Add primer-guided deck improvements and fix commander/import reliability - #85
Open
LlamaAdam wants to merge 10 commits into
Open
Add primer-guided deck improvements and fix commander/import reliability#85LlamaAdam wants to merge 10 commits into
LlamaAdam wants to merge 10 commits into
Conversation
40 community primers (21 Moxfield + 19 Archidekt, harvested 2026-08-29) distilled to per-deck structured records, shipped as package data with a fail-quiet, immutable, offline loader. Profiles-not-one-truth (§13): profiles_for_commander returns every build of a commander; win-line card names carry per-card mainboard-verification flags; budget swaps flatten into a §10 function-preserving swap corpus; prompt_block_for_commander renders clipped LLM context via primer.clip_for_prompt.
CONSISTENCY_TARGETS names the section-1 numbers that converged across independent primer authors (85% third land drop, 85% commander on curve, 90% card advantage by T5, the 13-enabler free-mulligan rule, at most 2 unconditionally-tapped fetchables) and evaluate_consistency_targets grades a deck against them, reusing the Monte-Carlo projection deck_health already computes plus closed-form hypergeometrics. Conditional floors (t1_enabler_plan / proactive_t2_plan) report their value but stay unjudged until a caller declares the plan. Wired as an additive-only deck_health tile under the standard outage contract; the health grade keeps ignoring it.
Renders met/evaluated with per-check tooltip, mirroring the consistency tile's three-state null handling. Display-only; the letter grade ignores it. Key-set pins in the audit tests extended for the additive consistency_targets key.
…e classifier contextual_role_targets adjusts ROLE_TARGETS by archetype, commander role, average mana value, and bracket, per the primer synthesis (Edgar aggro spec, Gishath resolve-engine deltas, Winota trigger-density trade, bracket-scaled interaction floors). One documented delta table is the tuning surface; unknown labels degrade to the flat table. role_target_report gains a keyword-only context= that is byte-identical to the old report when absent. infer_commander_role classifies only the regex-trustworthy roles (cost cheater, trigger multiplier, resolve engine by mana value) and returns None when unsure.
Four narrow, oracle-signature modifiers from the primer synthesis: commander_dependence (-8, effect gated on controlling the commander), tempo_fail (-6, value a turn-cycle late under an aggro plan), capped_engine (-4, printed once-each-turn limiter), tutor_top_delta (-5, Vampiric-class top-of-library tutors outside combo shells). Patterns are deliberately conservative — a miss costs nothing, a false positive would down-rank a fine card on every audit. Same flag and validation gate as the rest of FP-015: these refine the ranking prior, they do not enable it.
nonbo_lint encodes the primer synthesis section-14 table as data-driven pairwise checks (Skullclamp vs own anthems, Cursed Totem vs own dorks, forced draw vs Thassa's Oracle, Heartless Summoning vs one-toughness creatures, mode-exclusive Asceticism/Everlasting Torment, own shroud vs own targeting, symmetric-effect notes) — the first anti-synergy detector in the tree. Selectors match by exact name or oracle/type predicates; a card never pairs with itself; unresolvable cards match nothing. Wired as the additive-only 'nonbos' deck_health tile with the standard degrade contract.
…, budget swaps build_judge_prompt gains a COMMUNITY PRIMER CONTEXT block (clipped, attention-steering only, identical across A/B orderings) when the bundled KB covers the deck's commander. The Claude advisor payload carries community_primer_consensus under the same omit-when-absent policy as bracket peers, and its system prompt gains the distilled card-evaluation principles. Budget mode surfaces author-documented function-preserving swaps for the deck's own commander as paired add/cut recommendations (supplemental source, same post-processing as lift picks, fail-quiet).
LlamaAdam
commented
Sep 3, 2026
LlamaAdam
left a comment
LlamaAdam
left a comment
Owner
Author
There was a problem hiding this comment.
Round-3 adversarial review of this PR (maximum-effort pass; every item below was verified against source with file:line and executed reproductions, then cross-examined from the defense side — the cross-examiner's corrected version is what's stated here). Full report with evidence: docs/ollama-analysis/NEGATIVE_MODE_ROUND3.md on LlamaAdam/mtga-advisor#5, §1a and §6.
Owner decisions this PR needs before merge (recorded as R3-D1/R3-D2 in DECISIONS_FOR_REVIEW.md on that PR):
- Auto-Protect reversal (PR-06, major). This PR removes FP-018.3's primer card-link auto-protection and inverts the test that pinned it (
tests/test_adopt.py:239-250). The rationale in the commit is sound engineering — which is exactly why it's a product trade for the owner, not a bug fix; no decision record or review exists for it. - Knowledge-base provenance (PR-05 B3).
data/primer_kb.json,consistency_targets.py,staples.py,card_score.py,nonbo_lint.pyand_advisor_claude.pyciteprimer_harvest/deckbuilding_heuristics.md§1–§16 and two harvest JSONs that exist in no tree, branch or commit; 35 of the KB's 40 decks have no list or prose anywhere in the repo. Not a code defect — a question of whether an unreproducible data asset ships. - Duplicate
/api/deck_commander(PR-08, major). #84 adds the same endpoint with a different contract;git merge-treeof the two conflicts in 7 files / 18 hunks, and both declare the same top-levellets inapp.js(a SyntaxError for the bundle if both land). Each PR ismergeable_state: cleanalone, so GitHub warns nobody.
Fixes for this PR, in priority order:
- PR-02 (major) — the import "hardening" now 400s whole pastes master accepted:
Maybeboard/Commanders/Tokensheadings and Archidekt's default[Commander{top}]tail. Recognise them innormalize_card_line/deck_text_opsrather than rejecting. - PR-03 (major) —
quoted_win_linestreats any paragraph mentioning "maybeboard" / "Cons:" / "Updates" as a section heading and silences every later paragraph; one in-sentence mention blanks a primer's win lines. Heading heuristic in_primer_tokensmust not fire on in-sentence words. - PR-05(A) (major) — the KB budget-swap table emits non-card strings as advisor adds (14/36 rows, e.g.
"Kamahl, Heart of Krosa / End-Raze Forerunners","$200-tier builds");_safe_ci_lookupskips rather than drops them, so they reach output. Validate rows at load; pin every shipped row as one resolvable card name. Also PR-17key_cardsparses to empty for all 40 profiles (data shape ≠ loader), PR-18Bruce Banner // The Hulk (gift build)commander key, PR-S2 Ur-Dragon "d10 tokens" (d20). - PR-07 (major) —
_primer_kb_blockputs community-consensus card names into the judge prompt, with noprompt_versiononJudgeReport, so the Phase-1 agreement study now pools two prompts with no column to split them — the exact G3 confound the scope doc pre-registered against. Stamp a prompt version; render the block without card names or gate it off by default. - PR-09 (major) — partner decks structurally never reach
legalon the dashboard (only the hero commander gets the guarded lookup); fetch the ≤2 command-zone cards the same way. - PR-04 (major) — three of the six FP-019 slices (context-aware roles/quotas, plan-conditional floors, nonbo advisor pre-filter) have zero production callers; either wire them or reword the PR body/CHANGELOG to the
future-plans.mdSTATUS line. - PR-12 / PR-21 — the commander editor and
import_deckaccept"2 Krenko, Mob Boss"as a card name (executed: the import route writes1 1 Krenko, Mob Bossand returns 200); the editor also demotes aProtect=-locked card without touching the lock. Reject^\d+\snames; warn on a locked choice. - PR-01 (minor after cross-exam) — "tests isolate Forge paths" is true for the corpus-reading paths but 13 modules bind
VENDOR_FORGE-derived constants at import time, so the autouse fixture cannot reach them; resolveDECK_DIR-style constants at call time and softenCHANGELOG.md:32. PR-13 reportnormalized_lines; PR-20main_countchanged meaning silently; PR-S1 the Asceticism/Everlasting Torment nonbo "why" states the wrong mechanism; PR-14 five modules now exceed the documented 800-line ceiling; PR-19 the "independent backend/Python/frontend reviews", wheel-build and manual-smoke claims have no artifact (GitHub shows zero reviews) — label or drop. - Mirrors of master findings this PR carries: F-02
adopt.py:165-166compares card-link names (Front // Back) against front-face.dcknames, so every linked DFC is reported "NOT in the list"; W-03 a UTF-8 BOM should be accepted, not 400'd; W-07 thecreate_appconsumer of an unvalidateddeck_dir; W-10 PUT line endings.
What held on Linux: the "4,590 passed / 26 Playwright" numbers are substantiated by this PR's own CI; the --run-live gate resists -m/-k/node-id selection; the deck-dir precedence table (8×3 combinations) matches the description; the desktop-lock test isolation is a real fix.
Generated by Claude Code
This was referenced Sep 3, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
--run-liveconsent for real-service tests, independent of slow-test selection.Verification
--run-slow): 4,590 passed, 1 skipped in 636.85 seconds. The sole skip is the explicit live-service opt-in; all offline slow tests ran. Existing positional-maxsplitdeprecation warnings remain.Limits
The manual/UI tests use stub card data and simulation reports. No live Forge gameplay or successful model-service request is claimed. The initial full run exposed a legacy live-Claude test missing its optional SDK (4,582 other tests passed); the live lane now requires an explicit opt-in and configured optional dependencies. User deck libraries and raw primer harvests were not modified.
See
docs/negative-mode-audit-2026-09-01.mdfor the original findings and resolution details. This PR includes the nine existing FP-019 commits; it does not merge or deploy them.