@aistastudio/myc 0.1.1 → 0.2.1

package/README.md

@@ -11,6 +11,12 @@ machines through plain git.
 Design docs live in `docs/design/` (start with `00-brief.md`); the measurements
 quoted below are reproducible from `bench/` and `scripts/`.
 
+**Site: <https://aistastudio.github.io/myc/>** — the same measurements as charts,
+in English and Russian, with the command that reproduces each number printed
+next to it. Source in `site/`; `bun run site/build.ts` re-checks every figure
+against the measurement artefacts in this repository and refuses to build on a
+mismatch.
+
 ## Requires Bun — this is not fine print
 
 The runtime is bound to `bun:sqlite` (SQLite and `sqlite-vec` ship inside Bun,
@@ -25,7 +31,7 @@ Once published, installation is one command — the package is built and verifie
 from a tarball today, but nothing has been pushed to the registry yet:
 
 ```bash
-bun add @aistastudio/myc     # 3.16 MB, 9 files, no models pulled at install
+bun install -g @aistastudio/myc   # 3.17 MB, 10 files, no models pulled at install
 bunx myc --version
 ```
 
@@ -63,15 +69,15 @@ Full command list: `./dist/myc --help`.
 
 **Speed is a constraint, not an optimisation.** Every hot path has a budget
 enforced in CI; a p95 regression over 15% fails the build. Measured on 100 000
-nodes (`bun run scripts/bench-latency.ts`):
+nodes, 2026-09-07, darwin-arm64-14 (`bun run scripts/bench-latency.ts`):
 
 | operation | p99 | budget |
 |---|---|---|
-| `prime` (session context) | 0.70 ms | 30 ms |
-| read | 0.011 ms | 3 ms |
-| search | 9.1 ms | 25 ms |
-| write | 0.5 ms | 5 ms |
-| cold start | 24 ms | 60 ms |
+| `prime` (session context) | 0.755 ms | 30 ms |
+| read | 0.012 ms | 3 ms |
+| search | 10.354 ms | 25 ms |
+| write | 0.460 ms | 5 ms |
+| cold start | 23.820 ms | 60 ms |
 
 **Ranking is measured, not asserted.** Two labelled corpora with graded
 relevance, each containing a *control group that gets worse* when the feature
@@ -83,9 +89,38 @@ works — so a gain cannot be manufactured by shaping the corpus:
 
 **Caching that cannot go stale silently.** Result, embedding and hydration
 caches are invalidated by `MAX(oplog.seq)` read *from the database*, so a write
-by another process invalidates them too. A cache hit is 162–198× cheaper than a
-miss (≈25 000× for embeddings) and the ranking is bit-identical: same MRR to
-three decimals, zero rank differences.
+by another process invalidates them too. A cache hit is two orders of magnitude
+cheaper than a miss — 252× in the run of 2026-09-07, ≈27 000× for embeddings;
+the ratio is wall-clock and moves with the machine. The ranking does not: same
+MRR to three decimals, zero rank differences.
+
+**Memory survives context compaction.** `myc wire` installs a pre-compact hook,
+so the moment before an agent's context is squeezed the session episode is
+written to disk — raw, `L0`, `acl private`, secrets masked — and a rescue
+packet is printed back into the context that survives. Distillation is queued,
+never done on the write path. The episode is on disk before anything else is
+attempted, so exceeding the hook's timeout costs the summary, not the record:
+
+```
+$ myc absorb-session --reason manual --transcript … --agent claude
+# myc: контекст сжимается — вот что нельзя потерять
+эпизод sess-5jh8je4g050m сохранён (265 Б)
+ДАЛЬШЕ   myc show sess-5jh8je4g050m · myc ready --claim
+```
+
+**Memory is separated by session, and the separation is visible.** Every note
+carries a reach: `session` (this conversation) or `project` (everyone). The
+automatic context packet — `prime` — only carries the current session's notes;
+another agent's session does not leak into yours. An explicit `myc recall`
+still finds them, because hiding knowledge is not the same as scoping it, and
+marks each row for what it is: `ses` own session, `ses*` someone else's, `prj`
+project-wide.
+
+```
+$ MYC_SESSION_ID=s1 myc recall "ретраи"      $ MYC_SESSION_ID=s2 myc recall "ретраи"
+1.30 … ses  сессионное: ретраи…              1.30 … ses* сессионное: ретраи…
+1.10 … ses* в сессии один: ретраи…           1.10 … ses* в сессии один: ретраи…
+```
 
 **Memory has three independent axes**, and the surface says what it hid:
 tier (project vs personal), session reach, repository reach. `prime` prints
@@ -103,9 +138,12 @@ the first machine's priority and the second's tags. Nothing is lost to
 last-writer-wins over whole records.
 
 **Migration from beads is real, not a demo.** A working project imported in
-684 ms: 796 tasks, 972 dependencies, 265 notes, 41 memories — with unknown
+889 ms: 796 tasks, 972 dependencies, 265 notes, 41 memories — with unknown
 issue types carried over verbatim and named, and out-of-range priorities
-clamped and named, instead of one odd row aborting the import.
+clamped and named, instead of one odd row aborting the import. On that same
+graph `myc ready` offers 195 tasks and `bd ready` offers 144: beads inherits
+blockers down the parent chain and myc does not yet, so beads is right about
+those 51 (open bug `memory-atcm254ry6c7`).
 
 **Guards are proved by mutation.** Every refusal and every invariant is
 accompanied by a mutation that removes it; a guard whose removal breaks no test
@@ -114,24 +152,32 @@ is treated as absent.
 ## Roadmap
 
 Numbers are closed/total subtasks per milestone (`myc show <epic-id>`), as of
-2026-09-07. Done and not-done are shown the same way on purpose.
+2026-09-08. Done and not-done are shown the same way on purpose. Totals grow
+when work uncovers work: M0 went 33 → 39 because measuring it found four real
+defects, not because the plan changed.
 
 | milestone | status |
 |---|---|
-| **M0** core and tasks | 30 / 33 |
+| **M0** core and tasks | 35 / 39 |
 | **M0.5** self-hosting (myc developed through myc) | **4 / 4 — closed** |
-| **M1** memory | 20 / 23 |
-| **M2** semantics | 15 / 19 |
-| **M7** human interface (board, cards, threads, routing panel) | 11 / 13 |
-| **M3** code intelligence | 0 / 7 |
-| **M4** team: `myc serve`, ACL, network sync, Postgres, containers | 0 / 11 |
+| **M1** memory | 21 / 23 |
+| **M2** semantics | 17 / 20 |
+| **M7** human interface (board, cards, threads, routing panel) | 13 / 14 |
+| **M3** code intelligence | 5 / 9 |
+| **M4** team: `myc serve`, ACL, network sync, Postgres, containers | 3 / 14 |
 | **M5** swarm self-learning: routing by cost and outcome | 0 / 12 |
 | **M6** distillation | 0 / 7 |
 
 What that means in practice: **today myc is a single-user local tool over files
-in git.** There is no server, no ACL, no team mode, and no code↔knowledge
-anchors yet. Those are designed (`docs/design/03…`, `04…`, `05…`) and tracked,
-not implemented.
+in git.** There is no server, no ACL and no team mode. Those are designed
+(`docs/design/03…`, `04…`, `05…`) and tracked, not implemented.
+
+Code↔knowledge anchors now work: `myc task "…" --anchor src/file.ts:10-20`
+binds a task to a span, and the anchor follows the code as it moves. The
+binding is language-agnostic — it was verified on Python as well as
+TypeScript. What is *not* there yet is symbol-level understanding: parsing
+functions and classes covers `ts/tsx/js/jsx` only, and the symbol index is
+built but not yet wired to any command (tracked, not hidden).
 
 ## Syncing between machines
 
@@ -154,3 +200,4 @@ sqlite-vec) is permissive too.
 ---
 
 Russian version of this document: [`docs/README.ru.md`](docs/README.ru.md).
+Both languages, with charts: <https://aistastudio.github.io/myc/>.