Truncates sources, curriculum, content, quiz banks/results/cache, topics
and relations in dependency order so AI-generated state can be wiped
between smoke runs without leaving dangling references. Handbook sync
state is cleared by default (otherwise re-sync is a no-op); user
progress and leaderboard are opt-in. Team members, settings, and LLM
telemetry are preserved.
UI lives in Admin → Settings → Danger Zone and requires typing RESET
before the button enables. Per-collection deletion counts are reported.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Handbook sync ran files sequentially under a 5 req/min limiter with a
hardcoded 60s LLM timeout, causing long syncs and AbortError timeouts on
large files. Now: limiter at 20 req/min, files processed with concurrency
4, handbook extraction timeout raised to 180s, and near-empty files skip
the LLM call.
callLLM gains a timeoutMs option; passing a signal no longer silently
disables the per-request timeout.
llm_calls telemetry self-disables after the first 404 so deploys without
the migration applied don't spam the console.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Add dependency-free TF-IDF retrieval (src/lib/retrieval.js) with NL+EN
stopwords and a WeakMap-cached index.
- Rewrite buildKbContext to ship the top-K relevant topics + verbatim-
mentioned ids only, filter relations to the included set, and append a
[kb_hash: <8 hex>] suffix so the ephemeral prompt cache busts when the
graph changes. Returns { context, retrievedTopics, allTopics }.
- Add LOOKUP_TOPIC_TOOL and drive useChat through callLLM directly with a
multi-hop tool_result loop capped at 3 hops; preserve Anthropic-provided
tool_use ids through callLLM so the loop can echo correct tool_use_id.
- Truncate R42 history to the last 12 turns and prepend a single
"(earlier conversation truncated)" assistant message.
- Set R42 chat defaults: temperature 0.3, maxTokens 2048.
- Add pb_migrations/1780500002_created_llm_calls.js (the best-effort
logger in callLLM was already wired) and a new Admin → Diagnostics
view showing the last 100 calls with token usage, cache-hit rate, and
USD cost from a local Anthropic price table.
- Finalize AI_PIPELINE_HARDENING_PLAN.md: mark Phases 1–5 shipped and
Phase 6 (eval harness) explicitly out of scope.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- src/lib/random.js: Fisher–Yates shuffle/sample/pickInt; replace every
biased .sort(() => 0.5 - Math.random()) site in testService.
- testService: debias correctIndex via prompt + runtime re-roll (up to 2x
when one position holds >50%); quality gate rejecting <4 distinct
options, banned filler ("all of the above" etc) and explanations
shorter than 20 chars; dedup new questions against the existing bank
via normalised question text.
- Quiz schema/tool/prompt require difficulty ('easy'|'medium'|'hard');
db.getQuizBank defaults legacy records to 'medium' on read.
- learningService.generateCustomTopic: kebab-case slug ID from the
polished label with collision suffixes; default learning_relevance
'standard' when the model omits it.
- Tests for random helpers, dedup/quality-gate behaviour and the
extended quiz schema.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When a topic's quiz bank is empty (or below the requested count), we
previously seeded it with a fresh batch of 10 questions. That meant the
first weekly quiz for any new topic triggered a 10-question LLM call —
heavy for what's ultimately a 1-question sample for review topics, and
overkill for the typical 5-question primary topic.
- forceGenerateTopicQuestions default count: 10 → 5
- getOrGenerateTopicQuestions seed amount: 10 → 5
- TestManager "Generate" defaults + empty-state button copy: 10 → 5
- QUIZ_SYSTEM difficulty hint: rewritten for a 5-question batch (2 easy
/ 2 medium / 1 hard) with explicit "scale proportionally for larger
batches" so admins can still generate 10+ via TestManager when they
want more depth.
Tests 61/61 pass, lint clean (0 errors), build clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace stateless one-shot extraction with a stateful, paced, cancellable
pipeline. Six subtasks:
- 3.1 Sentence-aware chunking with 800-char overlap (was paragraph-only
at 4000 chars). Hard-split fallback for runaway sentences.
- 3.2 Stateful extraction: chunks 2+ receive an "already-extracted topic
IDs" hint capped at 200 IDs, so the model reuses IDs instead of
inventing variants like software-developer vs software-engineer.
- 3.3 Token-bucket limiter in llmRetry.js (extractionLimiter, 5 req/min).
callLLM awaits the limiter before fetch; 429+Retry-After calls
pauseUntil. Replaces hard setTimeout(12000) and setTimeout(15000).
- 3.4 relevance_locked column on topics — admin edits to relevance are
sticky across re-extraction. Migration + merge respects the flag +
unlock checkbox in KnowledgeGraph edit form.
- 3.5 Unify relation vocabulary — handbook prompt no longer mentions
legacy "executes"; one-shot migration rewrites existing executes rows
to executed_by with source/target swapped.
- 3.6 Cancellation — Cancel button on UploadZone wired to an
AbortController threaded into callLLM; aborted runs persist status =
"cancelled" rather than "failed".
Tests: 16 new unit tests for chunkText, buildKnownIdsHint, and
createLimiter. All 61 tests pass, 0 lint errors, build clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Anthropic has deprecated the temperature parameter for their reasoning
models (claude-opus-4-7). This was causing a 400 error when analyzeGraph
called callLLM with tier: 'reasoning'.
Solution: conditionally exclude temperature from the request body when
tier === 'reasoning'. Fast and standard tiers retain their temperature
parameter.
This unblocks the "Analyse and Optimize" button in the Knowledge Graph
admin panel post-Phase-2 deployment.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Every structured-output call now uses an Anthropic tool instead of
parsing JSON out of free-form prose, and stable system prompts are
sent as cacheable blocks. Behaviour-equivalent to phase 1 from the
caller's point of view; the savings show up in token usage and in the
absence of "AI returned non-JSON response" failure modes.
* src/lib/llmTools.js — single source of truth for tool definitions:
emit_knowledge_graph, emit_handbook_delta, emit_learning_article /
_slides / _infographic / _all, emit_custom_topic, emit_quiz_questions,
emit_graph_actions, plus five article-patch tools (set_intro,
set_section, add_section, remove_section, replace_takeaways).
* src/lib/articlePatches.js — pure applyArticlePatches +
applyAndValidate; rebuilds the article from a sequence of patch tool
calls and re-validates against learningArticleSchema. set_section
falls back to appending when no matching heading exists so the
model's intent is preserved rather than silently dropped.
* src/lib/llmSchemas.js — Zod schemas for the five patch ops,
registered in toolSchemaRegistry so callLLM validates them
automatically.
* src/lib/llm.js — simulation mode now returns a tool_use stub matching
toolChoice.name, so the UI keeps working with Simulation Mode on
after the structured-output migration.
* src/lib/extractionPipeline.js — processSourceText and
analyzeHandbookDelta migrated to callLLM + tool use. System prompts
sent as { cache_control: ephemeral } blocks. Handbook results pass
through normalizeHandbookResult to collapse legacy "executes"
relations into executed_by with swapped source/target.
* src/lib/learningService.js — generateLearningContent picks the right
tool per selectedType; generateCustomTopic uses emit_custom_topic;
refineLearningContent now drives the five patch tools with
toolChoice 'any' and rejects the whole turn if the patched article
fails validation. Article-only refinement is intentional for phase 2;
refining a topic without an article surfaces a clear error.
* src/lib/testService.js — quiz generation via emit_quiz_questions.
* src/components/admin/KnowledgeGraph.jsx — analyzeGraph routed through
the reasoning tier (Opus) since graph-wide consolidation benefits
from a stronger reasoner.
* src/components/chat/prompts.js — buildSystemPrompt now returns three
text blocks: stable preamble (cached), KB context (cached, hash-bust
deferred to phase 5), per-turn user/admin tail (uncached).
* src/lib/__tests__/ — 13 new tests covering each patch op, multi-op
sequencing, post-patch validation failure, and tool/registry shape.
Acceptance: lint and 45/45 tests green; build succeeds; no
`match(/\{[\s\S]*\}/)` JSON extraction left in src/. Live verification
of cache hits on a second extraction within 5 minutes is deferred to
manual smoke testing — needs real `/api/anthropic` traffic.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a BuildStamp component pinned to the bottom-right of every page
(desktop only, dim monospace text) so it's obvious at a glance which
build is currently running.
The git short SHA and an ISO build timestamp are injected at build time
via Vite's `define`, falling back to 'unknown' if git is not available.
ESLint config declares the two compile-time constants as readonly
globals so no per-file disable comments are needed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds Claude Code's per-user .claude/ directory (settings.local.json
allowlists, etc.) to .gitignore so it doesn't get accidentally
committed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Implements phase 1 of AI_PIPELINE_HARDENING_PLAN.md. Every Anthropic call
now goes through one module that owns retry, timeout, abort, structured-
output parsing, schema validation, and best-effort call telemetry.
* src/lib/llm.js — single callLLM entry point. Resolves model per tier
(fast / standard / reasoning) with admin:model legacy fallback for the
standard tier; 60s default timeout via AbortController; balanced-brace
JSON extraction; LLMHttpError, LLMTruncatedError, LLMOutputError, and
LLMValidationError surface clearly distinct failure modes.
* src/lib/llmRetry.js — exponential backoff with full jitter, retries
only on transient HTTP statuses, honours Retry-After up to 60s, never
retries on AbortError.
* src/lib/llmSchemas.js — Zod schemas for every structured task plus
normalizeHandbookResult (collapses legacy "executes" relations into
the canonical "executed_by" vocabulary).
* src/lib/api.js — thin shim over callLLM so existing callers (extraction
pipeline, learning, quiz, R42, knowledge graph) keep working unchanged.
* src/lib/__tests__/ — 32 Vitest cases covering parse paths, error
surfaces, simulation mode, model resolution, and schema validation.
* src/pages/Admin/index.jsx — three model inputs (fast / standard /
reasoning) replacing the single legacy field; legacy value falls back
for the standard tier so existing overrides survive.
Adds Zod and Vitest, plus an "npm run test" script.
Also cleans up the pre-existing repo-wide ESLint failures so phase 1's
"npm run lint passes" acceptance criterion can be checked: drops unused
React imports across the JSX tree (React 19 JSX runtime auto-imports),
attaches cause to rethrown errors in the service modules, ignores
pb_migrations in the ESLint config (PocketBase JSVM globals), and
removes one dead handleCreateCustom function in Leren.jsx. A real
behaviour bug surfaced in Testen.jsx — the quiz timer captured a stale
finishQuiz via setInterval closure; now updated via finishQuizRef so the
timer always invokes the latest callback.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds AI_PIPELINE_HARDENING_PLAN.md — a phased, self-contained plan an AI
agent can execute to harden the Anthropic integration (central LLM
client, tool-based structured outputs, prompt caching, retrieval-based
R42 context, eval harness).
Renames src/lib/giteaService.js to src/lib/githubService.js. The module
calls api.github.com and raw.githubusercontent.com; the previous name
was misleading. No behaviour change. Updates the single import site in
src/components/admin/KnowledgeGraph.jsx.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
