You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 515defb
Browse filesBrowse the repository at this point in the historyBrowse files
当前版本已经提供 Run 级结构化 Work Board 与 Notes:工作项负责可执行计划,Note 负责观察、假设、决策、摘要和来源引用,所有变更都与 Run 事件在同一事务提交。Supervisor 使用 8192 token 的独立记忆预算选择摘要、活跃工作项和当前 root 可见 Note,并把来源 ID 与 token 估算记录在 `model.started`;Note 正文不会写入模型事件。仍有活跃工作项时,模型不能自行 `finish`。
@@ -291,6 +293,8 @@ The v56 local release gate passed full ordinary/race suites (178.5s/181.3s), vet
291
293
Schema v57 adds a separate `sandbox_docker_host_input_staging.v1` rehearsal while the v56 container remains stopped. On Linux, `openat2` pins the workspace root and every read-only source tree with no-symlink, no-magic-link, beneath-only, and no-cross-device resolution. Each entry is preflighted with `O_PATH`, then only a matching regular file or directory is opened for content, so FIFOs and other special files fail before a potentially blocking open. Directory and single-file mounts are supported; symlinks, hard links, nested mounts, traversal, excessive depth, and bounded-resource violations fail closed. After all descriptors are pinned, inode, size, timestamps, and link counts are rechecked before a deterministic tar is assembled from sanitized source entries and exactly revalidated Run/Session/Workspace Artifacts. Filesystem-specific directory inode sizes are excluded from the content digest. The archive exists only in a `memfd`, receives write/grow/shrink/seal kernel seals, and is read back for digest verification. SQLite first stores an immutable intent bound to the current v56 attempt, container-ID fingerprint, plan, input digest, and lease generation, then stores counts and digests without paths, content, file descriptors, or raw container IDs. SQL blocks attempt completion while a v57 intent lacks evidence. `--stage-host-inputs` requires a separate `--confirm-host-input-staging`; failure performs best-effort stopped-container cleanup, and takeover can resume without another create. The sealed bundle is not yet handed to the Docker daemon, so `daemon_consumed=false` and `execution_evidence=false`: v57 closes the descriptor-capture race but is not production proof of bytes consumed by a future container and grants no start, exec, network, export, or Artifact authority.
292
294
The v57 local release gate passed the final code's full ordinary/race suites (155.0s/168.1s), vet, zero-warning staticcheck, module verification, zero-finding govulncheck, 17 frontend tests, OpenAPI drift, production build, zero-vulnerability npm audit, credential/runtime-artifact/process-entry scans, diff checks, and an isolated real-binary schema-v57 smoke. Ten Application staging rounds, twenty two-Store independent-ID contention rounds, and ten race contention rounds pass; the Linux test binary cross-compiles. Focused coverage includes rename/replacement/deletion/symlink/hard-link/FIFO/single-file mounts, bounded directory enumeration, cancellation, replay, lease takeover, migration, and privacy. The audit fixed single-file report/SQL constraints, cross-filesystem directory-digest stability, special-file pre-open blocking, pre-limit directory allocation, read cancellation latency, random-ID idempotency conflicts, and failure-ledger consumption after omitted confirmation, with no unresolved high- or medium-severity issue. One known low-risk recovery gap remains: v57 opt-in is not durable before v56 stage, so a process exit after that checkpoint but before the v57 intent requires the staging flags to be resubmitted and can otherwise finish an optional v56-only rehearsal. This cannot start a container or widen authority; v58 must close the window first.
293
295
296
+
Initial GitHub Actions run `29395980413` exposed only an invalid Linux single-file test fixture: it changed the sole mount target to a file while retaining a directory working directory, which the Manifest correctly rejected; production code did not fail. The corrected fixture retains one read-only directory mount covering the working directory and adds a separate file mount, exercising the intended `directory_count < read_only_mount_count` case. Run `29396264276` passed commit `8719dff`; Go/Linux and TypeScript completed in 3m55s and 23s.
297
+
294
298
The v47 release audit also hardens Supervisor concurrency recovery and budget accounting. Phase-transition replay rechecks the idempotency operation after observing the target state, cancellation tests wait for the Provider call to begin, and every root model call with a durable `model.started` fact is charged at least 1ms so a `999/1000ms` deadline cannot be entered repeatedly. This changes no schema, grants no capability, and preserves the interpretation of historical Specialist ledgers. GitHub Actions run `29325171043` passed every Go and TypeScript gate.
295
299
296
300
The current build includes a structured, Run-scoped Work Board and durable Notes. WorkItems hold executable plans, while Notes hold observations, hypotheses, decisions, summaries, and source references. Every mutation commits with its Run event. The Supervisor selects summaries, active work, and root-visible Notes under a separate 8,192-token memory budget, then records source IDs and token estimates in `model.started` without persisting Note bodies there. Model-driven `finish` remains blocked while active work exists.
Copy file name to clipboardExpand all lines: docs/PROJECT_MEMORY.md
+2Lines changed: 2 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -78,6 +78,8 @@ Application verifies returned Artifact bytes and payload digest, stores only bou
78
78
79
79
Focused tests cover separate confirmation, default-disabled and unsupported behavior, deterministic replay, rename/replace/delete detection after pin, symlink/hard-link/FIFO rejection, single-file mounts, bounded directory enumeration, cancellation, report mismatch, cleanup-first failure, restart/takeover without a second create, stale-generation fencing, SQL completion gating, immutability, privacy, and v56-to-v57 migration. Final ordinary/race suites pass in 155.0s/168.1s; vet/staticcheck/module/govulncheck, 17 frontend tests, OpenAPI/build/npm audit, repository scans, isolated schema-v57 binary smoke, focused repetition, and Linux test-binary cross-compilation pass. GitHub Actions supplies the Linux runtime proof. The audit tightened root-parent symlink rejection, public report construction, Artifact byte/digest revalidation, stage-to-intent chronology, resource-limit errors, ambiguous confirmation, file-mount SQL/report constraints, filesystem-independent directory digests, special-file preflight, bounded/cancellable reads, independent-ID semantic convergence, and pre-acquire rejection of missing resume confirmation. No high/medium issue is currently known. The bundle is deliberately not passed to Docker, so this slice adds no execution usability and does not satisfy a future start gate.
80
80
81
+
GitHub Actions run `29396264276` passed commit `8719dff` with Go/Linux in 3m55s and TypeScript in 23s, providing the Linux runtime proof. The preceding run `29395980413` failed only because the single-file test fixture no longer covered its directory working path; the corrected mixed directory/file fixture now exercises the intended report constraint.
82
+
81
83
## Next Slice
82
84
83
85
Continue P6 through a separately gated schema-v58 daemon-owned immutable input handoff; do not reinterpret v57 capture evidence as bytes consumed by Docker:
Copy file name to clipboardExpand all lines: docs/PROJECT_STATUS.md
+2Lines changed: 2 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -692,6 +692,8 @@ Expected context behavior:
692
692
693
693
- Schema v57 adds a separately confirmed, default-disabled descriptor-pinned host-input capture between v56 stage and cleanup. Linux `openat2` pins no-symlink/no-magic-link/beneath/no-cross-device trees; `O_PATH` preflight rejects FIFO/special files before content open, and both directory and single-file mounts are supported. Descriptor metadata is rechecked before a deterministic sanitized tar is sealed in `memfd` and reread. Exact Artifact payload metadata is independently rechecked by Application. SQLite binds immutable intent/result rows to attempt, container fingerprint, plan, input digest, requester, and generation; generated row IDs are excluded from semantic fingerprints, pending intent blocks completion, and missing resume confirmation is rejected before lease acquisition. Failure cleans the stopped container before releasing the lease, and takeover resumes without another create. Tests cover rename/replacement/deletion/symlink/hard-link/FIFO/single-file, bounded enumeration, cancellation, replay, report mismatch, cleanup-first recovery, stale-generation fencing, migration, immutability, privacy, CLI, and Linux cross-compilation. Final ordinary/race suites pass in 155.0s/168.1s; vet/staticcheck/module/govulncheck, 17 frontend tests, OpenAPI/build/npm audit, repository scans, focused repetitions, and isolated schema-v57 binary smoke are green. The audit additionally fixed single-file report/SQL constraints, filesystem-dependent directory digest input, special-file pre-open blocking, unbounded directory allocation, cancellation latency, independent-ID convergence, and failure-ledger consumption on omitted confirmation. No unresolved high/medium issue is known. The sealed bundle remains local and daemon-unconsumed, so no process or production authority has been added.
694
694
695
+
- GitHub Actions run `29396264276` passed commit `8719dff` with Go/Linux in 3m55s and TypeScript in 23s. Initial run `29395980413` exposed only an invalid single-file test fixture whose file target did not cover the retained directory working path; the corrected mixed directory/file fixture now validates the intended report constraint. Production behavior was unchanged.
696
+
695
697
## Recommended Next Slice
696
698
697
699
Continue P6 with a separately gated schema-v58 daemon-owned immutable input handoff. First persist the staging choice before v56 daemon stage, closing the low-risk crash window where a post-stage/pre-v57 exit currently requires the operator to resubmit both staging flags and can otherwise finish an optional v56-only rehearsal. Then choose a closed Docker mechanism that consumes the exact v57 bundle without making the target input writable, broadening the fixed endpoint/API allowlist, accepting arbitrary archive destinations, pulling/building unreviewed images, or starting the target container. Bind a daemon-consumption receipt to the v57 digest and current generation, prove crash/retry/cleanup/collision behavior, and keep every execution/backend/Artifact flag false. Only after that gate should termination, running-container orphan recovery, output collection, and atomic Artifact commit be designed. HTTP/React mutation, Rust analyzers, broader network/secret support, real end-user execution, and CTF solving remain deferred.
0 commit comments