feat: TraceDecay V2 delivery (PR8–PR13) + daemon performance, memory, and reliability overhaul - #707
feat: TraceDecay V2 delivery (PR8–PR13) + daemon performance, memory, and reliability overhaul#707ScriptedAlchemy wants to merge 3666 commits into
Conversation
🦋 Changeset detectedLatest commit: 9ca6759 The changes in this PR will be included in the next version bump. Not sure what this means? Click here to learn what changesets are. Click here if you're a maintainer who wants to add another changeset to this PR |
|
Post-reopen correction pushed in a23f86a: source-neutral background reconcile no longer advances the scheduler epoch and cancel in-flight text activation; mounted hook overflow uses the explicit invalidating path. Evidence: RED reproduced 0/1, GREEN 1/1 for ordinary_background_reconcile_does_not_supersede_in_flight_text_work; cargo check --lib --locked passed. |
💡 Codex Reviewtracedecay/.github/workflows/release.yml Lines 408 to 411 in fc20c4f  Install the bundled Linux runtime beside the binary
For Linux targets, this archive now contains tracedecay/release-please-config.json Lines 22 to 25 in fc20c4f Advance the SDK version in release PRs
The new npm publication job derives its version from the packed tracedecay/dashboard/src/workspaces/delivery/DeliveryPage.tsx Lines 80 to 82 in fc20c4f Scope Delivery queries to the selected project
When the scope bar selects a project other than the active project, this query keeps the same cache key and continues fetching the unprefixed active-project route. The dashboard already provides tracedecay/src/tracedecay/lifecycle/registry.rs Lines 118 to 122 in fc20c4f Include Git remote identity in the registration digest
When AGENTS.md reference: AGENTS.md:L159-L161 tracedecay/plugin/cursor-native-extension/src/extension.ts Lines 187 to 189 in fc20c4f  Batch startup diagnostics instead of dropping the tail
At activation or language-client restart, ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
fix(clippy): clear the workspace clippy failures on #707
Removing the re-export shims left their module doc comments orphaned in `context` and `graph`. The rest are pre-existing lints that only fire under `--all-features`, which compiles the `test-transport` support surface as non-test code: - the registered-test-server constructor gets the file's existing explicit `expect_used` allow, since `lib.rs` denies it only outside `cfg(test)`; - `http_application_registry` is read by the `cfg(test)` capacity journey and only reads as dead under `test-transport` alone, so the allow is gated to `not(test)` rather than renamed to a discard; - the hotpath cadence helpers keep `#[inline(always)]` with an explicit allow, preserving the author's intent instead of silently downgrading it. The `daemon_suite` restart wait is now the original helper recovered from the PR #707 head rather than a reconstruction: it additionally proves the runtime readiness receipt authorizes the exact configuration the SDK selected.
The relocation fix that anchored this test above the package allocated a `PathBuf` only to hand out a reference, which `-D warnings` rejects as `unnecessary_to_owned` and which failed the Clippy gate on #707. `validate_requested_workload` takes `&Path`, and `CARGO_MANIFEST_DIR` is a literal, so `Path::new` yields a `&'static Path` with no allocation and no temporary to borrow from.
|
#733 landed Clippy-red. Helper branch (not merged, shared checkout untouched):
Clears the #733 Clippy denials:
Adopt when you want: I will not land this on #707. |
|
Clippy helper is now on the Codex worktrees (still not landed on this PR head):
Same change as |
|
Isolated post-#721 verification (worktree Green (non-vacuous) on merge
Not green
Clippy helper remains on this tree at |
Ports the missing pieces of commit b1e41b712 ("fix(build): land the
batching commit's missing consumers") onto #707: HookCompletedReadinessDistributions
moved into tracedecay-agent-hosts with pub(crate) fields, so
input_rows_processed, input_rows_dropped_at_cap, and events_considered
are no longer reachable outside the crate. Adds the three public
accessors.
source_event and input_rows_received (also added by the source commit)
are not ported: nothing on this branch reads them through an accessor
today - the root-crate benchmark still compares via
serde_json::to_value(&readiness_distributions)["field"], which does
not need them.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Zack Jackson <ScriptedAlchemy@users.noreply.github.com>
5d8f2a1 to
c7fcf3f
Compare
…d6ed chore(vendor): stacked vendor-only PR for #707 — generated vendor churn, land first
…rules Co-authored-by: Zack Jackson <ScriptedAlchemy@users.noreply.github.com>
…plan-reopened' into fable/root-lib-tests # Conflicts: # crates/tracedecay/src/lib.rs
…plan-reopened' into codex/tracedecay-total-redesign-plan-reopened
`02cd8b31f` derived `ApplicationSurfaceOperation::from_tool_name` from the transport spelling alone, so `tracedecay tool diagnostics_read` stopped resolving: `diagnostics_read` is the operation's canonical identity, but its CLI/MCP binding is spelled `diagnostics`, and the schema lookup then searched for an unadvertised `tracedecay_diagnostics_read` definition and failed with `unknown tool`. Resolve `tracedecay tool` names through the catalog's two spellings — the CLI binding name and the canonical identity — then select the one advertised MCP definition and the shipped request adapter by the operation's transport spelling, so per-key flags, `--help`, and `--args` behave identically whichever name the operator typed. The MCP transport keeps its spelling-only contract.
`628006360` retired the in-process analysis owner for `tracedecay_diagnostics`; the tool now dispatches to the application surface, which reports typed transport unavailability instead of awaiting the verified graph query. The absent-query-port probe still listed it and expected the graph refusal.
`02cd8b31f` derived the parity set from every current dashboard binding, sweeping in `native_integration_status`, whose read serves the dashboard alone by design and has no HTTP handler to agree with. Derive the set from dashboard bindings whose capability HTTP also serves, and assert the diagnostics read stays in it.
`run_native_capture` measured its single synchronous budget from hook
start and spent it on both bounded lock waits: spool admission and the
delivery-receipt writer. The analytics row, enrolled-layout lookup, and
decode that precede admission consumed it on a loaded runner, so
`lock_until` refused an uncontended lock and the hook exited 1 after
answering `{}` — the `hook-cursor-stop exited with Some(1)` failure of
`hook_replay_test`. `95095d932` fixed the same anchoring for the
response hooks' output write; give each capture-path wait one budget
measured from its own attempt.
…lan-reopened # Conflicts: # crates/tracedecay-contracts/src/doctor/adapters.rs
…al-redesign-plan-reopened
…ign-plan-reopened
…an-1073-1088 refactor(runtime): replace project facade traits
…plan-reopened' into codex/tracedecay-total-redesign-plan-reopened
02cd8b3 made every application tool advertise its canonical schemars request body verbatim. The default-profile tools/list grew from 573,502 to 737,884 bytes and tripped the 640 KiB regression tripwire: the three CAS-gated configuration writes alone added ~75 KiB by re-embedding the complete ConfigurationValueV1 / ProtectedChange unions (86 `$defs` each), and 153 tools now carried a `$schema` marker, a root `title` (the Rust type name), and rustdoc rationale paragraphs in every description. Every canonical body now passes through `mcp_input_schema` before it is advertised: `$schema`, the root `title`, and the paragraphs after each description's summary are dropped (validation semantics unchanged; SDK generation reads the Rust type path, not the title). configuration_set, configuration_batch, and configuration_protected_preview bound their value unions to the canonical discriminator tags plus read-first guidance, since the required `expected_revision` comes from a read that already shows the typed value; HTTP and SDK keep the full schema. The per-tool `format` description is shortened. Measured default-profile payload: 737,884 -> 626,799 bytes, 4% under the unchanged ceiling. The canonical-projection tests compare against the same projection instead of the raw body.
Stock Hermes ranks deferred tools with BM25 over the tool name, description, and parameter names. def5bc9 tripled the tracedecay_search description without saying what the tool is, so its length normalisation dropped the tool from first to fifth for the query "semantic code search"; 02cd8b3 then re-derived the callable code tool descriptions ("… code symbol search …", "… phrase search …"), which pushed it to seventh and out of Hermes' five-result default. The description now opens with what the tool does — semantic code search over the code graph, fusing semantic and lexical routes with a lexical fallback — and keeps the freshness, lexical_anchors, and prefer_symbol guidance in fewer words. tracedecay_search ranks first again (BM25 9.19 vs 7.60 for the runner-up) and scripts/hermes_stock_check.py passes against the pinned upstream ref.
Feature gates (run 34309279328): the transcript_ingest bench's run_provider_bench overflows rustc's query depth under --features hotpath,hotpath-mcp; boxing the ingest pass future keeps its layout out of the loop body. Clippy on 3e9c635: map_or_else for the kill/reap messages, a heap buffer for the 64 KiB pipe-echo helper, and a lint-free byte count in the drain test.
pr_context folds each changed config file into one config_summary entry that is returned but never analyzed as a seed, so seed_symbols_analyzed is legitimately below symbol_page.returned whenever the PR touches config (run 34309279328: 193 seeds + 6 summaries = 199). The strict validator demanded equality; this branch was never reached before the context phase started passing. It now requires every returned entry to be a seed or a config summary, exactly.
observation_authority_reset_recovers_the_retained_temporal_authority failed deterministically (CI runs 34296614024 and 34303181887, and locally) at its `observed_converging` assertion: the describe it polled after reopening the reset store never answered the converging refusal, so the loop broke on evidence at the first read. The transient is still there and still typed. Since 40c7841 the CLI (`call_tool_with_project_open_retry`, behind the `tracedecay tool` path every probe in this journey uses) honours a completed tool result's `problem.retry: after_delay` and re-sends the request until the tool deadline can no longer hold the delay. The converging refusal is exactly that state: `RetainedSurfaceExecutionErrorV1::unavailable` renders as an authority-class unavailable, pre-admission, after-delay, 250 ms, so `tracedecay tool tracedecay_lcm_describe` now rides out the whole convergence and hands back the converged answer. Measured on the 24-rollout fixture: the first describe after reopen took 30-42 s and returned evidence at generation 2, while a doctor read 0.7-0.8 s after reopen reported partial / stale / historical_convergence. Requiring the journey to catch the refusal through a client whose contract is to wait it out is a stale expectation, not a regression: MCP clients still receive the typed state, and the CLI still surfaces it once its deadline is exhausted. Restate the journey around what it can assert without racing the worker. The first doctor read after reopen (evidence, which the CLI never re-sends) must be the typed converging state or already current, and current only with generations rebuilt. Describe is read as the converged answer and must serve the session's preserved LCM content unchanged. The rebuilt authority must re-record the swept corpus and provider coverage. No fixture growth, no sleeps, no deadline changes.
…al-redesign-plan-reopened
…design-plan-reopened
Both run_with_timeout harnesses polled try_wait without reading stdout or stderr, so a child blocked on a full pipe could never exit and was killed as a hang. branch_list_reads_profile_sharded_branch_meta hit this at 90 s with the daemon's answer already written (~300 stderr lines under machine-wide pipe pressure). Each pipe is now drained on its own thread from spawn and joined when the child exits.
Master-bound pull requests kept their runs to completion so each landed tip would get a verdict. In practice GitHub holds one pending run per concurrency group: while the aed0b8d run spent 100+ min on the macOS job, three later heads were queued and cancelled by each other without ever scheduling a job, and the surviving head would have been judged ~100 min late. Cancelling the superseded run gives the current head the earliest verdict, which is the only one anyone can act on.
The macOS job built `--workspace --bins --tests` in 80.2 min and ran the suite in 19.9 min on the 3-vCPU runner (run 34303181887; run 34309279328 built for 74.9 min and was cancelled in its tests at 85 min), landing ~80 min after every other lane: it was the critical path of every pull request run. rust-cache carries dependency artifacts only, so that build is the workspace crates plus ld64 linking ~130 test binaries, every run. macOS now runs the Linux partitions in four groups, one `macos-14` job each, and `Test macOS` folds their verdicts and junit reports as `Test Linux` does. A group runs each partition in turn against one target directory with the selection `linux-test-partitions.py` resolves for the Linux job (`build-args`, then nextest under the `ci` policy with `--no-tests=fail`), so `root-sessions` and `root-journeys`, which share one feature resolution, compile the chain beneath the root crate once. Grouping, from the warm Arm partitions of run 34309279328 (compile + tests, minutes): root-lib 14.7 + 7.5 alone, its root chain, library test target and 1675 tests being the floor no split lowers; root-sessions 15.3 + 3.1 with root-journeys, which adds 2.7 + 2.7 on the shared resolution; runtime 15.6 + 2.4 with core-storage 10.4 + 1.9; core-contracts 13.9 + 0.6 with root-dashboard-api 12.7 + 0.8. Every pairing that puts a second partition beside a root group is heavier than the heaviest of these, and none of the other four shares a resolution, so no other grouping saves compile. Expected wall: 28/31/40/38 min at the 3-vCPU floor (compile x4/3), 38/43/57/56 at the 2x the one macOS measurement showed (the chain beneath the root crate: 19.0 min on macOS against 9.3 on Arm). The budgets take the 2x plus the 8 min cold-dependency allowance and 25 % headroom the Linux budgets carry, 60/65/85/80, and the manifest records the arithmetic beside each group; the first hosted run replaces them with measured time plus headroom. The manifest is the one authority: every partition names its `macos_group`, `check` proves every partition names a listed group and every group runs a partition (beside the exactly-one-partition proof for test targets), and `MACOS_GROUP_CAP` keeps a run to four of the account's five concurrent macOS jobs so another run can start. `scope-gate` derives the macOS matrix from it as it does the Linux one.
…esign-plan-reopened
…plan-reopened' into codex/tracedecay-total-redesign-plan-reopened
1 participant
Replacement review for #421 after its accidental merge was reverted from master in 52a9aab. This commit has the exact tree of #421 head d4c67d4. Do not merge until explicitly authorized.
Issues this PR fixes
Refs #753 (source-identity + restart journey tests landed on branch at eae9328; keep draft until full product journey/CI)
Refs #792 (journey acceptance pending)
Refs #837 (journey acceptance pending)
Refs #838 (journey acceptance pending)
Refs #855 (journey acceptance pending)
Fixes #862
Fixes #861
Fixes #810
Fixes #845
Fixes #856
Fixes #853
Fixes #799
Fixes #836
Fixes #830
Partially addressed here, tracked to completion on their own issues: #852, #800, #863, #843, #842, #848.
🤖 Generated with Claude Code
#887 dead-surface deletion inventory
build_statecompatibility wrapper was deleted; the liveget_gc_metasuppression was removed.clean_c_comment/clean_c_doc_commentare retained because feat: TraceDecay V2 delivery (PR8–PR13) + daemon performance, memory, and reliability overhaul #707 now has production callers in C, C++, GLSL, Go, and Objective-C extractors.cargo check --workspace --all-targets;cargo clippy --workspace --all-targets -- -D warnings; both pass on the integrated branch.#888/#890/#891/#892 workflow cleanup
using-tracedecaybootstrap was deleted, detailed mechanics moved to live tool help, and unique safety/domain rules remain narrowly routed.no_skill_needed, and keep only actual host size/syntax limits.