Land the trace projection and snapshot endpoint #22

Activity

1. Dillpickleschmidt

   added

   wayfinder:taskWayfinder task ticketWayfinder task ticket

   on Aug 1, 2026
2. Dillpickleschmidt

   added a parent issue on Aug 1, 2026

   • 3D surfaces: port mindwalk's Tree and Terrain into the web app #1
3. Dillpickleschmidt

   marked this as blocking on Aug 1, 2026

   • Project the agent graph and per-agent traces #21
4. Dillpickleschmidt

   self-assigned this

   on Aug 1, 2026
5. Dillpickleschmidt commented on Aug 1, 2026

   

   Dillpickleschmidt

   on Aug 1, 2026

   OwnerAuthor

   More actions

   Answer

   Landed on branch feat/trace-projection-endpoint (commit f70e59227, pushed, no PR). GET /api/mindwalk/threads/:threadId?lens= returns { trace, citymap }, built in mindwalk's order (server.go:834-841): trace with paths only → citymap with those paths as ghostPaths → backfill Target.fileId → recompute ComputeStats with the real file count. Code lives in apps/server/src/mindwalk/; the wire contract is packages/contracts/src/trace.ts, shaped byte-identically to apps/web/src/mindwalk/types.ts.

   Differentially verified against the real Go binary

   The citymap builder set the bar at zero diffs against mindwalk. That bar isn't reachable here — the inputs genuinely differ — so here is the characterisation instead. Thread c4e71440 ("Redesign plant growth system", the largest in the store, 1,287 rows) ran as Claude session 21d3c20e; mindwalk trace was run over that session's JSONL and diffed against our projection of the same thread.

   • Action classification: 1204/1204 identical. Every event the two traces share got the same verdict. Not one divergence in the ported classifier.
   • Targets: 1284/1284. Every one of our events has an exactly matching target set — same paths, same touch levels, same weak flags — somewhere in mindwalk's trace for the same tool.
   • Final per-file touch map — the thing the colour ramp actually renders — 146/150 paths identical, 0 differing, 0 extra. The 4 gaps are all files touched only by events T3 never recorded.

   Event count differs: 1,284 ours vs 1,819 mindwalk's, and ours is the more correct number. The JSONL has 1,819 tool_use blocks but only 1,348 unique ids — Claude Code replays history into the file on resume and compaction, and mindwalk emits an event per replay. T3's activity log is append-only and records each call once. The remaining 65 (1,348 unique − 1,283 tool.completed) are calls the JSONL has and T3's store does not, spread evenly at ~5% across every tool (Bash −23, Edit −15, Read −13, Write −2 …) rather than concentrated in any one type — the fingerprint of dropped ingest events, not of a projection gap. Recorded as fog; it is a T3 ingest-completeness question, not this ticket's.

   Ran over all 21 real threads: 3,805 events, no crashes, largest thread projected in 76 ms.

   Deviations from mindwalk, all deliberate

   1. Codex's commandActions targets are not weak — as the design decided. Typed discriminator with an absolute path from Codex's own sandbox parser beats a shell string.
   2. Codex's error signal is exact, not estimated — a second deviation of the same character, not previously enumerated. mindwalk sniffs Codex failures out of output text because rollout JSONL is all it has; T3 stores status and exitCode structurally. Reporting "estimated" here would be exactly the lie that grade exists to prevent.
   3. No exec / js / js_repl branches in actionFor/targetsFor. Those parse Codex's free-form JavaScript orchestration wrapper out of rollout JSONL. T3 talks to the Codex app-server, which delivers a commandExecution item with a plain shell string, so the ~200 lines of JS-source parsing behind them are unreachable. Everything else in the copy list is verbatim.

   What the port confirmed

   • The scraping is load-bearing, as predicted. Across the whole store: 2,120 exec, 724 edit, 543 read, 265 search. Of 3,971 targets, 3,030 are weak — scraped out of Bash. Without the scraping, search and the "seen" level would be near-empty.
   • Observability.Reads is estimated on essentially every thread, and unavailable on the few that only ran shell commands. Honest and correct.
   • verify fires almost never in practice: 1 event across 3,805. mindwalk's patterns are go test / npm test / pnpm test / pytest / cargo test / make test; this environment runs vp test run and CMake. Faithful to upstream and left alone — recorded as fog, since editsAfterLastVerify and the verify bar are dead here until the list grows.

   Two things the build settled that the design had not

   • Provider comes from the session row, not the thread shell. getThreadShellById only returns active threads, so an archived thread would have projected with the wrong adapter and rendered a dark, empty city. Caught by the service test. Now reads ProjectionThreadSessionRepository.
   • An unnamed provider is read by payload shape, not assumed. Defaulting an unknown provider to one adapter fails silently — a Claude payload read as ACP finds nothing and looks like "the agent touched nothing" rather than like a failure. The fallback now dispatches on whether data.toolName / data.item / data.tool is present.

   Cache

   Exact, keyed on COUNT(*) of the thread's activities (per lens), no TTL — the log is append-only, so the count is a perfect fingerprint. A repository test pins that an upsert of an existing row does not move it. Read is SQL-filtered to the three kinds the projection wants, which skips the majority of a long thread's payload bytes before parsing. Ghost re-layout on a cache miss measures 40–80 ms on a 15.7k-file repo; the expensive walk stays cached per root in CitymapBuilder.

   Degradation

   A failed citymap build returns an empty citymap and the trace anyway — you lose the buildings, not the timeline. Verified by test.

   Tests

   107 in apps/server/src/mindwalk/ plus 4 for the new repository queries. TestCommandReadPaths, TestBuildEventClassifiesBashSearchCommands, TestSummarizeToolTruncatesCommandAtRuneBoundary and all three of stats_test.go are ported case for case from the Go suite. Claude and Codex adapter fixtures are trimmed copies of real payload_json rows. Nothing asserts ACP or OpenCode behaviour beyond the graceful fallback, per the ticket.
6. Dillpickleschmidt

   closed this as completedon Aug 1, 2026
7. Dillpickleschmidt

   added 6 commits that reference this issue on Aug 1, 2026

   fix(server): address Macroscope review on the trace projection

   ...

   d25b725

   fix(server): fold repo state and message content into the snapshot key

   ...

   011138a

   fix(server): never cache a degraded mindwalk snapshot

   ...

   7ad10e3

   fix(server): address Macroscope review on the trace projection

   ...

   7c7802d

   fix(server): fold repo state and message content into the snapshot key

   ...

   b46b044

   fix(server): never cache a degraded mindwalk snapshot

   ...

   2b2c9ea