[Core] Feat: cross-quest experience distill#60
Open
droidlyx wants to merge 77 commits intoResearAI:mainfrom
Open
[Core] Feat: cross-quest experience distill#60droidlyx wants to merge 77 commits intoResearAI:mainfrom
droidlyx wants to merge 77 commits intoResearAI:mainfrom
Conversation
Implementation plan for the persistent experience distillation skill: opt-in companion skill plus post-record routing hook in ArtifactService.record(). Experience is a knowledge kind subtype. Retroactive CLI ds distill-quest produces drafts for human review. 8 TDD tasks with complete test code, no placeholders. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Imports will be re-added in Tasks 3 and 7 when they're actually used. Avoids F401 lint warnings on the Task 1 commit.
…ience_distill Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…k 6) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Fix 1: Use json.dumps() to safely quote slice titles in the YAML frontmatter (e.g. 'effect of "warm-up"' → valid YAML quoted scalar). Prevents injection if a title contains double-quotes. Other fields (quest, run, campaign) remain unquoted since they're internal IDs without special characters. Fix 2: Capture emit_experience_drafts' return value (written file paths) and report len(written) instead of len(records), ensuring the JSON output reflects the actual number of files created (some records may be filtered). Test: Added test_emit_experience_drafts_escapes_title_with_quote to verify titles with quotes parse correctly; extended test_iter_analysis_slice_records_filters_index to verify malformed JSONL lines and missing files are skipped gracefully. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 1 added the distill-quest subcommand to deepscientist.cli but forgot the bin/ds.js launcher allowlist, so users got "Unexpected launcher argument" instead of reaching Python. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Past failed installs left .staging.* dirs lying around, and stale src/ui/dist or src/tui/dist could fool the freshness check into reusing an obsolete bundle (or get the build into a wedged state that surfaces as Rollup-cant-resolve-deps errors). Sweep both at the start of every install so users do not have to clean caches by hand. Source tree node_modules are intentionally left alone because they support local npm run dev workflows; --clean-style flag can be added later if observation justifies it. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
git_readiness() previously surfaced missing user.name/email as warnings + guidance only, so ds startup printed a hint and continued. The first agent commit then died with "empty ident name" deep into quest execution, far from the actual cause. Move both checks from warnings to errors. The existing ds.js block on payload.git.installed === false is extended to also bail when payload.git.errors is non-empty, surfacing the configuration step the user needs to run. ok now reflects whether errors is empty so doctor and other callers get a single source of truth. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Quest 009 confirmed Phase 1 distill never fires when the agent skips analysis_campaign and routes main_experiment results straight to write. The spec adds a finalize-gate: when decision(action='write'|'finalize') lands, scan completed runs (analysis.slice / main_experiment / experiment) against a reviewed-set derived from new distill_review artifacts; if anything is undistilled, route to the distill skill before allowing write/finalize. Cursor is set-difference, not timestamp, so the gate is idempotent and stateless. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Nine TDD tasks: distill_review schema, candidate iterator, gate evaluator,
overlay function, service.py wire-in, list_distill_candidates MCP tool,
SKILL.md rewrite, retroactive CLI alignment, end-to-end lifecycle test.
Two spec deviations called out in plan (both net simplifications):
1. Hook moves from guidance.py to experience_distill.py overlay + service.py
wire-in - preserves build_guidance_for_record() purity.
2. Replaces "extend list_runs filter" with a new dedicated
list_distill_candidates MCP tool - list_runs does not currently exist.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…pe cases Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…uard Apply .strip().lower() normalization to both 'kind' and 'action' checks in maybe_inject_distill_finalize_gate to absorb upstream casing variations, matching the pattern used elsewhere in guidance.py. Add test covering uppercase and whitespace-padded action values. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Also extends maybe_inject_distill_finalize_gate to handle the gate-clear case: when a previously gate-injected guidance_vm is re-evaluated and the gate no longer fires, the previous_recommended_skill is restored. This ensures suppressed-equivalent decision records don't carry a stale distill redirect after a distill_review has covered all candidates. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds experience_distill to START_SETUP_FORM_FIELDS so the field round-trips through the start-setup MCP patch surface, and coerces the value as bool in _sanitize_start_setup_form_patch. Pairs with the UI toggle in CreateProjectDialog. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…or symmetry Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…st helper Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…/keywords/neighbor_decisions conventions
…le and complete_quest
Quest 011 validation showed the Phase 3 finalize gate only fires on
`record(kind='decision', action ∈ {write,finalize})`. Real closure paths
(`submit_paper_bundle` records `report`; `complete_quest` is a separate
MCP tool) bypass that surface entirely, so an agent with experience_distill
on can ship and complete without ever invoking distill.
Two complementary changes:
- B (prompt cue): _priority_memory_block now injects `distill_required_rule`
whenever distill is on, the gate has pending candidates, and the active
skill is a stage skill — naming both closure tools so the agent learns
the constraint before acting.
- C (hard guards): submit_paper_bundle raises ValueError and complete_quest
returns `{ok: false, status: 'distill_required', ...}` when the gate has
pending candidates, regardless of which artifact route the agent picked.
Tests: 4 new prompt-builder cases (cue fires/skips for stage/companion/off);
7 new closure-guard cases covering each gate state per tool. 187 distill +
prompt + paper-bundle + complete-quest tests green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…fact dirs Quest 011 re-validation showed the closure guards from the previous commit only looked at quest_root/artifacts. Real `main_experiment` records typically live in the active idea worktree at `quest_root/.ds/worktrees/<name>/artifacts/` until the git-graph merge promotes them — so a worktree with pending runs would still slip past the prompt cue and both hard guards. Adds `evaluate_distill_gate_for_quest(quest_root)` which walks every workspace artifact dir (canonical + each worktree subdir), dedupes candidates and reviews by artifact_id, and returns the same payload shape as `evaluate_distill_gate`. Switches the prompt cue and both service guards to the new aggregator. Tests: 4 new closure-guard cases (worktree-only candidate; dedupe across workspaces; review in quest_root covers worktree run). 133 distill + prompt-builder tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…yable Quest 011 audit found 137 401 errors across multiple quests in the last 2 days. Each failed turn triggered the runner retry loop (default 4 attempts), and each attempt re-tried claude's OAuth refresh, eventually rate-limiting the refresh endpoint and locking the account out for hours even after credentials are restored. Adds a runner_failures pattern that matches "Failed to authenticate" / "API Error: 401 ... authentication_error" / "Invalid authentication credentials" only when runner=='claude', returns retriable=False, and points the user at `claude login` or the ANTHROPIC_API_KEY override. This stops the retry storm at the daemon's _non_retryable_failure_diagnosis gate before it reaches the retry policy. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ce immediately Reuses the same `probe_runner_bootstrap` call that `ds doctor` runs for each enabled runner, but kicks it from `serve()` in a background thread on daemon start. Probe results land in daemon.jsonl as either `daemon.runner_probe_ok` (info) or `daemon.runner_probe_failed` (error if a deterministic FailureDiagnosis matches, warning otherwise) — so a stale claude OAuth token shows up the moment the daemon comes up rather than waiting for the first quest turn to discover and pause itself. Probe runs once per daemon lifetime, off the HTTP serving thread, with a 90s per-runner cap inherited from the existing probe code. Combined with the previous claude_authentication_failed diagnosis, a dead token now fails fast in two distinct surfaces (startup probe + per-turn diagnose) without any retry amplification. Tests: 5 new cases covering enabled/disabled gating, deterministic diagnosis routing, no-diagnosis fallback, exception handling, and default runner-list resolution. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…est worktrees The MCP wrapper sets DS_WORKTREE_ROOT once at server start (= quest root) and does not refresh it when activate_branch swaps the active worktree, so service.record(...) is invoked with workspace_root=quest_root even when the run lives only in the active worktree's artifacts/_index.jsonl. The distill_review run-id validator scanned only write_root/artifacts and therefore rejected perfectly valid reviews with "unknown run artifact_ids", while list_distill_candidates and evaluate_distill_gate_for_quest already aggregated across every workspace. Quest 012 hit this in production and worked around it by manually mirroring the run entry into the quest-root index. Make the validator collect known_run_ids across _quest_workspace_artifact_dirs(), matching the closure-side aggregation. write_root and the actual on-disk write location are unchanged, so this only widens the acceptance set. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…lution flakes vite/rollup intermittently fails to resolve hoisted deps (clsx, framer-motion, react-router) on the fuse-overlayfs container even though npm install reports success. Retry the build (not the install) until it converges; failure rate observed ~10-30%, max 5 attempts keeps p(all-fail) below 1%. Could not reproduce in isolation (11/11 attempts succeeded in a probe), so the underlying filesystem/cache cause is left for a future bisect — this is a workaround, not a fix. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Without this check, passing a wrong quest_id silently returns drafts=0 exit 0, since read_distill_reviews and iter_distill_candidate_records both early-return when their target dir does not exist. Print a clear "Quest not found" message to stderr and return 1 instead. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sync the two paragraphs added on the English side (07_MEMORY_AND_MCP.md section 8 and 14_PROMPT_SKILLS_AND_MCP_GUIDE.md section 13) so the Chinese docs cover the experience_distill toggle, list_distill_candidates MCP tool, and distill_review artifact at the same level of detail. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Move the experience-distillation pipeline out of 07_MEMORY_AND_MCP.md section 8 and 14_PROMPT_SKILLS_AND_MCP_GUIDE.md section 13 into a new standalone document 34_EXPERIENCE_DISTILL_GUIDE.md (English and Chinese), covering the experience_distill / recall_priors toggles, the finalize gate, hard guards on submit_paper_bundle and complete_quest, the distill skill flow, the distill_review schema, the cross-quest recall surface, the new MCP tools, the retroactive ds distill-quest CLI, and operational notes / debugging. The original 07 / 14 sections now point at the new doc as the single source of truth; index entries added to docs/en/README.md and docs/zh/README.md under the architecture section. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
These plan and spec files were development-process notes produced by the local superpowers plugin while iterating on the distill pipeline. They are not user- or maintainer-facing documentation; the canonical reference is now docs/en/34_EXPERIENCE_DISTILL_GUIDE.md (and the Chinese mirror). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… specifiers Vite/rollup occasionally exits 0 on this filesystem but leaves bare specifiers like `import"scheduler"` in the main bundle (also seen with clsx, framer-motion, react-router). The browser can't resolve those at runtime, producing `Failed to resolve module specifier "scheduler"` and a blank page. The previous retry loop only triggered on non-zero exit and so missed this mode. After a successful `npm run build`, scan dist/assets/*.js for any of these bare imports; if any are found, drop dist/ and retry the build. Same max-5 attempts. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
droidlyx
changed the title
…edicated kind) - DECISION_ACTIONS gains 'distill_review'; review payload fields (reviewed_run_ids, cards_written, neighbor_decisions) validated under the decision kind when action=='distill_review' - Drop 'distill_review' from ARTIFACT_DIRS; one less public kind - Closure gate and MCP error messages updated to suggest artifact.record(kind='decision', action='distill_review', ...) - Tests + en docs + skill examples updated to the new shape Per RFC ResearAI#64 slim: a distill review is a decision about coverage; folding it into the existing kind keeps the public surface narrower without losing any validation. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Remove maybe_inject_distill_routing and iter_analysis_slice_records; the per-slice trigger forced every analysis slice through a distill detour even when no batch was ready - Closure gate (maybe_inject_distill_finalize_gate) stays as the sole discipline mechanism — its coverage check already catches every uncovered run at quest end - Drop tests/test_experience_distill_routing.py and per-slice references in SKILL.md / cli + integration tests Per RFC ResearAI#64 slim: the closure gate alone provides enough discipline; forcing distill per slice is overreach. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- validate_experience_metadata now enforces only `claim` (non-empty) and `lineage` (non-empty list with quest+run per entry) - Drop required enforcement of subtype / mechanism / conditions / confidence; agents tend to game shallow taxonomy fields and the cross-quest patch invariant only ever needed lineage - Cards may still carry these fields; we just don't reject when absent - validate_cross_quest_patch unchanged — still enforces lineage chain on patched cards Per RFC ResearAI#64 slim: shallow metadata depth was overreach; claim+lineage carries real semantic load and the patch invariant. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- src/skills/distill/SKILL.md: remove per-slice trigger framing, subtype/mechanism/conditions/confidence as required, and any remaining `kind='distill_review'` shape examples; keep finalize-gate protocol, neighbor decisions, claim+lineage minimum, decision-shape call example - docs/en/34_EXPERIENCE_DISTILL_GUIDE.md: same trim; drop the ds memory normalize-knowledge mention (CLI doesn't exist) and per-slice analysis-slice routing references - docs/zh/34_EXPERIENCE_DISTILL_GUIDE.md: deleted; zh readers follow the en guide pointer instead Per RFC ResearAI#64 slim: docs match the actual surface after S1-S3. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The S4+S5+S6 cleanup removed the `ds distill-quest <quest_id>` section on the assumption it was a non-existent CLI like `ds memory normalize-knowledge`. It is real (cli.py registers `distill-quest` and the `distill_quest_command` handler). Restore a tight section documenting it. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The S4 SKILL.md trim collapsed the description and dropped the word 'experience'. test_experience_distill_skill_bundle asserts the description contains 'experience' as a discoverability anchor; restore the keyword so the bundle remains the canonical experience- distillation skill. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…lows - mcp.memory.write: default `scope` flips quest -> global. Explicit scope='quest' still works for quest-internal task tracking - Stage skills (decision/finalize/idea/baseline/experiment/ analysis-campaign/write/etc) update memory.write examples and prose to land reusable lessons globally by default; quest-scope is reserved for quest-internal tracking - Memory + MCP en/zh docs updated to reflect the new default - New regression test: memory.write without explicit scope writes global; existing quest-scope tests pass scope='quest' explicitly Per RFC ResearAI#64 slim section 6: quest 014 wrote 9 quest-scope cards and 1 global; flipping the default and aligning stage prose makes global the path of least resistance for reusable lessons. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- ArtifactService.list_recent(quest_root, kind=None, limit=20): newest-first aggregation across quest_root/artifacts/_index.jsonl and every .ds/worktrees/*/artifacts/_index.jsonl - MCP `artifact.list_recent` exposes it under the existing namespace; read-only annotations; codex auto-approves - Tests cover service-level (main + worktree aggregation, kind filter, limit cap) and MCP-level shape - en/zh docs updated; experience_distill guide cross-references it Per RFC ResearAI#64 slim: stage skills can now read 'what did I just do' from artifacts directly, freeing memory.list_recent from being co-opted as a runtime-state surface. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ures - Fold 7 small/topical files into 3 broader topical files plus a new config_cli file: - schema (+ artifact_schemas_distill_review) - gate (+ finalize_gate + closure_guards) - integration (+ candidates) - config_cli (config + cli + skill_bundle) - Lift shared setup helpers (write_quest_yaml, make_quest_root, seed_runs, seed_pending_run, seed_worktree_pending_run, seed_distill_review, seed_distill_review_in_quest, make_artifact_quest, enable_distill_in_quest, make_decision_record, make_baseline_guidance) into tests/_distill_fixtures.py — eliminates ~200 lines of fixture duplication across 10 files - All 102 test function names preserved; assertions byte-identical to pre-consolidation behavior; no test dropped - Net: 10 files -> 4 + 1 helper; 2002 -> 1992 lines (line trim is small because savings from collapsed helpers were offset by the helper module + per-file docstrings; the visible-file-count drop is the win the PR reviewer was asking for) Per PR feedback: 10 distill test files with duplicated fixtures was test-file proliferation. Same coverage in fewer files. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
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
This PR adds an opt-in discipline + recall layer on top of DeepScientist's existing memory subsystem so each quest can review its own completed runs, write reusable knowledge cards into global memory, and surface those cards as priors to future quests. Paper-bundle submission is gated on the review. The whole feature can be toggled on or off when creating a quest.
Main already provides cross-quest memory storage (
MemoryServicetwo-scope,knowledgecard kind, MCPmemory.write/read/search/list_recent/promote_to_global,read_visibility_mode). What it lacks is a discipline that ensures lessons get written globally before a quest closes, and a recall hook that surfaces relevant priors at stage entry. This PR closes that gap.Changes included:
experience_distillandrecall_priorsfields to the startup contract and project-creation form (both default off)distill_reviewas adecision(action='distill_review')validation path under the existingdecisionartifact kind (not a new kind)distillstage skill implementing per-batch summary-scan + neighbor decisions over candidate runs; cards require onlyclaim+lineagemetadataartifact.list_distill_candidates,artifact.list_recent, andmemory.list_knowledge_summariesMCP tools (all under existing namespaces; no new public namespace)memory.writeMCP defaultscopefromquesttoglobalso reusable lessons land in the right place by default; explicitscope='quest'still worksartifact.submit_paper_bundleandartifact.complete_questblocking submission until every completed run is covered by adecision(action='distill_review'); aggregate the check acrossquest_root/artifactsplus every.ds/worktrees/*/artifactsdecision(write|finalize)guidance to thedistillskill while the gate is openrecall_priors_rulecue into stage-skill prompts whenrecall_priors: onds distill-questCLI for retroactive draft emission on existing repos401as non-retryable and probe enabled runners on daemon startup so auth/config failures surface immediatelyWhy
DeepScientist already accumulates per-run artifacts and per-quest paper bundles, and main's
MemoryServicealready supports global cards with cross-quest read visibility. But knowledge stayed locked inside the producing quest in practice — agents wrote to quest scope, didn't query global at stage entry, and skipped distillation at quest end. Quest 014 made this concrete: 9 quest-scope cards and 1 global card; useful lessons stayed locked inside the producing quest.This PR makes "review and externalize what you learned" an explicit gate without hard-coding which conclusions are worth keeping. The knowledge is saved as global cards and read by future quests through a tag-overlap recall MCP. Three small defaults (write→global, stage prose preferring global,
artifact.list_recentfor runtime view) make global the path of least resistance instead of a road less travelled.What changed since the original draft
The original draft introduced a separate
distill_reviewartifact kind, a per-slice analysis-routing trigger, deep card-metadata enforcement, and a Chinese distill guide. The slim revision (this PR's current shape) drops those:distill_reviewis no longer a separate artifact kind — it'sdecision(action='distill_review'). One less public kind to maintain.submit_paper_bundle/complete_quest) is the sole discipline mechanism; forcing a distill detour after every analysis slice was overreach.claim+lineage. Previous required depth (subtype,mechanism,conditions,confidence) was overreach — agents would game shallow taxonomy fields, and the cross-quest patch invariant only ever needed lineage.docs/zh/34_EXPERIENCE_DISTILL_GUIDE.mddeleted; zh readers follow the en guide. zh07_MEMORY_AND_MCP.mdand14_PROMPT_SKILLS_AND_MCP_GUIDE.mdcross-reference the en distill guide.memory.writeMCP default scope flipped fromquesttoglobal.scope=global; quest-scope is reserved for quest-internal task tracking.artifact.list_recent(kind, limit)exposes "what did I just do" via the artifact namespace, somemory.list_recentis no longer co-opted as a runtime-state surface.Tracking: #64.
Public surface
memory.writescopeflipsquest→global; explicitscope='quest'still worksmemory.list_recent/read/search/promote_to_globalartifact.record(kind='decision', action='distill_review', ...)decisionkindartifact.submit_paper_bundle/artifact.complete_questexperience_distill: onand any completed run is uncoveredmemory.list_knowledge_summaries(domain_tags, kind, limit)artifact.list_recent(kind, limit)artifact.list_distill_candidates()distill_reviewdecisionNo new public MCP namespace.
Scope
This PR intentionally does not:
experience_distill: off(the default)Both new toggles are opt-in advanced fields. There is no migration impact: existing quests keep
experience_distill: offand behave as before; the gate only fires when the toggle is explicitly turned on. Thememory.writedefault-scope flip changes the default only; call sites passingscope=explicitly are unaffected.Tests
pytest tests/test_experience_distill_*.py tests/test_mcp_servers.py tests/test_memory_and_artifact.py tests/test_prompt_builder.py— greenschema,gate,integration,config_cli) plus a shared fixtures moduletests/_distill_fixtures.py. The original 10 distill files were merged to reduce file-count proliferation while preserving every test name and assertion.Docs
docs/en/34_EXPERIENCE_DISTILL_GUIDE.mdcovering the toggles, finalize gate, hard guards, distill skill flow,decision(action='distill_review')payload, cross-quest recall, MCP tools, retroactiveds distill-questCLI, and debuggingdocs/en/07_MEMORY_AND_MCP.md,docs/en/14_PROMPT_SKILLS_AND_MCP_GUIDE.mdand their Chinese mirrors so existing readers land on the en distill guideRelated
Related: #64
AI Assistance
This PR was prepared with AI assistance and reviewed locally before submission.