PaperClipAI/doc/agent-files.md

Persistent agent files

Each managed agent has one current directory, scoped by company and agent. The Instructions Editor reads and writes this directory. AGENTS.md (or the configured entry) is one file in it. Agents may create ordinary files and nested folders for notes, memory, and other personal working material. Files in the task working directory remain task files.

Agents can edit their own managed files under the responsible user’s current target permissions. Access to another agent’s files additionally requires the caller’s own target-scoped configuration permission; shared company membership or a responsible user alone does not grant peer access.

Layout

The canonical host directory keeps its existing physical location:

<instance>/companies/<company>/agents/<agent>/
  instructions/                 current agent files (editor)
    AGENTS.md
    notes/
    any-supported-file
  file-sync/                   controller-only operational state
    adopted.json
    canonical-manifest.json   metadata/hash cache, no file contents
    runs/<initial-run>/
      owner.json              current run owning the writable copy
      live/                   writable copy for the provider lifetime

The process starts in its existing task workspace. AGENT_HOME points to the registered writable agent copy. Adapter HOME and CODEX_HOME keep their existing meanings and are not personal-file storage. Local copies are outside the task workspace. Remote providers currently confine file sync to their workspace: their independent agent copy therefore lives under the excluded .paperclip-runtime/agent-files/<agent>/<run>/ area. It is not included in task workspace sync, Git staging, or task deliverables.

Claude CLI runs keep the working-copy location in each run's prompt, separate from the cached system instructions. A new copy path alone does not reset the task session. Each turn names the current copy for relative file references; instruction or enabled skill content changes still invalidate the prompt bundle. Sessions saved with the older path-bearing bundle start fresh once after upgrade.

Regular files (including binary bytes) and directories are supported, up to 100,000 entries (files and folders), 256 MiB per file and 2 GiB total. Symlinks, hardlinks, and special files are rejected, rather than followed or silently skipped. The instruction entry remains valid UTF-8, at most 1 MiB, and cannot be deleted. The editor edits text up to 1 MiB and offers downloads for binary or larger files. The reserved .paperclip-runtime directory and the compatibility-only virtual file promptTemplate.legacy.md are not user storage. Task cache and Git ignore exclusions do not apply to this directory.

These storage limits are separate from the 1 MiB instruction/editor limit. Large files are hashed and downloaded as streams; listings bound concurrent reads, and only editor-sized text is buffered. Storage counts uncompressed file bytes, not allocated disk blocks. These are sync validation limits, not live filesystem quotas: an agent can write beyond them while running. Storage limits never pause an agent, fail a provider run, or block future task admission. A folder at or above a limit produces a warning on each run until enough files have been removed or shrunk. Existing saved files are restored even when already over quota, so the agent can continue working and clean them up with ordinary filesystem tools. Unsafe paths and links still fail validation; bypassing a storage quota does not bypass those checks.

An API save above a storage limit returns 422 without changing the saved files. If a stopped run exceeds a storage limit, none of its agent-folder changes are saved. The run shows a nonblocking storage warning and its save receipt reports AGENT_FILES_LIMIT_EXCEEDED with the specific limit and, for an oversized file, its path. The previous saved folder is used on the next run. The temporary run copy is discarded, including on a limit failure; there is no retained recovery archive or partial-save option. Transient sync failures get up to three attempts at the stop boundary before cleanup and an explicit failure receipt. Individual file writes are atomic, but an I/O failure partway through a sync can leave some files updated; a failed receipt does not claim whole-folder success.

Sync failures are diagnostics for the affected run, not errors on the current files in the Instructions Editor. Historical failures remain in the run log; the run detail also shows warnings from its save receipt. The editor only shows preserved instruction-only candidates that may need review, alongside errors from the current browser edit. Later successful saves do not erase run history.

Initial restoration copies the canonical folder once. Warm native Codex turns reuse that working directory. Checkpoints enumerate file metadata, hash files whose identity/size/mode/mtime/ctime changed, and temporarily copy and transfer only changed file contents. Deletions and empty directories travel as manifest entries. An unchanged image is neither rehashed nor recopied after its first checkpoint. A modified file is transferred in full; this is a file-level delta, not block-level deduplication. Canonical hash caches and manifests contain no file contents. Temporary checkpoint payloads are removed after application. The operator still provisions storage for the canonical folders, active working copies and changed-file payloads; these are not aggregate disk quotas.

Other providers use full-directory collection after verified retirement. A successful warm ACP turn can retain its unchanged copy with the still-running provider; restart recovery preserves files until retirement is proven.

Run lifecycle

1. Under the agent lock, restore current files into a private run copy and save a baseline of paths, kinds, modes, and hashes. This is sync metadata, not a revision history.
2. Stage the copy through the existing workspace transport. Point AGENT_HOME and instruction guidance at that registered root.
3. For warm native Codex, capture a manifest and changed-file payload at each terminal turn boundary before admitting another turn. Check file metadata before and after streaming and hash the captured payload independently on the host. Retry an unstable checkpoint up to three times; if it cannot be validated, close the owned provider and perform the stopped collector. Other execution paths retain their stopped-provider collection boundary.
4. Recheck the responsible user's current authorization. Under the same agent lock used by editor writes, apply only files changed or deleted relative to the last acknowledged baseline. For a competing edit or deletion of the same file, the last synchronization to acquire the lock wins. Unchanged files do not overwrite another run's changes; newly added unrelated files survive.
5. Record the save receipt and advance the baseline only after application. A warm session keeps its directory and hands ownership to the next run using a controller-owned marker. Old callbacks cannot collect or remove the next owner's files. On session retirement, collect any later writes and remove the private directory. No per-run file versions or conflict copies accumulate.

These are validated per-file checkpoints, not an atomic snapshot of arbitrary background writers across an entire directory. Writes after a checkpoint remain pending until the next checkpoint or verified session retirement. A save receipt acknowledges only the captured bytes. Lost remote bytes or missing stop proof cannot become a successful save.

Warm reuse requires the actual live session, the same remote environment and provider lease, and unchanged canonical files since its last checkpoint. An editor or another task changing canonical files retires that session before a fresh copy is restored. A directory-path mismatch also forces retirement. Only the loaded instruction entry participates in the runtime instruction digest; ordinary memory/image edits do not change it. Changes to loaded instructions, policy, credentials or provider configuration may still replace the process. Relative supporting files are read from AGENT_HOME, not the read-only prompt snapshot. External instruction bundles remain read-only and use their existing lifecycle.

For warm ACP providers, unchanged whole-directory retention uses the following additional ownership and lease fences:

A successful warm turn may retain its complete, unchanged materialized directory and the exact AGENT_HOME root with the same provider process. A bounded read-only probe must verify the full baseline and root identity without unsafe paths or concurrent changes. Local hashing runs in an isolated child; remote observation runs at the registered remote root. A failed or uncertain probe requires stopped collection. The next authorized run claims that same materialization within the company, agent, workspace, environment and configuration scope. Projectless runs use the stable workspace descriptor (cwd, repository URL/ref and branch), rather than the per-run workspace placeholder. A remote handoff verifies both DB leases belong to the same company, environment and provider allocation, then binds the successor run's active lease and collector without re-uploading the host mirror. Heartbeat explicitly permits a collection-only successor handoff before checking warm reuse. A changed or uncertain directory records durable retirementRequired; that receipt cannot authorize reuse. The successor capability is installed only after the ownership transaction commits. Configuration changes and abandoned preparation then stop the provider and collect through that current capability, never an expired prior lease. Ordinary adoption remains unchanged-only. Failed authorization or lease validation retires the prior owner without borrowing an unverified successor; if its lease is unusable, the bytes remain pending. Before initial admission, the remote transfer's two empty scratch directories are removed with exact-path, identity-checked rmdir operations. Any content, link or uncertain identity fails preparation; the complete probe never excludes them. An exact recorded lease that is missing or no longer matches the run/environment blocks retrieval as well as deletion. Legacy receipts without a lease ID require exactly one company/run/environment lease, including when a transport is still cached. Missing or ambiguous ownership preserves pending files without remote commands. No SSH exception bypasses this fence, and no-ID receipts grant no warm adoption authority. The original materialization root remains unchanged. The prior run loses collection authority; stale cleanup cannot remove the current owner's root.

Edits, additions, removals or changed configuration retire the owner before collection and fresh preparation. A configuration or copy change found at the final dispatch fence fails that run after retirement; it does not replay the request automatically. Idle expiry, restart and abandoned preparation also retire before collection. Immediate collection retries have a fixed call budget and must advance; a deferred ownership state ends the callback without inventing attempts. A failed close leaves an unstopped pending receipt through generic run cleanup. Later owner retirement or independent current-run stop proof can still collect it; a prior run's stop proof cannot authorize collection. The whole-directory contract closes the provider process, including child processes, to establish this safe collection boundary. It preserves the provider's resumable conversation. Only the loaded instruction entry participates in the new runtime instruction digest; adding or editing another file does not change that digest. Relative supporting files are read from AGENT_HOME, not from the read-only prompt snapshot.

The editor supplies the hash of the file it read. A stale browser save returns 409 and retains the user's unsaved draft. Run synchronization itself uses per-file last-sync-wins: a later run can overwrite a saved browser edit to the same file. There is no text merge or historical copy to recover the overwritten version. Ordinary task files continue using their existing workspace contract.

Upgrade and recovery

Migration 0287 creates the preview tables idempotently after master’s 0285/0286. Existing preview receipts, rows, constraints, and pending captures are retained. On first use, while holding the agent row lock, import any deployed revision heads into the existing managed directory once. A controller-owned marker outside agent files prevents any later replay of those heads. Existing revision rows remain readable for recovery; new saves never append to them. Old UUID-based clients receive content tokens and can still submit their previously recorded revision IDs, which are checked against the corresponding bytes before a write.

Working-copy receipts and native runtime inputs record the new file contract. A restored native session with no contract field keeps the old instruction-only copy shape, prompt digest, paths, and collector. Its writes use the compatibility bridge into current files, with the original baseline fence. Existing pending legacy candidates remain resolvable. Neither old task workspaces nor arbitrary external instruction roots are imported as agent directories.

Stock-agent and plugin resets update their declared files while retaining unrelated personal files and formerly configured entries. Automatic stock upgrades first record baseline hashes in the existing resource binding, then apply and finalize under the agent lock. A failed file write or database commit retries against those hashes and already-applied bytes. Removed, unchanged stock files are removed; intervening personal edits stop the retry. This pending operation metadata is cleared on success and does not retain file revisions.

External bundles retain their existing behavior. Their migration to managed storage is an explicit configuration action. Historical task cwd, provider-home, checkpoint, and workspace restoration formats are not rewritten.

Backups must include the persistent instance filesystem as well as the database. New current-file bytes are not database revision rows. Old instruction-only candidates are retained solely for upgrade compatibility.

Crash recovery can collect a stopped working copy without starting a model. Cleanup does not wait for a directory lock before process-stop proof exists, or after the copy is superseded or cleanup is complete. An unavailable copy keeps its failed-save receipt. Recovery can later clean an unavailable remote copy after destruction of its exact lease and executes no remote command. An unavailable local copy can still contain uncollected edits; this cleanup path preserves those bytes even if local stop proof arrives later. Deferred cleanup retries after a delay so one blocked copy does not prevent other copies from being cleaned. If releasing a run's instruction copy fails, the run records a cleanup warning and leaves the durable copy for the recovery sweep. Cleanup does not replace the provider's result, discard usage accounting, or prevent environment lease release. It does not claim that unsaved agent-file changes were saved; collection failures keep their separate failed-save receipts. A run attempts failed cleanup only once before handing it to recovery, rather than repeating the lock wait in teardown. Re-preparing an existing run uses the same lock as cleanup and rechecks its receipt under that lock. Preparing a new run keeps its separate admission path. Missing stop proof or lost remote bytes produce a visible diagnostic, never a save receipt. An interrupted apply can replay its changed files with the same last-sync-wins rule. Cleanup resumes for terminal runs; no copy is retained as an archive after cleanup succeeds.

An unchanged-turn observation is neither a save nor proof that a provider stopped. If retirement fails, the controller records unresolved ownership and preserves the materialized root. A crash after ownership transfer can leave the new run without independently recorded stop proof; terminal run status or the old run's alias is not enough to collect that root. Recovery leaves it preserved until the required proof is available. Exact destruction of the current owner's remote allocation permits an explicit unavailable/no-save outcome and owned local cleanup. A stopped-but-retained allocation instead stays pending with its files preserved. The current remote execute API may restart a stopped sandbox even when it bypasses a persistent session, so recovery cannot use it for retrieval or deletion. Safe retrieval from a stopped allocation remains a transport gap; live owner retirement still collects before the environment stops. No save is claimed for the pending case, and stale prior-run leases cannot authorize cleanup.

Verification

agent-directory-working-copies.test.ts exercises nested/binary files, directory isolation, last-sync-wins edits and deletions, terminal cleanup, link rejection, old-head adoption, stable prompt digests, warm ownership transfer, concurrent stale claims, and stop-proved recovery. agent-directory-probe.test.ts covers stable snapshots, unsafe paths, mutation races and bounded child failures; composed native executor tests cover retained roots and retirement before collection. The legacy working-copy and native-tool suites exercise compatibility. Workspace merge tests exercise preflight and interrupted replay.

The explicit Product E2E instruction-persistence suite creates a file through the browser editor, runs an agent that changes instructions and supporting files, checks exact binary bytes via the public download route, restarts the server, and asks a fresh task to prove restored contents using an independent nonce. A third task edits its entry while the browser saves that same file; the later run sync wins while a separate browser-created file survives, with no conflict candidate or manual resolution. Three more tasks save a sparse file at its 256 MiB boundary, exceed that boundary with a nonfatal save rejection, then remove it and save a new small file. All tasks must succeed, with warnings visible in run details while full and cleared after cleanup. Run results, including unavailable credentials, must be reported separately from unit or matcher results; a passing matcher does not prove a live run.