feat: consolidate TraceDecay V2 PR8-PR13 delivery - #421
Draft
ScriptedAlchemy wants to merge 6772 commits into
Draft
feat: consolidate TraceDecay V2 PR8-PR13 delivery#421ScriptedAlchemy wants to merge 6772 commits into
ScriptedAlchemy wants to merge 6772 commits into
Conversation
|
ScriptedAlchemy
force-pushed
the
codex/tracedecay-total-redesign-plan
branch
from
July 10, 2026 05:54
e50f8f5 to
4f5ab66
Compare
Owner
Author
|
@codex review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 986f25cca6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
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".
ScriptedAlchemy
force-pushed
the
codex/tracedecay-total-redesign-plan
branch
4 times, most recently
from
July 11, 2026 16:52
5862bd6 to
53502de
Compare
ScriptedAlchemy
force-pushed
the
codex/tracedecay-total-redesign-plan
branch
4 times, most recently
from
July 23, 2026 17:06
642a33a to
5f2f610
Compare
ScriptedAlchemy
changed the title
ScriptedAlchemy
force-pushed
the
codex/tracedecay-total-redesign-plan
branch
from
July 24, 2026 17:16
76dee0f to
0258866
Compare
ScriptedAlchemy
force-pushed
the
codex/tracedecay-total-redesign-plan
branch
from
August 1, 2026 18:32
062285a to
b73f8bf
Compare
This was referenced Aug 3, 2026
ScriptedAlchemy
force-pushed
the
codex/tracedecay-total-redesign-plan
branch
from
August 7, 2026 08:32
ef68e05 to
f92b7f9
Compare
ScriptedAlchemy
force-pushed
the
codex/tracedecay-total-redesign-plan
branch
2 times, most recently
from
August 7, 2026 21:43
1486ceb to
f44f5c5
Compare
The plugin bundle is the roster agents actually read, and a tool nothing references is a tool nothing calls. Add the mount audit to the Cursor mcpAllowlist, the code-health auditor agent, the audit-safety, check-health and clean-dead-code commands, and the code-health and reviewing-changes skills — alongside dead_code and unused_imports, whose findings it precedes: a file no build root reaches is dead wholesale, and every symbol in it reads as live to the other three tools. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The README workflow groups and the USER-GUIDE code-quality table are where a reader learns the tool exists at all, and the Rust probe matrix is where it gets exercised against real repositories. Describe it by what it answers rather than by one ecosystem's mechanism, so the entry stays true as more ecosystems gain a reachability model. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
BREAKING CHANGE: cursor transcript records now write their session-location
metadata under `cursor_session_*` like every other provider. The previous
`cursor_event_*` namespace existed only so readers of already-persisted
metadata resolved the same keys — a persisted-shape compatibility exception
the V2 fresh-store cutover forbids. There is no dual-read fallback: readers
resolve the canonical key only, and an old store fails typed at admission
through the existing reset-required machinery, which this change does not
touch.
Because the namespace is now provider-uniform, `metadata_namespace` and its
cursor constants carry no decision and are gone; the reducer formats
`{provider}_session` at its one call site and the descriptor honestly
documents two provider-shaped decisions instead of three.
Also renames the misnamed "compatibility metadata hook": appending
`tool_calls`/`tool_events`/`tool_use_id` is live V2 normalization of cursor
records into the cross-provider shape, not legacy support. The type, the
selector, and the cursor implementation are now `ToolMetadataNormalizer`,
`tool_metadata_normalizer`, and `normalize_cursor_tool_metadata`.
The canonical_projection characterization tests that pinned `cursor_event`
are re-pinned to `cursor_session`. That is an intentional behavior change
under the cutover mandate, not a test repair.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…tal-redesign-plan
Ingest failure classification named a reason code and a retryability verdict but not the admission status, so the status had to be recovered downstream by matching the reason-code string. One match now owns all three: `classify_transcript_ingest_disposition` returns the new `TranscriptIngestDisposition`, and `ClaudeObservationFailureClass` carries a `status` alongside its reason code. `TraceDecayError::hook_runtime_with_status` records that status on the error so it survives the crate boundary. The status enum lives above the runtime kernel, so the value travels as its canonical serde wire spelling and `HostAdmissionStatus::as_wire`/`from_wire` round-trip it without anyone re-deriving it. The spool open paths, which already hold a `HostAdmissionOutcome`, report its status directly. `classify_transcript_ingest_failure` keeps its signature and its `TranscriptCatchUpFailure` shape, so catch-up telemetry is unchanged. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The evaluation journey requires 7 projection cases (REQUIRED_PROJECTION_CASES) at 2 scales (REQUIRED_RESOURCE_SCALES = current, 10x), so it performs 14 real publishes, each re-hashing the whole recovered generation. Measuring only prepare understated the per-case cost. Measured (768-d, prepare + publish + settle): 2048 chunks 2036ms, 4096 2766ms, 8192 5557ms, 22016 15011ms. Flat at 675-682us per chunk beyond the smallest scale: linear in corpus, paid once per case.
The digest probe timed only prepare+publish, so the per-pass corpus build cost was invisible. At real width (43 pages / 22016 chunks, 768-d) the corpus build is 28119ms against a 15629ms publish case: the build, not the digest, dominates a pass. Reported as corpus_build_ms and pass_total_ms.
measure_evaluation_projection_cases stands up a fresh TempDir, grafeo graph, and metadata store, then re-projects the whole corpus and runs five real FastEmbed batches. It was called unconditionally once per evaluation pass, while the driver runs one pass per profile x partition x scale -- so the identical measurement was rebuilt six times per scale. The measurement reads only the clean generation and the three mutation sources, never the profile or partition, so memoize it on exactly those four generation ids, mirroring the prepared_native cache beside it. The lock is released across the measurement so a race duplicates work once instead of blocking.
generate_candidate_outputs_with_native called generate_candidate_outputs, which publishes the 1x and 10x corpora, and then published byte-identical 1x and 10x corpora again for the native phase -- four full builds per evaluate-and-publish where two suffice. publish_corpus_with_scale is a pure function of its four arguments: its content_identity hashes the corpus digest and copy count, captured_at and sealed_at are constants, files are sorted before building, and the publication store is created and dropped inside the call. Equal arguments therefore yield byte-identical generations down to the generation id, so sharing one build between the two phases is unobservable. Cache the corpora by scale for the span of one call and let both phases read the same builds.
79211a2 added the required AutomationRunResultV1::request_digest but left two fixtures behind it: the curation terminal passed a bare 64-hex string the domain has always rejected as noncanonical, and the untagged variant-selection fixture omitted the field entirely so it matched no RetainedSurfaceResultV1 variant. Both now take the digest from AutomationRunRequestV1::input_digest, reusing the helpers the sibling automation tests already use, so no digest is hand-written. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Prices one publish_corpus_with_scale build against the real packaged workload, which is exactly what sharing the corpora between the fallback and native phases saves. Measured: 1x = 2170 chunks in 2312ms; 10x = 21700 chunks in 23249ms. Linear at 1065-1071us per chunk, so the removed duplicate pair costs 25.6s per evaluate-and-publish.
Message search and the direct LCM anchor query carried the multi-MiB ExecutionLimits defaults, which the admitted application binding refuses terminally — every such retrieval answered Saturated instead of searching. All admitted-path queries now size against one shared APPLICATION_RETRIEVAL_MAX_BYTES via admitted_execution_limits, which narrows only the four checked budgets and leaves ranking pools at their defaults so recall is not silently traded away. Completes 11184ac; lcm_claude_recall passes 4/4. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
measure_incremental_evaluation_projection reloads the semantic artifact and re-projects the incremental generation's changed chunks. It reads only the prepared generation and the incremental generation -- never the profile or the partition -- so the twelve passes the driver runs (3 profiles x 2 partitions x 2 scales) rebuilt one identical observation per scale six times over. Memoize it on exactly that generation pair, mirroring the projection-case cache beside it. Every field it returns is a property of the pair, each pass records them as its own single-element sample vectors with no averaging across passes, and the existing post-measurement binding check re-validates the cached resources against the pass context, so a stale or mismatched entry is rejected rather than reported. The lock is released across the measurement so a race duplicates work once instead of blocking.
The planner-scale root-retrieval test seeded 100k occurrence and summary rows with ordinary exact-SQL batches, which carry the 30s per-statement deadline enforced by the exact SQL guard's progress handler. Chunking at 20k rows only widened the margin; under CPU contention a seeding chunk still crossed the deadline and SQLite returned SQLITE_INTERRUPT, so the fixture aborted before a single asserted query ran. Reproduced at load average 104 on 96 cores. Bulk seeding is exactly the work authorized long-lease transactions exist for, alongside schema installation and full-index replacement. Route the chunks and the trailing ANALYZE through execute_authority_revalidated_batch, which drops the ordinary statement deadline while keeping shutdown, idleness, and authority revocation as cancellation conditions, and renews the lease on each committed chunk. No assertion changes: row counts, index-keyset checks, and query-plan expectations are untouched, and the fixture stays planner-scale at 100k rows per table. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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
1 participant
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.
Summary
This draft PR is the consolidated delivery branch for TraceDecay V2 PR8 through PR13.
The branch also contains the supporting rusqlite runtime cutover, worktree-aware incremental indexing, storage retention/Doctor integration, and direct product tests required by the canonical V2 plans.
Verification
cargo check --all-featurescargo test --all-featurescargo clippy --workspace --all-targets --all-features -- -D warningscargo fmt --all -- --checkCurrent status: implementation is still being consolidated from the shared checkout and CI is red. This PR remains draft until the working tree is committed, broad verification is stable, and the historical conventional-commit blocker is resolved. It must not be merged in its current state.
Safety boundaries
GitHub review ingestion is read-only. Semantic indexing never blocks exact/lexical/graph retrieval and publishes only complete compatible generations. Local coordination logs, private transcript evidence, secrets, and ad hoc artifacts are excluded from commits.