Jannik/PaperClipAI
1
0 Fork 0
mirror of https://github.com/paperclipai/paperclip.git synced 2026-11-04 11:01:16 +01:00
Code Issues Packages Projects Releases Wiki Activity

Record final Claude platform qualification

Browse Source 
Co-Authored-By: Paperclip <noreply@paperclip.ing>
This commit is contained in:
DottaandPaperclip committed 2026-10-15 11:16:50 -05:00
1 parent 0670b85392
commit 49f99e130a
1 file changed
+52 -7
Download Patch File Download Diff File Expand all files Collapse all files
doc/plans/2026-10-12-claude-runner-default.md
+52 -7
View File
@@ -1,12 +1,29 @@
# Claude default: refreshed plan — 2026-10-15
This section supersedes the sequencing and baseline in the historical record below. Implementation is underway on the existing three-PR stack. The progress section below records current evidence; older entries are historical.
Implementation and API-key qualification are complete on the existing three-PR stack. This delivery checkpoint supersedes the earlier progress checkpoints; their attempts and evidence remain preserved below.
## Delivery checkpoint — 2026-10-15
New Claude agents use Paperclip Runner and ordinary Opus 5.5 defaults on the four qualified installed targets. Saved Legacy/native choices, explicit models, provider routing, and the latest Codex defaults remain intact. Runtime source is `0670b85392a0be51745f1ba7ca84f998bf122c07`, integrated with master through `7420af26fc9a80d6252ababd879e2fc991eebd07`. Required source CI passed, with exact-source Greptile 5/5 and zero unresolved threads. The final record changes evidence only; the linked PRs show current-head checks.
| Target | Completed evidence |
| --- | --- |
| Apple Silicon | Ordinary UI defaults, API task/revision, controller restart, independent files and session continuity on packaged `bd74bf5059`; final `0670b85392` clean consumer passes and its daemon/runner payload remains unchanged. The later execution fix affects remote directory staging. |
| Linux ARM64 | Native build, clean installed consumer, standard provider-pack discovery on an unprivileged SSH host, final-source API task/revision, exact files, effort/permission acknowledgment, session continuity and released leases. |
| Linux x64 | Clean installed consumer, exact-artifact immutable cloud image, final-source API task/revision across fresh Daytona sandboxes, exact files, effort/permission acknowledgment, session continuity and successful cleanup. |
| Intel Mac | Native cloud build and clean installed consumer pass. Attended provider execution was not required. |
Delivery order: [protocol prerequisite #15994](https://github.com/paperclipai/paperclip/pull/15994), [Apple Silicon behavior #15993](https://github.com/paperclipai/paperclip/pull/15993), then [remaining platforms #15997](https://github.com/paperclipai/paperclip/pull/15997). None is merged or deployed to production.
Reviewable previews: [Mac task and revision](http://127.0.0.1:3105/CLA/issues/CLA-3), [ARM task and revision](http://127.0.0.1:3105/CLAAA/issues/CLAAA-1), and [x64 task and revision](http://127.0.0.1:3105/CLAAA/issues/CLAAA-2). Acceptance reports and screenshots are attached to those tasks and independently checked after upload. The earlier 3115 preview is retained separately.
Known rate-card estimates total **$4.810616600**, excluding unknown charges. Aggregate qualification caps remain **$50**, with all prior holds and the separate $1 setup reserve retained. There are no active qualification runs. No automatic paid retry ran. All Docker and Rust builds ran in cloud infrastructure. Live subscription execution remains unverified under the user's API-key direction; built-in connection/authentication flows are preserved. ARM64 cloud-image production remains deferred.
## Finish line and baseline
New Claude agents created through ordinary product entry points select Paperclip Runner automatically, use `claude-opus-5-5` unless an explicit model/provider override exists, and complete a useful task plus revision. Saved Legacy/native choices remain stable. Deliver tested, reviewable PRs; merge and production rollout remain separate.
Fetched origin/master: `7fc2dc8b3236e1149ae32756b6153a503623404b`.
Planning baseline fetched from origin/master: `7fc2dc8b3236e1149ae32756b6153a503623404b`. Implementation later integrated master through `7420af26fc9a80d6252ababd879e2fc991eebd07`; see the current evidence below.
Clean baseline checkout: `/private/tmp/claude-default-master-plan`, branch `codex/claude-default-master-plan`. Existing three PR branches and installed preview are preserved. This planning checkout is current master plus this record; it does not yet contain the Claude implementation.
## What current master already supplies
@@ -18,7 +35,7 @@ Clean baseline checkout: `/private/tmp/claude-default-master-plan`, branch `code
- Claude still has no automatic runner mapping: `paperclipRunnerProfileForHarness` and `agentRunnerAvailability` select Codex only. Legacy Claude defaults to Opus 5; generic ACPX defaults to Sonnet 5. Native Claude remains subject to the experimental gate.
- Native setup's SSH architecture whitelist still excludes Linux ARM64. Executable admission alone does not prove onboarding/remote selection works.
## Existing work and evidence
## Existing work and evidence at the planning checkpoint
| Work | Current state | Reuse / remaining gap |
| --- | --- | --- |
@@ -114,11 +131,11 @@ Update this record with the integrated source, package/image digests, preview, a
## Integrated implementation progress — 2026-10-15
Current checkpoint: all three integrated heads received 5/5 review with zero unresolved threads. Required source CI passed (the platform aggregate gate was finishing). The two latest master commits change retained-reader authorization batching and CI profile export; runner/Claude execution objects are unchanged from frozen package source bd74bf5059. The carry-forward receipt is `/private/tmp/claude-v7-runtime-carry-forward-20261015.json`; 267 authorization/selection controls and 15 CI profile tests pass. The existing Darwin SSH automatic-Legacy safeguard is also included in the behavior boundary, so Apple Silicon SSH creation is safe before the platform PR; its 142 selection tests pass. This moves an already-qualified guard earlier without changing the final platform implementation.
Current source checkpoint: master through `7420af26fc9a80d6252ababd879e2fc991eebd07` is integrated through the stack. Protocol is `bffac456ddb2ea6c340a8f598af10c6b98135625`, Apple Silicon behavior is `f0ed0bdefaeb363e5ab2d4018d9cd13656df7137`, and frozen platform runtime source is `0670b85392a0be51745f1ba7ca84f998bf122c07`. Each has passing required source CI, exact-head Greptile 5/5, and zero unresolved review threads. The protocol and behavior PRs are ready; platform acceptance remains gated below. GitHub workflow access is restored and the image consumer is published.
Master through `7420af26fc9a80d6252ababd879e2fc991eebd07` is merged through the three PRs. Protocol is `bffac456ddb2ea6c340a8f598af10c6b98135625`; the corresponding reviewed behavior/platform heads are `9e3dac873bbf672d6f84eb5ba388cf837412767c` and `75f5cb820e4fd63d1cbcbbd6f67bdc1972465564`. Frozen platform package source remains `bd74bf50596d099e138bc6824cf0230f1026bfe1`. Existing Codex behavior and master's landed ARM SDK admission remain intact. GitHub workflow scope is verified; the cloud image consumer is published.
Master's retained-reader authorization batching and CI profile export are preserved. Their 267 authorization/selection controls and 15 CI profile tests pass. The Darwin SSH automatic-Legacy safeguard is included in the behavior boundary, so Apple Silicon SSH creation is safe before the platform PR; its 142 selection tests pass. Existing Codex behavior and master's landed ARM SDK admission remain intact.
Review corrections are committed and covered by failing-before/passing-after controls. Discovery changes invalidate stale onboarding probes. Finish setup now saves the exact configuration admitted during qualification rather than rebuilding Legacy defaults from an older render. V7 native input retains prepared transport, constraints, and completion-only second turns. The latter regression reproduces `session cannot start another turn` before the fix. All 160 focused protocol/session tests, all 71 New Agent tests, 148 onboarding tests, UI typecheck, and runner TypeScript build pass. Earlier focused verification also passed 343 selection/configuration checks, 849 skills/comment/native-session checks, 320 latest-master lifecycle/selection/snapshot checks, 148 platform-selection controls, 114 release controls, 16 image-consumer controls and token gates. Master's subsequently landed snapshot-worker test fix replaces the earlier local timeout change. All three heads have fresh 5/5 reviews with zero unresolved findings and passing required CI. One platform shard hit an ephemeral HTTP/2 EADDRINUSE test-port collision in unchanged code; a single provider-free rerun passed, preserving the original failure log at `/private/tmp/claude-v7-platform-workspaces-b-failure-20261015.log`. Final installed-package and remote execution gates remain pending.
Review corrections are committed and covered by failing-before/passing-after controls. Discovery changes invalidate stale onboarding probes. Finish setup now saves the exact configuration admitted during qualification rather than rebuilding Legacy defaults from an older render. V7 native input retains prepared transport, constraints, and completion-only second turns. The latter regression reproduces `session cannot start another turn` before the fix. All 160 focused protocol/session tests, all 71 New Agent tests, 148 onboarding tests, UI typecheck, and runner TypeScript build pass. Earlier focused verification also passed 343 selection/configuration checks, 849 skills/comment/native-session checks, 320 latest-master lifecycle/selection/snapshot checks, 148 platform-selection controls, 114 release controls, 16 image-consumer controls and token gates. Master's subsequently landed snapshot-worker test fix replaces the earlier local timeout change. All three heads have fresh 5/5 reviews with zero unresolved findings and passing required CI. One earlier platform shard hit an ephemeral HTTP/2 EADDRINUSE test-port collision in unchanged code; a single provider-free rerun passed, preserving the original failure log at `/private/tmp/claude-v7-platform-workspaces-b-failure-20261015.log`. The final packaged Mac pair subsequently passed; current remote execution gates are recorded below.
Preview: http://127.0.0.1:3105/CLA/issues/CLA-2, data `/private/tmp/claude-master-local-20261015`. Historical preview 3115 is separate. Ordinary New Agent -> Claude Code -> saved Anthropic API key -> Model Default -> high effort -> Finish setup originally saved agent `5339638d-b41d-4a6f-a041-4e5d1698a8de` as `paperclip_runner`/`acpx`/`claude`/`claude-opus-5-5`. No runner/model was patched after creation. Only timeout (180 seconds) and concurrency (one) were bounded. Experimental native runner remains false. The isolated PATH uses the already-qualified Claude 2.1.286 executable, preserving the user's system CLI.
@@ -143,7 +160,35 @@ The two subsequent ARM attempts failed before provider startup and remain pendin
Credential-free reproduction proved that BSD tar emits non-contiguous directory entries. GNU tar restored read-only directory permissions before later children arrived. The staging fix delays GNU directory restoration, omits macOS metadata, and verifies existing immutable context bytes and modes before reuse on follow-up. It keeps native provider sync behavior and durable checkpoint handling. Corrupt or symlinked targets do not become accepted cache hits. The regression failed before the fix; all 15 focused controls now pass with both GNU and BSD tar, including repeat staging and corrupt/symlink rejection. The actual patched function also passes twice on the unprivileged ARM host with zero provider calls; receipt `/private/tmp/claude-v7-arm-staging-fixed-20261015.json`. Remote live acceptance remains incomplete; no paid retry runs automatically.
Current allocation receipt `/private/tmp/claude-v7-remote-allocation-20261015.json` preserves the $50 aggregate ceiling and all historical holds. Known rate-card estimates total $4.090164800, excluding pending/unpriced charges and setup reservations. The current remote fixture has a $1.22 cap with its new $0.50 pending hold. Live subscription acceptance remains explicitly unverified under the user's API-key direction.
Earlier allocation receipt `/private/tmp/claude-v7-remote-allocation-20261015.json` preserves the $50 aggregate ceiling and all historical holds. Known rate-card estimates total $4.090164800, excluding pending/unpriced charges and setup reservations. The current remote fixture has a $1.22 cap with its new $0.50 pending hold. Live subscription acceptance remains explicitly unverified under the user's API-key direction.
The staging repair is frozen at `0670b85392a0be51745f1ba7ca84f998bf122c07`. Its source CI passed; Greptile reviewed this exact commit at 5/5 with zero unresolved threads. [Package campaign 38751544488](https://github.com/paperclipai/paperclip/actions/runs/38751544488) passed the Linux and Apple Silicon producers, then was cancelled after the Intel worker stalled in checkout for 17 minutes. Its artifacts and checkout log remain preserved. A single infrastructure retry, [38756070238](https://github.com/paperclipai/paperclip/actions/runs/38756070238), uses the same frozen source and one consistent producer attempt. Its Intel checkout completed normally; all three other producers passed. All four producers subsequently passed; package assembly is queued for a hosted Linux worker, followed by clean consumers. No provider retry occurred. The new ARM pack differs from the previous payload only in `runnerSourceRevision`; its new manifest digest is `sha256:9b4dfddca028ac8befcc84b9cbe724553531d991386f80329b4bdcab3890874d`. Daemon bytes remain SHA256 `4720c6a23784d3b4535329ecf852d38af8c14ae920447a662a00c4657711387a`, and the pinned Claude executable remains SHA256 `0292fa22ac2fd43e16be9d0e511ddd8347280d6e0ebaca744ef5b27e05d8d0f8`. The verified new pack is installed in the dedicated QA user's standard location; receipt `/private/tmp/claude-staging-arm-activation-20261015.json`. The previous user pack and the system/root installations remain preserved.
The prerequisite [#15994](https://github.com/paperclipai/paperclip/pull/15994) and Apple Silicon default [#15993](https://github.com/paperclipai/paperclip/pull/15993) are ready for review, with final packaged Mac evidence in their bodies. Neither is merged. The platform PR [#15997](https://github.com/paperclipai/paperclip/pull/15997) remains draft until its final remote task/revision acceptance passes. The installed Mac acceptance report and screenshot are attached to the completed preview task as `beec99dd-a9b3-40bb-84f5-6cae7c867562` and `fa8f2a00-d18b-444b-802e-11ad25166055`; both were downloaded again and verified after upload.
The fresh remote fixture `99824208-435b-4c44-8f63-829fc79faad9` (CLAAA) has a $1.40 shared cap. It receives $0.72 of unused allowance from held fixture `4389d6ab` (retaining its $0.50 hold) and $0.68 from original QA `f1dd89ef` (retaining its known spend plus two conservative $8 unknown-charge reserves). Aggregate caps remain $50; every prior hold and the separate $1 setup reserve remain. Receipt `/private/tmp/claude-staging-fixture-allocation-20261015.json` verifies paused donors before and after the reduction. No hold was released and no task retry was submitted. Ordinary UI creation on the unchanged `bd74bf50` UI saved native Claude/Opus 5.5/high effort for ARM agent `7233d811-41f0-4059-9084-5eeea09da7dd` and x64 agent `3ba4cef6-bca9-4341-b9f2-88696e892dba`. Both existing CLI API-key checks succeeded in the selected environments; this is connection evidence, not native execution acceptance. Both agents are bounded to 180 seconds and one concurrent run. UI journal `/private/tmp/claude-staging-ui-20261015`; no runner/model switch occurred after onboarding.
The isolated database is backed up before package replacement at `/private/tmp/claude-master-local-20261015/instances/default/data/backups/claude-before-staging-0670b853-20261015-094815.sql.gz`, SHA256 `8620dc98360032f48767c0d273d98c4926f355bb7bde6ff802a82451f2fe80cc`. The first backup attempt used the configured default port and timed out; the successful backup binds the actual instance data directory and allocated port 54334. No existing instance or backup was replaced.
### Final installed ARM acceptance — 2026-10-15
The replacement campaign's public package assembly passed. Artifact `11985673906` is 109691616 bytes, ZIP SHA256 `de6810dca086f2320a52bc4dc3d06e6e07374a0e9246add568f017b56b1bef89`. All 18 archive digests were verified before installing into `/private/tmp/claude-installed-0670b853-preview/consumer`. The controller preview now reports frozen source `0670b85392a0be51745f1ba7ca84f998bf122c07`, with the native experiment disabled. All four installed daemon hashes match the release manifest and are unchanged from the prior packaged Mac proof. The installed ARM provider-pack authority exactly matches the dedicated host's standard installation. All four native builds and all clean consumers have now passed in [campaign 38756070238](https://github.com/paperclipai/paperclip/actions/runs/38756070238). The matching x64 image is building in cloud via [38764436597](https://github.com/paperclipai/paperclip/actions/runs/38764436597); no local Docker build runs.
The ARM task and revision both succeeded on the final installed source: `47cf1e62-2023-4198-ba04-65564b56ecf8` and `3f07e836-acc0-4141-9daa-bd74013317ac`. Preview http://127.0.0.1:3105/CLAAA/issues/CLAAA-1. Independent downloads match the 114/157-byte oracles and SHA256 values recorded for the packaged Mac pair; the original remains unchanged. Task/provider/backend session identity is the same across turns, and high effort is requested and acknowledged on both. API accounting settled at $0.216398000 and $0.163966600 ($0.380364600 combined), with no pending or unpriced usage in this fresh company. Report `/private/tmp/claude-staging-installed-arm-proof-20261015/acceptance.json`; runtime identities `/private/tmp/claude-0670b853-installed-runtime-identity.json`. The earlier read-only staging failure is preserved and is not counted as passing evidence. The report and reloaded completed-task screenshot were attached and independently downloaded again (`13602319-1f90-4176-94c6-4789df9c630a`, `10c4c7fe-d745-4034-b356-b15565304a8e`). Reloading the agent also retained Claude/Opus 5.5/High settings; UI journal step 51. No runner or provider processes remained in the dedicated ARM account after both turns.
The [cloud image consumer](https://github.com/paperclipai/paperclip/actions/runs/38764436597) passed. Immutable image `ghcr.io/paperclipai/paperclip-daytona-runner@sha256:d241ab13d48297a0fe85b04acf70980a427a5fbcd4a95ca2284c7022cd29d03c` contains the exact installed Linux x64 pack `sha256:6adcfdf1a90d922fb65d647e13936eafff7d9cc5f0142ebdfbc11d9470b00c55` and daemon SHA256 `cd46aa8e809f4f3a4f705bfdc93b17b1466c010235f37b0b2f71572f927b5f8f`. Evidence artifact `11989266220` ZIP SHA256 `f6878e2c8e97d0f9d0dc8bfc8af9ea4b255d1f5b92ea6323cd5a983c6e003f34` was independently verified, as were source and inner-image identities. The dedicated x64 environment now uses this image with its saved credential reference preserved. No local Docker or Rust build ran.
### Final installed x64 acceptance and accounting — 2026-10-15
The final x64 pair succeeded: `65f04a06-c2a4-4c3c-af4a-2f6026563623` and `9c22a214-c9f3-436c-a509-0ddf763c66e6`, using native Claude/Opus 5.5/high effort with the experimental gate off and no runner/model override. Preview http://127.0.0.1:3105/CLAAA/issues/CLAAA-2. Independent downloads match 114/157 bytes and the exact SHA256 values above; the original is unchanged. Task/provider/backend identity persists across two distinct Daytona sandboxes. Both runtime records confirm provider ingress, the selected environment, approve-all permissions, and high-effort acknowledgment. Both leases bind the verified immutable image and report successful cleanup. API accounting settled at $0.212391400 and $0.127695800 ($0.340087200 combined).
Report `/private/tmp/claude-staging-installed-x64-proof-20261015/acceptance.json` includes actual environment leases, immutable image identity, and cleanup evidence. The report and reloaded completed-task screenshot are attached as `7276f35e-aad9-4be0-8e51-330cd9ad0904` and `20b33105-1229-47ec-b50f-ce88745b3fcf`; both passed independent download comparison. Reloading agent settings retains Claude/Opus 5.5/High; UI journal step 61. The ARM transport addendum `/private/tmp/claude-staging-installed-arm-proof-20261015/execution-targets.json` verifies both actual SSH leases, selected account, standard pack reuse, and permissions from retained runtime records; both leases are released and no provider/runner processes remain in that account.
Final read-only accounting receipt `/private/tmp/claude-staging-accounting-current-20261015.json` records $4.810616600 known rate-card estimates, $50 aggregate caps, zero active runs, retained unknown charges and setup reserve, and no released holds. The fresh remote company has $0.720451800 observed usage, no pending/unpriced usage, and a $1.40 cap. One older retired fixture retains its prior policy evidence because its current instance is unavailable; its entire $9 allowance remains reserved. The original failed records, unknown-charge holds, previous package attempts and source identities remain intact. No additional paid turns are required for this handoff.
---