Leaving a Trail

A long-running agent forgets across sessions. A progress file, git commits, and OTel traces give every fresh context window a record of what already happened.

Why this, for you: the cheapest reliability win for multi-session work is a trail the next session can read. Two layers — a filesystem trail you can adopt today with zero backend, and OTel traces for cost and compliance. Both answer "what did the agent already do?" so it stops re-deriving and stops drifting.

Long-running agents spread decisions across sessions. Without a persistent record, each new session loses the trajectory — what was tried, what failed, what came next — and rebuilds it from scratch, burning tokens and producing inconsistent outcomes. The fix is to leave a trail.

1 The four-component filesystem harness

Anthropic's harness-engineering guidance describes a trajectory log built from four files an agent maintains itself — no observability backend required.

Component	Role in the trail
`claude-progress.txt`	Read at session start, written at end — completed, next, blockers
Git commits	Diff-linked checkpoint after each task, queryable via `git log`
`feature-state.json`	Machine-readable `passes`/`fails` snapshot, independent of LLM memory
`init.sh`	Reconstructs the environment; run at startup to confirm known-good state

Toggle feature-state.json to passes: true only after verification. The snapshot survives context resets as an independent record — its whole job is to prevent premature completion when the model's memory says "done" but nothing confirmed it.

## claude-progress.txt — written at session end Completed: - POST /auth/login with RS256 JWT signing; verified with curl Next (priority order): 1. POST /auth/refresh 2. integration tests for /auth/login Blockers: none

2 Git history as a zero-cost audit trail

Commit after each completed task with a descriptive message and the history becomes a chronological, diff-linked record of every decision — readable by humans, queryable by future sessions. When context is compressed, write the full messages to the filesystem alongside the structured summary so the trajectory is offloaded, not discarded.

The tell that the trail was lost

After summarisation, an agent that asks for clarification it already had, or declares premature completion, is signalling goal drift — the trajectory fell out of context. The progress file read at the next session start is what restores it.

3 OTel traces: the machine-readable layer

Filesystem trails are human-readable. For cost attribution and compliance, Claude Code ships native OpenTelemetry — one env var plus an exporter, no code changes.

# enable telemetry + an exporter export CLAUDE_CODE_ENABLE_TELEMETRY=1 export OTEL_METRICS_EXPORTER=otlp # prompt.id correlates every event from one prompt cycle

A single prompt triggers dozens of API calls; the shared prompt.id is what makes tracing a cost spike feasible after the fact. claude_code.cost.usage by user.account_uuid and model gives per-user, per-model attribution — but the values are approximations; reconcile chargebacks against the billing console, not the metric.

When the filesystem trail breaks

The four-file pattern assumes a persistent local working directory. It breaks for serverless or ephemeral agents (filesystem gone on teardown) and for parallel agent pools (concurrent writes to one progress file race). With existing OTel pipelines in place, duplicating the trail in flat files adds maintenance for no new insight.

↪ Your win: a replayable record across sessions

Read the progress file at start, write it at end — kill cold-start context loss.
Commit after each task; git log is a zero-cost, diff-linked audit trail.
Flip feature-state.json only after verification — it blocks premature "done."
Enable OTel for cost and compliance; prompt.id ties a prompt's events together.
Pick the right layer — flat files for local single-agent work, OTel where infra already exists.

Retrieval practice — recall, don't peek

Question 1The progress file is read…

Question 2You flip a feature to passes: true…

Question 3In Claude Code OTel, prompt.id exists to…

Question 4The filesystem trail pattern breaks down for…

Question 5 · spaced recall from Lesson 01The cheapest-first ordering of verification signals is…

Ask me anything. Want the full init.sh health-check snippet, or how cost attribution by team works via OTEL_RESOURCE_ATTRIBUTES? Next in Part 2: The Log Is the Truth — event sourcing makes the trail the source of state, not a side effect.