fix(claude-code): replace hardcoded PostToolUse hook with rich episodic logging

[aliirz]

What was broken

The PostToolUse hook in adapters/claude-code/settings.json called:

python3 .agent/tools/memory_reflect.py claude-code post-tool ok

Three hardcoded strings. Claude Code sends a rich JSON payload to every hook via stdin — tool_name, tool_input, tool_response, exit codes, error streams. The old command ignored all of it.

Every session produced identical records:

{"action": "post-tool", "detail": "ok", "reflection": ""}

The dream cycle clusters on the reflection field. Empty string → word_set() returns {} → Jaccard = 0 across all entries → zero clusters → zero candidates staged. Didn't matter how many sessions you ran. The brain never learned anything automatically.

Five cascading failures from this one line:

#	Problem	Consequence
1	reflection = "" always	content_cluster() has nothing to work with, stages zero candidates
2	pain_score = 2 hardcoded for all successes	max salience for a cluster of 6 = 5.4, never crosses the 7.0 promotion threshold
3	No failure detection	a failed deploy and git status logged identically
4	All tool calls importance = 3	salience formula produces near-zero for everything
5	Service names hardcoded in importance scorer	not portable across stacks

---

What this PR changes

.agent/harness/hooks/claude_code_post_tool.py (new)

Reads the JSON payload Claude Code sends via stdin on every PostToolUse event. Falls back to CLAUDE_TOOL_NAME / CLAUDE_TOOL_INPUT env vars for older versions.

• Derives a meaningful action label from tool_name + tool_input
• Scores importance from the operation, not the service brand — vercel deploy is 9 because of deploy, not vercel; supabase db push is 6 because of push
• Detects failures via exit_code, error stream, interrupted, is_error
• Generates a non-empty reflection the dream cycle can actually cluster on
• Sets pain_score = 5 for importance ≥ 8 successes so recurring patterns cross the 7.0 promotion threshold (10 × 0.5 × 0.9 × 3 = 13.5)

.agent/protocols/hook_patterns.json (new)

User-owned config for stack-specific service names. Ships with empty high_stakes / medium_stakes arrays and a _examples section covering common tools (supabase, vercel, stripe, railway, etc.). Users copy from _examples to activate. Universal patterns stay hardcoded in the hook — brand names do not.

.agent/harness/hooks/post_execution.py

Adds optional pain_score parameter to log_execution(). Default behaviour unchanged.

.agent/harness/hooks/on_failure.py

Removes the noisy str: type prefix from the reflection field when the error is already a string. Only adds the type name for actual Exception objects.

.agent/tools/memory_reflect.py

Exposes --pain CLI flag so manual reflect calls can set pain_score explicitly (e.g. --pain 5 for a significant successful discovery, --pain 9 for a production incident).

adapters/claude-code/settings.json

Points PostToolUse at claude_code_post_tool.py. Expands matcher to include MultiEdit, Task, TodoWrite.

Note: the original file had two deny rules: git push --force* and rm -rf /*. This PR keeps both unchanged. No new deny rules were added — service-specific constraints like supabase db reset belong in the user's own hook_patterns.json, not in the universal adapter.

adapters/claude-code/CLAUDE.md

Adds explicit guidance on when to call memory_reflect.py manually, with an importance guide table and concrete good/bad examples.

---

How to verify

git clone https://github.com/aliirz/agentic-stack.git
cd agentic-stack && git checkout fix/claude-code-hook-rich-episodic
./install.sh claude-code /tmp/test-project
cd /tmp/test-project
python3 /path/to/agentic-stack/test_claude_code_hook.py
# 54/54 passed

After a real Claude Code session:

tail -1 .agent/memory/episodic/AGENT_LEARNINGS.jsonl | python3 -m json.tool
# action should describe the actual command, reflection non-empty, importance > 3

python3 .agent/memory/auto_dream.py
# staged=N not staged=0

---

Out of scope

The three missing skills from the article (api-scaffold, test-writer, code-reviewer) and ACTIVE_PLAN.md are a separate PR.

[ujjwalbaliyan007]

It's on review