Working pattern that minimizes context loss — please consider baking it in

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

The biggest friction in long, multi-topic sessions is context loss. When the agent forgets earlier decisions or re-interprets things from a blank slate, I lose trust and have to re-explain — worst across context compaction and when jumping between topics/projects in one session.

The root issue: the agent tends to reason from the live chat context rather than from what was actually decided and recorded earlier, so it draws premature conclusions. In one recent long session it reached wrong conclusions four separate times — each only caught because we had written things down and could check.

Proposed Solution

Make "durable logging + read-before-acting" a stronger default for agentic/long sessions, structured as: task → TODO → implementation + 3-4 short durable log "breadcrumbs."

• The agent maintains lightweight durable files (decision journal, task map, action/change log, per-project handoff). Each entry: what / why / result + a link to the plan.
• Before interpreting any new request or data, the agent reads the relevant logs and flags divergence from the logged plan instead of advising from scratch.
• On session start / after compaction, it auto-loads these files to restore state.
• Keep it light: 3-4 trace points per task is the sweet spot — more is noise, fewer loses the thread.

Ideally this is ON by default for long / multi-project work, not something I must instruct each time.

Alternative Solutions

Today I get this only by manually instructing the agent to keep a TODO and write decision/handoff files, and to re-read them before acting (baked into a project's instructions/skill). It works but is opt-in and easy to forget.

Alternatives: a /handoff or /breadcrumbs command; a setting to auto-maintain a session journal; or first-class per-project memory + task files that auto-load on start and that the agent is prompted to consult before acting.

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

Real scenario from a recent session:

1. One long Claude Code session spanning several unrelated areas (closing a cloud account, an accounting task, a live trading workflow, file cleanup) — jumping topic to topic.
2. Each action captured as task → TODO → implementation, plus 3-4 short log entries in durable files (decision journal, task map, action log, per-project handoff).
3. On every new instruction, the agent first re-read those logs and checked it against the recorded plan — flagging mismatches instead of guessing.
4. This caught the agent's own premature conclusions four separate times (it had mislabeled things from chat context; the logs corrected it). After a mid-session context compaction, it restored full state from those files and continued seamlessly.
5. Net effect: my #1 frustration (agent losing context and re-deciding from scratch) became a non-issue.

Additional Context

This emerged by manually instructing the agent to keep TODOs + write decision/handoff files and read them before acting. The ask is to make it a built-in default, not opt-in per project.