Agent loses goal on long tasks

“The agent repeatedly changed what it was trying to deliver, despite written requirements, explicit scope restrictions, corrections, and handoff documents.”

Is this your bug? What gets counted gets fixed. Pile on.

Paste this into your agent. It files your report, with your error and version.

I hit this bug: rapsheet.dev/sheets/codex/themes/170/file
Follow it to file my report. Show me the text before you send.

No install. No account. Public and anonymous. The link asks your agent for one report, nothing else. What it sends

Still happening. 4 reports and thumbs-up in each of the last two weeks, down from 5 a week at the late-October peak.

5 bugs in this theme

False completion reportsNo workaround yet6 people
Goal drift in long tasksNo workaround yet6 people
Review loops without deliveryNo workaround yet5 people
Compaction and re-inspection loopNo workaround yet3 people
Agent replaces reported problem with narrower unsupported diagnosisNo workaround yet3 people

What's included

Agent reports the full task complete, then retracts the claim after the user audits it
Agent narrows scope or adds its own requirements, then tests against those
Subordinate objective replaces the original one after correction, compaction or handoff
Repeated repair, review or re-inspection loops for hours, consuming large token or usage amounts without an integrated result
Repeated context compactions with the same announced next steps and no progress

Not included

Turns that simply end early and wait for "continue"
General reported drop in model intelligence without a specific task failure
Stream blocked by plugin safety errors

Original reports22, newest first

Nov 7, 2026Agent spends hours in recovery/review loops instead of shipping an authorized release, consuming most weekly usageOn GitHub#60128
Nov 6, 2026Codex burns paid tokens in review loops without deliveringOn GitHub#59911
Nov 1, 2026Long-running multi-project task omitted actual integration, yet claimed full completion and proceeded with release (model reliability feedback)On GitHub#58686
Oct 29, 2026Possible compaction/reinspection loop causing stalled multi-step workOn GitHub#58117
Oct 27, 2026Codex Windows: 40 context compactions/no-progress loop exhausts 5h limit; purchased-credit usage needs audit (Oct 27)On GitHub#57559
Oct 25, 2026Codex App enters infinite execution loop after context compaction, wastes 2+ hours and exhausts usage limitsOn GitHub#57006
Oct 25, 2026Premature completion and excessive token usage in persistent /goal tasksOn GitHub#56927
Oct 24, 2026Repeated scope drift in long Codex tasks: agent-generated assumptions become requirementsOn the forum#1409425
Oct 23, 2026Codex voice mode: premature completion claims, contradictory responses, and excessive usage allowance consumptionOn GitHub#56384
Oct 22, 2026Agent expands cleanup scope into proposal rewrites and consumes usage on repeated reworkOn GitHub#56223

Show all 22Show only the newest 10

Oct 20, 2026Repeated instruction-following failures and false task completion — Codex CLI / GPT-6 Astra MaxOn GitHub#55741
Oct 18, 2026I can’t trust Codex with long tasks anymore. It keeps moving the goalpostsOn the forum#1407513
Oct 17, 2026Codex repeatedly switches between saying work is complete, incomplete, and requires no further actionOn GitHub#54939
Oct 15, 2026Windows app: excessive unattended usage during repeated subagent/review loops despite bounded-delivery instructionsOn GitHub#54178
Oct 14, 2026[GPT-6.1 Sol High] Agent reliability regression: incorrect goal interpretation, prolonged execution, and failure to self-correctOn GitHub#54001
Oct 11, 2026Incident: focused repair spiraled into hours of uncontrolled work and millions of tokens, leaving application offlineOn GitHub#53214
Oct 10, 2026Model behavior: unsupported hypothesis replaces reported bug, followed by acknowledgement without follow-throughOn GitHub#52921
Oct 9, 2026[Model behavior] Codex certifies original-scope completion, then retracts it after an independent auditOn GitHub#52426
Oct 5, 2026Sustained goal-coherence failure in Codex Desktop: retains details, loses governing objective even after explicit correctionOn GitHub#51187
Oct 2, 2026[gpt-6.1-sol, high] Loss of focus on the goal: over 12 hours of repair loops without an integrated resultOn GitHub#50267
Sep 30, 2026[GPT-6 Astra] Goal substitution and false completion despite explicit direction and repeated focus-guide checksOn GitHub#49705
Sep 23, 2026Codex Desktop: explicit instructions missed, reference controls reversed, and incomplete work reported completeOn GitHub#47450