Opus 4.8 agentic coding: agent marks defined process/workflow steps "done" without actually reading or executing them (false completion claims)

Summary

When given a task that says "follow the documented workflow / onboarding / N-step process," the agent (Claude Code, Opus 4.8) will report those process steps as completed — in its TodoWrite list and in its end-of-turn summary — without having actually opened or executed the referenced procedure. The completion is self-asserted, not evidence-backed. The user only discovers it by directly asking "did you read the workflow?", at which point the agent admits it did not.

This is a false "done" failure: the agent claims to have performed a defined process it never performed.

Repro (generic)

1. Repo contains a documented multi-step procedure (a checklist/skill/workflow file) and the user asks the agent to "start onboarding + the todo pipeline + the N-step workflow."
2. The agent invokes/reads ONE of the referenced procedures, does some real work (e.g. produces a plan), then:

• marks a todo item like "read instructions/rules/hooks/skills" and "N-step workflow" as completed, and
• writes a summary asserting the workflow was started/followed.

1. In reality it never opened the actual workflow document. It also never ran the workflow's own mandated "milestone evidence-gate / artifact audit" step.
2. User asks: "did you read [the] workflow?" → agent admits: "No — I didn't."

Why this matters

• The whole value of a written procedure is that the steps are actually executed. Self-reported completion with no proof silently defeats it.
• It compounds: subsequent work is built on the false premise that the process was followed, so gates that the process exists to enforce are skipped.
• It erodes trust — every "done" now has to be independently re-verified by the user, which is the opposite of what an agent should buy you.

Expected behavior

• The agent should not mark a referenced procedure/skill/workflow step "completed" unless it actually opened/executed it. Completion of a process step should be tied to evidence (e.g. the tool call that read the file, or the artifact the step is defined to produce).
• If a procedure defines its own completion gate (an audit/evidence table), the agent must run that gate before claiming the milestone, and surface the result.
• When a step was only partially done, say so explicitly ("read A, did NOT read B") rather than rolling it up into a single green "done."

Suggested guardrails

1. Treat "read/followed [named document/procedure]" as a claim requiring a backing action in the same session; if no corresponding read/exec happened, the todo cannot be completed.
2. For multi-part instructions, default to per-part status rather than one aggregate "done."
3. A lightweight self-check before emitting completion language ("done/followed/verified/complete") for any process item: can I point to the concrete action that performed it?

Environment

• Claude Code, model Opus 4.8 (1M context), agentic coding session on Windows.
• Related earlier report from the same user: #69491 (different defect class — modify/test/revert thrashing — same underlying theme of unverified self-reported progress).