feat(client): Agent Skills — self-healing reconciles, and keys no filesystem can hold - #57
Merged
Merged
Conversation
…esystem can hold Two gaps from the Agent Skills security design review, both in `skills_fs.py`: AV-3 (partial reconciles are unrecoverable) and DV-2 (the key grammar admits Windows device names). **AV-3 — adopt a file whose bytes already are the resolved content.** `_write_all` reconciles every skill and only then rewrites the manifest, once, last. A process killed in that window leaves a skill file at a managed path with no manifest entry — exactly the condition `_write_one` treats as an unmanaged-file collision, so the skill was wedged permanently: every later reconcile took the same refusal branch. Boot-time execution under a ten-second budget makes the crash window realistic. `_write_one` now reads and hashes first and decides from the bytes. Content byte-identical to what LaunchDarkly resolved is adopted — manifest entry recorded, reported `skipped_current` — and anything else falls through to the same refusal as before. This cannot weaken the clobber guarantee: differing unmanaged bytes are never overwritten, and the existing clobber tests pass unchanged. Three details carry the safety: - A read that fails is a refusal, never an overwrite, with a message distinguishable from the byte-mismatch refusal — it is the comparison that would otherwise authorize the write, and a file that could not be read has not been shown to be ours. - The read stays on `_read_regular_file`. Adoption widens it to genuinely foreign files, so its refusal of FIFOs and other non-regular files is now load-bearing rather than defensive. It gains a `max_bytes` bound of `len(content) + 1` — enough to prove inequality for anything longer, and what keeps a foreign file of arbitrary size out of memory. A bound of exactly `len(content)` would adopt every file that merely begins with the resolved content. - `skipped_current` is reused rather than adding an `adopted` action kind, so `ReconcileActionKind` — public, and owned by an approved PR — does not change. Its documented meaning already fits. Adoption also makes the file prunable later. That is correct rather than a weakening: only byte-identical LaunchDarkly content is ever adopted, so a later prune removes content LaunchDarkly delivered anyway — what would have happened had the crash not occurred. The review also floats a write-intent journal. Assessed as over-engineered; not built. **AV-3, secondary — sweep orphaned temp files.** `atomic_write` unlinks its temp file on any exception but not after a `SIGKILL`, and `_prune` walks manifest entries, which an orphan never has, so nothing would ever notice one. The second-order effect is worse than the disk: `_prune_one`'s `rmdir` only succeeds on an empty directory, so a single orphan pins a skill's directory forever. The sweep runs on both the write and the prune path, and is the one place this SDK removes a file the manifest does not list, so it is bounded on every axis: inside `<root>/<key>/` only, for a key that passes `_key_rejection_reason`; only names `safe_fs` itself recognizes, via a new `is_temp_name` beside the naming code rather than a copy of the format string that could drift from it; only regular files, with the type read off the descriptor; unlinked through the pinned descriptor. It never raises and never aborts a run. **DV-2 — reject the 22 Windows reserved device names.** `con`, `prn`, `aux`, `nul`, `com1`–`com9`, `lpt1`–`lpt9` are all valid skill keys and none can be a directory name on Windows. Rejected in `_key_rejection_reason`, which the write and prune paths already share, and *not* in the key grammar: `parse_ai_config` fails closed, so a grammar-level rejection would invalidate an entire AI Config for a Linux customer over a Windows-only constraint, and would silently shrink `skill_refs` — which is what authorizes a prune, turning "fails to write on Windows" into "gets deleted on Linux". The 255-byte component bound is in this layer for the same reason. Unconditional, not platform-gated: a root written from a Linux container is routinely read from a Windows host, and neither repository has a Windows CI runner, so a gated branch would be untestable — the condition that produced the gap. No suffix stripping and no case folding: the grammar admits no `.` and no `$`, so `con.txt` and `CONIN$` are unreachable, and keys are lowercase-only. `com0` and `lpt0` are not reserved and are not included. The trade is real and belongs in the release notes: a customer who legitimately names a skill `aux` now gets a reported `error` action on Linux where it previously worked. Neither gap emits the integrity-failure log record. A key rejection and a clobber refusal are `ReconcileAction` errors, not integrity failures. Tests cover crash-mid-reconcile recovery end to end (adopted, reported, recorded, and the next reconcile an ordinary no-op), byte-differing content still refused untouched, an unmanaged FIFO refused without hanging, a read failure refusing rather than overwriting, the `len + 1` off-by-one, the sweep and the `rmdir` it unblocks, lookalike temp names and a symlink wearing one left alone, all 22 reserved names through both destructive paths, and — the point of the layer choice — each reserved name still valid to `is_valid_skill_key`, `parse_ai_config`, and `skill_refs`. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…failure ``get_skill`` returns ``None`` for four unrelated outcomes: no such skill, the store raised, the requested version is not the one held, and content that failed hash verification. A caller cannot fail closed on suspected tampering while tolerating a merely-absent skill, so no automated customer-side response is possible — finding LA-2 of the Agent Skills security design review. The information already existed internally, as prose in ``Resolution.error``. This gives it a token: ``Resolution`` grows a typed ``reason``, set explicitly at every construction site and declared without a default so a sixth outcome added later has to choose which public token it maps to. ``get_skill_result`` maps that straight through to a frozen ``SkillOutcome`` (``skill``, ``reason``, ``detail``). Deriving the public reason by matching the error string is the fragility LA-2 is about, so the mapping is readable in one table. ``get_skill`` is untouched — its ``None``-for-every-failure contract is documented in its docstring and in the README, and a test now pins that all four failures still collapse to ``None`` and still never raise. Nothing new is emitted: Gap 1's integrity record already fired inside verification before ``resolve_from_store`` returned, and a test asserts one failed retrieval still produces exactly one record and one signal. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… remainder The four items left open in the response to the Agent Skills security design review. Three are documentation, one is tests; no behavior changes, and the code halves of rows 9 and 2 are deliberately untouched. **Privilege separation** (row 9 docs half, and the agreed counter-proposal for row 26). The recommended deployment runs the reconcile as a different identity than the agent, which is the whole reason the ``0644``/``0755`` modes deny anything: the agent reads its instructions and cannot rewrite them, or the manifest. That is the mitigation for AZ-1, a prompt-injected agent editing its own skills. Write access to the manifest is the worse half — it is what tells the *next* reconcile which paths the SDK may delete — which is why ``_prune`` re-validates every entry from scratch rather than trusting it. The README documents the pattern and hands the operator the check to run, because the SDK cannot run it: it knows only its own identity, which trivially has write access, having just written there. So ``ReconcileReport`` grows no writability field — the review asked for one and we declined, since any check the SDK could make would answer a different question than the one asked and manufacture false confidence exactly where caution is wanted. ``agents.md`` records that reasoning so the field is not added later by someone reading its absence as an oversight. **Three hostile-manifest prune tests** (row 12 remainder): a well-formed manifest listing ``/etc/passwd``, ``../../../etc/passwd``, and a path under a parent that has since become a symlink. ``_prune`` already refuses all three, so these turn asserted into verified. Two things make them worth more than their line count. They are deliberately *well-formed* — the corrupt-manifest suite above them proves nothing here, because a corrupt manifest suppresses every destructive action wholesale, whereas these manifests give the implementation everything it needs to prune. And "deleted nothing" is asserted through an unlink spy rather than by checking that ``/etc/passwd`` still exists: the test process cannot delete that file anyway, so the obvious assertion would pass against an implementation with no path check at all. **One sentence on** ``"*"`` (row 16 remainder). It materializes the whole project library, so every skill's ``description`` enters the agent's context — including skills no AI Config references and skills belonging to other teams. **The Windows platform bound is now explicit** (row 2 residual), in ``safe_fs.py``, ``agents.md`` and the README. Reparse-point checks (``GetFileAttributesW`` / ``FILE_FLAG_OPEN_REPARSE_POINT``) are not implemented, by decision: Windows is not a supported or tested platform for this release, neither repository has a Windows CI runner so the checks would ship unverified, and the TypeScript SDK could not match them in any case because Node exposes no ``*at()`` family on *any* platform. Implementing them in Python alone would break cross-language parity and trade a documented bound for an unverified one. Two consequences are recorded rather than left to be rediscovered: on Windows, write permission on the managed root is the only boundary, which is what makes privilege separation the mitigation and not merely advice; and this retroactively lowers the priority of the row 25 reserved-device-name work, noted where that code lives. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
write_skills is a one-shot reconcile, so a revocation takes effect at the next process restart. watch_skills runs that reconcile now and again whenever the configured store reports a change, so a revoked skill's files leave the disk within a debounce interval of the store learning about it. It is wired to the SkillStore interface, not to any one transport: it needs a store that implements add_listener and nothing more, and refuses loudly when the store does not, since a watcher that silently never fires looks exactly like one whose skills never changed. Above the interface, remove_listener joins add_listener as the optional second half of change notification, on the SkillStore contract and on InMemorySkillStore. SkillWatcher.close needs it to detach; without it a store held every watcher ever created for the rest of its life. The watcher probes for it, so a store without it keeps working. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…network The half of the delivery transport that has no I/O: identifying a skill object on the wire by kind inline-resource plus category skill, translating it into the raw object shape the SkillStore interface defines, holding it by (key, objectVersion), and applying a payload's events as one commit at payload-transferred. The store that puts a connection underneath this follows separately, so the three decisions that matter most can be reviewed on their own: - objectVersion is the skill's version; version is the payload's. The translation happens in one place and TestVersionTranslation asserts it in both directions, because confusing them fails silently. - Changes commit at payload-transferred, not per object. A half-applied full transfer would briefly empty the store, which with pruning on is the difference between a reconcile and deleting a customer's files. - A hashless object is held, not dropped, so verification withholds it with a reason code rather than the transport reporting it absent. Flag and segment objects share the connection and are skipped and counted, not rejected. Nothing here is exported yet; the store exports it. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…oncile watch_skills awaited write_skills and only then constructed SkillWatcher, which is where the store listener attaches. The reconcile snapshots the store as its first step and then spends the rest of its time on the filesystem, so every write, fsync, prune and manifest rewrite in that first pass ran with nothing listening. A change delivered in that window was never seen, and since nothing re-reconciles on a timer, a revocation that landed there waited for the next unrelated change — on a quiet root, the next restart. Exactly the gap watch_skills exists to close. The watcher now attaches its listener before the initial reconcile and starts its worker after. notify only sets an event, so a change arriving mid-reconcile is recorded and picked up by the worker's first pass, while holding the thread back keeps write_skills's one-root-one-reconcile contract: the worker cannot race the caller's own reconcile over the same manifest. A reconcile that raises detaches the listener on the way out, since the caller is handed an exception rather than a watcher to close. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The protocol reader took payloads[0]'s intentCode and applied it to the skill object set, which is what the delivery protocol requires — one payload per credential, read the first intent, tolerate the rest — but it left the assumption behind that rule undocumented and unguarded. If the one-payload guarantee ever widens, an xfer-full for another payload would start an empty pending set and the next payload-transferred would publish it: every skill reported revoked, and with pruning on, a customer's files deleted. The first payload is still the payload that is read. What is new is that the reader now knows which payload skills actually arrive on — learnt from the intent's id, or from the (p:<id>:<version>) selector, since no object or transfer event carries a payload id of its own — and declines to apply a transfer of any other, holding last known good, warning once, and counting it in diagnostics.payloads_ignored. An intent describing more than one payload warns once on its own, because that is the one case the comparison cannot catch: another payload's transfer arriving before any skill has been seen has nothing to be compared against. Behaviour under one-payload delivery is unchanged, and a full transfer of the skill payload still empties it — every skill deleted is a real state the guard must not mask. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The SDK-facing FDv2 channel now delivers skills the way streamer #4681 and gonfalon #70638 spell them: object kinds are open strings, the agent-skill payload is classified `generic`, and every generic object carries only `key`, `kind`, `version` and `object`, exactly like a flag. A skill arrives under kind `skill` with its own version folded into the key as `<key>:<version>`. There is no `category` field and no `objectVersion` field; both came from an earlier streamer draft that never shipped. Identification is now the kind alone. The wire key is split in one place, `_split_wire_key`, and both the put and the delete translation go through it. A key that will not split cleanly is held rather than dropped — version-less, or with the offending text as its version — so verification withholds it with `invalid_version` under a key the caller recognises; only a key with nothing before the delimiter is dropped, since there is no identity to hold it under. `SDK_DATA_MODEL_VERSION` goes with it: the connection's `mv` parameter only accepts flag model versions, and generic payloads ignore it. The transport stops sending it in the following change. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
FDv2SkillStore puts LaunchDarkly's SDK-facing FDv2 channel underneath the protocol layer: GET /sdk/poll and GET /sdk/stream, authenticated with the environment's server-side SDK key, streaming by default. It carries basis across requests, sends If-None-Match and treats 304 as a current answer, retries with capped jittered backoff, honours Retry-After only up to max_backoff, gives up after a bounded run of consecutive failures where a committed payload resets the count, and keeps serving last known good through every failure. A mobile key or client-side environment ID is refused in the constructor. Standard library only. close interrupts the socket rather than only setting a flag, because the delivery thread lives in a read no flag can reach; without that every shutdown of a healthy stream waited out the full join timeout. The no-store message now names FDv2SkillStore first, and watch_skills points at it as the store with a delivery transport. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
_Requester.stream wrapped only the connect as recoverable, so a read timeout, reset or truncated chunk in the body reached the delivery loop as whatever the socket raised. The loop read that as a bug and gave up: delivery stopped for the process lifetime, taking updates and revocations with it, the first time a socket died. read_timeout exists to bound a stream that has gone quiet so the loop can reconnect, and tripping it did the opposite. The body now carries the same promise the connect already did. Wrapping the line source rather than the whole read keeps protocol reader errors out of it: those are raised from the consumer's loop body, where they still surface as the bugs they are. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…n promptly Three faults in the delivery loop, all of which left the store reporting itself healthy while doing less than it claimed. **An up-to-date stream tripped the failure cap.** The consecutive-failure count reset only at a commit, and an environment whose skills are not changing answers every reconnect with `intentCode: "none"` and transfers nothing. A stream only ever ends by being dropped, so each recycle of a perfectly healthy idle connection counted as a failure — announced with a `goodbye` or not — and `max_consecutive_failures + 1` of them stopped delivery for the process lifetime, revocations included. The reset on commit covered only the case where content had changed, which is the case that was easy to test and not the case that runs in production. `_TransferOutcome` now reports `up_to_date`, and a complete answer that transfers nothing breaks the row of failures exactly as a commit does. An intent this module does not recognise is still not an answer. **`close` could not interrupt a poll.** The interrupt reached the streaming connection only, so polling parked in its request with nothing to reach and `close` returned when its join timed out — on a 300s-class request, long after the process meant to exit. `_Requester` now tracks the response of a poll in flight and offers `interrupt`, which `close` calls alongside the stream's own. A request still inside its connect has no response to reach; that one is bounded by `read_timeout`, and `start` no longer leaves the store inert when a join times out around it. An interrupt we asked for is no longer recorded as a delivery failure. **`close` left a waiter parked.** `wait_for_skills` waited on the first payload alone, so a shutdown racing a waiter added the waiter's whole timeout to it. Delivery ending is now its own event: a waiter is released by a payload, a give-up or a close, and reports whether a payload actually arrived rather than merely that it was let go. That also settles what `_give_up` had been quietly asserting — it set the first-payload flag to unblock waiters, which made `wait_for_skills` answer `True` for a store holding nothing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A stream only ever ends by being dropped, and LaunchDarkly — and any proxy in between — recycles a long-lived one. Every reconnect therefore logged "Skill delivery failed" at WARNING, for as long as the process ran. Until the previous commit that noise was bounded, because an idle stream gave up after eleven recycles and went quiet; now that delivery correctly survives them, it would run forever and describe a healthy store as failing. A connection that got a complete answer before it ended — a committed payload, or an up-to-date intent — delivered everything it was asked for, so its reconnect is now DEBUG and says so. A connection that ended without answering is the case the warning exists for and still gets it: a connect that never landed, or a transfer that died part-way through. Filling a customer's logs with a fault they do not have is not merely untidy; it teaches them that the level which means something can be ignored. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
andrewklatzke
approved these changes
Sep 14, 2026
andrewklatzke
left a comment
andrewklatzke
left a comment
Contributor
There was a problem hiding this comment.
Probably worth getting another review on this one from someone more colocated with the skills stuff, but code looks good to me
Stacked on the protocol-layer PR (`xie/skills-fdv2-protocol`). Third of three PRs split out of #69, and the one that makes the feature real: the network underneath the protocol reader, exported as `FDv2SkillStore`. ## Why this shape Skill content arrives over `GET /sdk/poll` and `GET /sdk/stream`, authenticated with the environment's server-side SDK key. These are the SDK-facing endpoints the base SDK's FDv2 data source uses, and the channel that payload signing will eventually cover. No private route is involved, and no credential other than the environment's own SDK key ships to a customer host. Standard library only, so the content path adds no dependency to a package whose sole runtime dependency is `opentelemetry-api`. ## What's here **`FDv2SkillStore`.** Authenticates, streams (default) or polls, carries `basis` across requests, sends `If-None-Match` and treats 304 as a first-class current answer, and serves `get_object` / `all_objects` / `add_listener` / `remove_listener` from what the protocol reader has committed. Capped jittered backoff; `Retry-After` honoured but clamped to `max_backoff` and rejected when non-finite; bounded consecutive-failure retries, where a committed payload resets the count. One network timeout, `read_timeout`, whose default follows the mode: 10s for a whole poll, 300s between reads on a stream. A mobile key or client-side environment ID raises from the constructor. Last known good survives every failure; `diagnostics` and `failed` report the degradation. **`close` interrupts the socket.** The delivery thread parks in a read no flag can reach, and closing a urllib response from another thread does not unblock CPython's buffered reader, so `_interrupt_read` shuts the socket down underneath it. Without that every shutdown of a *healthy* stream blocked for the full join timeout. **Above the interface**, two strings: `NO_STORE_MESSAGE` now names `FDv2SkillStore` first, since it is the first thing a user sees on a missing store and offering only the development store was wrong once a production transport existed; and `watch_skills`' refusal message names it as the store with a delivery transport. ## Bugs found and fixed while testing the loop Five, all sharing one shape: the store stopped delivering while continuing to report itself healthy. - **The consecutive-failure counter never reset in stream mode.** `_stream_once` always ends by raising, so a reset on return was unreachable and `failures` grew for the whole process lifetime. Eleven *fully successful* payload transfers were enough to trip `max_consecutive_failures` and stop delivery for good, revocations included. A commit now resets the count, in `_apply`. - **A non-finite `Retry-After` killed the delivery thread.** `float("inf")` parses, and `Event.wait(inf)` raises `OverflowError` from inside the recoverable-error handler. Non-finite values are rejected and every honoured delay is clamped to `max_backoff`. - **`close()` during the initial connect waited out its full join timeout.** `self._connection` was assigned after the connect returned, so a `close()` in that window found nothing to interrupt. The stop flag is re-checked immediately after the assignment. - **`close()` blocked for the full join timeout on every healthy stream.** See `_interrupt_read` above. - **A stream interrupted by our own `close` was reported as a delivery failure.** Also: `connect_timeout` was accepted and never used, so a poll against a black-holed host hung for 300s rather than 10. It is gone, with the request timeout now chosen by mode, and `TestTimeouts` measures the bound against a socket that accepts and never answers. ## Tests > **Rebase note.** The previous push of this branch had silently reverted the protocol PR's last commit (payload identity: `payloads_ignored`, `_is_foreign_payload`, `TestPayloadIdentity`). Rebasing onto the updated protocol branch restored it; the full suite passes with it present. `_FakeFDv2Endpoint` is an in-process `ThreadingHTTPServer` implementing the wire contract, so request construction and header handling are exercised over real sockets rather than mocked. Covers skill put/delete over the wire, mixed payloads, 304, `basis` round-tripping, reconnect/backoff in both modes, `Retry-After` including non-finite and oversized values, bounded retries and the reset on commit, prompt shutdown during connect and during a healthy stream, hashless envelopes end to end through the accessors, server-side-only credentials, timeouts, and `watch_skills` over the transport: a wire-level revocation pruning a file without a restart. Full suite 1629 passing, 11 skipped; `ruff`, `ruff format`, and `mypy` clean. ## Open items, none in this PR's scope - 🔴 **`contentHash` is not on the wire yet.** Against a real environment today every skill resolves to nothing. This PR makes that loud (an error per hashless object, a summary per wholly-hashless payload, `diagnostics.hashless_objects`) rather than surviving it. - 🔴 **Server-side skill delivery is not deployed.** The wire shape this store reads — kind `skill`, key `<key>:<version>`, generic payload — is what [streamer #4681](launchdarkly/streamer#4681) and [gonfalon #70638](launchdarkly/gonfalon#70638) emit; both are still open. No account can receive skill objects until they ship and the producer is enabled. - 🟡 **FDv2 is opt-in per account.** A real environment returns 403 today; the store reports it as fatal and explains what to do. - ~~🟡 **`mv` is a guess.**~~ Resolved: the request sends no `mv`. That parameter selects the *flag* data model and the connection rejects any value but the flag default, while the generic agent-skill payload is served regardless of it. The `data_model_version` constructor argument is gone with it. - 🟡 **No payload signing** on this channel yet, so Beta is TLS-only. - 🟡 **The connection also carries the environment's flags.** Skipped and counted; a transport property, not fixable here. - 🟡 **`ld-relay` does not speak the FDv2 endpoints**, so relay-only deployments cannot receive skills in Beta. **Nothing here has touched a real LaunchDarkly environment**, because it cannot yet. 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- CURSOR_SUMMARY --> --- > [!NOTE] > **Overview** > Adds **`FDv2SkillStore`**, a production `SkillStore` that pulls agent skills over LaunchDarkly’s SDK FDv2 **`/sdk/poll`** and **`/sdk/stream`** endpoints (stdlib HTTP, background delivery thread, stream-by-default). It implements **`SkillStore`** (`get_object`, listeners, etc.) on top of the existing protocol reader, plus **`StoreDiagnostics`**, **`wait_for_skills`**, capped backoff with **`Retry-After`**, and **`close`** that interrupts blocked socket reads so shutdown is prompt. > > **Public surface:** `FDv2SkillStore` and `StoreDiagnostics` are exported from the package; README documents production setup with `init_client` and `watch_skills`. Server-side SDK keys only; mobile/client credentials are rejected. Outages keep last-known-good content; accessors above the store are unchanged. > > **Delivery-loop fixes** bundled here: reset consecutive-failure counts on successful commits / up-to-date answers (so healthy stream recycling does not stop delivery), safe handling of non-finite **`Retry-After`**, and not treating intentional **`close`** interrupts as transport failures. Removed unused **`connect_timeout`**; **`read_timeout`** is the single knob with mode-specific defaults. > > **Tests:** in-process fake FDv2 server exercises poll/stream, basis/ETag/304, revocations, retries, timeouts, hashless payloads, and **`watch_skills`** over the transport. > > <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit 88c225e. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup> <!-- /CURSOR_SUMMARY -->
…network (#82) Stacked on the `watch_skills` PR (`xie/skills-watch`). Second of three PRs split out of #69; the transport that puts a connection underneath this follows. The half of the delivery transport that has no I/O: identifying a skill object on the wire, translating it into the raw object shape the `SkillStore` interface defines, holding it by `(key, objectVersion)`, and applying a payload's events as one consistent commit. Splitting it out lets the three decisions that matter most be reviewed without a socket in the way. ## Three things worth reviewing closely **1. The skill's version is in the object's `key`; `version` is the payload's.** Each version of a skill is its own object on the wire, identified as `<key>:<version>` (`pdf-extraction:3`), and that is the only place the skill's version appears. The event's `version` is the payload's, and confusing the two fails *silently*: the object verifies, the hash matches, and the caller gets content under a version number that means nothing. The wire key is split in exactly one place (`_split_wire_key`), both the put and the delete translation go through it, and `TestVersionTranslation` asserts it in both directions. A key that will not split cleanly is *held* rather than dropped — version-less, or with the offending text as its version — so verification withholds it with `invalid_version` under a key the caller recognises; only a key with nothing before the delimiter is dropped. **2. Changes commit at `payload-transferred`, not per object.** A payload version is the unit of consistency. A half-applied full transfer would publish a state the server never described and would briefly empty the store, which, with pruning on, is the difference between a reconcile and deleting a customer's skill files. An interrupted transfer leaves last known good intact, and listeners fire once per commit. **3. A hashless object is held, not dropped.** Dropping it at the transport would report `absent`, indistinguishable from "no such skill", and would let a prune delete the last known-good copy on disk. Holding it means verification withholds it with `missing_content_hash`, which is diagnosable: an ERROR per `(key, version)`, deduped per reader rather than per process so two stores never quieten each other, a summary per wholly-hashless payload, and a `StoreDiagnostics.hashless_objects` counter. There is deliberately no fallback that synthesises a hash from the delivered content. ## Also here - **Skills are `kind == "skill"`; everything else is ignored, not rejected.** Object kinds on the SDK-facing channel are open strings and the agent-skill payload is classified `generic`, so a skill arrives under the kind its producer registered — the bare category name — with no `category` or `objectVersion` field ([streamer #4681](launchdarkly/streamer#4681), [gonfalon #70638](launchdarkly/gonfalon#70638)). An environment's assignment carries its flag payload alongside its agent-skill payload, so flag and segment objects arrive as a matter of course. Erroring on them would turn a normal payload into a permanent reconnect loop. - **`_SkillObjectSet`** holds several versions of one key, with lookup semantics identical to `InMemorySkillStore` down to the fall-through to a version-less entry. Its opaque snapshot keys are spelt `<key>:<version>`, the same as the wire, and a test pins that round trip. `TestInterfaceParity` asserts the two resolve identically. - **`_require_server_side_credential`** refuses a mobile key or client-side environment ID. Its tests arrive with the store constructor that calls it. Nothing here is exported yet; the store exports it. The module imports nothing from the feature but the version validator. ## Tests `test_skills_fdv2.py` drives `_ProtocolReader` directly: identification, version translation, full and change transfers, interruption, revocation, tombstones, mixed payloads, unknown kinds and events, error and goodbye, and the hashless dedupe across readers. The wire builders it introduces are shared with the transport PR's fake endpoint. 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- CURSOR_SUMMARY --> --- > [!NOTE] > **Overview** > Adds **`skills_fdv2.py`**, a stdlib-only layer below `SkillStore` that parses LaunchDarkly FDv2 events without sockets. It maps wire `put-object`/`delete-object` into the raw store shape, keeps skills in **`_SkillObjectSet`**, and applies updates atomically in **`_ProtocolReader`** at **`payload-transferred`** (not per object). > > The critical wire rule is **`key` = `skillKey:objectVersion`** while the event’s **`version` is the payload revision** and is dropped—confusing them would silently serve the wrong pinned version. Non-`kind == "skill"` objects (flags, segments) are **ignored**, not errors. **Foreign payload** full transfers are declined once skills’ payload id is known, so a flag `xfer-full` cannot wipe held skills (and trigger prune). **Hashless** skills are retained for verification to withhold with `missing_content_hash`, with **`StoreDiagnostics`** and loud logging. **`_require_server_side_credential`** rejects mobile/client credentials (for the upcoming networked store). > > **`agents.md`** documents the transport contract; **`skills_core`** clarifies `SKILL_OBJECT_KIND` vs wire kind. **`test_skills_fdv2.py`** (~900 lines) exercises the reader, version translation, payload identity, and parity with `InMemorySkillStore`. Nothing is exported or wired from accessors yet—that’s the follow-on transport PR. > > <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit efc4ca7. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup> <!-- /CURSOR_SUMMARY -->
…ls (#81) Stacked on #66 (`split/skills-review-closeout`). First of three PRs split out of #69; the FDv2 protocol layer and the transport follow on top of this one. `write_skills` is a one-shot reconcile, so a revocation takes effect at the next process restart. `watch_skills` runs that reconcile now and again whenever the configured store reports a change, so a revoked skill's `SKILL.md` leaves the disk within a debounce interval of the store learning about it, rather than at the next restart. ## What's here **`skills_watch.py`** — `watch_skills` / `SkillWatcher`. Wired to the `SkillStore` interface, not to any one transport: it needs a store that implements `add_listener` and nothing more. It refuses loudly when the store does not, because a watcher that silently never fires looks exactly like one whose skills never changed. The listener itself only sets an event; the reconcile runs on a single worker thread, debounced, so a burst of changes coalesces into one pass and a slow disk never stalls the store's delivery thread. A reconcile that raises is logged and the watcher continues. `on_unavailable="keep"` stays the default: an outage must not read as "everything was revoked". **`remove_listener(kind, fn)`** joins `add_listener` as the optional second half of change notification, on the `SkillStore` contract and on `InMemorySkillStore`. `SkillWatcher.close()` needs it to detach; without it a store held every watcher ever created for the rest of its life. The watcher probes for it and skips detaching when a store does not implement it, so a customer's own store keeps working. ## Tests `test_skills_watch.py` drives the watcher through `InMemorySkillStore`, whose `put` notifies synchronously, and through small store doubles: initial reconcile, coalescing, refusal of a store without `add_listener`, detaching on close, and a store without `remove_listener`. Three tests on `InMemorySkillStore.remove_listener` join `test_skills.py`. The end-to-end case, a `delete-object` arriving over a live connection and pruning a file, lands with the transport PR where the fake endpoint lives. 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- CURSOR_SUMMARY --> --- > [!NOTE] > **Overview** > Adds **`watch_skills`** and **`SkillWatcher`** so agent skill files stay in sync with the configured skill store without waiting for a process restart. The API runs an initial `write_skills` reconcile, then listens for store delivery changes and re-reconciles on a **debounced background thread** (listener only wakes the worker; disk I/O never blocks the delivery thread). Revocations can prune `SKILL.md` within the debounce window. **`on_unavailable="keep"`** remains the default so transport outages are not treated as mass revocations. > > Extends the optional **`SkillStore`** change-notification contract with **`remove_listener(kind, fn)`**, implemented on **`InMemorySkillStore`**, so **`SkillWatcher.close()`** can detach; stores without it still work but keep listeners registered. **`watch_skills`** raises if no store is configured or the store lacks **`add_listener`**, and cleans up the listener if the initial reconcile fails. > > Public exports and README/agents docs are updated; new tests cover watcher behavior (coalescing, mid-startup revocations, close/detach) and **`remove_listener`** semantics. > > <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit 29ad3fe. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup> <!-- /CURSOR_SUMMARY -->
… remainder (#66) Stacked on #58. Closes the four items left open in [our response](https://launchdarkly.atlassian.net/wiki/spaces/PD/pages/5293965360) to the [Agent Skills security design review](https://launchdarkly.atlassian.net/wiki/spaces/PD/pages/5264506969). Three documentation items and one test item. **No behavior changes** — the code halves of rows 9 and 2 were already done and are deliberately untouched; the only change to `safe_fs.py` is its module docstring. ## 1. Privilege separation (row 9 docs half + counter-proposal for row 26) The proposed mitigation for **AZ-1**, the review's most serious finding. The README now documents reconcile-identity ≠ agent-identity as the recommended deployment, which is the whole reason the `0644`/`0755` modes deny anything: the agent reads its instructions and cannot rewrite them, or the manifest. Write access to the manifest is the worse half — it is what tells the *next* reconcile which paths the SDK may delete — which is why `_prune` re-validates every entry from scratch rather than trusting it. The operator gets an explicit check to run, because the SDK cannot run it. **`ReconcileReport` deliberately grows no writability field.** Row 26 asked for one; we declined. The SDK knows only its own identity, which trivially has write access, having just written there. Any check it could make would answer a different question than the one asked and manufacture false confidence exactly where the review wants caution. `agents.md` records that reasoning so the field is not added later by someone reading its absence as an oversight. ## 2. Three hostile-manifest prune tests (row 12 remainder) A well-formed manifest listing `/etc/passwd`, `../../../etc/passwd`, and a path under a parent that has since become a symlink. `_prune` already refuses all three; these turn asserted into verified. Two things make them worth more than their line count: - They are deliberately **well-formed**. The corrupt-manifest suite above them proves nothing here — a corrupt manifest suppresses every destructive action wholesale, whereas these manifests give the implementation everything it needs to prune, and it must refuse anyway because the recorded *path* is not one this SDK could own. - **"Deleted nothing" runs through an unlink spy**, not through checking that `/etc/passwd` still exists. The test process cannot delete that file anyway, so the obvious assertion would pass against an implementation with no path check at all — permissions would be doing the work. The spy proves the removal is never *attempted*. ## 3. One sentence on `"*"` (row 16 remainder) It materializes the whole project library, so every skill's `description` enters the agent's context — including skills no AI Config references and skills belonging to other teams. ## 4. Windows platform bound (row 2 residual) — deferred, explicitly **Decision: defer, with the bound written down.** Reparse-point checks (`GetFileAttributesW` / `FILE_FLAG_OPEN_REPARSE_POINT`) are not implemented, and that is now recorded in `safe_fs.py`, `agents.md` and the README rather than left implied. Rationale: - Windows is not a supported or tested platform for this release, and **neither repository has a Windows CI runner** — every matrix job is `ubuntu-latest` — so the checks would ship unverified. Both codebases already cite that absence as a reason not to add `os.name`/`process.platform` branches. - **The TypeScript SDK could not match them in any case.** Node exposes no `*at()` family on *any* platform, so its racy `lstat` floor is universal rather than Windows-only. Hardening Python alone would break the cross-language parity the two SDKs are held to. Two consequences are recorded rather than left to be rediscovered: on Windows write permission on the managed root is the *only* boundary, which is what makes privilege separation the mitigation and not merely advice; and this **retroactively lowers the priority of the row 25** reserved-device-name work, noted where that code lives. That code stays — it keeps a root written on Linux usable when read from Windows — but it is not evidence that Windows is hardened. ## Parity The Node half is launchdarkly/js-ai-sdk#48, case for case. No customer-visible surface moved: the `ld.skills.integrity_failure` event name, the eight `reason_code` tokens, the five `SkillOutcomeReason` tokens, and the log-record serialization are untouched. ## Verification - `uv run pytest` — 1484 passed, 11 skipped - `uv run ruff check .` / `ruff format --check .` — clean - `uv run mypy packages/*/src` (the CI gate) — no issues in 43 source files 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- CURSOR_SUMMARY --> --- > [!NOTE] > **Overview** > Closes the remaining **Agent Skills security review** items with **documentation and tests only** — no runtime behavior changes beyond an expanded `safe_fs.py` module docstring. > > The **README** and **`agents.md`** now spell out deployment and platform limits: run **`write_skills` under a different identity than the agent** (with a shell check for agent writability), **decline a `ReconcileReport` root-writability field**, warn that **`"*"`** materializes the full skill library into context, and document the **POSIX-only** descriptor-pinned guarantee versus Windows’s racy `lstat` floor (reparse-point hardening explicitly deferred). > > **`test_skills_fs.py`** adds **`TestHostileManifestPrune`**: well-formed manifests that would prune `/etc/passwd`, traversal paths, or paths under a symlink-swapped parent must error without **`os.unlink`** being attempted (unlink spy), matching existing `_prune` defenses. > > <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit 6a0e6aa. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup> <!-- /CURSOR_SUMMARY -->
…failure (Gap 2 / LA-2) (#58) ## Gap 2 (finding LA-2): a distinguishable outcome for integrity failure versus absence **This PR adds public API and needs review as such.** Three new exported names: `get_skill_result`, `SkillOutcome`, `SkillOutcomeReason`. ### The defect `get_skill` returns `None` for four distinct outcomes — no such skill, the store raised, the requested version is not the one held, and content that failed hash verification. A customer therefore cannot fail closed on suspected tampering while tolerating a merely-absent skill, so no automated customer-side response is possible. LA-1 (shipped in #51) gave the *operator* a structured log record; this gives the *application* the same distinction. The information already existed internally. `Resolution` in `skills_core.py` distinguished all four cases — but only as prose in its `error` string, which `get_skill` discarded with `.skill`. ### The API ```python SkillOutcomeReason = Literal["absent", "integrity_failure", "ok", "store_unavailable", "wrong_version"] @DataClass(frozen=True) class SkillOutcome: skill: Skill | None reason: SkillOutcomeReason detail: str | None async def get_skill_result(key: str, *, version: int | None = None) -> SkillOutcome: ... ``` Chosen over a typed exception, an opt-in strict mode, and an init-time callback. `detail` is the existing human-readable reason string — already safe to surface (skill key and failure mode only, never content, never a filesystem path), and a test pins that it stays that way. `get_skill_result` raises `RuntimeError` only when no store is configured, with the same message as `get_skill`. ### `get_skill` does not change Its contract — `None` for every failure, never raises for one — is documented in its docstring and in the README, and every existing caller treats `None` as "no skill". A test drives all four failures through both accessors: the reason is distinguishable *and* the collapsed form still collapses. ### Not matched on the error string `Resolution` grows a typed `reason`, set explicitly at every construction site. Deriving the public reason by pattern-matching `Resolution.error` is exactly the fragility LA-2 is about, so the mapping is a table a reviewer can read: | `resolve_from_store` outcome | `reason` | |---|---| | the store raised (`unavailable=True`) | `store_unavailable` | | `raw` is not a dict | `absent` | | `verify_raw_skill` returned `None` | `integrity_failure` | | `skill.version != wanted_version` | `wrong_version` | | success | `ok` | `reason` is declared **first and without a default**, so a contributor adding a sixth internal outcome has to decide which public token it maps to rather than inheriting one. That made the two `Resolution` sites in `skills_fs.py` explicit as well — both are the "could not retrieve" path, both already `unavailable=True`, both now say `reason="store_unavailable"`. `Resolution.unavailable` is unchanged. It stays load-bearing on the prune path — only a raising store suppresses pruning, because deleting managed files after a failed lookup would turn an outage into data loss — and `store_unavailable` stays distinct from `absent` for the same reason. ### Three things deliberately not built - **No new telemetry or log record.** Gap 1's `ld.skills.integrity_failure` record already fired inside verification before `resolve_from_store` returned. A test asserts one failed retrieval still produces exactly one log record and exactly one signal. - **Gap 1's 8-token `IntegrityReasonCode` is not threaded into `SkillOutcome`.** `verify_raw_skill` returns `None` and does not surface which token fired; plumbing it up would change that function's return type for a detail the operator already gets from the log record. The five-token public reason is the actionable surface. - **No batch equivalents.** `get_skills` and `all_skills` keep omitting failed entries and keep logging the run-level WARN count. A `get_skills_result` / `all_skills_result` would double the accessor surface for a case nobody has asked for — **possible follow-up** if a customer needs per-key reasons from a batch. ### Tests `packages/client/tests/test_skills.py`, one new class plus three additions to existing export/immutability tests: - one case per reason token — `reason`, whether `skill` is populated, non-empty `detail` - `store_unavailable` is distinct from `absent`: a raising store and an empty store - every non-`ok` outcome carries a detail (swept, so a fifth failure path with no message is caught by a test whose name says what it is about) - `get_skill` still returns `None`, and still never raises, for all four failures - `get_skill_result` raises `RuntimeError` with no store, asserted equal to `get_skill`'s message - no second integrity record or signal from `get_skill_result` - `detail` never carries the skill content - `SkillOutcome` is frozen; `SkillOutcomeReason`'s tokens and the three new exports are pinned ### Docs - `packages/client/README.md` — `get_skill_result` in the Agent Skills API table, a fail-closed example under the Gap 1 observability material (exit on `integrity_failure`, tolerate `absent`), the reason table, and an explicit statement that `get_skill` is unchanged and the two accessors differ only in what they report. - `packages/client/agents.md` — the five-token vocabulary, the `Resolution` → reason mapping, and the instruction that a sixth internal outcome must choose a public token rather than defaulting to `absent`. ### Lockstep with TypeScript The type name, the accessor name, and the five reason tokens are fixed across both SDKs; a sibling PR implements the identical shape in `typescript/`. Tokens are alphabetical, matching how `IntegrityReasonCode` was written on both sides in Gap 1. ### Placement Based on `split/skills-self-healing` (#57), which is based on #54. A new PR at the top of the stack so it disturbs none of the existing reviews: Gap 2 spans `types.py`, `skills.py`, `skills_core.py` and `__init__.py`, which #50–#53 own, and #52/#53/#54 are approved while #50/#51 have live reviewer conversation. ### Checks `uv run pytest` (1481 passed, 11 skipped), `ruff check`, `ruff format --check` all clean. `uv run mypy .` fails repo-wide on a pre-existing duplicate-`conftest` error before checking anything; scoped to the files this PR touches, mypy reports only two pre-existing `unused-ignore` warnings that are present on the base commit too. 🤖 Generated with [Claude Code](https://claude.com/claude-code) <!-- CURSOR_SUMMARY --> --- > [!NOTE] > **Overview** > Adds **`get_skill_result`** and frozen **`SkillOutcome`** / **`SkillOutcomeReason`** so applications can tell **integrity failure** apart from a missing skill, store outage, or version mismatch—without changing **`get_skill`** (`None` for all failures, same path, no extra telemetry). > > Internal **`Resolution`** now carries a required **`reason`** mapped explicitly at each **`resolve_from_store`** site (including **`skills_fs`** store-unavailable paths); **`get_skill_result`** exposes **`skill`**, **`reason`**, and safe-to-log **`detail`**. > > README and **`agents.md`** document fail-closed handling; tests cover all five reasons, immutability, exports, no double integrity signals, and unchanged **`get_skill`** behavior. > > <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit eda4ae1. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup> <!-- /CURSOR_SUMMARY -->
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using default effort and found 3 potential issues.
Reviewed by Cursor Bugbot for commit 7b59680. Configure here.
| logger.debug("Skill payload unchanged (HTTP 304)") | ||
| # A 304 counts as a first payload, so a boot that reconnects with a | ||
| # cached basis is not blocked on a transfer the server will not send. | ||
| self._publish_first_payload() |
There was a problem hiding this comment.
Up-to-date intent never unblocks waiters
Medium Severity
wait_for_skills only releases on a committed transfer or HTTP 304, not on xfer-none. That intent is documented as the same complete answer as a 304 — the server has nothing to send — and _apply already treats up_to_date as success for retries. Boot still waits out the full timeout and then returns False on a healthy connection, so callers proceed as if delivery failed.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 7b59680. Configure here.
| committed=True, | ||
| changes=changes, | ||
| basis=state if isinstance(state, str) and state else None, | ||
| ) |
There was a problem hiding this comment.
Ignored payload still advances reconnect basis
Medium Severity
A foreign payload-transferred correctly leaves last known good skills in place, but still returns committed=True with that payload's state. _apply then overwrites _basis. The next reconnect asks LaunchDarkly for changes from the ignored payload's selector, so later skill updates and revocations can be skipped.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 7b59680. Configure here.
| target=self._run, name="ld-ai-skills-fdv2", daemon=True | ||
| ) | ||
| self._thread.start() | ||
| return self |
There was a problem hiding this comment.
Restart after give-up keeps failed state
Medium Severity
start() after delivery has given up does not clear _failed_reason or _failures. failed keeps reporting the old error after a new thread is running, and the first recoverable drop on that thread increments a counter that is already past the cap, so delivery gives up again immediately unless a payload commits first.
Additional Locations (2)
Reviewed by Cursor Bugbot for commit 7b59680. Configure here.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Sixth PR in the Agent Skills stack, and the first that is not part of the original
five-PR split. Implements Gap 3 (AV-3) and Gap 4 (DV-2) from the Agent
Skills SDK security design review. Both live in
skills_fs.py, so they shiptogether.
Stack: #50 references → #51 retrieval → #52 safe-fs → #53 materialization →
#54 fs-hardening → #57 (this).
Why this is a new PR rather than an edit to #53
The code Gap 3 changes lives in
skills_fs.py, which #53(
split/skills-materialization) owns. Editing #53 would be the wrong move:APPROVED. Editing feat(client): Agent Skills — materialize onto disk under a manifest (4/5) #53 revokes that approval on thesecond-largest PR in the stack and forces test(client): Agent Skills — the filesystem abuse matrix (5/5) #54 to rebase again — it has already
been rebased twice this week (the str→bytes pivot, then the Gap 1 log record).
in the whole gap set, and it needs to be reviewable in isolation rather than
buried in a rebase of an already-approved PR.
Being a new branch on top, nothing below needs rebasing and no approval is
disturbed.
Gap 3 — self-healing partial reconciles (review priority: Medium)
_write_allreconciles every skill and only then does_rewrite_manifestwritethe manifest, atomically, once, last. A process killed after a skill file lands
but before that write leaves the file at a managed path with no manifest
entry — exactly the condition
_write_onetreats as an unmanaged-filecollision. The skill was then wedged permanently: every later reconcile took the
same refusal branch. Boot-time execution under a ten-second budget makes the
crash window realistic.
_write_onenow reads and hashes first, and decides from the bytes:This cannot weaken the clobber guarantee, because only byte-identical
LaunchDarkly content is ever adopted. Differing unmanaged bytes are still
refused, untouched. The existing clobber tests pass unchanged — including
test_unmanaged_file_is_never_overwritten, which writes"user authored\n".Three details carry the safety:
distinguishable from the byte-mismatch refusal. It is the comparison that would
otherwise authorize the write, and a file whose bytes could not be read has not
been shown to be ours.
_read_regular_file, notPath.read_bytes. Adoptionwidens it to genuinely foreign files, so its refusal of FIFOs and other
non-regular files is now load-bearing rather than defensive.
skipped_currentis reused rather than adding anadoptedkind, soReconcileActionKindintypes.py— public, and owned by approved feat(client): Agent Skills #50/feat(client): Agent Skills — materialize onto disk under a manifest (4/5) #53 —does not change. Its documented meaning ("the bytes on disk already are the
resolved content") already fits, and the reuse is documented in a comment.
One decision the brief did not settle: a bound on the read
Adoption reads files the manifest does not list, so
_read_regular_filegained amax_bytesparameter and reads at mostlen(content) + 1bytes. Without it,adoption would be a memory amplifier: a multi-gigabyte file at a managed path
would be slurped whole and then refused, where before the change it was refused
without being read at all. The
+ 1is load-bearing — a bound of exactlylen(content)would adopt (and later prune) any customer file that merelybegins with the resolved content. There is a test for that off-by-one, and it
fails if the
+ 1is removed.The caveat, stated plainly
Adoption creates a manifest entry, so the adopted file becomes prunable on a later
reconcile. That is correct and not a weakening: adoption only fires when the bytes
are byte-identical to LaunchDarkly-resolved content, so a later prune removes
content LaunchDarkly delivered anyway — exactly what would have happened had the
crash not occurred. The guarantee that matters is preserved: differing unmanaged
bytes are never overwritten. There is a test pinning the caveat too, so it
cannot change silently.
The review also floats a write-intent journal as an alternative. Assessed as
over-engineered; not built. No case was found where the adoption rule fails.
Gap 3, secondary — orphaned temp files (landed, not deferred)
atomic_writenames its temp file.SKILL.md.<16 hex>.tmpand unlinks it onexception — but not after a
SIGKILL._prunewalks manifest entries, which anorphan never has, so nothing would ever notice one. The second-order effect is
worse than the wasted bytes:
_prune_one'srmdironly succeeds on an emptydirectory, so a single orphaned temp pins a skill's directory forever.
A sweep now runs on both the write and the prune path (before the
rmdir, so itactually unblocks it). It is the one place this SDK removes a file the manifest
does not list, so it is bounded on every axis:
<root>/<key>/only, for a key that passes_key_rejection_reason;safe_fsitself recognizes — a newis_temp_name/temp_name_prefixpair beside the naming code, matched with
fullmatchso both ends are anchored.The pattern is derived from the producers rather than copied:
atomic_writeand
_mkstemp_atnow build their names from the same constants the recognizerreads, and the recognizer covers both producers (
secrets.token_hex(8)on thedescriptor path,
tempfile.mkstemp's own sequence on the Windows fallback);unlink_file— no newpath-based unlink;
Gap 4 — Windows reserved device names (review priority: Low)
The key grammar
^[a-z0-9][a-z0-9-]*$admitscon,nul,aux,prn,com1–com9,lpt1–lpt9— 22 names, none of which can be a directory name onWindows. Rejected in
_key_rejection_reason, afteris_valid_skill_key, matchingthe existing ordering discipline in that function.
It belongs in the filesystem layer, not the shared key grammar
types_validation.skill_key_rejection_reason/is_valid_skill_keyaredeliberately untouched. Please do not "correct" this:
parse_ai_configfails closed on a badskillsentry, by design. Agrammar-level rejection would therefore invalidate the entire AI Config —
model, provider, instructions, tools — for a Linux customer, over a constraint
that only exists on Windows.
skill_refswould silently shrink, and its own docstring explains theconsequence: a dropped reference lets
write_skillsprune the skill'son-disk copy. That converts "this skill fails to write on Windows" into "this
skill gets deleted on Linux."
exactly this reason, documented in both codebases in those words.
_key_rejection_reasonis already shared by the write and prune paths, so oneedit covers both destructive paths.
A test asserts the layer choice directly: for each of the 22 names,
is_valid_skill_keyis stillTrue,parse_ai_configstill accepts a configreferencing it, and
skill_refsstill projects it.Unconditional, not platform-gated
No
if os.name == "nt".ordinary deployment — so the on-disk result must not depend on the writer's OS.
ci.ymlisubuntu-latest), so a platform-gated branch would be untestable in CI — theexact condition that produced this gap.
The trade, stated honestly and belonging in the release notes: a customer who
legitimately names a skill
auxnow gets a reportederroraction on Linux whereit would previously have worked.
The set is exactly 22 names. No suffix stripping and no case folding are needed,
and a comment says why: the key grammar admits no
.and no$, so thecon.txtand
CONIN$/CONOUT$forms are unreachable, and keys are already lowercase-only.com0/lpt0are not reserved and are not included — there is a test forthat, and for
con1/nul2/conx.Residual documented, not implemented
MAX_PATHoverflow. The 255-byte bound is per component and does not bound thetotal path —
<customer root>+<key>+/SKILL.mdcan still exceed Windows'260-character
MAX_PATHwith a legal key. The SDK cannot validate this becausethe root is the customer's, so it is one README sentence rather than a check.
Scope and non-goals
clobber refusal are
ReconcileActionerrors, not integrity failures — noreason_codetokens, norecord_integrity_failurecall._unsafe_path_reason,_key_rejection_reason, orsafe_fs.pywasrelaxed.
agents.mdmarks these non-relaxable, and the new material there saysthe same about the adoption rule and the sweep, so a later contributor cannot
widen either.
types.pyis untouched.here — it is a separate public-API decision, planned next. Gap 1 landed in feat(client): Agent Skills — retrieval through an injectable store (2/5) #51
and was not modified.
Docs
packages/client/README.md: the adoption rule under the existingwrite_skillsconservatism paragraph, the reserved-name rejection with the honest trade, and
the
MAX_PATHresidual.packages/client/agents.md: the 22-name list, the adoption rule and itsnon-widenable details, the layer-choice reasoning, and the sweep's bounds — all
in the non-relaxable-checks material.
Tests
packages/client/tests/test_skills_fs.py, +388 lines, 91 new cases:suite): a byte-identical file with no manifest entry is adopted, reported
skipped_current, recorded in the manifest, and the next reconcile is anordinary no-op. Plus: adoption writes nothing (
os.replacespy sees no call),emits
skipped_currenttelemetry, and the adopted file is prunable afterwards.len + 1off-by-one: a longer file sharing the content prefix is not adopted.distinguishable message, and adds no manifest entry.
rmdiron prune; eight lookalikenames left alone; a symlink wearing the temp name is not removed and its target
survives.
write_skillsand the prune path, plusthe grammar-level assertion above and the non-reserved neighbours.
Every new test was mutation-checked — reverting each of the four behaviours
(reserved-name check, sweep, adoption, the
+ 1) makes the suite fail.Checks
uv run ruff format --check .,uv run ruff check .,uv run mypy packages/*/src(CI's command),
uv run pytest— all clean: 1469 passed, 11 skipped.Note:
uv run mypy .(theMakefile'stypechecktarget) fails on apre-existing duplicate-
conftestmodule-resolution error inpackages/ai/tests/packages/claude-agents/tests, unrelated to this change andpresent on
main. CI runsmypy packages/*/src, which is clean.🤖 Generated with Claude Code
Note
Overview
Agent Skills grows a production delivery path and clearer failure semantics, while
write_skillsbecomes more recoverable and cross-platform-safe on disk.Delivery and live materialization: New
FDv2SkillStorepulls skills over LaunchDarkly’s FDv2 poll/stream channel (server-side SDK key only), holds last-known-good through outages, and exposesStoreDiagnostics.watch_skillsdebounces store change listeners into automatic re-reconciles so revocations can prune disk without a restart.InMemorySkillStore(and the store contract) gainremove_listenerfor clean watcher teardown.Retrieval:
get_skill_resultreturns a frozenSkillOutcomewithSkillOutcomeReason(ok,absent,integrity_failure,store_unavailable,wrong_version) on the same path asget_skill— so callers can fail closed on tampering whileget_skillstill collapses failures toNone.Filesystem reconcile (
skills_fs/safe_fs): If a managed path exists with no manifest entry but bytes match the resolved hash, the file is adopted (skipped_current) instead of wedging after a crash between writes and manifest update; reads are bounded (len(content) + 1). 22 Windows reserved device names are rejected at materialize/prune time on all platforms. Orphan temp files from killedatomic_writeruns are swept via sharedis_temp_name/temp_name_prefixinsafe_fs.Docs: README and
agents.mddocument FDv2 usage,*vs pinned refs, privilege-separated deployment, and POSIX vs Windows filesystem guarantees.Reviewed by Cursor Bugbot for commit 7b59680. Bugbot is set up for automated code reviews on this repo. Configure here.