oxidus — technical notes

Source: gh repo clone gesslar/oxidus-mudlib (cloned 2026-08-28/30). Number 954, port 40256. Status: done — boots clean natively, full registration-to-gameplay playthrough verified live and repeatedly.

Hosting (2026-09-17) — real submodule-patch

work/ is a git submodule of gesslar/oxidus-mudlib (gitlink mode 160000 — the 2026-09-15 cutover had left a vendored tree in the parent index despite .gitmodules; corrected here so the rebase timer updates a pin, not 1600 tracked files). upstream.auto_rebase is true — the twice-daily rebase timer advances the pin to upstream main when catalog patches still apply cleanly.

Native play requires FluffOS promise phase 1.5 combinators (promise_all / promise_race / promise_cancel). Local ~/src/fluffos/build-debug was rebuilt from origin/master (2c272875, 2026-09-17) and confirms those symbols. Site Pages WASM still ships FluffOS v2026.0901.0, which lacks combinators, so wasm_status is limited until a newer published WASM release includes phase 1.5.

Catalog FluffOS-compat lives in patches/ (applied by python3 scripts/apply_lib_patches.py oxidus / source-zip pack):

• 0001-PACKAGE_UIDS-stubs-on-master.patch — get_root_uid / get_bb_uid / creator_file stubs for the shared PACKAGE_UIDS build.
• 0002-OLD_ED-classic-ed-for-shared-driver.patch — classic ed() for shared OLD_ED.
• 0003-WASM-hash-stub-without-PACKAGE_CRYPTO.patch — hash() sefun stub so WASM rooms compile without PACKAGE_CRYPTO.
• 0004-fix-compile-warnings.patch — every compile warning fixed (2026-10-13, see the dated section at the end).

Driver headers (include/driver/*.h, including promise.h) and the adm/custom/ first-user seed live in overlay/ (rsync'd onto zip staging).

What this lib is

A genuinely modern, actively-developed (pushed within the last day as of this session), from-scratch FluffOS mudlib — not a 1990s/2000s legacy archive like nearly everything else in this corpus. Already .lpc throughout, already UTF-8, ships its own AGENTS.md/.claude/skills/ (37 skill files) documenting its own coding conventions, has a real unit test suite under tests/ (STD_TEST framework), and targets a specific, current FluffOS build via its own adm/dist/rebuild pipeline (clones fluffos/fluffos as a submodule and builds it from source with a custom local_options). Ran scripts/convert_lib.sh per §2.3 convention (cheap and safe even though 0 lossy conversions were needed) — it also mechanically rewrote dual-extension .c compatibility guards and English “static” in MSSP comments (fixed 2026-09-10, issue #3). The 46 local angle-bracket includes converted to quotes stay: the shared driver include path does not search the including file's directory (scripts/convert_lib.sh).

8 authored content areas confirmed real (not stock/copied): a village (Olum) with a bakery, tailor, financier, manor interior, an arcanist NPC/spell tree, forest/cavern/maze/tunnels/wastes zones. Account-based login (register an ACCOUNT with a password, then create/select a CHARACTER under it) rather than the straight-into-a-name flow most libs in this corpus use — matches the README's own description. The first character ever created becomes the mud's owner/admin automatically (confirmed live).

Driver-build mismatches found and fixed (see AGENTS.md for the full writeup)

This lib assumes a considerably newer/differently-configured driver than this corpus's shared build had, in three independent ways. All three are documented as new AGENTS.md catalog entries since they're real, generalizable classes other modern-lib onboardings could hit:

1. async/await support was simply not yet built into this session's shared driver — the checked-out ~/src/fluffos was on a local branch (claude/async-phase15) 3 commits ahead of its own origin/master, and the async/await feature itself (858d5da9, merged 2026-08-27) postdated the driver binary's last build (2026-08-21) by 6 days. Not a driver bug — confirmed via a clean minimal repro (async mixed foo() { ... } failed identically in a throwaway single-file test, succeeded after a plain rebuild) before concluding anything. Fixed by rebuilding ~/src/fluffos/build-debug (cmake --build . --target driver lpcc -j8) from the current checkout — a corpus-wide driver rebuild, not an oxidus-specific change, so every other lib in this corpus benefits/is affected by it too going forward. 2. **include/driver/*.h didn't exist** — oxidus's own build pipeline (adm/dist/rebuild) copies the driver's own src/include/*.h (type.h, socket.h, function.h, etc. — "headers shared between driver and mudlib" per their own README) into the mudlib tree at include/driver/ as a build step; this repo's libs/oxidus/work never got that step run. Fixed by copying ~/src/fluffos/src/include/*.h into libs/oxidus/work/include/driver/ once, by hand — this is a one-time bootstrap step, not something that needs to happen on every boot. 3. PACKAGE_UIDS stub applies missing — adm/obj/master.lpc had no get_root_uid()/get_bb_uid()/creator_file() at all (this lib doesn't use the UID/euid privilege model — "Oxidus uses the FluffOS privs model, not PACKAGE_UIDS" per its own AGENTS.md), but this corpus's shared driver build has PACKAGE_UIDS compiled in, which exit(-1)s the whole driver at boot without the first two apply. This is the well-catalogued AGENTS.md §7.2 pattern (also just seen again on majik4, §7.160) — fixed with the standard flat "everyone is Root"/"BACKBONE" stub, harmless since oxidus's real security model (path-based ACL + role groups in adm/obj/master/security.lpc) never consults these. 4. OLD_ED vs the modern ed_start()/ed_cmd()/query_ed_mode() trio — new AGENTS.md §7.161. This corpus's shared driver has OLD_ED defined (hundreds of other libs' own wizard editors depend on the classic ed() efun), which compiles the modern efun set out entirely; std/living/ed.lpc's ed_edit() had been rewritten to use them, breaking std/living/body.lpc's ENTIRE inheritance chain (every player/NPC class) at compile time. Fixed by reverting ed_edit() to the classic ed(file, write_fn, exit_fn, restricted) API — simpler than the session-style version it replaced, since the driver owns the whole input loop once ed() is called (no input_to() loop needed).

Real programming bugs fixed (beyond the driver-mismatch items above)

• New AGENTS.md §7.162: ed_write()/ed_exit() (the write- authorization/session-end callbacks passed to ed() as plain string arguments) were declared private, following this lib's own correct convention that fixed-name driver applies (create, heart_beat, valid_read, etc.) should be private since "the driver ignores visibility" for those. But ed_write/ed_exit are a DIFFERENT mechanism — a mudlib-chosen callback name resolved via the driver's generic by-name apply(), the same path call_out()/socket callbacks use, which explicitly does NOT bypass visibility (function_exists()'s own docs: private/protected functions are "treated as not defined" for any external caller). A private ed_write() is silently never invoked: ed's w/x commands always print the generic "Failed command.", with nothing in any log file anywhere (safe_apply() swallows "function not found" silently). Isolated with a throwaway command object before touching the real fix: identical ed()+callback code worked instantly as public, failed identically as private. Fixed by making both public.
• New AGENTS.md §8.21: adm/obj/login.lpc's create_body() called body->restore_body() a second, redundant time immediately after BODY_D->create_body() had already restored the body AND run std_setup() (which computes role-based command search paths — /cmds/adm/, /cmds/dev/, or just the standard set, depending on adminp()/devp()). Since __command_paths isn't nosave, the second restore silently reverted it to whatever was saved BEFORE the character's last promotion — every admin/dev who reconnects (not just their very first post-promotion login, since makeadmin/makedev add the paths directly to the live body) gets full normal gameplay but every /cmds/adm/, /cmds/dev/, /cmds/file/ command — including the in-game ed/edit editor — silently doesn't exist, with the driver's generic "What?" fail message giving zero signal this is a permissions bug rather than a typo. Root-caused via targeted debug instrumentation showing std_setup()'s own get_path() result (6 entries, correct) diverging from command_hook()'s live __command_paths (3 entries, stale) moments later. Fixed by deleting the redundant call in login.lpc.
• std/living/body.lpc's std_setup() called adminp()/devp() with NO argument, which defaults to previous_object() — at that call site, previous_object() is std_setup()'s OWN caller (body_d, or the driver's create() apply), never the body itself, so the role check was silently checking the wrong identity's privileges (always false in practice). This bug was masked by the §8.21 restore-clobber bug above during testing (fixing one alone wasn't enough to observe correct behavior; both needed fixing together to actually see role- based paths survive a reconnect) but is independently real and worth fixing on its own merits. Fixed by passing this_object() explicitly.
• std/cmd/reporter.lpc's set_git_hub_label() (shared by the bug, idea, and todo commands, both the std and ghost/dev variants — 5 command files total) called mud_config("GITHUB_REPORTER") directly; CONFIG_D->get_mud_config() error()s on any unset key rather than returning null/0, and GITHUB_REPORTER has no default and ships unconfigured in a fresh self-hosted install (nothing in adm/etc/default.lpml, no example config anywhere, no mention in README/DOCKER.md's setup steps). Since this ran from setup() during the object's own create(), the uncaught error() took the WHOLE command down for every player, not just the optional GitHub-issue- filing half — even though finish_report() already has a graceful git_hub_label != "" skip-if-unconfigured path built in, meaning the intended "GitHub filing is optional, local-log-only reporting always works" behavior (explicit in three of the five commands' own help text) was entirely broken by this one setup-time validation. Fixed by wrapping the mud_config() call in a catch() and returning early (leaving the label unset, matching the already-correct downstream skip) when the config key genuinely doesn't exist, while keeping the strict error() for a config that DOES exist but names a label the caller passed that isn't in its types list (a genuine caller bug, should stay loud). Issue #3: a caught Oxidus error() is still logged and announced, so 2026-09-10 replaced the catch() with a CONFIG_D->get_all_config() presence check (no throw).
• std/object/include/object.h had a stale duplicate/conflicting void set_real_name(string str); prototype (a second, correct string set_real_name(...) prototype already existed 4 lines later) and a stale void remove(); prototype conflicting with the real int remove() implementation in object.lpc — both produced "Previous function prototype ... does not match" warnings on every single boot (cosmetic, no functional impact, but real dead/wrong declarations). Removed the two stale lines.

lpcc-only artifact — NOT a real bug (cross-checked against a live boot)

lpcc_check.sh's batch compile-sweep segfaults deterministically at /d/village/bakery (confirmed reproducible: identical crash, same file, same C++ stack trace, on two separate runs). Root cause: FluffOS's main_lpcc.cc unconditionally calls Tracer::start("trace_lpcc.json") at the top of lpcc_main() — lpcc ALWAYS runs with the internal profiling Tracer active, unlike a normal driver boot (mainlib.cc only starts it when a --trace-log-style flag is explicitly passed, which nothing in this pipeline does). The crash itself is inside the Tracer's get_trace_id() (a null/garbage function_table[...].funcname pointer reaching std::string's constructor → strlen() segfault) and is reached via a nested call_other → simul_efun evaluate() → call_other chain during bakery.lpc's create() (it's the only file in the village using EXT_SHOP_MENU). This looks like a real driver- level bug in the Tracer/frame-bookkeeping code on this session's WIP claude/async-phase15 branch (a code comment elsewhere in the driver already flags a concern about FRAME_ASYNC/FUNC_VARARGS bit overlap in exactly this kind of frame-kind check) — but it is a compile-sweep- tool-only artifact, not a real gameplay bug: bakery loads and plays correctly under the real driver boot (confirmed live — visited the room via a fresh player character, no crash, normal shop room). Per the established "lpcc-vs-live-driver artifact" pattern in this project, flagged here rather than "fixed," and NOT reported as an oxidus bug. The rest of the corpus (482 other files, bakery excluded from that one run) compiles 100% clean via lpcc_check.sh.

Intentionally not "fixed" — real, environment-dependent, by design

• adm/daemons/discord/chatter.lpc (explicitly labeled "Chatter Discord Bot Example" in its own file header) fails to load with error("Bot not setup") — it needs a real Discord bot token via DISCORD_BOT config, which nothing in this sandboxed environment can provide. It is not preloaded and not reachable from normal gameplay; this is example/reference code for admins who want to wire up a real bot, not a broken feature. Matches this corpus's existing precedent for imud's real Intermud-3 network dependency (§2.3).

Playthrough verification

Full session via scripts/tmux_mud.sh, native driver (~/src/fluffos/build-debug/src/driver config.fluffos, cwd libs/oxidus), port 40256:

• Registered a new account (elarion/password, with confirmation) → became the mud's first character/owner automatically → landed in the admin-only "Developer Workroom" starting room (not the normal player start) → look/score both correct.
• Registered a second, independent account (wisteria) → normal character creation (lowercase-only name validation confirmed) → landed in the real starting room, "Village Square of Olum" → look/ score/inventory all correct → walked west twice into the village content (tavern, a "gathering place for newcomers" room) → confirmed real, distinct room descriptions, not placeholders.
• Reconnect/restore path verified independently for both accounts (quit, disconnect, reconnect, select existing character — correct each time).
• Promoted wisteria to developer rank (makedev, run as the admin account) to test wizard-only commands on a non-admin identity; used dest/reconnect cycles (destroying and recreating the live body object) to confirm each fix against a genuinely freshly-compiled character rather than a stale in-memory clone predating a given fix — this was essential for finding and confirming both the §8.21 restore- clobber bug and the §7.162 private-callback bug, neither of which would have been distinguishable from "already fixed, just testing a stale object" without it.
• Full ed (in-game line editor) session verified end-to-end as a freshly-promoted developer: ed <file> on a nonexistent file → a (append) → typed a line → . (end input) → x (write and quit) → "Exit from ed." with no error, and the file's actual on-disk content confirmed correct afterward (home/w/wisteria/final.txt). Also confirmed the permission gate itself works correctly: a write: "self" path rule (/home/*/*/**) let a developer write their own home directory but correctly denied a write to mudlib root (/tmp_final_verify.txt → "Permission denied.", not a crash).
• quit clean both times (proper goodbye message, connection closed).

Minor unfixed observation (not root-caused, doesn't block anything)

An empty directory named after the character (e.g. wisteria, elarion) appeared at the mudlib ROOT (not under /home/<letter>/<name>/, the correct home-directory location) after the first admin bootstrap login and again after a makedev promotion. Harmless (empty, never referenced by anything observed during the playthrough) and not chased down to a specific line — cwd defaults to /doc (std/living/player.lpc), not a bare name, so the stray mkdir isn't there; likely somewhere in login.lpc's admin/dev promotion path or body_d. Deleted both before committing (test-session cruft, not shipped content). Worth a closer look in a future pass if it turns out to recur for every ordinary character too, rather than just the admin/dev promotion paths tested here.

Issue #3 — conversion corrupted compatibility guards (2026-09-10)

scripts/convert_lib.sh treated every ".c" and every static as a migration token. Restored upstream Oxidus meaning (gesslar/oxidus-mudlib main):

• Dual-extension guards / .c fallback: source_file() in adm/simul_efun/file.lpc, plus vnum.lpc, master.lpc, travel.lpc, showtree.lpc, trace.lpc strip_ext(), error_tester.lpc, std/cmd/cmd.lpc.
• MSSP English “static values” (not the LPC modifier): mssp.lpc.
• Optional reporter: get_all_config() lookup instead of catch(mud_config("GITHUB_REPORTER")).
• tests/.../string.safe.test.lpc “keeps interior dots” back to a.b.c.

Quoted includes and collection-driver comments are left as shared-build adaptations (documented above), not reverted. 2026-09-15 follow-up: softened the three LPC comments that cited AGENTS.md section numbers (std/living/ed.lpc, adm/obj/master.lpc, adm/simul_efun/hash.lpc) so they say "required by the collection's shared driver build" without implying those constraints are upstream Oxidus policy. Issue #3 closed.

WASM

Not yet WASM-packaged for the browser gallery (§1.4 pass) — this session's scope was native onboarding + bring-up per the standard pipeline; WASM triage left for a follow-up pass, matching the same "native-boot verified, WASM pending" status other very recently onboarded libs in this corpus (e.g. rifts2) currently carry.

WASM measurement (2026-09-03)

meta.json was already playable from the 2026-08-31 deploy-unblock; the README still said "not yet packaged." Cold-boot under the shared ~/src/fluffos/build-wasm reached the account prompt, but every room failed to compile: std/room/room.lpc gmcp_room_info() calls hash("md4", ...) (a PACKAGE_CRYPTO efun the WASM build does not have). Added adm/simul_efun/hash.lpc (compiled only when __PACKAGE_CRYPTO__ is absent; native keeps the real efun) and included it from adm/obj/simul_efun.lpc. After that, scripts/wasm_client.js created account wasmxid / character wasmox / Mud@2026 and landed in Village Square of Olum (five exits), look reprinted the square, score showed "You are Wasmox, a level 1 human." HP 100, quit printed "Y'all come back, now, y'hear?" Shop/combat/death were not exercised this pass. The MEMFS copy does not write the throwaway account back to the host.

§10.7 deep functional test (2026-08-31)

Round-two deep-functional-test pass, going beyond the onboarding playthrough into actual game systems: combat, skills/spells, equipment, shop economy, consumables, death/respawn, and a real quit+reconnect gap. Native driver (~/src/fluffos/build-debug/src/driver config.fluffos, rebuilt since onboarding — confirmed still the correct build, async/await and ed() fixes both still in effect), port 40256, via scripts/tmux_mud.sh. Three accounts used: thornwick/elmswood (first-ever character on this driver boot, auto-admin, used for makedev promotion and eval-assisted mob spawning), briarwyn/willowmere (second character, used only to verify the stray-directory fix), and cinderfall/moonshade (main playthrough character, promoted to developer for gimme/healup).

Upstream movement check (informational, not acted on)

Per the task brief, checked whether gesslar/oxidus-mudlib has moved since this project's clone (2026-08-28/30). It has, substantially: gh api repos/gesslar/oxidus-mudlib/commits?since=2026-08-28T00:00:00Z returns 21 commits between 2026-08-29 16:01 and 2026-08-31 14:02 (today, a few hours before this test pass — gh repo view confirms pushedAt: 2026-08-31T17:44:59Z). Notable commit subjects suggest the areas most likely to have diverged from this snapshot are exactly the two this project's onboarding had the most friction with: the login/account system (login object updated, promise-driven login handshake) and the async/promise infrastructure (gmcp_await, promises header file, request_deadline never fulfils; type it that way, LPCDoc unwraps promises; document the payload). This is informational only per the task brief — not re-synced, not investigated further, and none of the bugs found in this pass were re-checked against the newer upstream code. A future re-onboarding pass should expect the login/async areas in particular to have changed shape.

Bug found and fixed

adm/obj/login.lpc:449 — new_character()'s home-directory setup call passed the bare character name to assure_dir() instead of a real path, creating a spurious empty directory named after the character at the mudlib ROOT on every single character creation (not just admin/dev promotions) — this closes the "Minor unfixed observation" left open in this lib's own onboarding NOTES.

• The function's own doc comment says it "Validates length, format, and uniqueness, then creates the character, body, and home directory" — but the actual call was assure_dir(str); where str is the bare lowercase character name (e.g. "elmswood"). assure_dir() (adm/simul_efun/file.lpc) builds a directory tree by explode(path, "/") and mkdir()-ing each successive segment; a bare name with no slashes just does one mkdir("elmswood") relative to the mudlib root, not the character's real home directory. home_path(name) (the correct helper, already used two lines further down inside first_admin_login() for exactly this purpose) returns "/home/<first-letter>/<name>/".
• Confirmed the stray directory is genuinely harmless-looking but 100% reproducible for EVERY new character, not just the first-ever admin bootstrap as the onboarding note speculated: created a third, ordinary (non-admin, non-dev) character moonshade on a fresh driver boot before applying the fix and got work/moonshade/ (empty) at the mudlib root exactly as before.
• The real home directory (/home/<letter>/<name>/) is otherwise never created for any character except the very first (first_admin_login() only runs once, gated by the FIRST_USER marker file) — so every ordinary player's home directory silently doesn't exist until something else happens to create it (e.g. ed's own write-permission path, which per onboarding's own testing DID succeed writing to a non-admin's home dir — the file-write efun apparently tolerates a missing intermediate directory here, or creates it, so this was latent rather than an active second bug).
• Fix: assure_dir(str) → assure_dir(home_path(str)). Verified live end-to-end: reverted the stray-directory symptom by creating a fresh character (briarwyn/willowmere) BEFORE restarting the driver and confirmed the bug still reproduced (objects don't recompile from a live edit — driver restart required, per AGENTS.md §10.3), then restarted the driver, applied the fix, created a THIRD fresh character (cinderfall/moonshade) and confirmed: no stray directory at mudlib root, and /home/m/moonshade/ now exists with the correct .keep sibling structure. A final full long-sit idle boot watch (~220s, native mudclient.py, nobody logging in) also confirmed no new stray directories appear from boot alone.

Confirmed working, no bugs found

• Shop economy, both architectures (currency-and-shops skill): the bakery's EXT_SHOP_MENU (list/buy muffin/eat muffin, correct price and change-in-copper math, item cloned fresh and consumed) and the tailor's EXT_SHOP (list/buy tunic/wear tunic/remove tunic/sell tunic, correct sell_factor payout and exact-coin change on both legs of the transaction). wealth command, gimme dev command, and adjust_wealth()'s currency-denomination math all verified correct across several buy/sell round-trips.
• Equipment (equipment skill): wear/remove on a clothing item worked correctly end-to-end including the shop round-trip above.
• Consumables (consumables skill): eat muffin correctly depleted the single-use food item and removed it from inventory with the documented "You have eaten the last of the muffin." message.
• Skills (skills-and-advancement skill): use_skill()'s auto-creation-on-first-use fired correctly — punching for the first time printed "You have gained a new skill: combat.melee.unarmed.", taking a hit for the first time printed the matching combat.defence.dodge message, and casting a spell for the first time printed arcane.discipline.lightning. The skills command (cmds/std/skills.lpc) correctly prints nothing for a skill-less character (its own recursive walk over an empty query_skills() mapping legitimately produces an empty string) — this is expected behaviour given initialize_missing_skills() (the function that would pre-seed the config's default skill tree) is dead code, never called from anywhere; not a bug, since the use-based auto-creation path is the one actually exercised by every real gameplay action.
• Combat (combat-system skill): fought a level-1 field mouse (admin-spawned into the village's field virtual zone via add_inventory("mob/field_mouse") — the correct virtual-object path form; see gotcha below) with the punch ability end-to-end across a full multi-round fight — hit/miss rolls, cooldown gating ("You must wait before you can use that again."), mutual damage, and HP loss all behaved correctly and consistently with the formulas in the skill doc.
• Death and revival: died for real (HP reached 0 from the field mouse's counter-attacks) — "You have perished." fired, a the rotting body of moonshade corpse appeared, wealth converted to loose coin objects in the room (not inside the corpse — a small presentation detail, not chased down further), and the player auto-revived 1 second later at 1 HP/1 SP/1 MP via ghost.lpc's call_out("revive", 1) — a deliberately forgiving, near-instant revival design, not a bug. The revived player's own move_cost()-gated movement correctly refused a distance-2 exit ("That location is too far away to travel to right now.") until MP regenerated — also correct, not a bug (confirmed by reading std/room/room.lpc's move_cost()).
• Arcanist spell tree (skills-and-advancement + combat-system skills): found Madame Brille in the manor's salon (see navigation gotcha below), learn arcane correctly granted /cmds/spell/arcanist/ path access, shield (self-buff, costs SP) and shock <target> (offensive lightning bolt, costs MP, trains arcane.discipline.lightning) both worked correctly — the shield's damage-absorption message ("Your inertial shielding absorbs some of the damage.") correctly appeared on subsequent hits taken.
• Reconnect with a real gap: quit from the moonshade character, waited (a ~4-minute gap covering the long-sit boot-watch pass below, which required a driver restart in between), reconnected with character@account form — full state correctly restored (room, HP/SP/MP, wealth all exactly as left).
• Long-sit idle boot watch (~220s, native mudclient.py --idle 220 --timeout 220, no login attempted): clean, only the same cosmetic compile warnings already documented as pre-existing/harmless; the connection was closed by the login object's own legitimate 60-second login timeout, not by anything unexpected.

Not a bug: two same-id doors in the manor's Foyer

d/village/manor/foyer.lpc has two add_door() calls both using id: "foyer door" (one north-facing to the porch, one east-facing to the salon) with no distinguishing adjective. A bare open door (or open foyer door) in that room correctly reports "There are multiple doors with that name." — looked like a content bug at first (the room's only two exits both blocked), but std/room/door.lpc:add_door() already auto-appends a direction-specific alias (door.name = "<direction> door") to each door's id array specifically to handle this case: open north door / open east door disambiguate correctly and are the intended way to interact with a room that has multiple same-named doors. Confirmed live. Not fixed, because it isn't broken.

Flagged, not fixed: benign first-boot-only messaging race during the very first character's world-entry

Every fresh driver boot's first-ever new_character() call (which is also always the auto-admin bootstrap path) logs 2-3 caught, non-fatal *Object cannot be loaded during compilation. errors from std/ext/messaging.lpc:172 (receive(message)), reached via enter_world() → move_living() → force_me("look")'s own command_hook() chain. Confirmed genuinely reproducible, not a one-off: it happened identically for elarion (the original onboarding session's first character, 2026-08-30 19:51, per log/runtime) and again for elmswood (this pass's first character, 2026-08-31 12:24) — but did NOT recur for any subsequently-created character on either driver boot (wisteria, briarwyn/willowmere, cinderfall/ moonshade), confirming it's specific to the very-first-ever compile of the full player class hierarchy on a cold boot, not a general messaging bug. enter_world() already wraps the relevant calls in catch(), so nothing crashes, and the player-visible outcome was correct both times (the new admin lands in the Developer Workroom with a correct room description). Root cause not fully chased down (plausibly a COLOUR_D/LINES_D daemon still finishing its own first-ever lazy compile at the exact moment receive() tries to use it) — flagged here per this round's "document honestly rather than guess" instruction rather than risking a speculative fix to first-boot compile ordering.

Flagged, not fixed: defence/defense skill-path spelling split (zero live impact)

Every live combat/spell code path (std/living/combat.lpc, std/living/damage.lpc, cmds/ability/punch.lpc, weapon procs, all five arcanist spell files) consistently uses the Canadian-English spelling "combat.defence.dodge" / "combat.defence.evade" — correct per this project's own mandated Canadian-English house style (AGENTS.md's own example list literally includes "defence"). Live gameplay in this pass confirmed this spelling is what actually gets trained and queried ("You have gained a new skill: combat.defence.dodge."). However, adm/etc/default.lpml's configured default skill tree, and this lib's own .claude/skills/ documentation (combat-system and skills-and-advancement SKILL.md), both use the American spelling "combat.defense.dodge" / "combat.defense.evade" instead. This is a real spelling inconsistency, but has zero live functional impact today: initialize_missing_skills() (the function that would seed the config's "defense"-named default tree onto new characters) is dead code, never called from anywhere in the codebase, so nothing ever tries to reconcile the two trees — every skill node in practice is created lazily by use_skill()'s own auto-creation using whichever spelling the calling code happens to use, and combat code is internally self-consistent. Not fixed: which side is "canonical" here is genuinely ambiguous (more call sites use "defence", but the config schema and generated docs use "defense"), and there is no live bug to reproduce/verify a fix against — flagged for whoever eventually resolves it (initialize_missing_skills() dead code, or the doc spelling, or both) rather than guessed at.

Gotchas hit while testing (useful for future passes on this lib)

• Virtual mob paths are NOT filesystem paths. Spawning a monster for testing requires the routed virtual form add_inventory("mob/<name>") (no leading slash, "mob/" prefix), which master()->compile_object() routes through adm/daemons/modules/virtual/mob.lpc to /d/mobs/<name>.lpml. Calling add_inventory("/d/mobs/<name>") directly (a real absolute path, since only the .lpml data file exists there, no .lpc) fails with "Something went wrong." broadcast to the room and logs *Too deep recursion. to log/VIRTUAL — an artifact of feeding the wrong path form, not a mudlib bug (self-inflicted during this pass, verified the correct form works cleanly on the next attempt).
• this_body()/this_player() inside the eval dev command's scratch object is unreliable for chained calls. eval compiles the given statements into a fresh, unrelated /tmp/eval_<name>.lpc object and runs them with the caller's privs, but this_body() (efun::this_player()) inside that context returned 0 when chained (this_body()->environment()). Use find_player(name) to get a reliable reference to a live player object from inside eval, and the bare environment(ob) efun (not ob->environment(), which calls a same-named method that doesn't exist) to get their room.
• The village's own two north/south "A Path Through the Village" rooms (village_path1/village_path2, reached from the square's north and west exits respectively) share the same room title and a very similar opening sentence structure but lead to different content (bakery vs. tavern) — easy to misnavigate by title alone; check the room's own listed exits, not just the heading, when route-planning through the village by hand.

深度功能测试（§10.7，2026-09-16）— post submodule-patch cutover

Re-verified against the cutover tree: work/ = gesslar/oxidus-mudlib @ b5beb8ab + patches 0001–0003 + overlay/ (driver headers incl. promise.h, adm/custom/first_user). Port 40256. Cutover wiped prior seeds — first admin this pass was olumadmin / account deepox / Mud@2026 (wipe first_user + data if re-testing first-admin on a dirty tree).

2026-09-17 follow-up: upstream.auto_rebase re-enabled; pin tracks main. Local native driver rebuilt with promise phase 1.5. Site WASM release v2026.0901.0 still lacks combinators → wasm_status: limited until the next published FluffOS WASM includes them. The 2026-09-16 WASM pass below was on the pre-combinator pin b5beb8ab.

Native (scripts/tmux_mud.sh)

Continuous play on the patched+overlay work tree: shop (list / buy tunic at the tailor — exact item name required), combat (environment(this_body())->add_inventory("mob/field_mouse") then punch, not kill), halt, clean quit, reconnect as olumadmin@deepox. No new programming bugs vs the 2026-08-31 pass; cutover patches/overlay hold.

WASM

node scripts/wasm_client.js … libs/oxidus with olumadmin@deepox / Mud@2026 → Developer Workroom look / score ("You are Olumadmin, a level 1 human.") / quit ("Y'all come back, now, y'hear?"). Playable. Same first-boot messaging-race noise as §10.7 2026-08-31 (caught, non-fatal).

Cleanup

git -C work reset --hard + apply_lib_patches.py oxidus + overlay rsync; removed throwaway play dirs (olumadmin/, etc.). Patches stay out-of-tree in patches/ / overlay/ — do not commit dirty submodule work tree.

深度功能测试（§10.7，2026-10-05）— 主控 author_file()/domain_file()

共享驱动的 mudlib_stats 包在主控载入时各调用一次 author_file(<主控文件>) 和 domain_file("/")。补丁 0001 加的 stub 里没有这两个 apply，所以每次启动都打印两行 ... in the master file does not work。0001 的 stub 块补上 author_file() / domain_file()，返回 get_root_uid() / get_bb_uid()（ROOT / BACKBONE），正是驱动缺省时用的值（KB 06 §7.222）。写进 0001 而不是新补丁：rebase 脚本单独检查每个补丁，新补丁的上下文会是 0001 加的行（KB 02 §2.3）。site 同构树（pin + 补丁 + overlay）跑 lpcc config.fluffos /nonexistent.lpc 前后对比，只少了这两行。

Deep functional test (§10.7, 2026-10-13): compile warnings 64 -> 0 (catalog patch 0004)

work/ is the upstream submodule, so the in-place warning campaign never scanned it. Scanned in a scratch repo (scripts/submodule_patch_scratch.py build oxidus); it needs the overlay's include/driver/*.h (now laid by scripts/pristine_tree.sh) and an lpcc with the promise combinators (build-debug/src/lpcc, not the older build-debug/bin/lpcc). 490 files compile, 64 warnings -> 0 after the fixer and hand fixes; exported as patches/0004-fix-compile-warnings.patch (byte-for-byte round trip; all four patches apply on the pin). 26 changed objects loaded before and after in fresh drivers: 25 -> 26, no regressions; the patched tree boots to the login prompt.

Bugs behind the warnings:

• d/cavern/cavern_daemon.lpc: its private helper set_path(z, y, x) shared a name with the public set_path(mixed path) that std/object/command.lpc provides, so the carving loop's calls never carved a tile, !is_path() stayed true and the daemon died on the eval-cost limit every time it loaded (three fresh runs each way: fails at the pin, loads with the fix). Renamed carve_path(); d/maze/maze_daemon.lpc had the same helper and got the same rename.
• adm/daemons/modules/gmcp/Char.lpc: the Char.List filter (: $1 != $2, $(who) :) is a comma expression that always returns who, so the player was never excluded from their own list; now (: $1 != $(who) :).
• adm/simul_efun/lpml.lpc: "\x01BACKSLASH\x01" — \x takes every hex digit after it (\x01B), so the placeholder was not the intended bytes; now "\x01" "BACKSLASH" "\x01".
• Rest: english.lpc case "none" : "its"; without return (5, same result through default), adm/obj/login.lpc inherited EXT_MESSAGING twice (STD_ITEM carries it; identical initialisers) and a private __name like its base's, std/ext/proc.lpc __cooldowns vs std/object/cooldown.lpc (two private maps, renamed __proc_cooldowns), prototypes made to match (set_wealth, generate_map, query_value now mixed because lib/coin.lpc returns an array), (mixed) casts on deliberately mixed ?: results, /* inside doc comments.

• 2026-10-14 (FluffOS 970d9f09 rescan): new catalog patch patches/0005-float-truncation.patch, 13 Float value truncated to int sites (unicode bar widths, food/drink percentages, XP adjustment, heart rate clamp, master read-size limit) spell their truncation to_int(); same values. Each of the five patches applies alone on the pin and in sequence; the 6 changed objects load before and after.