| name | convergent-bug-triage |
| version | 1.4.0 |
| description | Triage when a single campaign (or single LLM/prompt surface) has accumulated multiple sibling bug investigations — the prompt-discipline family where "LLM emits right structured field, render/persistence/propagation path drops it", the now-merged NPC-status persistence family where "LLM omits/writes-wrong `state_updates.npc_data[NPC]`", AND the canonical-state-anchor family where "LLM forgets what canonical `core_memories` / faction / NPC already established" (the LLM re-derives or contradicts canon). Trigger on /repro requests where session_search reveals 3+ sibling issues on the same campaign within 24h OR 5+ within 9 days, "is this a duplicate", "this feels like the same bug as |
| tags | ["worldarchitect","repro","prompt-discipline","bug-class","npc-status-persistence","canonical-state-anchor","canonical-dead-npc-revival","3-layer-durable-fix","sibling-campaign-cluster","deployed-vs-branch-state"] |
Convergent Bug Triage
When a single campaign (or single LLM/prompt surface) has accumulated 3 or more sibling bug investigations within a 24h window, the dispatcher is not looking at a single bug — they are looking at a bug class with N surface symptoms. Filing each as an isolated PR creates cascade churn; bundling them creates mixed-scope PRs that stall on review. This skill covers the triage decision tree.
When this skill fires
- A
/repro request lands on a campaign ID where gh issue list --repo <OWNER>/<REPO> --state all --limit 50 shows ≥3 issues opened against the same campaign in the last 24h, OR ≥5 within 9 days.
- The user asks "is this related to other bugs?" AND "what's in core memories?" — this is the canonical-state-anchor question pattern (see Pitfall 9); route here after confirming ≥3 siblings exist.
- The user describes a symptom where the LLM "forgot", "ignored", or "contradicts" canonical state —
core_memories, npc_data.<NPC>.status, faction control, NPC co-presence, scene history — even when the served prompt context likely contains the anchor.
- The issues share a family pattern even when their surfaces differ — e.g. "Aemond capture persistence", "queen-level-14 directive", "scene event not rendered", "NPC status propagation", "lineage canonicalization" are 5 different surfaces but they all reduce to "LLM emits correct structured field; render/persistence/propagation path drops it."
- The operator's instinct is "is this a duplicate?" — the answer is usually NO, but a sibling that should cross-reference.
The five-step triage
1. Classify the new symptom into the family
Load the candidate symptoms for the campaign:
gh issue list --repo <OWNER>/<REPO> --state all --limit 30 \
--json number,title,state,labels,createdAt \
--jq '.[] | select(.createdAt > (now - 86400 | strftime("%Y-%m-%dT%H:%M:%SZ"))) | "\(.number) [\(.state)] \(.title)"'
For each issue in the last 24h, ask: does the bug reduce to "structured field present, downstream path drops it"?
| Symptom surface | Field written correctly by LLM? | Drop happens in... |
|---|
Aemond capture persistence (status=captured) | Yes | npc_data persistence path |
| Queen-level-14 directive | Yes (or never written by parser) | state_updates.god_mode_directives save path |
Scene event companion_request in narrative | Yes | render path (narrative_integration vs narrative) |
| NPC narrative status propagation | Sometimes | state_updates.npc_data write path |
| Lineage canonicalization | Sometimes | player_character_data_extras move path |
| NPC status persistence — god-mode-retcon missing-write (added 2026-07-12, #8335) | No — god-mode agent emits admin-retcon narrative but never writes npc_data because it is locked into "narrative freeze" mode | prompt layer — god-mode agent needs a structured ## NPC Status anchor in its system prompt too, not just the gameplay agent |
| NPC status persistence — wrong-key write / key-confusion (added 2026-07-12, #8335 scene 628) | Yes but to wrong NPC key (e.g. Queen Rhaenyra instead of Queen Rhaenyra Targaryen) | prompt layer + state layer — fuzzy-key writes accepted by merge layer; canonical state retains both keys with diverging statuses |
If the new symptom also fits the family: it is a sibling, NOT a duplicate. File a new issue with explicit cross-references.
Note on the 9-day / N-instance cluster pattern (added 2026-07-12): the original trigger of "3+ in 24h" misses the most common shape — a campaign that accumulates 1-2 issues per day across many days because the underlying bug class is structural, not per-scene. The trigger threshold should be ≥5 issues against the same campaign in any 9-day window in addition to the original 24h rule. The 2026-07-12 repro of "queen is supposed to be dead but LLM forgot" was the 7th sibling on xK3fp5XrV24oarIINTF7 over 9 days; ruling it out as "just another scene bug" because no other issues opened today would have been wrong.
Note on the 7+ sibling floor (added 2026-07-19): the canonical-state-anchor family on Cg2m2TkGFFez7XBynEah accumulated 8 siblings in 24h (scenes 78/82/171/318/335/367/369/386/392). At this density, the durable fix is already in PR branches but unmerged — see Pitfall 8's three-check pre-flight. Filing per-scene issues beyond the 5th sibling is almost always wrong; the correct action is to flag the deployment gap (existing PRs not on origin/main) and request merge approval.
2. Decide sibling-vs-bundle vs duplicate
| Pattern | Decision | Why |
|---|
| Same root mechanism, different surface, different files | Siblings — N parallel PRs | Mixed-scope PRs stall on review; bundling conflates measurement-integrity and structural changes |
| Same root mechanism, same surface, same file | Duplicate — fold into existing PR | One fix covers both; second PR is churn |
| Different root mechanism, same surface | Siblings — parallel PRs | Different root = different fix; bundling forces reviewer to accept unrelated trade |
| Different root mechanism, different surface | Different class entirely — file normally | Not in scope for this skill |
The default for the prompt-discipline family is siblings, one PR per symptom. Each PR owns its own file/section.
3. Cross-reference in the new issue body
Open the new issue with explicit cross-references in a "Sibling investigations" section:
### Sibling investigations on the same campaign (today)
- #8160 (closed dup) — character-creation spells/abilities
- #8266 / PR #8267 (closed) — Aemond capture persistence
- #8275 / PR #8276 (open scaffold) — queen-level-14 god-mode directive
- PR #8271 (open) — NPC narrative status propagation in $PROJECT_ROOT/prompts/game_state_instruction.md
- PR #8265 (open) — lineage canonicalization
- #8277 (this issue) — scene_event not rendered into narrative
All five share the same family: LLM emits correct structured field, downstream path drops it.
Without these refs the next dispatcher will re-investigate from scratch.
4. Dispatch ONE AO worker per sibling, not a fan-out
Even when the family is clear, dispatch one AO worker per sibling issue — do NOT fan out N workers for N siblings in one shot. Why:
- Each PR's GREEN fix is in a different file (
living_world_instruction.md vs game_state_instruction.md vs world_logic.py vs game_state.py). Parallel workers will step on each other's edits.
- Each PR has its own
/es evidence requirement (real-LLM replay with the new prompt on a copy campaign). Two parallel /es replays on the same campaign compete for the same test copy.
- The operator can read the first PR's GREEN before deciding whether the second sibling's fix recipe is even right.
Order of dispatch: file the issue → spawn the worker → wait for the worker's PR to reach a verdictable state → then dispatch the next sibling. Use a 1-worker babysit cron, not N workers in parallel.
5. In the AO worker brief, force a "step around siblings" clause
When the new symptom is a sibling, the worker MUST NOT touch files owned by sibling PRs. Encode this in the brief:
## Files to NOT touch (sibling PRs own these)
- $PROJECT_ROOT/prompts/game_state_instruction.md — PR #8271 owns
- $PROJECT_ROOT/game_state.py — PR #8265 owns
- $PROJECT_ROOT/world_logic.py — PR #8276 owns the god-mode save path
- Any branch in the Sibling investigations list above — DO NOT rebase onto, cherry-pick from, or merge with
The worker reading this brief won't waste cycles asking "should I also fix the lineage bug?" — the brief already says no.
The canonical-state-anchor family (added 2026-07-19)
A third bug family has emerged distinct from prompt-discipline and NPC-status persistence. Triggered when the LLM emits prose that contradicts canon already established in core_memories, faction rules, NPC status, or location/state co-presence — even when those facts are present in the served prompt context.
| Sub-class | Symptom | Canonical anchor the LLM forgot |
|---|
| Magic-sensor / invention | "Vaelaros-tuned Blood-Scent focus", "frequency-sensitive ward", "draconic resonance detection", "Reaver-Hounds", "Ghost-Hunter" | custom_campaign_state.rule explicitly forbids magical tracking; setting is non-/low-magic |
| NPC status erased | "Aenar would sabotage your success" while Aenar is status="disgraced_front_line" / status="dead" | npc_data.<NPC>.status |
| NPC co-presence violated | "Rejoin the Host" choice presupposes Aegon at Mander mouth while canonical state has him co-present at Highgarden | entity_tracking.active_entities[].status.location |
| Faction-control anchor | "the Iron Bank in control" emitted without validating who actually controls the Iron Bank per core_memories section 5 | core_memories section anchor |
| Argella-suspicion continuity | Scene 386 should reference Argella's earlier "you are more than you seem" suspicion but the planning_block drops it | narrative_history.remove of earlier scenes that established the suspicion |
| Canonical-dead-NPC revival (added 2026-07-20, #8472) | "Archon Jaenor Vaelaros (Lvl 20) stands at the base of the dais, his heavy signet ring extended in a trembling hand" while npc_data["Archon Jaenor Vaelaros"].status = "dead" and hp_current = 0 | npc_data.<NPC>.status == "dead" + hp_current == 0 + core_memories[N].rule re-affirms death |
Canonical-dead-NPC revival recipe (verified 2026-07-20, scene 492, campaign EROaUnSbmDhqBedTbJMg, issue #8472, PR #8473):
User verbatim: "My dad is dead why did you forget" (turn 493 after LLM narrated Archon Jaenor Vaelaros as physically present extending his signet ring at the Consular Coronation).
Model self-admission (scene 493): "Administrative Correction: Archon Jaenor Vaelaros is canonically deceased. I have corrected the game state and added a persistent directive to ensure he no longer appears in the narrative. The previous turn's description of him at the coronation was a continuity error."
Canonical contradiction: npc_data["Archon Jaenor Vaelaros"].status = "dead", hp_current = 0, core_memories[363] + [392] re-affirm "Archon Jaenor is dead; Sariel has ascended as God Empress".
The 7 forbidden-pattern categories the durable fix must enumerate (each maps to a scene-492 example):
- Physical presence — "stands at the base of the dais" while dead
- Hand-object action — "extends his signet ring" / "presses his signet into the wax"
- Voicing dialogue — direct line from a dead NPC
- Vehicle/dragon/mount operation — "Jaehaerys rode Vermax toward..." when rider dead
- Physical posture — kneeling / bowing / leaning / sitting
- Multi-NPC group activity with present-tense verbs — "feared your father's indifference" presupposes father's current existence
- Artifact custody transfer — "Jaenor yields the signet ring" (transfer of one's own artifact requires physical agency)
Resurrection exception (verified 2026-07-20): the rule must explicitly enumerate when a canonical-dead NPC's physical presence IS permitted. Verified shape from PR #8473:
- Same-turn
state_updates.npc_data.<npc_id> write must (a) remove "dead" from the status list, (b) restore hp_current > 0, AND (c) provide explicit narrative justification (resurrection ritual, confirmed mistaken-death reveal, spiritual return).
- The justification MUST appear in the same turn's
state_updates AND the same narrative block — a deferred resurrection is not a resurrection.
- The planning_block mirror must DROP (not hedge) any choice whose premise requires a canonical-dead NPC's presence without the resurrection write.
Verified 8 instances on Cg2m2TkGFFez7XBynEah over 24h (2026-07-18 → 2026-07-19):
- #8438 / PR #8439 — Blood-Scent focus / Vaelaros-tuned silver vial (scene 78, magic-sensor invention)
- #8440 / PR #8441 — Inquisition choice emits despite retcon (directive vs prose)
- #8442 — MBTI / Alignment leak (internal-only vs player-facing)
- #8444 / PR #8445 — Aegon Mander-mouth while co-present at Highgarden (NPC co-presence)
- #8451 / PR #8452 — "frequency-sensitive ward" / "Ghost-Hunter" / "Reaver-Hounds" (scene 171, magic-sensor invention)
- "Aenar is dead" follow-up — dispatched via AO worker
worldarchitect-65, bead rev-gkl06 (NPC status)
- "Argella suspicion scene 386" — repro session
20260719_160921_3faa1909 (continuity)
- "Iron Bank in control scene 392" — repro session
20260719_160921_... continuation (faction-control anchor)
Durable-fix shape: add a §"NPC CANON ANCHORING (MANDATORY)" section to $PROJECT_ROOT/prompts/narrative_system_instruction.md that mandates anchoring every NPC behavior/prop/equipment to core_memories, the entity manifest, and npc_data; AND add §"Canonical-State Anchor" to $PROJECT_ROOT/prompts/planning_protocol.md covering NPC Co-Presence / God-Mode Directive Compliance / NPC Reachability / NPC Status Alignment.
Two PRs already contain the durable fix but NEITHER is on origin/main:
| PR | Branch | Status | Contains |
|---|
| PR #8443 | fix/narrative-anti-tracking-prompt-rule | OPEN, NOT DRAFT, mergeable=true | NPC CANON ANCHORING + Anti-Invented-Artifact + 17 follow-up commits (incl. b7b2620671 "prioritize explicit canon over stale memories", dd5f94b279 "preserve NPC canon rule through compaction") |
| PR #8445 | fix/aegon-rejoin-co-presence-8444 | CLOSED (branch alive, head ff419d7a7) | Commit 4524525569 adds §"Canonical-State Anchor" to planning_protocol.md |
This is the deployed-vs-branch-state gap — the deployed DEV bundle is still serving prompts WITHOUT either rule because neither commit is on origin/main (HEAD 444c83d825). The LLM is forgetting canonical state because the fix exists in PR branches but hasn't shipped. See Pitfall 8 for the pre-flight check.
Pitfalls
Pitfall 1: Treating siblings as duplicates
Bug-ref: 2026-07-08 — issue #8160 on campaign xK3fp5XrV24oarIINTF7 was closed as "duplicate of #7885" but the two symptoms (character-creation spells vs scene-event rendering) are different root mechanisms. The cross-campaign audit never happened because the duplicate closeout skipped the investigation. Rule: if the user describes a NEW user-visible symptom on the same campaign, file a new issue. The triage step above decides if it's truly duplicate, not surface similarity.
Pitfall 2: Bundling siblings into one PR
A common anti-pattern is "since these are all prompt-discipline, let me fix all five in one PR." This stalls because:
- Reviewer accepts the smallest fix and asks to split the rest
- /es evidence covers only one symptom, not all five
- Green-gate /er verdict is per-symptom, mixed verdicts block merge
- Conflict surface grows — each sibling's GREEN rewrite touches adjacent prompt fields
Pitfall 3: Conflating the prompt-discipline family with state-persistence family
Two distinct families share surface similarity:
- Prompt-discipline family (this skill): LLM emits the right structured field; downstream path drops it. Fix is in
$PROJECT_ROOT/prompts/<X>.md field definitions and ESSENTIALS rules.
- State-persistence family: LLM emits correct structured field; server-side persistence re-asserts a stale value. Fix is in
$PROJECT_ROOT/world_logic.py or $PROJECT_ROOT/game_state.py save paths.
If the bug surface looks like prompt-discipline but the actual root is a backend override (e.g. Factor D from issue #7453 — "LLM commits level=14 but state shows different"), the fix is in backend code, not the prompt. The triage step that distinguishes them: read the corresponding BQ llm_forensics.llm_payloads row — does the LLM request actually contain the corrected value? If yes, it's prompt-discipline. If the response contains it but Firestore state doesn't, it's state-persistence.
Reference
references/2026-07-08-xK3fp5XrV24oarIINTF7-sibling-table.md — worked example: 4 sibling investigations + 2 sibling-PRs on the same campaign in one day, all reduced to one root family with 5 different surfacesreferences/2026-07-12-xK3fp5XrV24oarIINTF7-queen-death-sibling-8335.md — worked example: 7th sibling on the same campaign over 9 days; introduces the god-mode-retcon missing-write and wrong-key write sub-classes; full scene-by-scene npc_data table + diagnostic recipereferences/2026-07-19-Cg2m2TkGFFez7XBynEah-canonical-state-anchor-cluster.md — worked example: 8 canonical-state-anchor siblings on the Sariel Valyria campaign in 24h (Blood-Scent / Vaelaros-tuned / frequency-shield / Reaver-Hounds / Aenar-dead / Argella-suspicion / Iron-Bank-in-control); introduces the "durable fix exists in unmerged PR branch" pre-flight (Pitfall 8), the "what's in core memories" answer shape (Pitfall 9), the custom_state_keys == [] architectural-gap diagnostic (Pitfall 10), and the deployed-vs-branch-state gap (PR #8443 + PR #8445's commit 4524525569 not on origin/main)references/2026-07-20-EROaUnSbmDhqBedTbJMg-canonical-dead-npc-revival-8472.md — worked example: 5th sibling on EROaUnSbmDhqBedTbJMg (scene 492 Archon Jaenor Vaelaros at coronation, issue #8472 / PR #8473); introduces the canonical-dead-NPC revival sub-class (Pitfall 11's 3-layer architecture: state-update + narrative-emit + planning-block), the 7 forbidden-pattern categories, the resurrection exception, and the twin-clone workflow for scene-by-scene regression evidence~/.hermes/skills/repro/SKILL.md — thin pointer to the canonical /repro workflow~/.hermes/skills/repro/references/npc-status-persistence-bug.md — 6 sub-classes of the NPC-status persistence family; canonical worked examples A/B; scene-by-scene table diagnostic; Option C fix shape (structured ## NPC Status block in system prompt)~/.hermes/skills/repro/references/repro-llm-invented-lore-artifacts-2026-07-18.md — the bug-class-4 (LLM-prose invention) durable fix (NPC Development section in narrative_system_instruction.md); the canonical-state-anchor family's magic-sensor sub-class is the same root cause~/.hermes/skills/repro/references/repro-planning-block-and-campaign-cluster-2026-07-18.md — the 5-anchor taxonomy for planning_block canonical-state violations (NPC co-presence / God-Mode Directive Compliance / NPC Reachability / NPC Status Alignment); the planning-block emit side of the canonical-state-anchor family~/.hermes/skills/hermes-imports/dispatch-task/SKILL.md — AO worker dispatch mechanics (now patched with 20-session-cap and GitHub-rate-limit preflight gates)
Pitfall 4: Confusing campaign convergence with daily-cron regression
The campaign-convergence pattern (this skill) and the daily-cron-failure pattern (sibling skill wa-daily-cron-failure-diagnosis) overlap on the SAME campaign but are different classes of investigation:
- convergent-bug-triage (this skill): user-visible behavior on the campaign is broken — directives ignored, lineage erased, scene events dropped. Fix surface is in
$PROJECT_ROOT/prompts/*.md or $PROJECT_ROOT/world_logic.py.
- wa-daily-cron-failure-diagnosis: GCP cron job exits 1 because the audit script can't parse the LLM's dice rolls. Fix surface is in
scripts/audit_dice_rolls.py + structural fix #7695.
Both classes hit the same campaigns (e.g. xK3fp5XrV24oarIINTF7 / Visenya v7 has been touched by 6+ PRs in BOTH classes in the last 24h). When diagnosing, classify first, then route to the right skill. Cross-references the sibling investigations don't substitute for one — you must apply each skill's protocol separately, in sequence.
Rule: if the trigger is a [GCP Cron] ... - FAIL email or a FAIL in ~/.cache/wa_daily_test_watcher/<job>/YYYY-MM-DD.posted, load wa-daily-cron-failure-diagnosis first. If the trigger is a /repro request about user-visible campaign behavior with 3+ sibling issues, load this skill first. Don't try to handle both classes with one protocol — they have different fix surfaces and different dispatch templates.
Pitfall 5: gh api GraphQL rate limit vs REST API (added 2026-07-12)
When the diagnostic step tries to enumerate sibling issues via gh issue list --repo X --search "...", the GraphQL-backed gh CLI frequently returns GraphQL: API rate limit already exceeded for user ID <N> even when REST endpoints still have budget. Verified 2026-07-12: rate-limit fired at the start of a 6-issue triage on $GITHUB_REPOSITORY while REST API still worked.
Workaround (Python, no gh dependency):
import urllib.request, json
req = urllib.request.Request(
"https://api.github.com/search/issues?q=<KEYWORD>+repo:<OWNER>/<REPO>+type:issue&per_page=30",
headers={"Accept": "application/vnd.github+json"})
with urllib.request.urlopen(req, timeout=15) as resp:
items = json.loads(resp.read().decode()).get("items", [])
For PR creation, use the REST API directly with a token from ~/.config/gh/hosts.yml (parse the oauth_token: line). This bypasses the gh CLI's rate-limited path entirely. Use this fallback in any triage that needs to file issues + create draft PRs in one batch.
Pitfall 6: Branch must exist on origin before gh pr create will accept it (added 2026-07-12)
The "branch-not-on-origin" failure mode is subtle and has two distinct error messages that look unrelated:
- 422 Validation Failed
{"field": "head", "code": "invalid"} — the head ref doesn't exist on the remote.
- 422 Validation Failed
{"message": "No commits between main and <branch>"} — the branch exists but is at the same SHA as main.
The fix is the same sequence: git worktree add with -b <branch> origin/main → git push -u origin <branch> (even if it's empty — push just creates the ref) → commit evidence files → push again → THEN call the PR API. Don't try to skip the empty-push step — the API checks origin/<branch> before it accepts the request, even for a no-commit branch.
Pitfall 7: Confusing the NPC-status persistence family with prompt-discipline (added 2026-07-12)
The NPC-status persistence family (worked example: campaign xK3fp5XrV24oarIINTF7, 6+ siblings over 9 days) superficially looks like prompt-discipline (LLM emits wrong narrative) but is structurally different:
| Family | LLM behavior | Fix surface |
|---|
| Prompt-discipline | Emits correct structured field; downstream path drops it | $PROJECT_ROOT/prompts/<X>.md field definitions + ESSENTIALS rules |
| NPC-status persistence | Emits narrative but no/wrong state_updates.npc_data[NPC] write, OR writes to wrong NPC key, OR god-mode agent doesn't write at all | Prompt layer — needs structured ## NPC Status block as system-prompt anchor (Option C from ~/.hermes/skills/repro/references/npc-status-persistence-bug.md "Fix shape") |
The 6 known sub-classes of NPC-status persistence (as of 2026-07-12): missing-write, wrong-write, prompt-anchor hallucination, confused-state with/replace, god-mode-retcon missing-write, wrong-key write. All six have the same fix shape — pre-render every known NPC's current status as a structured block in the system prompt. The god-mode sub-class is the trickiest because the god-mode agent treats its turn as "narrative freeze" and doesn't even consider writing npc_data; it needs its own system-prompt anchor.
Cross-reference (sibling references): ~/.hermes/skills/repro/references/npc-status-persistence-bug.md holds the worked examples and the diagnostic (scene-by-scene npc_data[NPC] write table).
Pitfall 8: Durable fix exists in unmerged PR branch — check before filing N+1 sibling (added 2026-07-19)
When the cluster signal fires (≥3 siblings on same campaign), the first check before filing a new issue is to scan existing PRs/branches for the durable fix. If the fix is sitting in an open or closed-but-branch-alive PR, the correct action is merge the existing branches, not file a new issue.
Three-check pre-flight (run BEFORE filing):
cd <repo>
git log --all --oneline --grep="Canonical-State\|NPC CANON\|canon-state\|NPC canon\|NPC.*anchor" --since="<campaign-first-sibling-date>"
git log --all --oneline --grep="magic-sensor\|anti-invent\|anti-tracking\|forbidden invention"
COMMIT=<candidate-sha>
git branch --contains $COMMIT -a
gh pr view <candidate-PR-number> --repo <OWNER>/<REPO> --json number,title,state,draft,headRefName,baseRefName,mergeable,url
git log origin/main --oneline | grep -E "<short-sha>|<commit-subject-words>"
Decision matrix:
| Check 1 (commit found) | Check 2 (PR found) | Check 3 (on main) | Action |
|---|
| ✅ | ✅ OPEN | ❌ NOT on main | Comment on PR requesting merge review + cross-link this N+1 sibling |
| ✅ | ✅ CLOSED (branch alive) | ❌ NOT on main | Reopen PR OR cherry-pick onto fresh feat/<topic> from origin/main |
| ✅ | ✅ MERGED | ✅ ON main | This is a NEW bug — file issue normally; the existing fix doesn't cover this sub-class |
| ❌ | ❌ | n/a | No existing durable fix — branch a fresh worktree for root-cause-first prompt fix |
Why this matters: filing N+1 per-scene issues when the durable fix is sitting in an unmerged PR is churn. The user will keep hitting the same canonical-state-anchor bug class because the deployed prompts haven't changed. The fix has been waiting — sometimes for days — for someone to merge the PR branch to origin/main.
Verified case 2026-07-19 (campaign Cg2m2TkGFFez7XBynEah, scene 392 "Iron Bank in control"): the LLM forgets canonical state because PR #8443 + PR #8445's commit 4524525569 exist in branches but neither is on origin/main. Filing a 7th per-scene issue would have been wrong; the correct response is to flag the deployment gap and request MERGE APPROVED on the existing branches.
PR-closed-but-branch-alive detection (added 2026-07-19): a PR can be closed without merge, leaving the fix branch alive and reachable via git fetch origin <branch>. Always check the branch state alongside the PR state:
gh pr view <N> --json state returns CLOSED (the PR was closed)git branch --contains <fix-commit> -a shows remotes/origin/fix/<topic> is still alive- → the fix is reachable; reopen the PR or cherry-pick onto a fresh branch
Wrong JSON field-name pitfall (re-verified 2026-07-19): gh pr view --json changed_files returns Unknown JSON field: "changed_files" — the correct field is changedFiles (camelCase, per gh pr view --json schema). Use gh pr view <N> --json changedFiles or skip the field and use gh pr diff <N> instead.
Pitfall 9: User asks "is this related to other bugs? what's in core memories?" — answer directly (added 2026-07-19)
When the user's question contains "related to other bugs?" + "what's in core memories?", they want a contextual answer grounded in canonical state, not a verdict. The /repro skill's cluster-recipe answers the bug-class question, but the user is also asking for the direct read of canonical state to confirm the LLM is contradicting established canon.
Answer shape (run these in parallel):
- Sibling enumeration:
gh issue list --repo <OWNER>/<REPO> --state all --search "<entity-name> OR <campaign-id>" --limit 30 --json number,title,state,createdAt
- Core-memories direct read: pull the test subject's
_game_state.json export, search for the named entity (Iron Bank, Aenar, Argella, etc.) in core_memories and custom_campaign_state.rule
- Canonical anchor validation: for each entity the user named, check
npc_data.<entity>.status, entity_tracking.active_entities[].status.location, and core_memories section content
- Deployed-vs-branch gap check: see Pitfall 8 — is the durable fix in a merged PR or in an unmerged branch?
Format the answer with 4 sections:
| Section | Content |
|---|
| Related? | Yes/No + sibling table (issue#, scene#, symptom, anchor violation, status) |
| In core memories? | Direct quote from core_memories for the named entity + any custom_campaign_state.rule field |
| Root-cause direction | One-line: same family as N prior siblings, durable fix in PR #X but not yet merged |
| Recommended next step | Either "merge the existing branches" or "branch fresh prompt-fix worktree" |
Anti-pattern: answering only the bug-class question ("yes, this is the canonical-state-anchor family") without showing the direct canonical-state evidence. The user explicitly asked "what's in core memories?" — give them the verbatim canonical anchor, not just the cluster signal.
Pitfall 10: Forgetting the custom_state_keys == [] architectural gap (added 2026-07-19)
The canonical-state-anchor family has an architectural root cause beyond prompt wording: most campaigns have empty custom_state — there is no per-campaign surface for the user to enforce "no magic detection", "Iron Bank is faction X-controlled", or "Argella knows Y". The LLM re-derives these constraints from core_memories archetype descriptors every turn.
Diagnostic: in the test subject's _game_state.json, check:
state = json.load(open(test_subject_state_path))
custom_keys = list(state.get('custom_campaign_state', {}).keys())
Verified case 2026-07-19 (Cg2m2TkGFFez7XBynEah): custom_state_keys == [] despite the custom_campaign_state containing a rule: "Explicitly forbid magical 'Iron-Scent' or 'Shadow-Signature' tracking of..." directive. The rule lives in the prompt but is not materialized as a structured state field the LLM can check on every emit.
Recommended durable-fix extension: add custom_campaign_state.no_magic_detection_zone: bool, custom_campaign_state.faction_control: {<faction>: <owner-npc>}, and npc_data.<NPC>.lore_origin = "user-introduced" | "LLM-invented" provenance fields to $PROJECT_ROOT/game_state.py. This converts the prompt-only canon rules into structured state the LLM can validate against on every emit, not just the scenes where the user actively retcons.
This extension belongs in a follow-up prompt-fix PR, not in the existing PR #8443 / PR #8445 branches — those fix the prompt layer; the custom_state extension fixes the data-model layer.
Pitfall 11: Durable-fix shape lives at the prompt layer, but in THREE places (added 2026-07-20)
A canonical-state-anchor durable fix requires three coordinated prompt-layer sections, not one. The 2026-07-20 PR #8473 / issue #8472 (canonical-dead-NPC revival, scene 492) landed the missing two halves after PR #8352's state-update-only half was insufficient for 4 weeks.
The three layers:
| Layer | Where it lives | What it governs | Bug it catches |
|---|
| State-update layer | $PROJECT_ROOT/prompts/game_state_instruction.md | state_updates.npc_data writes | "LLM resurrects dead NPC via state write" |
| Narrative-emit layer | $PROJECT_ROOT/prompts/narrative_system_instruction.md (NEW §NPC Presence at Canonical Status) | The prose the LLM produces | "LLM narrates dead NPC physically present in scene" — what PR #8352 alone did NOT catch |
| Planning-block layer | $PROJECT_ROOT/prompts/planning_protocol.md (NEW §NPC Presence at Canonical Status (Choice Premise Validation)) | planning_block.choices[] premise validation | "LLM emits choice whose premise requires dead NPC's presence" |
Why all three must land together:
The state-update layer alone is insufficient because the LLM's narrative-emit pass can re-introduce a canonical-dead NPC into the prose even when the game-state pass correctly suppresses the resurrection write. Verified 2026-07-20: PR #8352 / commit 31d8b452c5 (June 28) shipped §"Narrative Revival of Canonical-Dead NPCs" into game_state_instruction.md. The rule said "do not resurrect canonical-dead NPCs in state_updates.npc_data". On 2026-07-20, scene 492, the LLM correctly did NOT emit a state_updates.npc_data resurrection write — but the narrative-emit agent still narrated Archon Jaenor Vaelaros as physically present at the coronation, "stands at the base of the dais, his heavy signet ring extended in a trembling hand". The state-update-layer half held; the narrative-emit-layer half did not exist.
The architecture discovery (added 2026-07-20):
Each canonical-state-anchor sub-class (NPC co-presence, NPC knowledge-of-PC, faction-control, canonical-dead-NPC revival, Argella-suspicion, magic-sensor invention) maps to the same 3-layer prompt architecture. When filing the durable-fix PR for a new sub-class:
- Check the state-update layer first (
game_state_instruction.md). If a rule already exists for the sub-class, you have the state-update half — the bug is in one of the other two layers.
- Check the narrative-emit layer (
narrative_system_instruction.md). This is the most-missing layer; PR #8352's commit history shows the team fixed state-update-only first.
- Check the planning-block layer (
planning_protocol.md). Often forgotten because the choices look "advisory" but they bind NPC presence in subsequent turns.
PR #8473 cross-reference map (verified 2026-07-20):
state-update half (PR #8352 / commit 31d8b452c5):
game_state_instruction.md §"Narrative Revival of Canonical-Dead NPCs"
→ Governs: state_updates.npc_data.<npc_id> writes
narrative-emit half (PR #8473 / commit f4bb3a6687):
narrative_system_instruction.md §"NPC Presence at Canonical Status (Forbidden Revival Patterns)"
→ Governs: the prose the LLM produces
→ Enumerates: 7 forbidden-pattern categories + permitted patterns + resurrection exception
planning-block half (PR #8473 / commit f4bb3a6687):
planning_protocol.md §"NPC Presence at Canonical Status (Choice Premise Validation)"
→ Governs: planning_block.choices[] premise validation
→ DROP-not-hedge rule for choices whose premise requires a canonical-dead NPC's presence
Anti-pattern: "fix it in the prompt file and ship":
If the new sub-class fix only touches game_state_instruction.md, it will land fast but the narrative-emit bug will re-emerge 1-4 weeks later (as it did with PR #8352 → #8472). Always ship the narrative-emit + planning-block halves in the same PR (or a chained PR with the state-update half on the same branch).
Anti-pattern: "extend PR #8469's anchor because it covers the same area":
PR #8469 / issue #8468 anchors NPC knowledge claims against mask-layer prohibitions. PR #8473 / issue #8472 anchors NPC physical presence against canonical-dead status. The two sub-classes are structurally distinct — keeping them as independent PRs lets reviewers merge in either order, and avoids coupling that complicates CodeRabbit review. Cross-reference via cluster-sibling mention in the PR body, NOT via branch merge.
Verification recipe for the 3-layer architecture:
After shipping the durable fix, verify with a fresh git worktree add -b feat/<topic> origin/main + copy campaign + replay LLM call. If the bug surfaces in any of the three layers, the corresponding half is missing. The test file pattern (tests/test_<subclass>_<issue#>.py) ships alongside each PR; PR #8473's test_canonical_dead_npc_revival_8493.py has 14 prompt-contract pins covering all three layers, with one skipped (Layer-C server-side post-LLM guard pending human approval per AGENTS.md).
Pitfall 12: Opt-in features — mirror the faction minigame pattern (added 2026-07-20)
When a user reports "X activates without my permission, I want it opt-in like faction mode", the canonical template is FactionManagementAgent + faction_minigame.enabled. The template has three pieces that ALL must ship together; shipping any subset leaves the user visible-bug intact:
| Template piece | Where it lives | What it does |
|---|
| Opt-in flag in state | custom_campaign_state.<feature>_minigame.enabled = False (default off for new campaigns) | The user-controlled toggle |
matches_game_state gates on the flag | <Feature>Agent.matches_game_state() returns False when flag is False | Stops state-driven hijack of ordinary turns |
| Enablement patterns | <FEATURE>_ENABLEMENT_PATTERNS list (e.g. enable faction mode, enable multiverse minigame) with matches_input accepting them unconditionally | Lets the user opt in WITHOUT a god-mode prefix |
Anti-patterns (verified on $GITHUB_REPOSITORY PR #8475, issue #8474 — multi-verse activating without consent):
- ❌ Implementing only the flag (state-only opt-in). The bug recurs whenever the threshold crosses during ordinary play because
matches_game_state doesn't check the flag.
- ❌ Implementing only
matches_input patterns. Enablement phrases still get blocked when the LLM happens to emit the right input but state flag is False.
- ❌ Adding a "backwards-compat shim" that auto-opts-in when a legacy flag is set (e.g.
multiverse_upgrade_available=True). This re-introduces the original bug for every existing sovereign-tier campaign. Delete the shim; existing campaigns explicitly opt in.
- ❌ Forgetting
get_pending_upgrade_type() (or the equivalent state-readable signal) when adding the flag to matches_game_state. The narrative-surfacing path still fires on the threshold even when routing doesn't — which is fine, but make sure the helper reads the flag too so downstream agent._upgrade_type and matches_game_state agree.
Test recipe for opt-in features:
- RED: with flag=False, ordinary gameplay input routes to
<DefaultAgent>, not <Feature>Agent, even when state threshold is crossed.
- GREEN: with flag=True, the same input routes to
<Feature>Agent and existing modal-lock / completed-tier guards behave identically to before.
- ENABLE PATTERN:
<enable phrase> without god-mode prefix flips the flag through god-mode directive processing and the next turn routes to <Feature>Agent.
- The test helper for
<Feature>Agent should accept the flag as a parameter, NOT derive it from other state flags (otherwise the test silently bypasses the new gate).
Verified case (PR #8475, 2026-07-20): the multi-verse fix shipped flag + matches_game_state gate + enablement patterns in one PR. 447 tests pass; 4 pre-existing tests had to be updated to opt-in (one had a backwards-compat-shim bug that the user would have hit immediately). The "ship all three pieces" rule was the lesson — even small partial implementations of this pattern will fail at runtime.
Pitfall 13: Static path unambiguous — ship the fix on the same draft PR (added 2026-07-20)
When the canonical /repro workflow opens the issue → draft PR → reproduction → report gates, but the bug class is unambiguous from static review alone (no live LLM turn can falsify the root cause), ship the fix on the same draft PR. The gate contract requires the issue be filed BEFORE any other work; it does NOT require the draft PR wait for a fresh session.
Triggers — static path is unambiguous when ALL of these hold:
- The user message names the exact symptom (e.g. "multi-verse activates without consent"), the desired behavior (opt-in like faction mode), and the canonical comparison target (faction minigame opt-in flag).
- Static review confirms the bug class:
is_<feature>_upgrade_available() returns true on threshold, matches_game_state does not gate on a flag, matches_input returns false.
- The fix shape is a known pattern (Pitfall 12's opt-in template, Pitfall 11's 3-layer prompt architecture, etc.) — the fix is a copy-and-adapt, not a discovery.
- A live production-ingress replay cannot falsify the static review because the routing decision is determined by code, not by the LLM's prose output.
Anti-pattern: holding the fix back because the canonical /repro skill says "Step 6: production-ingress replay". For unambiguously-static bugs, the replay is theater — you already know what it would show.
The "honest gap" requirement: the PR body MUST say "no live production-ingress replay was run; the local fix cannot be falsified by a static review alone. CI + green-gate + a 3-turn streaming lifecycle test on top of this branch will give the final verdict." Do not claim REPRO for the live verdict when no live turn was run; claim REPRO → FIXED for the static path only.
Verified case (PR #8475, 2026-07-20, campaign EROaUnSbmDhqBedTbJMg, issue #8474): static review of $PROJECT_ROOT/campaign_divine.py:is_multiverse_upgrade_available() + $PROJECT_ROOT/agents.py:CampaignUpgradeAgent.matches_game_state() was unambiguous — there is NO path where the LLM could emit a state_updates write that bypasses the routing code, because routing happens before the LLM is called. The fix landed on the same draft PR (fix/multiverse-opt-in-8474, head d72e990ed5) as the original repro scaffold. 447 tests pass. CI + green-gate verdict still pending but the static path is closed.
Cross-reference: Pitfall 8 covers the "durable fix exists in unmerged PR branch" pattern. Pitfall 13 is the inverse — the durable fix does NOT exist yet, the user wants it, and shipping it on the same draft PR is the right move.
