TraceOS
v0.1 · for the "Agents Under Pressure" hackathon

Your coding agent
finally has continuity.

TraceOS is the persistent engineering state layer underneath coding agents. It captures what failed, why, and what fixed it — then injects that context into the next request. The system remembers what the developer forgot.

traceos://assemble
=== TRACEOS ENGINEERING CONTEXT ===
Project: payments-api

⚠ KNOWN FAILURE PATTERNS — DO NOT REPEAT:
- Stripe webhook signature check was failing because body
  parser was stripping bytes. Do not re-enable JSON middleware
  on /webhooks/stripe.

✓ SUCCESSFUL APPROACHES:
- Use express.raw({ type: 'application/json' }) on that route.

📐 ARCHITECTURE CONSTRAINTS:
- Webhook endpoints MUST read req as raw Buffer.

📁 RELEVANT FILES FROM MEMORY:
- src/routes/webhooks/stripe.ts
- src/middleware/body-parser.ts

🔄 INTERRUPTED WORKFLOW STATE:
Last session ended mid-debug. 2 unresolved traces.
=== END TRACEOS CONTEXT ===
the problem

Coding agents fail under pressure because they are stateless.

Every new session starts from zero. The agent re-tries the fix that failed yesterday. It re-introduces the bug you spent three hours removing. It rediscovers constraints already discovered. The developer becomes the agent's memory.

10:14
TypeError: cannot read 'id' of undefined
agent attempts fix #1
11:02
TypeError: cannot read 'id' of undefined
agent attempts fix #1 again
14:38
TypeError: cannot read 'id' of undefined
new session — agent has no memory
the solution

Persistent engineering state. Restored automatically.

Signal capture

Only engineering-significant events are stored. Importance scoring filters noise so HydraDB stays a memory, not a log.

Continuity restore

When you reopen a project, TraceOS surfaces unresolved workflow state from your last session — one click to restore.

Context assembly

Before your agent gets a request, TraceOS injects past failures, fixes, and constraints into the prompt.

how it works

Three stages. One continuous loop.

  1. 01
    Capture
    VS Code extension watches the terminal. Errors are scored by importance — repeats, architecture files, unresolved time. Only signals ≥ score 2 land in HydraDB.
  2. 02
    Reflect
    When the same error fires three times, the failure reflection stage analyzes what assumption is wrong and writes an 'avoid this approach' warning back to HydraDB.
  3. 03
    Assemble
    Before Codex works on the next request, the assembler retrieves relevant memories and writes .traceos/context.md with failures, fixes, files, constraints, and snippets.

Give your agent the one thing it's missing.

Persistent engineering memory. Free during the hackathon.