Repository navigation
fix(rewind): plan closed tool exchanges and preserve notices - #1026
theaspirational wants to merge 12 commits into
Conversation
Co-Authored-By: OpenAI Codex <noreply@openai.com>
…d marker Review follow-up to the anima-research#737 notice-gap fix. - A tool result right after its call body is paired with it whatever the IDs, as before anima-research#737; IDs decide only when the walk crosses notices or other result bodies. - A rewind marker ends the walk, so it is never part of a rewound interval (the raw fallback removed it and the marker update then threw). - A result body in shards is paired as a whole body, at selection and at removal, not by the IDs of its first shards. - Tests: next-request regressions for mismatched, missing and mixed IDs (adjacent and across a notice, and adjacent on the raw record), an earlier result body's unknown ID, a marker between call and result, a passed-over sharded result body, and the adjacent sharded-body retract restored beside the notice form. - Comments and the changelog fragment state what is and is not paired. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of the 79 runs it could compare don't differ between the base and this head. In the files profiled hosts loaded, 5 changed source lines that map to code ran in no profiled drive: Parts left out where an arm's stack doesn't bring what they need: 2 runs couldn't be compared; the reasons are below. Usage: the figures of the agents' requests, subagents' included, in the 79 runs are the same in the base and this head: 210,343 tokens sent a request, 72.1% of them read from the cache, 139,465,910 written to it, 3,976,880 output, 24 thinking blocks dropped, USD 0.5622 at list price a priced request and 123 requests with no list price. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Runs that couldn't be compared (2)This head couldn't be compared at This head couldn't be compared at Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by file
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
|
Address Anima R1 selection, barrier, accounting and repeated shard-scan findings without changing retry or history-repair policy. Co-Authored-By: GPT-6.1 Sol <noreply@openai.com>
|
Addressed in commit
No retry/provider/history-repair policy change; adjacency still pairs corrupt/missing IDs, matching gaps still shed the complete signed call and results, notices within eligible intervals still leave, and presented rewinds still affect only the shedding window. No full suite or live provider verification; the exact commit and focused scope are stated above. Independent review is bounded source/synthetic acceptance, not provider or release qualification. |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of the 79 runs it could compare don't differ between the base and this head. In the files profiled hosts loaded, 4 changed source lines that map to code ran in no profiled drive: Parts left out where an arm's stack doesn't bring what they need: 2 runs couldn't be compared; the reasons are below. Usage: the figures of the agents' requests, subagents' included, in the 79 runs are the same in the base and this head: 210,343 tokens sent a request, 72.1% of them read from the cache, 139,465,910 written to it, 3,976,880 output, 24 thinking blocks dropped, USD 0.5622 at list price a priced request and 123 requests with no list price. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Runs that couldn't be compared (2)This head couldn't be compared at This head couldn't be compared at Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by file
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
|
Leave interval notice retention pending a maintainer decision; preserve legacy adjacent-ID compatibility. Co-Authored-By: GPT-6.1-sol <noreply@openai.com>
|
@nissa-seru — maintainer direction requested on The concern is valid: the current, documented call-to-result interval takes intervening notices with it. On CM 0.27.0, source review found selective presented Which contract should this PR target?
No choice has been implemented or presented as safe by default. The independent minor fixes are under separate review; I am not asking for merge while this should-fix remains unresolved. The question is about public source semantics, not a live-agent rollout. |
|
Implementation follow-up:
Checks: clean-commit build passed; focused caller/real-store regressions 442 pass / 0 fail / 0 skipped. Notice-consequence diagnostics 2 pass / 0 fail demonstrate the unresolved behavior, not a fix. No local full-suite, dependency-floor matrix or live-provider run on this candidate. Independent Astra-medium review accepted the delta after fresh build, 188 focused passes / 1 TODO and six extra contract controls; the PR remains held on finding 1, not approved to merge. No new notice-retention policy has been selected. The maintainer question remains open. Incidental effects: obsolete reasoning-loss rows now leave with matching legacy exchanges; hidden tool-call shards/unrelated mixed-body results now stop pairing and retain those intervals' notices. Neither resolves unrelated-notice retention. |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of the 79 runs it could compare don't differ between the base and this head. In the files profiled hosts loaded, 7 changed source lines that map to code ran in no profiled drive: Parts left out where an arm's stack doesn't bring what they need: 2 runs couldn't be compared; the reasons are below. Usage: the figures of the agents' requests, subagents' included, in the 79 runs are the same in the base and this head: 210,343 tokens sent a request, 72.1% of them read from the cache, 139,465,910 written to it, 3,976,880 output, 24 thinking blocks dropped, USD 0.5622 at list price a priced request and 123 requests with no list price. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Runs that couldn't be compared (2)This head couldn't be compared at This head couldn't be compared at Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by file
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
|
Keep notice retention and rewind budget/performance choices explicitly unresolved. Co-Authored-By: OpenAI Codex <noreply@openai.com>
|
Bounded follow-up:
Validation and exclusions: 206 pass / 0 fail / 1 TODO (207 tests, seven focused files), with a clean-HEAD build gate; The retained author evidence includes the initial traversal-bound failure and correction without loosening its threshold. No full suite, floor/platform matrix, live provider/private history/config/credential probes, CM API work, or general retention redesign. Independent Astra-medium review of this exact SHA accepted the bounded delta, separately reporting a fresh build, 206 focused passes / 1 TODO and eight complementary controls. It explicitly keeps the PR HELD on |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of the 79 runs it could compare don't differ between the base and this head. In the files profiled hosts loaded, 7 changed source lines that map to code ran in no profiled drive: Parts left out where an arm's stack doesn't bring what they need: 2 runs couldn't be compared; the reasons are below. Usage: the figures of the agents' requests, subagents' included, in the 79 runs are the same in the base and this head: 210,343 tokens sent a request, 72.1% of them read from the cache, 139,465,910 written to it, 3,976,880 output, 24 thinking blocks dropped, USD 0.5622 at list price a priced request and 123 requests with no list price. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Runs that couldn't be compared (2)This head couldn't be compared at This head couldn't be compared at Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by file
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
|
|
Additional feasibility information for the open notice-retention question, without changing the patch or assuming a maintainer decision: A colleague's bounded public-source check of Chronicle v0.6.0 ( Consequently, putting N existing redactions inside a CM Given that cost, my preference for this PR is option B: preserve unrelated notice bodies with selective presented retraction, and explicitly decline the affected legacy-raw operation when that preservation cannot be represented safely. Do not force presentation conversion after failed write-down or silently report a successful shed. Whole-body and malformed mixed-body limits still need to be stated. Raw deletion is an existing fallback, not the goal in itself. This remains a proposed contract for your direction, not an implemented resolution of the blocker or permission to merge. |
|
R4 dispositions for
A significant compatibility choice: a strict closed-exchange safety claim cannot quietly preserve the old adjacent-whatever-IDs best-effort exception. Under this B proposal corrupt/missing/ambiguous IDs decline; keeping that exception would need a separately explicit best-effort contract. Known complete normal-writer rounds are the intended recovery path, not universal repair of existing history. One exchange tombstone batch does not make refused-reply cleanup, markers and counters a cross-write transaction. Unknown/post-commit outcomes must stop rather than replay; later independent wakes are not durably suppressed by this small proposal. No new product code or tests, real-provider diagnostic/compression replay, storage-format change or live operation was run for this source-only plan. Maintainer direction on the reduced support/compatibility contract remains necessary before treating it as the PR's accepted behavior. |
Use selective presented-window removal, decline unsupported legacy or corrupt shapes, and reject operator recovery while a logical turn is already alive. Co-Authored-By: OpenAI Codex <noreply@openai.com>
Safe B round — R4 findings and the chosen reduced-support contractWe chose B rather than inclusive notice deletion or a new Chronicle format project. The implementation now declines unsupported exchanges explicitly; this includes retiring adjacent-whatever-ID best effort. Please request a different contract on review if preferred; I am not treating the earlier unanswered contract question as approval.
A subsequent independent review found a separate real Final implementation and independent reviewer each ran 169 recovery + 309 adjacent passes, with three separately identified reviewer-witness replays. The independent review reuses its identified full B review for unchanged code. These are bounded fake-provider checks, not a full-suite, live/provider, large-history, release or cross-write crash-atomicity claim. Unknown mutation outcomes stop rather than blindly replay; later independent wakes and arbitrary future writers are not durably fenced. |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of the 78 runs it could compare don't differ between the base and this head. Couldn't tell whether 41 changed source lines ran, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\ (the reason's first 263 of 1,341 characters). 37 changed source lines map to nothing the host runs: 37 of Parts left out where an arm's stack doesn't bring what they need: 3 runs couldn't be compared; the reasons are below. Usage: the figures of the agents' requests, subagents' included, in the 78 runs are the same in the base and this head: 210,168 tokens sent a request, 72.0% of them read from the cache, 139,145,027 written to it, 3,952,176 output, 24 thinking blocks dropped, USD 0.5656 at list price a priced request and 123 requests with no list price. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Runs that couldn't be compared (3)This head couldn't be compared at This head couldn't be compared at This head couldn't be compared at Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by fileCouldn't tell for any line or deletion not seen to run, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}","stack":"MembraneError: 400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}\n at errorFromProviderStatus (candidate/node_modules/
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
|
…lts; document manual recovery caveats Co-Authored-By: GPT-6.1-Sol <noreply@openai.com>
|
Answers to the automated review of
Build and typecheck pass; full suite 10,966 passed, 0 failed, 63 skipped, rerun by an independent review. Targeted recovery tests also pass on Membrane 0.8.0 and 0.9.0. Not run: the exact minimum-version CI install, hosted CI, Bun, and live providers. |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of the 78 runs it could compare don't differ between the base and this head. Couldn't tell whether 46 changed source lines ran, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\ (the reason's first 263 of 1,341 characters). 45 changed source lines map to nothing the host runs: 45 of Parts left out where an arm's stack doesn't bring what they need: 3 runs couldn't be compared; the reasons are below. Usage: the figures of the agents' requests, subagents' included, in the 78 runs are the same in the base and this head: 210,168 tokens sent a request, 72.0% of them read from the cache, 139,145,027 written to it, 3,952,176 output, 24 thinking blocks dropped, USD 0.5656 at list price a priced request and 123 requests with no list price. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Runs that couldn't be compared (3)This head couldn't be compared at This head couldn't be compared at This head couldn't be compared at Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by fileCouldn't tell for any line or deletion not seen to run, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}","stack":"MembraneError: 400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}\n at errorFromProviderStatus (candidate/node_modules/
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
Minor findings 8 to 25 (of 25)
|
…B rewind Co-Authored-By: GPT-6.1-Sol <noreply@openai.com>
|
Answers to the automated review of
Build and typecheck pass; full suite 10,982 passed, 0 failed, 63 skipped. An independent review of the whole change reran the build and the full suite and added seven probes (duplicate, undone, foreign and hidden IDs, unresolved reuse, window barriers); it reused the exact-commit typecheck. Not run: live provider traffic. |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of 1 of the 78 runs it could compare differ between the base and this head. Couldn't tell whether 64 changed source lines ran, or whether 1 deletion was reached, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\ (the reason's first 263 of 1,341 characters). 51 changed source lines map to nothing the host runs: 47 of Parts left out where an arm's stack doesn't bring what they need: 3 runs couldn't be compared; the reasons are below. Separately, 3 of the checks' flags differ between the base and the head, in 1 run; they're listed under "Checks' flags that differ", below. Usage: the agents' requests, subagents' included, in the 78 runs, base → head: 210,168 → 210,222 tokens sent a request, 72.0% of them read from the cache, 139,145,027 → 139,145,919 written to it, 3,952,176 → 3,950,756 output, 24 thinking blocks dropped, USD 0.5656 → USD 0.5658 at list price a priced request and 123 requests with no list price. The figures of 1 run moved, for any kind of request; they're under "Usage that differs", below. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Context in the 78 runs, base → head. Seams: 776 in both, moved in no run. Tokens in the parts they changed or took out: fewer in 1 run; 56,754,168 → 56,753,488 in all. The most one agent request held: lower in 1 run; 351,214 tokens at most in both. The summariser's requests: 1,841 in both, moved in no run. Their list price, leaving out 127 requests with no list price: USD 347.0521 in both, moved in no run. What these count. No model read these records for this view, which started on its own. A view asked for with a Runs whose records differ: 1 of 78
Runs that couldn't be compared (3)This head couldn't be compared at This head couldn't be compared at This head couldn't be compared at Checks' flags that differ (3)
Usage that differs: 1 of 78 runs, base → head
The figures of the other 77 runs are the same in the base and this head. Context that differs: 1 of 78 runs, base → head
Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by fileCouldn't tell for any line or deletion not seen to run, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}","stack":"MembraneError: 400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}\n at errorFromProviderStatus (candidate/node_modules/
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
|
…contained every selected body Co-Authored-By: GPT-6.1-Sol <noreply@openai.com>
|
Answers to the automated review of
Build and typecheck pass; full suite 10,998 passed, 0 failed, 63 skipped. A public probe of the late-result case failed in both presented and legacy modes before the fix and passes now. An independent review reran that probe and the earlier witnesses and checked the lines above. Not run: live provider traffic. |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of 1 of the 78 runs it could compare differ between the base and this head. Couldn't tell whether 105 changed source lines ran, or whether 1 deletion was reached, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\ (the reason's first 263 of 1,341 characters). 94 changed source lines map to nothing the host runs: 11 of Parts left out where an arm's stack doesn't bring what they need: 3 runs couldn't be compared; the reasons are below. Separately, 3 of the checks' flags differ between the base and the head, in 1 run; they're listed under "Checks' flags that differ", below. Usage: the agents' requests, subagents' included, in the 78 runs, base → head: 210,168 → 210,222 tokens sent a request, 72.0% of them read from the cache, 139,145,027 → 139,145,919 written to it, 3,952,176 → 3,950,756 output, 24 thinking blocks dropped, USD 0.5656 → USD 0.5658 at list price a priced request and 123 requests with no list price. The figures of 1 run moved, for any kind of request; they're under "Usage that differs", below. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Context in the 78 runs, base → head. Seams: 776 in both, moved in no run. Tokens in the parts they changed or took out: fewer in 1 run; 56,754,168 → 56,753,488 in all. The most one agent request held: lower in 1 run; 351,214 tokens at most in both. The summariser's requests: 1,841 in both, moved in no run. Their list price, leaving out 127 requests with no list price: USD 347.0521 in both, moved in no run. What these count. No model read these records for this view, which started on its own. A view asked for with a Runs whose records differ: 1 of 78
Runs that couldn't be compared (3)This head couldn't be compared at This head couldn't be compared at This head couldn't be compared at Checks' flags that differ (3)
Usage that differs: 1 of 78 runs, base → head
The figures of the other 77 runs are the same in the base and this head. Context that differs: 1 of 78 runs, base → head
Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by fileCouldn't tell for any line or deletion not seen to run, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}","stack":"MembraneError: 400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}\n at errorFromProviderStatus (candidate/node_modules/
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
|
…l-ID walk and teardown fixes Co-Authored-By: GPT-6.1-Sol <noreply@openai.com>
|
Answers to the automated review of
Typecheck and 477 relevant tests pass, run by an independent review that also checked each line above against the code and confirmed: only a finished stream counts, recovery stays tied to the same attempt, the witness 400 is one the policy admits with a live-stream negative control, and the widening is stated. A real streamed 400 fails on the previous head and passes now. Not run: live provider traffic. |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of the 78 runs it could compare don't differ between the base and this head. Couldn't tell whether 106 changed source lines ran, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\ (the reason's first 263 of 1,341 characters). 121 changed source lines map to nothing the host runs: 11 of Parts left out where an arm's stack doesn't bring what they need: 3 runs couldn't be compared; the reasons are below. Usage: the figures of the agents' requests, subagents' included, in the 78 runs are the same in the base and this head: 210,168 tokens sent a request, 72.0% of them read from the cache, 139,145,027 written to it, 3,952,176 output, 24 thinking blocks dropped, USD 0.5656 at list price a priced request and 123 requests with no list price. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Runs that couldn't be compared (3)This head couldn't be compared at This head couldn't be compared at This head couldn't be compared at Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by fileCouldn't tell for any line or deletion not seen to run, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}","stack":"MembraneError: 400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}\n at errorFromProviderStatus (candidate/node_modules/
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
|
|
Answers to the automated review of
Typecheck passes. The held-draft regression and the focused lifecycle and marker tests pass. An independent review checked each line above against the code, reran the held-draft test against the previous source (it fails at the notice assertion), and confirmed again: only a finished stream counts, recovery stays tied to the same attempt, the admitted 400 has a live-stream negative control, and the widening is stated. Full-suite and live provider runs are not claimed here. |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of the 78 runs it could compare don't differ between the base and this head. Couldn't tell whether 109 changed source lines ran, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\ (the reason's first 263 of 1,341 characters). 140 changed source lines map to nothing the host runs: 11 of Parts left out where an arm's stack doesn't bring what they need: 3 runs couldn't be compared; the reasons are below. Usage: the figures of the agents' requests, subagents' included, in the 78 runs are the same in the base and this head: 210,168 tokens sent a request, 72.0% of them read from the cache, 139,145,027 written to it, 3,952,176 output, 24 thinking blocks dropped, USD 0.5656 at list price a priced request and 123 requests with no list price. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Runs that couldn't be compared (3)This head couldn't be compared at This head couldn't be compared at This head couldn't be compared at Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by fileCouldn't tell for any line or deletion not seen to run, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}","stack":"MembraneError: 400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}\n at errorFromProviderStatus (candidate/node_modules/
Held the same in base and head: connectome-host |
…ine repair undone Co-Authored-By: GPT-6.1-Sol <noreply@openai.com>
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
Minor findings 7 to 24 (of 24)
|
…d the witness after failed speech delivery Co-Authored-By: GPT-6.1-Sol <noreply@openai.com>
|
Eighth automated-review pass, now at
An independent review ran the full suite at this head: 11,036 pass, 0 fail, 63 skipped. It ran the nine public-start regressions, all passing, and three of its four control cases fail on the previous head as expected. It also checked the four recovery bounds from the description: only a cleanly finished stream counts, the removal stays tied to the same failed attempt, the witness 400 is one the shed policy admits, and the widening is stated. Not tested: real provider streams, and the dependency-floor CI matrix. |
|
/review |
|
Automatic context view of What a context view runs and what its checks judge: CHECKS.md. The records of the 78 runs it could compare don't differ between the base and this head. Couldn't tell whether 128 changed source lines ran, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\ (the reason's first 263 of 1,341 characters). 154 changed source lines map to nothing the host runs: 12 of Parts left out where an arm's stack doesn't bring what they need: 3 runs couldn't be compared; the reasons are below. Usage: the figures of the agents' requests, subagents' included, in the 78 runs are the same in the base and this head: 210,168 tokens sent a request, 72.0% of them read from the cache, 139,145,027 written to it, 3,952,176 output, 24 thinking blocks dropped, USD 0.5656 at list price a priced request and 123 requests with no list price. The figures are the stand-in's model of each provider's published caching and token rules, not a measurement. Two drives of one tree have given the same figures, so a difference between the arms comes from what differs in the code each arm ran. A list price is what each arm's host prices those figures at with its own table (connectome-host's Runs that couldn't be compared (3)This head couldn't be compared at This head couldn't be compared at This head couldn't be compared at Upgraded runs beside their fresh twins
A run that writes its own history first, as a fresh twin does, is shown without that history's requests, so its figures are its workload's alone, as the upgraded runs' are. Changed source lines and deletions not seen to run, by fileCouldn't tell for any line or deletion not seen to run, since H6/continuation: its profiled drive failed at step c.debt: the turn's first inference failure (inference:failed, which agent-framework may still have retried): {"type":"inference:failed","agentName":"fixture","error":"400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}","stack":"MembraneError: 400 {\"type\":\"error\",\"error\":{\"type\":\"invalid_request_error\",\"message\":\"prompt is too long: 200634 tokens > 200000 maximum\"}}\n at errorFromProviderStatus (candidate/node_modules/
Held the same in base and head: connectome-host |
|
Automatic review of About these findingsThe findings below came from reading this PR and relevant code; not a holistic understanding of the project's functional intent and its architecture choices. Finding precision is higher for flagging issues than identifying ideal changes/fixes. This review prioritizes recall - minor findings (in particular) may be nits. Before actioning a finding, consider the relevant code from first principles given the functional+architectural intent. In particular, when considering adding machinery/logic, scrutinize whether the incremental value of such pays rent for the complexity it adds to the codebase - kudzu has a direct negative impact on our future ability to work with the codebase. Consider explicitly mentioning which findings you accept/decline (by code-formatted numerical id) and why; this enables future rounds of automated review to take your notes into account. Findings, most severe first:
|
Problem and scope
Fixes #737 using Safe B: remove only complete, unambiguous eligible whole
bodies/exchanges, not single tool endpoints or intervening notices.
This explicitly widens the supported path relative to the published
“active-stream terminal failure stops unchanged” behavior. A registered stream
may recover its own qualifying exhausted provider failure only after clean
iterator completion, successful queued delivery and terminal teardown. An error
event, elapsed silence, or caught delivery rejection is not that boundary.
Round-eight corrections
the initial served raw-body set is extended by each verified submitted tool
round and its supported injected messages, then snapshotted at terminal
failure before waiting. Later unsent arrivals are not witnesses. Unverifiable
hand-offs and missing compile provenance cannot authorize quarantine.
but uncertainty disables automatic exhaustion quarantine for that Agent
instance. Stop/recreate only after inspecting uncertain external/store effects.
Qualifying pre-stream failures also await earlier streams' queued deliveries.
consolidation states both outcomes without requiring agent access to host logs.
refusal decline guidance and remaining misleading marker/cap/cause comments.
Soma's four bounds remain
streams still block recovery.
forced session and branch; logical quiescence; settled guard work; all Safe B
ownership, audience, body-closure, visibility and atomic-removal conditions.
Every selected body head must be carried by that failed request. A late
uncarried tool endpoint declines the whole plan without sacrificing older
history or consuming a step.
HTTP 400. Existing live-stream negative controls remain, with new independent
held-iterator/held-delivery controls through the public start path.
stated here and in docs/recovery-b.md.
Verification and remaining boundaries
Both required public-start regressions fail against the previous head: the two window-mode
request-body witnesses and rejected-speech control fail; successful-delivery
control passes. In the implementation run, typecheck and build passed and the 13-file relevant
suite passed 887 tests with 0 failures (28 suites).
No durable writer fence, cancellation/deadline protocol, full reentrant recovery
transaction or settlement split is introduced. Hung waits, forced-session-only
settlement loss, teardown/removal accounting and notification-driven session
changes remain explicitly retained limitations, not claimed fixed. Existing
supplemental timing-based guard/bounce controls are not deterministic scheduler
proofs. The reply on this pull request gives one disposition per finding. An independent review ran the
full suite at this head (11,036 pass, 0 fail, 63 skipped). External CI and the
complete dependency-line matrix remain outstanding.
Recovery is step-bounded, not history-size/memory bounded; exclusive offline
repair guidance and transaction/no-replay limitations in docs/recovery-b.md
remain. This is Safe B compatibility widening, not permission to repair malformed
history, replay side-effecting tools, bypass caps, or mutate successor history.