Conversation
Strip growing <total_tokens> and TaskCreate footers from Claude Code system text before it becomes Responses instructions, and reattach the latest notice as a trailing input message so Muse/Go prefix cache can hit. Paper and measured rates live under docs/claude-instructions-cache-stabilize/. Co-authored-by: Warexpor <warexpor@gmail.com>
|
✅ Deterministic PR hygiene checks passed. |
The Claude Code path still uses metadata.user_id. Desktop has no session id, so the fallback key now fingerprints the cacheable prefix instead of raw systemParts that still contain <total_tokens> / TaskCreate footers. Co-authored-by: Warexpor <warexpor@gmail.com>
Restrict strip to standalone integer <total_tokens> footer lines and the exact Claude Code reminder paragraph so inline docs of those tags or TaskCreate stay in instructions. Co-authored-by: Warexpor <warexpor@gmail.com>
Identify Claude Code <total_tokens> / TaskCreate notices as a trailing suffix outside markdown fences. No match returns the original system text byte-for-byte so fenced examples and whitespace-only prompts stay untouched. Translator regressions cover those no-ops. Co-authored-by: Warexpor <warexpor@gmail.com>
Relocation runs only when metadata.user_id identifies Claude Code. Unmatched fence openers cover through EOF so an unfinished fenced example plus a numeric total_tokens line is not a harness suffix. Desktop keeps the raw systemParts cache-key fallback. Co-authored-by: Warexpor <warexpor@gmail.com>
Claude Code writes <total_tokens>N tokens left</total_tokens>, not a bare integer. Digit-only matching left the harness suffix in instructions and collapsed Muse/Go hit rate to baseline. Match that shape as a trailing unfenced suffix, run peel on every inbound translation (no-op when absent), and hash Desktop prompt_cache_key from the same string used for instructions. Co-authored-by: Warexpor <warexpor@gmail.com>
Keep the real N tokens left matcher and unclosed-fence-to-EOF handling. Relocation is now stabilizePromptCache (default false) on the translator, not metadata.user_id. Ordinary callers keep a matching suffix in instructions; /v1/messages passes true. Desktop prompt_cache_key hashes raw systemParts outside opt-in and the instructions string when opted in. Co-authored-by: Warexpor <warexpor@gmail.com>
|
@Ingwannu addressed on Kept the real
Verification: |
|
@Ingwannu addressed on Kept the real
Verification: inbound-cache-stabilize + inbound — 75 pass; |
Owner
Author
|
Duplicate of upstream lidge-jun#4052 — same head branch . Closing to avoid split review. |
lidge-jun
added a commit
that referenced
this pull request
Sep 15, 2026
…d hardened (lidge-jun#4285) * feat(devin): add experimental Cognition (Devin/Windsurf) adapter Carries the adapter from lidge-jun#4078 onto current dev and places the test in its layout domain (tests/providers/devin-adapter.test.ts) with the layout map and membership fixture updated. Co-authored-by: Sayo <hi@sayo.wtf> * fix(devin): harden the cloud-direct adapter before it ships Four independent reviews of the carried lidge-jun#4078 adapter found one credential-leak blocker, one abort blocker, and a set of routing and lifecycle defects. Credentials: RegisterUser and GetUserJwt copied raw upstream bodies into Error.message, which reaches CLI output, the adapter error event, and /api/logs. A Connect error can quote the request, and the request holds the sign-in token or the api_key; redactSecretString does not match a bare JWT. Every auth and chat error now reports status, an allowlisted Connect code, and a trace id only. The four credential-bearing POSTs stop following redirects, and the api-server host is checked against a Cognition allowlist before it reaches a URL - including on the way into auth.json, so an EU or FedStart tenant host survives a reload instead of being dropped by the Copilot-only validator. Routing: the adapter posted to the static registry baseUrl, so an EU or FedStart account signed in and then sent every RPC to a server it is not provisioned on. The signed-in account's tenant now decides the host. Cancellation: after headers arrived nothing observed the caller's signal, so a client cancel drained until the idle timer fired and then surfaced as truncated_stream, while the adapter emitted neither done nor error and left the bridge to synthesize adapter_eof. The body is cancelled on abort and the turn reports the cancellation. Also: a natural completion no longer reports stopReason "stop", which was costing every clean turn its final_answer phase; sampling options reach the cloud instead of its 128k/0.7 defaults; thinking stays out of replayed assistant content; usage survives an error; gzip frames are bounded on output as well as input; dotted model ids normalize to the catalog spelling; the session cache is bounded; and logout clears the cached user_jwt whose payload carries the api_key. Co-authored-by: Sayo <hi@sayo.wtf> * test(devin): assemble the userinfo URL so the privacy scanner does not read it as an email * feat(devin): pin the client version to the shipped release and accept the real token shape Evidence from a live free-tier account and the shipped Devin Desktop 3.9.19 bundle. Details in devlog/_plan/260911_devin_two_providers/003_live_evidence.md. The sign-in token is not a JWT. A real sign-in returns a 47-character ott$<base64url> one-time value and RegisterUser accepts it, so the JWT-shape gate would have rejected every real login. The paste parser now recognises one opaque credential-shaped word rather than a token format. RegisterUser returned api_server_url https://server.self-serve.windsurf.com for an ordinary free account, which is what the tenant-routing fix in the previous commit exists for: the hardcoded server.codeium.com was wrong for this account before anyone reached an enterprise tenant. The api-server allowlist gains the staging and beta hosts the shipped bundle names, and the client version default moves from 2.0.0 - which predates the Devin rebrand - to the 3.9.19 the desktop client reports, overridable through OPENCODEX_DEVIN_CLIENT_VERSION. Co-authored-by: Sayo <hi@sayo.wtf> * docs(devin): carry the upstream MIT notice on the derived cloud-direct files A similarity check against rsvedant/opencode-windsurf-auth puts wire.ts at 1.000, index.ts at 0.988, chat.ts at 0.912, metadata.ts at 0.863, auth.ts at 0.835 and catalog.ts at 0.753 - same module split, same comments, same field layout. These files are a derivative of that repository's src/cloud-direct/, which is MIT licensed, Copyright (c) 2026 Vedant, and the carry arrived with no notice at all. The full permission notice sits in the module entry point and the other five files carry a short attribution header pointing at it, which is what MIT asks for in a distributed source tree. Co-authored-by: Sayo <hi@sayo.wtf> * docs(devin): record that the cloud chat path is unverified on a free account * fix(devin): explain the opaque cloud denial instead of guessing at it * docs(devin): restore the cloud provider rows and adapter sections on the new dev devin-cli landed on dev as lidge-jun#4288, so this branch now carries only the cloud provider. The provider rows, the login line, and the adapters reference sections for devin are re-added on top of the current docs, and every locale still leads with the measured result that its chat path is unverified. * fix(devin): restore the invalid_argument matcher the rebuild dropped The rebuild re-applied the last commit's documentation by hand and its chat.ts hunk went with it, so the runtime explanation was back to keying only on permission_denied while the measured free account returns invalid_argument - the one trailer it needed to fire for. Also from the rebuild audit: the conformance test no longer names devin-cli in guards the RUN_TURN_ONLY_WIRES set already skips, the two layout maps list the devin test files alphabetically, and structure/adapters/registry.md records why the cloud devin wire is a direct registry entry alongside devin-cli. * feat(devin): fix the request encoding that made every chat turn fail The cloud provider could not complete a single turn on any account. A paid account settled what it was not: all 229 catalogue models came back enabled and GetChatMessage failed exactly as it had on the free tier, so entitlement was never the cause. Importing the working reference's zero-dependency builder and sending its request through our own transport returned HTTP 200 and a real stream, which put the fault in our encoder rather than the wire. Diffing the two encoded messages field by field left one difference: in CompletionConfiguration, #2 is the output cap and #3 is the context window, and we had them swapped. A caller asking for 32 output tokens wrote 32 into the context-window field, and Cognition answered with an opaque invalid_argument. Fields lidge-jun#6 and lidge-jun#11 are not part of the message at all. A temperature of exactly 0 is refused with that same opaque error. Deterministic output is the ordinary case for a coding client, so it is clamped to the smallest accepted value rather than silently replaced with the service default. Three transport facts had to hold together, which is why testing them one at a time looked fruitless: the credential is the session token doubled and dash-joined in an Authorization: Basic header while the body keeps one copy, the request envelope is uncompressed, and Metadata lidge-jun#31 carries a 732-character fingerprint whose length the service checks. The metadata identity is its own seven-field shape rather than the desktop client's telemetry set, the request carries the verified tag set, and the short-lived user_jwt is now opt-in because the chat path does not need it. Verified live on a paid account: six combinations, two hosts by three models, all returning PONG with a finish reason and usage. A regression test pins the tag map so the swap cannot return silently. Co-authored-by: Sayo <hi@sayo.wtf> * fix(devin): stop the opaque-denial message asserting a retracted explanation The error string still told the user entitlement was proved and that the request fields had been ruled out. That was the hypothesis this work retracted: the same sentence came back for every turn until the CompletionConfiguration tag map was corrected, and a temperature of exactly 0 still produces it. It now points at the request first and names the test that pins the accepted field layout, and only then at the account's model access. Also from the pre-merge review: the comments claiming the hosted chat path needs the user_jwt, the stale 128k output-default comment, the promptId that is now optional because lidge-jun#22 is omitted on a first turn, and an English docs line that claimed tool calls were verified when the live evidence is chat and usage across three models. --------- Co-authored-by: Sayo <hi@sayo.wtf>
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.
Summary
instructions. Growing<total_tokens>N tokens left</total_tokens>/ TaskCreate footers churn that prefix and collapse Muse/Go prompt-cache hit rates.stabilizePromptCache?: booleandefaults to false. Ordinary Anthropic callers with that exact unfenced suffix keepinstructionsandinputunchanged. The Claude Code/v1/messagesinbound path passes true. This is not inferred frommetadata.user_id.TRAILING_TOTAL_REfor<total_tokens>\d+\s+tokens left</total_tokens>, exact TaskCreate paragraph, trailing unfenced only, unclosed fence through EOF). No match returns the original string byte-for-byte.prompt_cache_keyfallback hashes rawsystemPartsoutside opt-in (previous key). When opted in, that fallback hashes the same string asbody.instructionsafter stabilize.SOCKS5 / Go session-header work is intentionally out of scope. Related Anthropic thinking-replay issue lidge-jun#3719 stays open (link only).
Upstream review PR: lidge-jun#4052
Verification
bun test tests/claude-integration/claude-inbound-cache-stabilize.test.ts tests/claude-integration/claude-inbound.test.ts— 75 passbun run typecheck— passChecklist