fix(subagents): pin model + strip plugins/MCP/thinking on every spawn#52
Merged
Conversation
The capture-extract / dream / graph_extraction subagents were spawning `claude -p` with no `--model`, no `--settings`, and no `--strict-mcp-config`. That meant each spawn inherited the user's full Claude Code stack: default model (Opus), `alwaysThinkingEnabled`, `effortLevel`, every enabled plugin, and every MCP server (including lore itself, recursively). Real cost on a near-empty extraction was ~$0.35 / spawn with 46k cache-creation tokens — most of it spent rebuilding a system prompt the subagent never used. This adds `lore.subagent_config`, which materializes a minimal MCP config (lore-only or empty) plus a settings override (no plugins, no thinking, low effort) under `~/.lore/subagent/` and returns the flag set for each role. Capture and graph_extraction default to Haiku 4.5; dream defaults to Sonnet 4.6 because it does multi-step reflection (and runs at most once per 24h). Overridable via `LORE_SUBAGENT_MODEL`, `LORE_DREAM_MODEL`, `LORE_GRAPH_MODEL`. Smoke-tested end-to-end: a one-shot extraction now costs ~$0.012 (28x cheaper) with cache-creation down from 46k to 7.7k tokens. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
5 tasks
amitpaz1
added a commit
that referenced
this pull request
May 9, 2026
…ions (#53) Each ``claude -p`` capture-extract / dream / graph_extraction subagent is itself a Claude Code session. The user's PostToolUse / Stop / SessionEnd hooks therefore fire for the subagent's *own* tool uses. Without a guard, this cascades: 1. Active session does N tool uses → PostToolUse hook → spawns capture-extract subagent. 2. Subagent's own tool uses (Read, Bash, Grep …) re-fire PostToolUse, which writes to *its* session_id buffer.jsonl. 3. After ``LORE_CAPTURE_N`` entries, the hook spawns *another* capture-extract for the subagent — and so on. Observed in production for one user: ~700 spawns/hour, ~$34/h on Haiku, 685 distinct sessions hit in 60 minutes despite only one interactive ``claude`` process running. ``--strict-mcp-config`` from PR #52 isolates MCP but does not isolate hooks. Two layers of defense: 1. ``subagent_config.settings_body()`` now writes ``hooks`` as empty arrays. ``--settings`` overrides the user's hooks for the subagent's session. 2. New ``SubagentConfig.env_overrides()`` returns ``LORE_AUTO_SAVE=false`` and ``LORE_DREAM_AUTO=false``. The hook scripts honor these as master kill switches and exit 0 immediately. All three spawn sites now pass ``env={**os.environ, **cfg.env_overrides()}`` so the guard survives any caching of ``settings.json`` by the running parent Claude Code process. 3 new tests: hooks-empty in materialized settings, env_overrides shape, env_overrides parity across roles. Existing TestSpawnClaudeArgs extended to assert recursion-guard env vars are passed to subprocess.Popen. 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
claude -pwith no--model, no--settings, and no--strict-mcp-config, so each spawn inherits the user's full Claude Code stack — default model (often Opus),alwaysThinkingEnabled,effortLevel, every enabled plugin, and every MCP server (including lore itself, recursively).lore.subagent_configmodule materializes a minimal MCP config (lore-only or empty) plus a settings override (no plugins, no thinking, low effort) under~/.lore/subagent/and returns the flag set per role. All three spawn sites now append--model,--strict-mcp-config,--mcp-config,--settings.LORE_SUBAGENT_MODEL,LORE_DREAM_MODEL,LORE_GRAPH_MODEL.Result
End-to-end smoke test on a one-shot extraction:
total_cost_usdcache_creation_input_tokensmodelTest plan
tests/test_subagent_config.py— 12 new tests covering model defaults, env-var override chain, MCP config bodies (lore-only vs empty), settings body,claude_flags()shape, lazy materialization, idempotent rewrite.tests/services/test_graph_extraction.py::TestSpawnClaudeArgs— extended to assert--model,--strict-mcp-config,--mcp-config,--settingsare present (regression guard against silently re-inheriting the user's stack).pytest tests/ --ignore=tests/integration→ 2746 passed, 1183 skipped, 0 failed.ruff checkclean on all touched files.claude -paccepts the flag set, model override took effect, cost numbers above.🤖 Generated with Claude Code