Repository navigation
Conversation
Keep existing lineage and every contributing producer root across response observation, detached delivery and shunt aggregation. Refuse unknown automatic context and retain unacknowledged reports. Native bootstrap and direct MCP publication remain unqualified. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Load the isolated home before response-test imports and apply the formatter pinned by CI. Preserve every test assertion and the existing native-client qualification requirements. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Recognize the existing Cline host marker and bind the two consumer fixtures before invoking real hooks. Check the current claim filename in the delivery cleanup assertion and state the raw session identity requirement. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep delivery classifications and reviewed source fingerprints aligned with the existing-family plugin wrappers. Preserve all sender admission assertions and unrelated inventory records. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Bind existing task families before parent capture and repository requests. Retain queued prompt roots and recheck them before context publication or acknowledgement. Preserve unknown and legacy evidence without assigning a clean family. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Observe every contributing root before returning a response, then stop automatic consumers and file lookup acknowledgements when protection changes during the request. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep both durable family marks while checking the response refusal instead of dereferencing the withheld body. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Recheck retained protection after awaited work and before primer cache writes. Keep ordinary primer output and cache refresh covered alongside marked-response and sibling-mark suppression. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Recheck retained authority before publishing the local decision. Preserve ordinary health primer controls and require malformed JSON health to protect the family and stay silent. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Resolve local records through the process token SID and protected owner ACLs. Bound native jobs contain compiler descendants across caller death, and strict JSON checks retain the existing record contract. Preserve native positive and refusal assertions within the existing time limits. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Native creation fails before its ownership assertions. Preserve fixed phase and numeric failure evidence through the existing generic refusal without exposing raw helper output or changing its deadlines. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
A startup phase alone cannot distinguish a captured marker from the parser default. Record fixed entry boundaries and bounded capture counts on the existing launches while preserving their deadlines and authority checks. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Hosted calls reach script entry but time out before the first encoding assignment completes. Keep the same construction and assignment while recording their separate fixed boundaries. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Use the same typed constructors and arguments at the measured first-construction boundary while retaining diagnostics, authority checks and deadlines. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Use an exclusive claim before publishing an immutable alias. Keep existing privacy state when startup hooks overlap or a producer exits. Account Windows timer setup against the original absolute deadline and distinguish native timer termination from other helper failures. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Select the JSON alias before corrupting it so the existing refusal assertions still exercise damaged resume metadata. Resolve the plugin version conflict with main while retaining the unused 1.2.40 behavior version in both manifests. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Allow an explicit initial SessionStart event to establish its host-scoped family while retaining refusal for inherited, damaged and protected sessions. Extend the real hook-entry controls to both native clients. Add fixed Windows setup phases to locate native deadline failures without changing operation budgets, authority checks, jobs or cleanup. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Avoid copying a security descriptor and attempting directory creation for each existing ancestor. Missing directories still receive the private descriptor and the same pinned-handle validation. Keep operation deadlines, authority checks, job containment and exact cleanup unchanged. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Drain fixed helper diagnostic frames in the existing real witnesses and include observer elapsed time when startup fails. Keep raw stderr out of the report and bound the retained frames and numeric fields. Preserve the native workload, deadlines, process identity checks, forced termination and cleanup assertions. No product operation or timing limit changes. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The second interpreter and duplicate native bootstrap consume the bounded setup window. A suspended compiler joins its private job before execution. The guardian stops compiler descendants and removes their owned files before record access, retaining the original absolute deadlines. The existing shunt regressions include bounded server and child observations in failure messages without changing their assertions. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Capture fixed numeric compiler diagnostics after owned descendants stop. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Preserve both hook regression groups and bind the local-query fixture to its existing family. Check ordinary and marked startup publication through the native session harness. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Establish the contract and recall fixture aliases before prompt delivery. Preserve all request and output assertions, and format the plugin manifest with the repository formatter. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep fixed parsing, construction and encoding markers inside the existing budget. Record request entry and failure-only queue metadata in the shunt fixture without changing its assertions or timing. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep the checked claim after publication so loss of an alias cannot create a clean replacement session. Cover both startup hosts and the retained claim ACL. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Use the existing strict reader and fixed primitive responses to remove general PowerShell JSON commands from the bounded helper path. Require exact string fields and reject duplicate request keys while preserving stored-record semantics and native authority checks. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The two integration cases share a path and namespace. A row written by one case can enter the other case's exact result assertion when test order changes. Keep every assertion and pass a separate registered org through each case. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
5 tasks done
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
…in-existing-session-lineage Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Report only the bounded sample count and prefix category after the existing reader closes. Keep those observations outside phase and error selection, and restore the marked aggregate fixture's original call order. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Observe the existing root creation method and both checked record writes before readback. Diagnostic I/O stays private and cannot interrupt authority or cleanup; all native deadlines remain unchanged. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Flush both response streams after owned compilation cleanup before ending this process. Keep native watchdogs armed and preserve failure when a terminal call returns. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Keep fresh directory and file authority checks while removing redundant parent-creation attempts. The two private calls run while CreateRoot retains every parent; an unexpected missing parent refuses. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
A missing clean delivery leaves an owned claim without proving its producer. Record bounded control-validity and row counts after the original failure while preserving every assertion and successful path. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Recognize the fixed fatal-error prefix within the existing bounded numeric diagnostic scanner. Exercise the real PowerShell scanner for fixed frames, deduplication, cutoff and rejected text without exposing compiler output. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Retain the child-process assertion while reporting only an allowlisted error code. Both Windows runs fail before scanner observations, and raw child errors must stay private. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The scanner harness emits only allowlisted labels and checked integers. Build that fixed JSON directly so its native check does not need the general PowerShell serializer; keep the fixtures, assertions and child limits. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The lifetime witness fails before announcing its child. Record only whether the fixed expected pipe path and compiler code occur on one bounded output line; retain the refusal and cleanup. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The native protection call can stop after record dispatch without exposing whether it opened and flushed the state file. Emit two fixed progress markers while retaining the checks, durable write, refusal and cleanup. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Witness the assigned suspended compiler under the existing native job deadlines, then qualify real resumed compilation with a separate fixed-input control. Retain the original lifetime assertions and limits. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Record four fixed test-only phases around the assigned-process check and receipt construction. Retain the serializer, deadlines and termination assertions; the current failure stops before any receipt is announced. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The witness reaches receipt construction before exiting without an announcement. Keep its fields and lifetime assertions while removing the general JSON serializer from this fixed test receipt. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Record a bounded primary stage and error category before rethrowing. Keep cleanup assertions and native deadlines unchanged. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Cleanup timeouts can mask the last completed observation. Report only fixed boundary labels and distinguish directory equality from existence. Keep the existing assertions, cleanup commands and deadlines. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The directory assertion and cleanup observer mask different failures. Retain bounded comparison facts and distinguish completed input reads from JSON conversion without changing the assertions or deadlines. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Resolve the existing owned directory before deriving the compiler path used by the exact directory assertion. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Default resolution retains the short spelling in the expected path, while the compiler receipt uses the long spelling. Resolve the existing fixture directory through the native path before appending its child. The exact assertion and all lifetime budgets remain unchanged. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
An empty observation conflates missing or exited processes, changed identities and swallowed process-state errors. Retain bounded outcomes from the existing reads and call timing against the operation deadline. Assertions, observer codecs, process calls and budgets stay fixed. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
Retain the cutoff already supplied to the witness and capped clock facts on either side of the existing lookup statements. Distinguish a thrown conversion from an unreturned interval. Assertions, process calls, conversion expressions and budgets stay fixed. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The pre-death check started a fresh interpreter after the native deadlines were armed and could miss the compiler's live window. Have the existing witness recheck its exact compiler handle after capturing both identities, then require observed parent exit before the same cutoff. Keep the post-death absence checks and all budgets. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The parent relays the guardian's stderr and loses later cleanup phases when the test kills it. Inherit that existing stderr handle and report the validated phase ring on post-death failure. Keep the original assertions, exceptions and deadlines. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
The parent-death trace ends at the startup receipt. Emit one fixed phase after the existing exact-handle wait returns, and accept only its zero code in the bounded collector. Preserve assertions, waits, deadlines and cleanup. Signed-off-by: zlareb1 <lareb.zafar@clickhouse.com>
This branch has not been deployed
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
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.
Related: #920
Changelog category (leave one):
Changelog entry (one or two sentences a user of Loom can read without the diff; the release pull request collects these into the release notes):
Automatic coding-agent hooks retain their session family's privacy state across responses and detached delivery. Explicit Claude Code and Codex startup establishes one durable family before automatic work; unknown, damaged or protected families suppress capture, signup and the session primer.
Why
A response can make a session private after a hook starts or queues work. Hooks must retain their original contributors, observe response privacy before their consumer, and check those contributors before sending or acknowledging queued content. Selecting a clean family later must never release protected work.
Scope and risk
Hook entry binds the raw
session_idbefore normalizing payload fields or running automatic work. Only the installed Claude Code or Codex host'sSessionStartevent with exacthook_event_name: SessionStartandsource: startupestablishes a family. Codex follows its published lifecycle contract. Known contributors disable fresh bootstrap. Concurrent startup hooks reuse the same immutable alias; existing privacy marks never reset. Resume, fork, clear, compact, subagent and other adapter events resolve existing state. A payload containing onlysessionIdstays silent.Response, callback and job contexts retain every contributing root across awaits and detached dispatch. Conflicting known anchors cannot select a clean replacement, and a failed privacy write does not leave another known contributor unmarked. The Cline adapter captures its task family before prompt capture or repository refresh, carries producer roots through its queue, and checks again before stdout publication and delayed acknowledgement. Unattributed prompt files remain unused and preserved.
Shunt batches retain immutable producer roots and restore their union across sessions. Unknown or protected reports are not sent; failed claims stay unacknowledged without automatic replay. Cleanup preserves unattributed files and isolates queue names from earlier consumers. Older plugin versions can still expire older-format files.
Aliases use a private, exclusive regular-file claim whose descriptor identifies the created object. Claims remain alongside immutable aliases; cleanup closes descriptors without deleting a pathname. A complete alias remains usable after producer death; an unpublished claim stays unavailable and is never taken over. This adds one content-free permanent marker per alias. Receipt checks detect observed replacements but do not promise continuous pathname ownership against an arbitrary process with the same user's directory access.
Windows records require the process token owner, protected owner-only DACLs, checked non-reparse handles, single-link files and pinned NTFS ancestors. Strict request parsing and fixed response encoding run in the same compiled assembly as the record methods. Requests require exact lowercase keys and operations, string fields and no duplicate decoded keys. Stored-record duplicate handling is unchanged. Privacy writes append and flush before success.
Each Windows operation retains its 1500 ms primary budget and 2000 ms total ceiling, including the existing cleanup reserve. A PowerShell guardian creates the fixed compiler suspended, assigns it to a native job, and checks the absolute deadline before resuming it. Native timers contain the guardian, compiler and descendants after caller death. Compiler descendants stop before assembly bytes are loaded and owned files are deleted. The record operation runs under the same earlier absolute cutoff. After checked compilation cleanup and record-method return, the helper flushes its streams and terminates only itself with the fixed status. A returned success-termination attempt becomes failure and is never repeated. Root-record creation still checks every parent and record, writes and flushes both records, and refuses an unexpected missing parent.
Bounded diagnostics retain fixed phase labels, numeric errors and compiler codes, counts and format categories. They never return raw helper output, requests, paths, identities or stacks. The compiler scanner reads at most 8192 bytes and retains at most four deduplicated codes from fixed
fatal error,errorandwarningprefixes. Root and protection progress markers identify observed checked-write boundaries. Missing markers prove nothing because diagnostic I/O errors are suppressed; observed markers do not prove subsequent cleanup or overall success.Compatibility, rollout and rollback
1.2.42; generated Codex copies match the canonical source. Version uniqueness must be checked again before merge.Tests
Changed JavaScript passes syntax and formatting checks. The canonical generator and its check pass; its generated mirror is byte-identical. The latest test-only observation preserves all 80 current Windows assertions and all 15 observer, process-spawn and kill expressions in the authority test. Nine helper and observer functions remain byte-identical; the native phase collector adds only the fixed wait-returned phase to its two closed allowlists. Original deadlines and timeouts remain. This includes the previously counted 13 synchronous-observer and kill expressions plus two asynchronous spawns. The earlier pre-death cold-observer assertion is replaced by six receipt and cutoff assertions. Latest source inverses account for exactly one witness marker and its closed allowlist entries across two test fixtures. Small dependency-free test-reporter probes check diagnostic visibility; no local plugin suite, PowerShell execution, native compilation or application execution ran. Full exact-head hosted CI and a fresh independent whole permitted-source review remain required. Restricted source hunks remain explicitly unreviewed.
The lifetime fixture observes the real suspended compiler after checked native-job assignment. Its already-running witness captures the compiler PID/start time and its own guardian identity, rechecks the exact created compiler handle is still live, and retains the observation time before the actual compiler cutoff. A second cutoff check precedes receipt output. The outer test validates that proof and its bounds, retains exact directory equality/existence, then deliberately kills the actual Node parent and requires its exit to be observed before that same cutoff. Late receipt or exit fails without retry. The original 1500 ms post-exit wait, exact PID-plus-start-time absence observer, owned-directory absence assertion and final cleanup remain, with unchanged native deadlines and observer/test budgets. The receipt comes only from the trusted test-launched witness; positive PIDs alone are insufficient. This replaces the fresh-interpreter round trip in the expiring pre-death window, changes assertion implementation and receipt shape, and requires hosted qualification. It tests startup containment and does not witness resumed compilation. The never-qualified pipe-input fixture and its co-occurrence observer are removed.
A separate positive control uses the same fixed ordinary source and response files, skips only the startup hold, and requires checked resume and real compiler success. It loads the produced assembly bytes, resolves the fixed type and verifies its literal integer field before normal owned cleanup. Only then may it emit fixed success and terminate itself. Its existing minimal child environment, 2000 ms timeout and 1024-byte bound stay fixed. Both controls still require hosted qualification.
The direct helper controls exercise escaped and reordered keys, a Unicode path, all six operations and fixed primitive results. Sixteen invalid requests must reach parsing and exit with the ordinary failure envelope before native authority construction; compiler failure or timer termination cannot pass. The protected state, claim and root entries must remain unchanged. At
dedf8df644, both Windows architectures pass the codec positive, malformed-request loop, real scanner and fixed-input compilation controls. These passes cover those controls in this invocation. Earlier deadline-coded setup failures occurred before the malformed-request loops; their causes remain unresolved.At
dedf8df644, the x86 helper-death test passes in full, including its final observer. The ARM helper-death test reports primary completion before its final observer times out, so the whole test fails. Both parent-death tests pass exact directory equality and existence, then fail at the live-identity check. The x86 check returns an empty array against the two expected exact identities, and its final observer returns. The ARM primary observer times out; its final observer also times out and replaces the primary exception in the visible stack. The selected final stack retains input-read; the primary observer's last boundary is unknown. Neither case reaches mainline parent termination, later deadline absence or directory absence; finally cleanup may still attempt caller termination. Empty results do not distinguish missing or exited processes, identity mismatches and swallowed process-state errors. Whole-test duration does not establish per-call timing. These results establish fixture-directory equality and existence in this invocation, with no broad path-equivalence or native lifetime guarantee.Test-only diagnostics retain directory type, capped lengths, exact base and compilation-ID component comparisons, and a closed difference category. Raw directory and environment values remain outside the added diagnostic. Normalization classifies differences and never replaces the exact equality assertion. A separate fixed label follows the input read; the same JSON-conversion pipeline then consumes the assigned string. This changes the combined expression and can change pipeline setup timing and scheduling. The added timeout-assertion message contains only a validated closed boundary label; existing error values and status reports remain. Labels locate compound intervals; missing labels prove nothing. No individual call latency, codec fix or native cleanup qualification follows. The fixed test receipt retains its original identity and directory fields, with quote and control-character refusal and backslash escaping for the path. It adds the numeric compiler cutoff already passed to the witness and the exact-live observation time; these change the concrete test receipt and observer input shape. The lifetime fixture uses native filesystem resolution for its existing owned temporary directory before appending the compiler child and creating the existing deadline. Default resolution retained the short spelling in the earlier x86 expected path. At
dedf8df644, native resolution advances both parent tests past the unchanged equality and existence assertions. This changes the concrete fixture and expected values and does not establish broad filesystem equivalence. Current-head hosted results remain required. Other Windows native-operation failures remain unresolved; no timeout, positive assertion or authority check is relaxed.At
bd33cee88b, the x86 parent test still returns an empty array; both existing lookup statements report caught argument exceptions before the exit and start-time properties. Its call begins before and ends after the operation deadline. Those stopwatch offsets do not locate either lookup against the earlier compiler cutoff or establish a termination cause. ARM records primary and final input-read boundaries separately before each observer times out. Neither parent test reaches mainline parent termination or later absence assertions; finally cleanup may still attempt caller termination. The x86 helper-death test passes in full; ARM reports primary completion and then a final-observer timeout. The direct codec control times out on x86, and fresh-root setup fails before its malformed-request loop; both controls pass on ARM. Scanner and fixed-input compilation pass on both architectures. These are historical results; candidate qualification remains open.At
c6185d0720, platform CI finishes with five passing jobs and three failures. The x86 parent fixture passes directory equality/existence, then its two cold lookups report endpoints 218/246 ms and 248/250 ms after the retained compiler cutoff. It fails the live premise before mainline parent death; these clocks do not establish timer firing or exit cause. ARM primary and final observers separately time out after input-read, with no parsed, parse-failed or slot frame retained. The missing interval remains unexplained. The x86 helper-death test passes in full; ARM reports primary completion before final timeout. The x86 escaped-key create and intervening alias checks return before later protection failure; the whole control fails. ARM passes that control, and both architectures pass the malformed-request loop, scanner and fixed-input compiler control. WSL fails its unattributed-batch control; its cause and other Windows failures remain unresolved. These are historical results and establish no native timer cause.At
7f2afd2b01, platform CI finishes with five passing jobs and three failures. Both parent tests pass the trusted live receipt, actual parent kill and observed exit before the retained compiler cutoff, then the original 1500 ms wait. On x86 the exact-identity observer reports no matching identities, then the owned-directory absence assertion fails because the path is present. Its final observer returns, preserving the original directory assertion. That sample does not establish a timer callback or unique exit cause. On ARM the primary post-death observer times out after input-read, followed by a separate final-observer input-read timeout that replaces the primary exception in the visible stack. No later parsed, parse-failed or slot frame is retained. Exact absence is unestablished and directory absence is unreached. The x86 helper-death test passes in full; ARM reports primary completion before final timeout. The x86 direct escaped-key create, acceptance and inspection return before alias binding fails; its malformed-request control fails in fresh-root setup before the loop. ARM passes those two controls. Both architectures pass the scanner and fixed-input compiler controls. macOS Intel passes; WSL fails its unattributed-batch control. Both Windows failure causes and other failures remain unresolved.At
b312cd024d, platform CI finishes with four passing jobs and four failures. Both parent tests pass the trusted live receipt, actual Node-parent death before the retained compiler cutoff and the original post-exit wait. The x86 exact-identity observer reports no matching identities before the owned-directory absence assertion fails; its final observer returns. ARM's primary post-death observer times out after input-read, followed by a separate final-observer timeout. Exact absence is unestablished and directory absence is unreached on ARM. Both new catch reports retain 24 validated startup frames through witness-announced and zero rejected frames. No later cleanup, compiler-stop or compiler-query frame is retained. These reports establish diagnostic visibility, not native frame delivery after parent death or a cleanup or timer cause. The x86 helper-death test passes in full; ARM reports primary completion before final-observer failure. Both architectures pass the codec, malformed-request loop, scanner and fixed-input compiler controls in this invocation. macOS Intel fails its marked-aggregate control, and WSL fails its unattributed-batch control. Those causes and other Windows failures remain unresolved.At
32a9bf54a9, platform CI finishes with six passing jobs and both Windows jobs failing. Both parent tests pass the trusted live receipt, actual Node-parent death before the compiler cutoff and the original wait. The x86 exact-identity observer returns no matching identities before directory absence fails; its final observer returns. ARM records a primary post-death input-read timeout and a separate final-observer input-read timeout; exact absence remains unestablished and directory absence is unreached. Both rolling catch reports retain 24 validated startup frames through witness-announced and no wait-returned, cleanup, stop or query frame. Their absence establishes no inverse cause or native delivery after parent death. The x86 helper-death test passes in full; ARM reports primary completion before its final observer times out. Both architectures pass the whole codec, scanner and fixed-input compilation controls. The x86 malformed-request control reaches its fifteenth input, id coercion: the ordinary failure status and envelope pass, then the required request-parse marker assertion fails. Its trace contains compiler exit zero followed by numeric errors under worker-assign and cleanup; those labels do not identify the failing statement or establish a parser defect. The sixteenth input and remaining assertions are unreached. ARM passes the whole malformed-request control. macOS Intel and WSL pass. Other Windows failures and these native causes remain unresolved.The latest diagnostic candidate passes the guardian's stderr directly through the Node parent's existing outer stderr handle and removes that JavaScript stderr relay. Stdout still relays the unchanged trusted receipt; the test still kills the actual Node parent and awaits its exit event. On the existing post-death observer or directory failure, a suppressible catch report publishes only the existing validated native phase ring under a fixed prefix. The ring keeps at most 24 allowed phase, numeric-code and capped observer-time frames, with its existing pending-input and rejected-count bounds. The collector's input-size, frame-count and observer-time bounds and the original exception remain. No raw helper streams, paths, identities or errors enter this added report.
This can retain positive cleanup boundaries that the killed relay could not forward. A reported compiler-stop or compiler-query error can locate that refusal. A cleanup marker after the before-cleanup callback can show that callback returned. A numeric error under the cleanup label can also be re-emitted by the outer catch; it does not locate enumeration or deletion without distinct evidence. It does not identify a particular file handle or prove a timer firing. Missing frames leave interruption, blocked work, scheduling and suppressed I/O unresolved. A synchronous observer can delay the outer listener, so the catch snapshot may omit queued frames; no wait is added. Native evidence delivery after parent death remains unqualified. Changed pipe topology can alter broken-pipe behavior and scheduling, so a future pass alone cannot diagnose the original residue. The retained startup frames establish no cleanup, ARM observer or codec repair. All exact identity and directory assertions, cleanup and budgets remain.
The witness reports the fixed witness-wait-returned phase with code zero immediately after its existing exact-compiler-handle wait returns successfully. The existing suppressible writer emits it, and the bounded rolling collector accepts the new phase only with code zero. A retained frame establishes that wait return and the following statement; it does not establish timer firing, termination cause, cleanup entry or completion, or delivery after Node-parent death. Arrival time is an observer time. An absent frame establishes none of the inverse claims. The extra call, string work, write and flush can consume the remaining cleanup interval and change scheduling or broken-pipe behavior. No wait, retry, warmup, deadline, assertion or cleanup change accompanies it. At
32a9bf54a9, neither catch snapshot retains this marker, so its delivery after parent death remains unqualified. It makes no cleanup or ARM observer repair claim.The parent-test observer retains closed phase and fixed guardian/compiler-slot outcomes from its existing reads and catches. Lookup, exit and exact-start checks keep their original short-circuit order, followed by the original conditional kill and wait. The witness now retains the compiler cutoff it already receives. Suppressible wall-clock reads immediately before and in a finally block after each existing cast-and-lookup statement report capped remaining time against that retained cutoff. Both slots use that same compiler cutoff. The output separately reports its lead relative to the supplied operation deadline; receipt, spawn and completion timing still concern that operation deadline. Observer elapsed time starts after interpreter startup. These values do not measure a timer callback, prove a termination cause or resolve clock changes. They do not bracket the later exit or start-time property reads.
The same JSON-conversion expression is wrapped in a catch that emits a closed parse-failed frame and rethrows. A missing parsed or parse-failed frame leaves completion unknown; termination, scheduling and suppressed diagnostic output remain possible. Within this observer function, input serialization, output conversion, process calls, spawn options and budgets remain. Its pre-death invocation is replaced by the witness receipt proof described above. The numeric test receipt adds the cutoff and exact-live observation time to concrete observer inputs without selecting authority or changing a deadline. At most nine fixed ASCII frames of 64 characters plus line endings fit alongside the existing labels within the unchanged buffer; this bounds these frames, not arbitrary preexisting stderr errors. Combined Node diagnostics emit only on failure. Missing clock values are unknown, and diagnostic failures preserve the original exception. Added clock reads, try/finally work, allocations and writes can change scheduling and stderr load. The separate helper-death final observer source is unchanged; it also receives the expanded receipt, so its concrete input shape changes.
The native CLI harness requires an ordinary marker to reach actual model input. A marked marker must stay out of model input, CLI streams and automatic HTTP bodies throughout the detached-flusher observation window. Failure output redacts markers after checking raw observations. Event, token, shell-turn and node-launch assertions remain. These checks do not inspect native first storage or rendering.
Queue diagnostics run only after original assertions fail and report bounded control validity and counts. At
dedf8df644, macOS Intel and WSL pass. Earlier WSL marked-aggregate and unattributed-batch controls failed with expected-one/actual-zero counts; the marked-aggregate diagnostic retained two claim payloads with no observed HTTP request. Their send and cleanup causes remain unresolved. Changed outcomes do not establish flakiness. Earlier cleanup, compiler, timing, acknowledgement and parser-growth failures remain recorded. No CI retry, fixture timeout increase, threshold change or skipped assertion is part of this change.The sender inventory retains 91 classified sites and 49 current source hashes. Native privacy before first storage and rendering, direct MCP publication, private source creation, and Windows timing, authority and cleanup remain unqualified. Other adapters require an existing family binding. This PR stays a draft until the required behavior, source review and hosted CI pass.
AI assistance in the code (leave one; writing this description with AI is fine):
AI-generated: An agent wrote the change.
I can explain every line of this diff and how it interacts with the rest of Loom.
Checklist