Repository navigation
Conversation
df4c766 to
078cf6f
Compare
hidb4ai
left a comment
hidb4ai
left a comment
There was a problem hiding this comment.
Two reproducible collector recovery issues remain: temporary outbox capacity pressure permanently stops reading the affected turn, and pending attachments from a disabled workspace can block delivery for other enabled workspaces. Details are inline.
| outbox.finish(row["id"], state="done", error="image_consent_changed") | ||
| except (ValueError, TypeError, KeyError, OSError, AttributeError, RecursionError) as error: | ||
| code = str(error) if str(error).startswith("image_") and len(str(error)) < 80 else "image_read_unavailable" | ||
| outbox.finish(row["id"], state="error", error=code) |
There was a problem hiding this comment.
[P2] Keep temporary outbox capacity pressure recoverable
When the shared 128 MiB staging budget fills, Outbox.advance() raises image_outbox_full, and this handler permanently marks the turn's reader as error. Outbox.jobs() subsequently selects only watching jobs, while reset_delivery() resets attachments but never restores the reader. Thus an offline backlog can cause all remaining images in that turn to be omitted even after uploads free space or the user explicitly retries. I reproduced this with a scaled capacity limit: the first attachment was staged and then uploaded, but the second remained unread after capacity was freed and delivery reset. Keep the saved cursor recoverable for capacity pressure so reading can resume without changing its consent boundary.
| for row in outbox.pending(): | ||
| try: | ||
| if not _permitted(json.loads(row["job"]), settings): | ||
| continue |
There was a problem hiding this comment.
[P2] Allow enabled workspaces to progress past disabled deliveries
Outbox.pending() selects the oldest eight due attachments globally before this consent check. If workspace A accumulates eight pending attachments while offline and capture is then disabled for A, every pass selects and skips those same rows; they remain pending and due forever. Images from an enabled workspace B are consequently never selected. I reproduced this with eight staged attachments for A and one for B: after disabling only A, three delivery passes made zero uploads and B remained pending. Preserve A's consent and staged data, but ensure ineligible rows cannot monopolize the delivery page, for example by suspending them or advancing through a bounded eligible selection.
9090294 to
f71748a
Compare
|
Before moving this capability toward merge, I would like us to review the design through an RFC:
The feature has a useful direction, but I would prefer to settle these architectural choices in the RFC before merging it. |
|
Thank you for the architectural feedback. I agree that this capability should go through RFC review first. I’ve opened RFC #1972, linked to #1932, for design review. It compares the Source/Artifact boundaries, evaluates existing binary storage and shared processing contracts, and gives a table-by-table analysis with a proposed three-table design. The proposal keeps idempotency, recovery, historical evidence and concurrency guarantees explicit. I’ll keep #1956 as a draft prototype and wait for the RFC to be reviewed and accepted before proceeding with further implementation changes. After that, I’ll align the implementation with the agreed design and split it into focused PRs where appropriate. |
|
please update & resolve conflicts |
eca666c to
072bcfc
Compare
072bcfc to
3fde6d5
Compare
| ExperimentalImageTaskPage, | ||
| ExperimentalListImageTasksRequest, | ||
| ) | ||
| from powercontext.paths import powercontext_data_dir |
There was a problem hiding this comment.
[P1] powercontext[builtin] 安装后无法导入内置运行时
builtin/runtime/application.py:151 在模块层导入 powercontext.builtin.images.service，该文件第 64 行导入 powercontext.paths，而 src/powercontext/paths.py 在模块层执行 from platformdirs import user_data_path。pyproject.toml 的 builtin extra 没有声明 platformdirs（只有 server 与 cli 声明），images extra 又只组合 powercontext[builtin] + pillow。
后果：pip install "powercontext[builtin]" 之后任何 import powercontext.builtin.runtime 都抛 ModuleNotFoundError: No module named 'platformdirs'，powercontext[images] 同样不可用。
证据：
- CI 同签名失败：
isolated-installations (3.11, builtin)job 119297345299 与isolated-installations (3.14, builtin)job 119297345844，栈为ci_role_smoke.py:31→builtin/runtime/__init__.py→builtin/runtime/application.py:151→builtin/images/service.py:64→paths.py:22→ModuleNotFoundError: platformdirs。 - 同一
Main工作流在 base2d11abd6上 success（run 39602208746），即该 job 在 base 是绿的。 - 静态核对模块级导入闭包：base
2d11abd6的powercontext.builtin.runtime闭包不含powercontext.paths，本 head 含（唯一新增路径就是builtin/images/service.py）。所以缺失依赖是本 PR 引入，不是既有形态。
修复方向（二选一）：
- 在
builtinextra 增加platformdirs>=4,<5（与server/cli一致）； - 或者让 images 不依赖
paths：把数据目录通过ImageConfig/BuiltinConfig注入，或在service.py内延迟导入。
无论选哪条，请保留 isolated-installations (x, builtin) 的覆盖——它正是这次唯一抓到该问题的门禁；tests (x, unit) 与 quality 都因为工作区已装 server 依赖而看不到它。
This branch has not been deployed
3 participants
Which issue or RFC does this PR close?
Refs #1932 — image evidence foundation and initial Codex integration. Managed-local model provisioning is a separate follow-up; this PR does not close the entire proposal.
Based on master
2d11abd6ab24592427cc4a42132101ef77884245. Image evidence remains experimental and disabled by default.Rationale for this change
Screenshots and diagrams contain useful project evidence that text capture misses. Preserve the original image, save a validated interpretation as immutable Source evidence, and let extraction and recall use the derived text with a link to the exact original material.
What changes are included in this PR?
setup codex/doctor codexintegration.setup/doctorinstaller contracts, including explicit--capture-images/--no-capture-imagesoptions. Background infrastructure apps retain their restricted route surface; image errors use the shared HTTP error boundary.Are there any user-facing changes?
Images remain disabled by default. After optional configuration and normal host setup, users start the existing Server; a separate image env file or manually launched collector is unnecessary. Plugin activation and Hook trust remain host-required first-use steps.
Image receipt, interpretation, Memory extraction, and fresh-session recall are separate outcomes. Existing Memory governance is retained; no image-specific confirmation UI is introduced. The initial backend is file-backed SQLite with static PNG/JPEG. Older experimental databases require an explicit upgrade and paired DB/assets backup. Exact historical evidence must remain readable. Other storage backends, image embeddings, and joint multi-image reasoning are outside this PR.
How was this change tested?
Validation on Windows / Python 3.14, against master
2d11abd6:pnpm --dir desktop lint,test, andbuild: passed, including 32 UI tests. The current Desktop OpenAPI generator ran with Java 21; generated Rust models match the base branch.git diff --check: passed.httpxdependency for the images extra; no additional package version upgrades.uv run --no-sync ty check --python-platform linux: passed. This is a Linux-targeted static check on Windows, not a claim that native Windows whole-tree type checking passes.GitHub CI runs independently for the pushed head. No workflow or acceptance assertion is disabled.
Prior manual evidence includes real Codex CLI execution with controlled providers, an actual Desktop file-picker upload interpreted by MiniCPM-V 35B, and an Atomic Memory migration preserving existing Source identities and image assets. These were not rerun as part of the rebase.
Remaining acceptance and review limitations: real Atomic Memory background publication and fresh Desktop recall; paired DB/assets restore; drag/paste and remaining lifecycle variants; native macOS/Linux collector lifecycle; and complete external Open Code Review. These remain documented follow-up validation and are not claimed complete.
Details: acceptance status, implementation and validation log, desktop testing, contracts.
AI usage statement
OpenAI Codex assisted with design, implementation, rebase/conflict resolution, CI repairs, documentation, and testing. Open Code Review was used for a partial external-model review and delegate file/rule selection. The author is responsible for the implementation and will complete the documented acceptance and review before requesting merge.