Repository navigation
Trace: keep the action verdict (res/reason) and add a spells oracle to trace-score #289
Description
Source: ~/TortoiseBots-audit/PLAN-SPELL-AUDIT.txt (spell audit of 33 dungeon runs RFC/WC/Uldaman/ZF, 2026-10-20/21, 1 Hz TBM:VIEW feed; [V]=verified in data/code, [I]=inferred, [U]=unknown). Run ids are tests/ folders under ~/TortoiseBots-audit. Module commit read: 9f929a13.
Gaps found while auditing spell use (all runs r1-r7):
T-a: trace frames drop the verdict letter and reason (internal/dungeon/runtrace.go:454-456 does not copy tbm.PartyRec.Result/Reason). The 1 Hz VIEW only carries the newest verdict, so impossible/failed rows cannot be analysed from the trace.
T-b: run_meta.server_trace was off in every r1-r7 run, so the module's full action record (name, res, rel, src_trigger, cast_interrupted, killer) was never collected. This is the same blocker as #249 (CheckCommand rejects '.bot trace start'); this issue asks that test rounds turn it on (run_meta.server_trace=true + driver word in the slot settings; AiPlayerbot.RunTrace.Enabled=1 on the test server, needs a server config change = ask the owner first) and that the .jsonl/.summary.json is copied back with the run folder.
T-c: no spell-usage oracle in trace-score. Add checks (thresholds via --set, unit tests on small synthetic traces, each with pass + fail case, fail-if-removed; the 1 Hz caveat goes into each Detail): spell_mix (report), idle_share (15 %), mana_at_death (report), oom_in_fight (1 per 10 fights), interrupt_cover (30 % on CC/heal casts), taunt_latency (median 3 s), cc_hold (50 %), pet_threat (5 %), impossible_spam (report). Details: plan sections 1 and 6.
Owner: OURS (turtle-agent). Unlocks verification of the module issues filed from this audit.
Activity
Merge pull request #291 from pjerra/feat/trace-spell-usage
pjerra commentedon Oct 21, 2026
Part done in PR #291: the verdict letter/reason (res/why) are kept in the frames and trace-score has a report-only 'spells' check with a Spells section. Still open here: collecting the server-side run trace (the config change on the test box) and the taunt-latency, cc-hold and pet-threat checks.
Merge pull request #297 from pjerra/fix/ta289-spell-checks
pjerra commentedon Oct 22, 2026
Progress: T-c is merged into dev via PR #297 (e48f459): the spells check reports taunt_latency (median over 3 s warns), cc_hold (long CC only, under 50 % held warns; Entangling Roots included) and pet_threat (over 5 % of mob-frames warns), all warning-only and settable with --set. Still open, so this issue stays open: T-b (server-side run trace on the test box needs the owner's config decision and round scripts copying it back) and oom_in_fight as a warning.
pjerra commentedon Oct 28, 2026
Status 2026-10-28
Done:
res/why), and a report-onlyspellscheck exists.e48f4591):taunt_latency,cc_holdandpet_threatchecks, warning-only.Still open:
oom_in_fightis not a check yet (the check ids ininternal/tracecheck/tracecheck.goend atpet_threat).fight-scorecommand that agrees with the dashboard fight debugger's checks.pjerra commentedon Nov 1, 2026
Fixed on dev: T-a in #291 (abd37a3), T-c in #297 (e48f459), oracle and fight-score in #316 (f1ff3fa), so
oom_in_fightexists (internal/tracecheck/oracle.go:19). The one leftover, T-b (turning server_trace on in the test rounds), is a TB-side config choice and not code here.