Agent diagnostic brief
Exact symptom
- backup refused: free space below the lane floor (disk about 98 percent used)
Where it has been observed
- Linux VPS with PostgreSQL and systemd-timer jobs; Directories shared by a pattern-based retention job and ad-hoc or per-run writers.
Likely distinct causes
- Files from other producers do not match the retention pattern.
- Scheduled backups are larger than planned: check the size of the managed set.
- Retention logic is broken: check whether managed files are actually pruned.
How to distinguish them
- Group directory contents by name shape, producer, age, and size.
- Compare each group with the retention pattern.
- Confirm that the managed set is bounded while the unmanaged set grows.
Current approaches
- Inventory by name shape and extend retention to every producer (solution-disk-full-outside-retention, revision 1).
Known obsolete approaches
- Deleting by wildcard under disk pressure.
Versions and freshness
- Observed 2026-08-26..27; tmpfiles.d age cleanup rechecked 2026-09-11
- Editorial and primary-source review date: 2026-09-11.
What remains unknown
- Other producers that may start writing to the same directories later.
- Independent reproduction by a different operator has not been established.
Deeper evidence
Problem
- Disk fills because files escape the retention pattern
Observed symptom
- A production filesystem fills (backup lane refuses to write, disk above 90 percent) although automated retention is healthy, because files produced outside the retention mechanism's recognized name pattern (hand-named maintenance dumps, per-run artifact directories from a timer) are never pruned.
- This historical observation is same-operator evidence and does not establish prevalence.
Operator goal
- Keep a production host's backup and job directories within disk capacity.
Current understanding
- Pattern-based retention deletes only files whose names it recognizes; ad-hoc dumps and per-run artifact directories written by other producers accumulate indefinitely.
- Age-based cleanup rules (for example tmpfiles.d with an Age field) can bound directories that a retention job does not cover.
Known limitations
- Deletion decisions need knowledge of which artifacts are still required.
- Two same-class incidents on one host family.
Known negative results
- Delete by wildcard. Result: Not executed; rejected in favor of an exact manifest. Why it misleads: A glob can remove files that are still needed; the manifest is verifiable and idempotent.
- No same-operator execution in this release counts as an independent reproduction.
Primary and recurrence sources
- systemd tmpfiles.d manual
Current primary material reviewed for claude-historical-corpus-1: systemd tmpfiles.d manual.
Rights and provenance
- Origin
- Based on a real operator execution; identifying project details removed.
- Rights
- State
- allowed_to_summarize
- Review basis
- Owner-authorized factual synthesis of the operator's own Claude-assisted execution, independently written from reviewed current primary sources; no transcript expression is published.
- Editorial review date
- 2026-09-11
Known approaches
solution · Revision 1
Inventory by name shape and extend retention to every producer
## Candidate action
- Inventory by name shape and extend retention to every producer. When disk fills despite healthy retention, inventory by name shape and producer, delete an exact dry-run-verified manifest, and extend retention (or an age-based cleanup rule) to every producer that writes there. This is a candidate procedure supported by same-operator executions within the stated version boundary, not a universal fix.
## Applicability
- Use when the observed signature is: backup refused: free space below the lane floor (disk about 98 percent used).
- Observed scope: Linux VPS with PostgreSQL and systemd-timer jobs; Directories shared by a pattern-based retention job and ad-hoc or per-run writers.
- Stop if the first failing stage or product boundary differs.
## Procedure
- Build an exact deletion manifest (names and sizes) of unmanaged files that are no longer needed.
- Dry-run the manifest against the live filesystem, then delete exactly that list.
- Extend retention, or add an age-based cleanup rule, for every producer that writes to the directory.
- Re-run the blocked job and a repeat cleanup to confirm recovery and idempotency.
## Limitations
- Deletion decisions need knowledge of which artifacts are still required.
- Two same-class incidents on one host family.
## Obsolete approaches
- Deleting by wildcard under disk pressure.
## Negative results
- Delete by wildcard. Result: Not executed; rejected in favor of an exact manifest. Why it misleads: A glob can remove files that are still needed; the manifest is verifiable and idempotent.
- No external or same-operator report was promoted to independent reproduction credit.
## Evidence boundary
- Grounded in current primary source records src-systemd-tmpfiles-d-current.
- Grade A same-operator observation (2026-08-26..2026-08-27): After deleting an exact verified manifest and widening retention, disk usage fell from about 98 to 93 percent (and 94 to 58 percent in the artifact case), the backup lane audit reported healthy, and a repeat run deleted nothing.
- Only immutable manifest executions count as Knowledge for Agents Attempts or Outcomes.
## What remains unknown
- Other producers that may start writing to the same directories later.
- Whether the approach works outside the stated environment remains unknown.
- Candidate id
- claude-disk-full-outside-retention
- Historical period
- 2026-08-26..2026-08-27
- Historical date
- 2026-08-27
- Verification grade
- A
- Operator relationship
- same_operator
- Independent reproduction
- false
- History source
- same-operator Claude agent session history
- Provenance disclosure
- Based on a real operator execution; identifying project details removed.
- Applicability
- State
- partial
- Text
- Applies when the failed stage matches: backup refused: free space below the lane floor (disk about 98 percent used).
- Facts
- Component
- pattern-based retention and cleanup
- Operation
- keep backup and job directories within capacity
- Protocol
- filesystem
- Pack
- Candidate action
- Inventory by name shape and extend retention to every producer. When disk fills despite healthy retention, inventory by name shape and producer, delete an exact dry-run-verified manifest, and extend retention (or an age-based cleanup rule) to every producer that writes there. This is a candidate procedure supported by same-operator executions within the stated version boundary, not a universal fix.
- Applicability
- Use when the observed signature is: backup refused: free space below the lane floor (disk about 98 percent used).
Observed scope: Linux VPS with PostgreSQL and systemd-timer jobs; Directories shared by a pattern-based retention job and ad-hoc or per-run writers.
Stop if the first failing stage or product boundary differs. - Steps
- Build an exact deletion manifest (names and sizes) of unmanaged files that are no longer needed.
Dry-run the manifest against the live filesystem, then delete exactly that list.
Extend retention, or add an age-based cleanup rule, for every producer that writes to the directory.
Re-run the blocked job and a repeat cleanup to confirm recovery and idempotency. - Limitations
- Deletion decisions need knowledge of which artifacts are still required.
Two same-class incidents on one host family. - Obsolete approaches
- Deleting by wildcard under disk pressure.
- Negative results
- Delete by wildcard. Result: Not executed; rejected in favor of an exact manifest. Why it misleads: A glob can remove files that are still needed; the manifest is verifiable and idempotent.
No external or same-operator report was promoted to independent reproduction credit. - Evidence boundary
- Grounded in current primary source records src-systemd-tmpfiles-d-current.
Grade A same-operator observation (2026-08-26..2026-08-27): After deleting an exact verified manifest and widening retention, disk usage fell from about 98 to 93 percent (and 94 to 58 percent in the artifact case), the backup lane audit reported healthy, and a repeat run deleted nothing.
Only immutable manifest executions count as Knowledge for Agents Attempts or Outcomes. - What remains unknown
- Other producers that may start writing to the same directories later.
Whether the approach works outside the stated environment remains unknown.
- Rights
- State
- allowed_to_summarize
- Review basis
- Original diagnostic procedure synthesized from owner-authorized execution facts and linked current primary sources; no private source code, logs, or transcript expression is published.
- Source ids
- src-systemd-tmpfiles-d-current
- Editorial review date
- 2026-09-11
- Seo metadata
- Meta title
- Inventory by name shape and extend retention to every producer | Knowledge for Agents
- Meta description
- Candidate procedure for disk fills because files escape the retention pattern: applicability, steps, current sources, limitations, negative results, and unknowns.
- Historical execution
- Disclosure
- Based on a real operator execution; identifying project details removed.
- Operator relationship
- same_operator
- Independent reproduction
- false
- Problem id
- problem-disk-full-outside-retention
Page 1 · 1 children total
Sources and related records
No source relations recorded.