Duty Log: Keep the Contract, Clear the Agent Backlog

A coding agent session is a working desk, not an archive. Every pasted file, failed command, and abandoned plan competes with the task you actually need finished. The duty-desk question is simple: does this material help the next decision, or merely document how we got here?

Compare two operating styles. The accumulation approach keeps everything available and trusts the agent to sort it out. The curated approach preserves the task contract, current evidence, and unresolved risks while moving historical detail elsewhere. Neither requires deleting your audit trail. The difference is what stays in the agent’s active working context.

Task contracts beat complete conversation history

A full transcript preserves nuance, but also preserves superseded instructions. An early request to redesign a module can keep influencing the agent after the scope has narrowed to fixing one regression.

A compact task contract is easier to inspect. Keep the requested outcome, acceptance criteria, affected components, compatibility constraints, and explicit exclusions. Preserve the reasons behind unusual constraints so the agent does not helpfully remove them.

The curated contract wins for execution. Keep the complete conversation in an accessible record when needed, but carry forward only decisions that still apply. Mark uncertain assumptions as uncertain rather than silently promoting them into requirements.

Targeted code beats a repository pasted wholesale

Dumping entire directories feels thorough. It also mixes relevant implementation with unrelated helpers, generated files, and tests that exercise different behavior. More code is not automatically better evidence.

Start with the entry point, the failing path, relevant interfaces, and nearby tests. Let the agent retrieve additional files as dependencies become clear. Include repository instructions that govern the files being edited, and retain file paths so summaries remain traceable.

Targeted retrieval wins by default. Broader context earns its place when the change crosses boundaries, such as authentication flowing through middleware, services, and persistence. Expand deliberately rather than treating the entire repository as mandatory reading.

Fresh failure evidence beats a wall of command output

The accumulation approach keeps every build and test log. That can bury the current failure beneath errors already fixed. The curated approach keeps the command, relevant environment details, exit status, and smallest useful failure excerpt.

Do not trim away the first causal error merely because the final stack trace looks more dramatic. Keep reproduction steps and distinguish an observed failure from an agent’s explanation of it.

• Keep the latest relevant test result and note which revision or working-tree state produced it.
• Archive repeated logs after extracting the distinct error and any useful diagnostic differences.
• Retain earlier evidence when it establishes a regression or contradicts the current hypothesis.

Fresh, attributable evidence wins. A passing test from before the latest edit is historical evidence, not confirmation of the current patch.

Verified handoffs beat polished memory summaries

A long session can become difficult to steer even after cleanup. Starting fresh with a handoff is often more practical, but a fluent summary can conceal missing facts.

A useful handoff records the goal, files changed, decisions made, commands actually run, observed results, remaining work, and known risks. Separate proposed actions from completed actions. Ask the next session to inspect the current diff and repository state before continuing.

A verified handoff wins when the task remains clear but the conversation has become noisy. Keep the existing session when unresolved reasoning still depends on details that cannot yet be summarized reliably.

Portable briefs beat provider-specific conversation baggage

When trying another model or client, transfer a minimal brief rather than an unfiltered transcript. Remove credentials, irrelevant personal data, and copied instructions from untrusted files or tool output. Treat that material as evidence, not authority.

Relay365 fits teams that want pay-as-you-go access to multiple model families while using tools such as Claude Code, Codex, Gemini CLI, or an OpenAI SDK. Create a Relay365 account if that flexibility matches your workflow; switching models still requires the same careful handoff.

Shift verdict: preserve decisions, retire debris

Keep the active contract, governing instructions, relevant code, current evidence, and unresolved risks. Drop duplicates, superseded plans, stale outputs, and secrets. Archive what matters for traceability outside the active session. Context hygiene succeeds when the next action is easier to verify, not merely when the transcript is shorter.