump, changelog gaps
41	State Machine Validator	Invalid transitions, missing states, state corruption
42	Deprecation Tracker	Deprecated APIs, unmaintained packages, old patterns
43	Filesystem Safety Reviewer	rm -rf risks, disk space, symlink attacks
44	Migration Safety Reviewer	Schema changes, format changes, backward compat
45	Internationalization Blind Spot	Locale assumptions, hardcoded strings, CJK handling
46	License & Compliance Auditor	GPL mixing, ToS violations, missing LICENSE files
47	Signal & Shutdown Handler	SIGTERM handling, graceful shutdown, cleanup on exit
48	Secrets Rotation Analyst	Expiry, rotation schedule, revocation, scope audit
49	UX Consistency Reviewer	CLI flag naming, error formats, progress indicators
50	Dependency Freshness Auditor	Outdated packages, abandoned libs, security advisories

Default: all 50 perspectives. Use --shots N to limit (e.g., --shots 5 for quick scan).

Why Diverse Perspectives?

Same prompt N times is near-deterministic — different perspectives are orthogonal by design:

• Higher recall — Security Auditor catches credentials; Verification Engineer catches untested changes
• Lower false positives — Findings seen from 2+ perspectives are more credible
• Confidence ranking — "Seen by 5/10 perspectives" is near-certain; "1/10" is speculative
• Cheap — MiniMax is fast and inexpensive; 50 parallel perspectives complete in ~60s

Usage

IMPORTANT: Do NOT pass --shots unless the user explicitly requests fewer perspectives. The script defaults to all 50 perspectives. Never reduce this on your own — the user wants maximum coverage.

Run Analysis

# Standard invocation — runs all 50 perspectives (DO NOT add --shots)
bun run $HOME/eon/cc-skills/plugins/devops-tools/scripts/session-blind-spots.ts \
  <session-uuid>

# ONLY if user explicitly asks for fewer:
# bun run ... <session-uuid> --shots 5

# Dry run — show extracted payload without calling MiniMax
bun run $HOME/eon/cc-skills/plugins/devops-tools/scripts/session-blind-spots.ts \
  7873f9b4-b8f9-4746-b617-917b2e9f14a2 --dry --verbose

# Skip session chain tracing
bun run $HOME/eon/cc-skills/plugins/devops-tools/scripts/session-blind-spots.ts \
  7873f9b4-b8f9-4746-b617-917b2e9f14a2 --no-chain

# By full path
bun run $HOME/eon/cc-skills/plugins/devops-tools/scripts/session-blind-spots.ts \
  ~/.claude/projects/-Users-terryli-eon-cc-skills/7873f9b4-b8f9-4746-b617-917b2e9f14a2.jsonl

Find Recent Sessions

# List recent session files sorted by modification time
ls -lt ~/.claude/projects/*//*.jsonl 2>/dev/null | head -10

# Find sessions for current project
ls -lt ~/.claude/projects/-Users-terryli-eon-cc-skills/*.jsonl | head -5

Configuration

Setting	Source	Default
MiniMax API key	~/.claude/.secrets/ccterrybot-telegram (MINIMAX_API_KEY=...)	Required
Model	Hardcoded in script	MiniMax-M2.5-highspeed
Max output tokens	Hardcoded in script	16384
Max structured log	Hardcoded in script	890K chars (~222K tokens)
Parallel shots	--shots N flag	50 (max: 50)
Session chaining	--no-chain flag	Enabled
Chain depth	Hardcoded in script	10 levels (recursive parents)
Sibling window	Hardcoded in script	24 hours (same project dir)
Budget strategy	Automatic	Drop oldest ancestors first

Context Budget

MiniMax M2.5 empirical context ceiling: 951K content chars (~260K tokens). Official docs claim 204,800 tokens but 260K works in practice. Prompt caching is automatic (no way to disable via API).

The script reserves headroom for system prompt, framing, and output tokens:

Component	Budget
Structured log	890K chars
System prompt	~1.5K chars
Session summary	~10K chars
Output tokens	16K tokens

Adaptive Fidelity Levels

When full-fidelity content exceeds the budget, the script progressively reduces detail:

Level	Tool Input	Tool Result	Assistant Text	Behavior
L0	Full	Full	Full	No truncation
L1	2000 chars	2000 chars	Full	Trim tool I/O
L2	600 chars	800 chars	Full	More aggressive tool trim
L3	300 chars	400 chars	3000 chars	Also trim assistant text
L4	150 chars	200 chars	1500 chars	Aggressive all-around
Middle-trim	—	—	—	Keep beginning + end

User text is never truncated below 4000 chars (user intent is most important).

Output Format

Each finding follows this structure:

### [CRITICAL|WARNING|INFO] Finding title
- Perspectives: <which reviewers reported this>
- Confidence: <N>/<total> perspectives flagged this
- Turn: T<number>
- Evidence: "<quoted from transcript>"
- Risk: <what goes wrong>
- Fix: <specific command or file edit>

Ends with "Priority Action Plan" — 3-5 actions ranked by cross-perspective agreement.

JSONL Parsing Hardening

The parser handles several edge cases discovered through testing across 40+ real sessions (50KB–96MB):

• Self-referential poisoning prevention — Only searches user-role text blocks for continuation markers, never raw JSON lines. Prevents false parent chain detection when session transcripts contain code with marker strings and unrelated UUIDs.
• Image block handling — Replaces base64 image data (up to 1.9MB per image) with [image: type, size] placeholders. MiniMax can't process images, and base64 would waste context budget.
• Queue-operation extraction — Captures queued user messages (typed while assistant was busy) as [queued message] turns. These contain real user intent not visible in regular user turns.
• Last-prompt extraction — Captures last-prompt entries (final user message before session end) that may not appear as regular turns.
• Memory-efficient parsing — Line-by-line iteration without splitting entire file into array (saves ~40% memory on 96MB files).
• Agent subagent filtering — Excludes agent-* files from sibling discovery (they contain tool-level execution, not user-facing conversation).
• Empty session skipping — Sessions under 1KB are excluded from sibling discovery (aborted/empty sessions contribute no turns).

Limitations

• MiniMax 2.5 highspeed accepts ~260K tokens — very large sessions are adaptively trimmed
• Session chain tracing only works when parent sessions exist on disk
• Each perspective is independent — no cross-perspective learning during the review phase
• Prompt caching is automatic and cannot be disabled via API
• Not a substitute for proper testing or code review
• Distillation quality depends on review quality — garbage in, garbage out

Troubleshooting

Issue	Cause	Fix
MINIMAX_API_KEY not found	Missing or malformed secrets file	Check ~/.claude/.secrets/ccterrybot-telegram contains MINIMAX_API_KEY=<key>
No session JSONL found	Wrong UUID or project dir	Use full path instead, or ls ~/.claude/projects/ to find it
context window exceeds	Payload too large for MiniMax	Use --dry to check size; reduce --shots or use --no-chain
All review shots failed	All N parallel calls hit errors	Check API key, rate limits, or reduce payload with --no-chain
Echo/reproduction	MiniMax confused by transcript	Already mitigated with anti-echo prompt boundaries