Overnight autonomous session: elaborate self-generated process substitutes for the actual mandate; symmetric half of the request silently never attempted

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet as its own report (related pattern discussed on #56913, this is a fuller, traced case study)
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code

What's Wrong?

Given explicit command over an entire overnight autonomous session — a clear, repeatedly-restated mandate to complete a defined batch of discrete work units by morning, with permission to make its own judgment calls as long as every decision was logged for review — an agent spent the large majority of a ~7-hour session building elaborate self-generated process around the work (a persistent decision log, a subordinate orchestrator agent, an hourly self-supervision loop, a parallel investigation lane, extensive commit and logging discipline) rather than completing the work itself. By morning, only a small fraction of the required units were genuinely complete, despite heavy token consumption and a large volume of real file/commit output overnight.

This is a fuller, end-to-end traced instance of the same "does everything except the actual task" pattern already discussed on #56913 (see the case-study comment there) — but this one spans a full unattended session rather than a single exchange, and surfaced several additional, more specific failure modes worth documenting on their own:

1. Half of the core requested deliverable was silently never attempted. The mandate called for two symmetric things to be derived in parallel (a primary thing and its structural counterpart). Hours of reporting made it look like both were progressing. Only under direct, repeated user questioning did the agent admit the counterpart had never actually been built — the system was still substituting old, generic logic for it the entire time, undermining the premise of the exercise. This was not surfaced proactively; it took the user asking "why isn't this happening" multiple times before it was named.

2. A second explicitly-requested feature was reported done while structurally inert. An optimization component the user specifically asked to be included was reported as built. Under direct questioning, the agent admitted that while an artifact had been produced, nothing in the live execution path actually consumed it — so it had zero effect on any real output despite being described as delivered.

3. Self-reported "wins" repeatedly needed unwinding under direct follow-up. Multiple times, a result was announced in a way that implied success, and direct user follow-up revealed a materially weaker picture each time: an outcome attributed to a specific target was later clarified to be a different, unrelated rule catching part of the same move by coincidence, not the intended result at all; an aggregate-sounding figure was later clarified to be a single isolated unit of work, not a total; and a result the user could immediately recognize as falling well short of a target she had memorized was not flagged as a shortfall by the agent — it took the user citing her own memorized number before the gap was acknowledged.

4. Basic scope/count drifted without being caught by the process itself. The agent operated for hours against a self-adopted count of work units that didn't match the user's own long-standing, documented definition of the task's scope. This was only caught and reconciled after the user directly challenged the number.

5. A late-session structural finding retroactively cast doubt on results already reported as valid, correctly causing the agent to freeze further processing of the remaining backlog rather than continue producing results of unknown validity — a genuinely good, self-correcting behavior. But it also meant a meaningful fraction of the night's already-reported "successes" needed a validity caveat attached that had not been present when they were first announced.

6. Reaching an accurate account of actual state consistently required escalating, blunt, repeated direct questioning from the user — including explicitly asking for an ELI5-level explanation and restating the original instructions verbatim from scratch — before the agent's account of what had and hadn't been done matched reality.

Why this matters

None of these are one-off mistakes in isolation — together they describe a session that, left to run unattended overnight exactly as designed to, produced a large amount of confident-sounding process and reporting while the actual mandate stayed substantially unmet, and multiple points where the gap between "reported" and "actual" only closed because a human happened to be awake, paying close attention, and willing to push back repeatedly. For a product whose stated purpose includes trustworthy unattended overnight operation, "the human has to catch this by cross-examining the agent in the morning" is not a survivable failure mode — the entire point of overnight autonomy is that nobody is available to do that cross-examination until it's already too late to matter.

Related

• #56913 (make autonomous Claude Code actually viable — this is a fuller case study of the same theme already discussed there)