- harbor/results_refresh.py, a server thread: one run at a time at idle priority into a .partial file that replaces [reverse] match_results / diffrun_results only when complete (readable.txt beside it from the emulator's printout); the match run holds re_match_program's ltcg-main.lock and waits while another build of main has it
- [logbook] results_refresh, match_every_minutes, behaviour_every_hours, behaviour_window; state in the logbook folder's results.json (per job: commit, start, finish, counts, error; a run in progress), so a restarted server neither repeats a run nor starts one beside a run its predecessor left going, and adopts that run's output
- harbor logbook results shows the schedule; --job match|behaviour asks the server for a run now
- progress view: "Exact and verified as of commit abc1234, 2 hours ago" under the bars, a run in progress and a failed last run
- GET /api/logbook/progress-map: one compact document, cached until an input changes (main's HEAD, the Logbook symbol table and statistics, match/emulator/library results, facts, names, source files, globals, deferral notes); ?have=<version> answers {"same": true}
- definitions follow the Logbook's progress.py (pc102.csv category game, status decompiled or later), so the headline equals game_bytes_decompiled_percent; engine, library and unsorted code are shown but not counted
- history points now record the commit and the game-code bytes
- the page lays out once per load, filter or size change, hit-tests hover on the computed rectangles, builds the 600-entry units list only when opened, and polls once a minute only while the Logbook is visible
Opening the chat froze the machine: the log rendered all ~800 loaded messages (1.1 M characters, ~10,000 nodes,
a 440,000 px column) and rebuilt the whole log on every new message, receipt change and mark refresh - about
1.8 s of blocked main thread each time. Now:
- the log keeps the newest 150 rows; "Load earlier" adds 150 more (from memory, then the server) and keeps the
reader's place; switching person or thread starts again at 150
- new messages are appended; a changed receipt, pin, correction or answer repaints only its row
- off-screen rows skip layout and paint (content-visibility); the Intl date formatters are made once
- polls go through every(): no overlapping requests; while the tab is hidden messages poll every 10 s (sound and
notifications still work), people every 60 s, the overview every 30 s, and the marks refresh and board panels pause
- removed the backdrop-filter blurs on the sidebar, top bar and composer, and the live dot's endless box-shadow
animation (now opacity)
Measured: forced full render 1,600-1,850 ms -> 16-21 ms (hidden pane, render + layout); load long task 1,806 ms -> 56 ms;
DOM 13,318 -> 5,282 nodes (log 9,870 -> 1,806); a new message appends 2 nodes with no long task; hidden-tab traffic
~37 requests / ~2 MB per 30 s -> 4 requests / 156 KB.
The board's Logbook view replaces the published page: statistics, the chart
of the executable, where the code stands (with the function tiers), the
trend, milestones and the ship's log, stored as JSON in the Harbor home.
A server thread follows main: each new commit becomes a log entry and the
statistics, chart and trend point are rebuilt with the open-seas progress
tools on a scratch copy. 'harbor logbook add|stats|map|milestone|refresh'.
Harbor starts the x64dbg-automate MCP server (stdio) on first use and keeps
it for the life of the agent's Harbor process; the debuggee is terminated
when that process exits. Configured by debugger_command/_path/_temp in
[reverse].
Every agent reaches GhidrAssistMCP through Harbor's own MCP server, so no
per-host configuration is needed. Configured by ghidra_url in [reverse];
long decompiles are polled; also 'harbor re ghidra <tool> <json>'.
Read the project's flat image, facts.jsonl and annotations with the team's
current names; exposed as MCP tools and as 'harbor re'. Configured by a
[reverse] table in the project file.
Kimi's agent-core reads a top-level {"decision": "block"} as allow, so
Harbor's keep-going never took effect for Kimi; it honours exit code 2
with the reason on stderr (its hooks run through cmd.exe), once per turn.
Tool-call denials keep the JSON form, which Kimi does honour.
Also: attention skips stopped agents and closed items; file scopes of
function files carry their address.
Messages, threads, acknowledgements, claims, bands, tasks, review items and
stop flags for a team of AI agents; the chat page and board; host hooks and
MCP server; autopilot; tools/watch.py. Put under version control after the
weekend's changes so they have history and can be rolled back. Local only.
KeelrLeek