Spool v0.6.2 has three entry points: a desktop app, an MCP server, and a CLI engine slot. (the CLI engine slot runs the user's own Claude Code, Codex or Gemini CLI as a subprocess; Spool has no command line of its own)
v0.6.2's MCP server exposes 20 tools: 13 read-only and 7 behind a separate write switch. (counted from the released build on 2026-09-23)
Release
Counted with a tools/list call against the notarized v0.6.2 release on 2026-09-23; the server source is not linked because it quotes private content.Checked
v0.6.2
769 tests passed at the v0.6.2 tag: 627 TypeScript, 129 main-crate Rust and 13 sidecar Rust, with 4 skipped by design. (run on 2026-09-23; two timing-sensitive Rust tests failed on the first of 11 main-crate runs)
Test suite
Run at the v0.6.2 tag on 2026-09-23 by my AI coding agent; the log is not public, and the suites rerun from the tagged source.Checked
v0.6.2
Before v0.6.2, six evaluation rounds made 557 real API calls. (2026-08-20 to 08-24)
Log
Evaluation log in the Spool repository; not linked because it quotes real library content.Checked
v0.6.2
–
The six pre-release evaluation rounds cost at least ¥29.37. (token usage times the vendor's price list; failed calls reported no usage, so this is a floor)
Log
Evaluation log in the Spool repository; not linked because it quotes real library content.Checked
v0.6.2
–
Compressing two projects by 30% and 38% kept 39 of 40 answers correct; the uncompressed baseline scored 40/40. (round six: 20 questions, each asked twice per arm; one model; answers graded by the coding agent; the gap is within noise)
Log
Evaluation log in the Spool repository; not linked because it quotes real library content.Checked
v0.6.2
Four release blockers found in pre-release testing were fixed and re-tested before v0.6.2 shipped. (found in round five on 2026-08-23; re-tested in round six on 2026-08-24)
Two envelope defects from the last round shipped unfixed: failed calls report no usage, and one empty reply was marked ok. (still open as of 2026-09-23)
A linear cost estimate landed within ±30% for only 46% of 89 real bills, so the app shows no price. (round five; the bar set in advance was 90%)
Log
Evaluation log in the Spool repository; not linked because it quotes real library content.Checked
v0.6.2
Medium reasoning gave no compression gain over low; it was 3.5× slower and 3.4× costlier. (n = 5 vs 10 runs on one project and one setting, round five; a 5-point gain was required in advance)
Log
Evaluation log in the Spool repository; not linked because it quotes real library content.Checked
v0.6.2
Each project lists what to watch; a follow-up runs the user's own Claude Code with web search and queues sourced findings for approval.
A fix passed its tests but covered one of two relation types; the next offline re-check found 20 exposed blocks. (round six, on a copy of the library; the complete fix landed before the v0.6.2 tag)
Log
Round-six re-check on a library copy and the follow-up fix commit; not linked because both name a real project.Checked
v0.6.2
Spool's main process links no HTTP or TLS library; the only network client in the app is an opt-in subprocess, spool-ai. (MCP clients and the user's own CLI make their own network calls)
Release
Main-crate dependency tree (cargo tree) at the v0.6.2 tag, checked on 2026-09-23; reproducible from the tagged source.Checked
v0.6.2
AI clients get no access until the user turns on the MCP server, and writing has a second, separate switch. (the write switch is on by default once the MCP server is on)
AI-written blocks are append-only and carry the client's name; an AI cannot edit or delete what the user wrote.
Screenshot
A real MCP write against the notarized v0.6.2 release in a demo library on 2026-09-23, shown on this page; the server source is not linked because it quotes private content.Checked
v0.6.2
A compressed draft that dropped a number or date cannot be saved, and the original text stays on each block. (compression is opt-in, off by default and uses the user's own API key; keeping originals can be turned off)
A follow-on study of project memory across AI clients is under way.
Log
Study protocol and harness records, kept private until results exist.Checked
v0.7.4's MCP server exposes 28 tools: 15 read-only and 13 behind a separate write switch. (counted from the released build on 2026-10-29)
Release
Counted with a tools/list call against the notarized v0.7.4 release on 2026-10-29; the server source is not linked because it quotes private content.Checked
v0.7.4
1,540 tests passed at the v0.7.4 tag: 1,237 TypeScript, 272 main-crate Rust and 31 sidecar Rust, with 11 skipped by design. (run on 2026-10-29 from a clean copy of the tagged source; the main Rust crate passed six runs in a row)
Test suite
Run at the v0.7.4 tag on 2026-10-29 by my AI coding agent; the log is not public, and the suites rerun from the tagged source.Checked
v0.7.4
Corrections, replacements, files and project splits an AI proposes wait on a review screen until the user clicks. (first shipped in v0.7.0; the AI's own new block is saved at once; what waits is its effect on existing blocks, files or projects)
A private project is hidden from AI clients connected through Spool, though they can still write into it. (first shipped in v0.7.0; covers the route through Spool only; nothing is encrypted on disk)
A workspace can link to a local folder; a SPOOL.md there tells AI tools which Spool projects the folder belongs to. (Spool cannot stop an AI client from reading the folder with its own file tools)
Weekly review, follow-up, compression and the out-of-date check each run on the user's own CLI or own API key. (both routes since v0.7.3; follow-up needs Claude Code, Codex or an API provider with its own web search)
Spool v0.7.4 has three entry points: a desktop app, an MCP server, and a CLI engine slot. (the CLI engine slot runs the user's own Claude Code, Codex or Gemini CLI as a subprocess; Spool has no command line of its own)
A compressed draft that dropped a number or date cannot be saved, and the original text stays on each block. (compression is opt-in, off by default, and runs on the user's own API key or local CLI; keeping originals can be turned off)
Confirmed in writing by my supervisor at Libiao Robotics; not publicly linkable.Checked
–
Worked on a simulation-based multi-robot control system for warehouse sorting, integrating task scheduling, path planning, traffic control, and robot communication.
Written confirmation
Confirmed in writing by my supervisor at Libiao Robotics; not publicly linkable.Checked
–
Reorganized an existing operational simulation into six layers and added seven architecture tests to improve maintainability.
Written confirmation
Confirmed in writing by my supervisor at Libiao Robotics; not publicly linkable.Checked
–
Validated the architecture by replacing GDI+ with SkiaSharp through changes limited to the rendering layer.
Written confirmation
Confirmed in writing by my supervisor at Libiao Robotics; not publicly linkable.Checked
–
Diagnosed deadlocks and rebuilt traffic control around spatiotemporal reservations, validating 20 robots over 90,000 simulation ticks without a collision.
Written confirmation
Confirmed in writing by my supervisor at Libiao Robotics; not publicly linkable.Checked
–
On my own initiative, I added a learned local layer: a PPO policy adjusts each robot's speed along its A* path. (simulation only; not covered by my supervisor's written confirmation)
Log
My own code and benchmark report in the internship codebase; company code, not public.Checked
–
With four robots, the trained policy cleared all conflicts and deadlocks and cut average task time from 35.4 s to 26.6 s. (one fixed-seed simulation run per scenario, against the same run with the learned layer off; task time also fell with six robots and in a head-on corridor; throughput never dropped below the baseline; with six robots, conflicts rose from 1 to 6; traffic control resolved them and no task failed)
Log
My own benchmark report in the internship codebase; company code, not public.Checked
The Phase A workbench edits and plays multi-track MIDI in the browser, detects chords and suggests a key; its core logic passes 102 unit tests. (automated tests; they check that audio renders, not how it sounds; six need a separately downloaded sound bank)
Six deterministic motif transformations pass 20 tests of exact timing, bar alignment, staying in key and leaving their input unchanged. (pure functions with no model calls; float times are rejected; each test uses fixed example motifs, not every possible input)
The instrument constraint masks logits to an allowed instrument set and is not a target-track guarantee. (2x2 pilot of the Anticipatory Music Transformer on one melody, one run per condition; neither constrained run produced a bass note, though bass was allowed; top_p also differed between free and constrained runs)
Log
Outputs of the July 2026 pilot, kept locally and not in the public FLUX repository; re-read on 2026-10-02. The masking code and the run settings are public.Checked
On unusual harmonic requests, writing the chords into the context as a guide track met 41% of requests; the built-in controls tested met under 1%. (development set of 25 eight-bar Lakh MIDI excerpts with the human bass removed; 332 requests, three seeds each; pinned MIDI-GPT weights; compliance scored by a rule-based verifier, not by listeners; the verifier's agreement with human judgement so far rests on my own blind ratings)
Log
R1 evidence notes and run records in the FLUX repository, on a branch that is not yet public; re-read on 2026-10-23.Checked
The Qingcong music video received 396 likes and 328 saves on Bilibili. (figures supplied by Hanze JIN on 2026-10-13; observation date unknown; this music video only; current platform counts were not independently verified)
Written confirmation
Hanze JIN supplied the Bilibili video link and counts in the M5 material handoff.Checked