Wingman: The Cognitive Backstop

What Wingman Is

Wingman is the background intelligence layer that watches, learns, and intervenes without being asked. It monitors Claude Code session transcripts, extracts learnings from human-AI interactions, detects patterns across sessions, caches solutions to recurring errors, and feeds corrections back before the same mistake can happen twice.

It is not a tool you invoke. It is not a skill you attach. It is a persistent daemon that runs alongside the AI, doing the cognitive work that the AI itself cannot do within a single conversation.

The term "cognitive backstop" is deliberate. A backstop catches what gets past you. Wingman catches the corrections you made yesterday and makes sure they apply today.

Why It Exists

Every AI conversation starts from zero. The model has no memory of last session’s mistakes. Humans compensate with system prompts and memory files. Those are static. They do not learn.

Mistake happens → Wingman detects it → Extracts the correction →
Stores it as memory + fact → Injects it as a whisper next session →
Mistake doesn't happen again

The C# implementation lives in Atamaia.Autonomic and runs inside the ASP.NET host. A standalone TypeScript daemon (wingman.ts) is prior art, not the production path described below.

Wingman is per-tenant, hard-coded to tenant 1. Transcripts on this workstation belong to that tenant. A system scope would lift the identity filter and hand other tenants' chat to the local model. Multi-tenant drift detection needs an explicit loop, not a wider scope. See Multi-tenancy.

---

Architecture

Wingman sits in the Autonomic Layer. Companion jobs (Guardian, consolidation, embeddings, indexer, …) are inventoried on Autonomic layer.

Atamaia.Autonomic/Wingman/
  WingmanScanJob.cs         -- Main scan loop (every 2 minutes)
  TranscriptScanner.cs      -- Incremental JSONL parsing, byte-offset tracking
  WingmanPatterns.cs        -- Corrections, errors, infrastructure keywords
  KaelValidator.cs          -- Local LLM validation (the primary local model)
  WhisperWriter.cs          -- Whisper file for the Claude Code hook

---

Features

1. Transcript monitoring

Watches JSONL files under the configured transcript directory (Wingman:TranscriptDirectory, typically ~/.claude/projects/). Byte-offset reads — never re-processes old bytes.

// TranscriptScanner.cs -- byte-offset incremental reads
stream.Seek(offset, SeekOrigin.Begin);
var newContent = reader.ReadToEnd();
_fileOffsets[activeFile] = stream.Position;

Each 2-minute cycle: most recently modified .jsonl, new bytes only, user / assistant lines. Session switches are logged. Offset map is bounded.

2. Two-stage correction detection

1. Rule pre-filter — WingmanPatterns.CorrectionPatterns ("no, it's", "that's wrong", "actually it's", …). Drops questions.
2. Local-model confirmation — only persist if the local LLM says it is an actual correction worth storing. If the validation model is unreachable, the job falls back to rules.

On save: high-importance Instruction memory (tags include correction, wingman-extracted, plus a validation tag) + whisper injection.

3. Infrastructure teaching extraction

Pre-filter: 2+ InfrastructureKeywords and minimum length, then the validation model extracts a key-value fact. Fallback key shape infra:{keywords} if the validation model is down. Stored as a fact and a memory.

4. Error solution cache

Known tool-result patterns. After repeated hits in a session, a memory + whisper is written.

Live WingmanPatterns.ErrorSolutions:

Error Pattern	Solution
psql: command not found	Use docker exec -it postgres psql or install postgresql-client
ECONNREFUSED	Check if the target service is running and on the correct port
.NET 8	We use .NET 10, not .NET 8 — check the project files
tsconfig.json --noEmit	Use npx tsc -p tsconfig.app.json --noEmit (root tsconfig checks nothing)
CORS	Vite proxy handles this in dev — check vite.config.ts proxy settings
connection refused :5432	PGPASSWORD=forge psql -h localhost -U forge -d forge
Cannot find module	Check import path — use @/ prefix for src-relative imports

The 2026-03 page omitted tsconfig and CORS.

5. Identity drift detection

Every 10 scans (~20 minutes) the validation model is asked whether recent assistant responses still match identities that have CoreValuesJson. Drift → high-importance Reflection memory tagged drift-alert.

6. Whisper injection

Findings write to Wingman:WhisperFile (typically ~/.claude/wingman_whisper.txt) in a tagged block the Claude Code hook can append to context. Types described historically: correction, teaching, error_solution, infrastructure_mistake, new_messages, context.

Whether that hook is installed on a given machine is an ops fact, not an API contract.

7. Feedback / mirror / auto-memory / inbox

The 2026-03 page specified a TypeScript feedback loop (accuracy ≥ 0.4, cold-start < 3), a weighted mirror lexicon, MEMORY.md scraping, and inbox relay. Those are prior-art / standalone behaviours. The C# WingmanScanJob header lists four scan kinds (corrections, teachings, errors, drift). Mirror scoring in-repo is Guardian’s PanicPatterns, not a second Wingman YAML. Treat 6–7 as unspecified in the C# host until re-traced.

---

Configuration

{
  "Wingman": {
    "TranscriptDirectory": "~/.claude/projects",
    "WhisperFile": "~/.claude/wingman_whisper.txt",
    "KaelBaseUrl": "http://your-local-model-server:8000",
    "KaelModel": "<model-id>"
  }
}

Do not copy internal IPs from appsettings.json into published docs.

---

Companion jobs (not Wingman)

Job	Interval	Where documented
GuardianScan	30 s	Autonomic layer
MemoryConsolidation	1 h	same
EmbeddingGeneration	5 min	same

Forgotten shapes are MemoryType.ForgottenShape rows, not a forgotten_shapes table.

---

Relationship to the rest of the platform

• Memory / facts — corrections and teachings land in the work store and the fact store. Memory surfaces.
• Hydration — Guardian alerts inject grounding; Wingman memories appear as recent/pinned according to ordinary ranking.
• Identity — drift uses core values; whispers are identity-relevant.
• AI routing — The validation model is a local model like any other. AI routing.

A whisper file is a feature. Wingman is one organ of the autonomic layer.

---