devflow-kit 1.5.0 → 1.6.1

package/plugins/devflow-specify/.claude-plugin/plugin.json

@@ -4,7 +4,7 @@
   "author": {
     "name": "Dean0x"
   },
-  "version": "1.5.0",
+  "version": "1.6.1",
   "homepage": "https://github.com/dean0x/devflow",
   "repository": "https://github.com/dean0x/devflow",
   "license": "MIT",

package/plugins/devflow-specify/agents/skimmer.md

@@ -1,6 +1,7 @@
 ---
 name: Skimmer
 description: Codebase orientation using skim to identify relevant files, functions, and patterns for a feature or task
+skills: knowledge-persistence
 model: inherit
 ---
 
@@ -20,6 +21,7 @@ You receive from orchestrator:
 3. **Search for task-relevant code** - Find files matching task keywords
 4. **Identify integration points** - Exports, entry points, import patterns
 5. **Generate orientation summary** - Structured output for implementation planning
+6. **Check project knowledge** - If `.memory/knowledge/decisions.md` exists, read its `<!-- TL;DR: ... -->` first-line comment and include active decision count in orientation under "### Active Decisions". Only the TL;DR is read here (not full entries) — this is intentional for token efficiency; agents that need full entries read the file themselves.
 
 ## Tool Invocation
 
@@ -64,6 +66,9 @@ Always invoke skim via `npx rskim`. This works whether or not skim is globally i
 ### Patterns Observed
 {Existing patterns to follow}
 
+### Active Decisions
+{Count and key decisions from `.memory/knowledge/decisions.md` TL;DR, or "None found" if file missing}
+
 ### Suggested Approach
 {Brief recommendation based on codebase structure}
 ```

package/plugins/devflow-specify/commands/specify-teams.md

@@ -57,7 +57,11 @@ Find: project structure, similar features, patterns, integration points
 Return: codebase context for requirements (not implementation details)"
 ```
 
-### Phase 3: Explore Requirements (Agent Teams)
+### Phase 3: Load Project Knowledge
+
+Read `.memory/knowledge/decisions.md` and `.memory/knowledge/pitfalls.md`. Pass their content as context to the exploration team below — prior decisions constrain requirements, known pitfalls inform failure modes.
+
+### Phase 4: Explore Requirements (Agent Teams)
 
 Create an agent team for collaborative requirements exploration:
 
@@ -122,7 +126,7 @@ Max 2 debate rounds, then submit consensus requirements findings.
 
 **Exploration team output**: Consensus findings on user needs, similar features, constraints, failure modes.
 
-**Team Shutdown Protocol** (must complete before Phase 5):
+**Team Shutdown Protocol** (must complete before Phase 6):
 
 ```
 Step 1: Shutdown each teammate
@@ -139,9 +143,9 @@ Step 3: GATE — Verify TeamDelete succeeded
   If retry failed → HALT and report: "Exploration team cleanup failed. Cannot create planning team."
 ```
 
-### Phase 4: Synthesize Exploration
+### Phase 5: Synthesize Exploration
 
-**WAIT** for Phase 3, then spawn Synthesizer:
+**WAIT** for Phase 4, then spawn Synthesizer:
 
 ```
 Task(subagent_type="Synthesizer"):
@@ -151,7 +155,7 @@ Explorer consensus: {team exploration consensus output}
 Combine into: user needs, similar features, constraints, failure modes"
 ```
 
-### Phase 5: Plan Scope (Agent Teams)
+### Phase 6: Plan Scope (Agent Teams)
 
 Create an agent team for collaborative scope planning:
 
@@ -164,7 +168,7 @@ Spawn planning teammates with self-contained prompts:
   Prompt: |
     You are planning scope for feature: {feature}
     Exploration synthesis (what we know):
-    {synthesis output from Phase 4}
+    {synthesis output from Phase 5}
 
     Your deliverable: User stories in "As X, I want Y, so that Z" format.
     Cover all actors, actions, and outcomes identified in exploration.
@@ -175,7 +179,7 @@ Spawn planning teammates with self-contained prompts:
   Prompt: |
     You are planning scope for feature: {feature}
     Exploration synthesis (what we know):
-    {synthesis output from Phase 4}
+    {synthesis output from Phase 5}
 
     Your deliverable: v1 MVP scope, v2 deferred items, explicitly out of scope,
     and dependencies on other features or systems.
@@ -186,7 +190,7 @@ Spawn planning teammates with self-contained prompts:
   Prompt: |
     You are planning scope for feature: {feature}
     Exploration synthesis (what we know):
-    {synthesis output from Phase 4}
+    {synthesis output from Phase 5}
 
     Your deliverable: Testable acceptance criteria for success cases,
     failure cases, and edge cases. Every criterion must be verifiable.
@@ -221,9 +225,9 @@ Step 3: GATE — Verify TeamDelete succeeded
   If retry failed → HALT and report: "Planning team cleanup failed. Cannot proceed to Gate 1."
 ```
 
-### Phase 6: Synthesize Planning
+### Phase 7: Synthesize Planning
 
-**WAIT** for Phase 5, then spawn Synthesizer:
+**WAIT** for Phase 6, then spawn Synthesizer:
 
 ```
 Task(subagent_type="Synthesizer"):
@@ -233,7 +237,7 @@ Planner consensus: {team planning consensus output}
 Combine into: user stories, scope breakdown, acceptance criteria, open questions"
 ```
 
-### Phase 7: Gate 1 - Validate Scope
+### Phase 8: Gate 1 - Validate Scope
 
 Use AskUserQuestion to validate:
 - Primary problem being solved
@@ -241,14 +245,14 @@ Use AskUserQuestion to validate:
 - v1 scope selection (minimal/medium/full)
 - Explicit exclusions
 
-### Phase 8: Gate 2 - Confirm Criteria
+### Phase 9: Gate 2 - Confirm Criteria
 
 Present specification summary, then use AskUserQuestion for final confirmation:
 - Success UX pattern
 - Error handling approach
 - Ready to create issue / Needs changes / Cancel
 
-### Phase 9: Create Issue
+### Phase 10: Create Issue
 
 Create GitHub issue with `gh issue create`:
 - Title: feature name
@@ -268,32 +272,35 @@ Report issue number and URL.
 ├─ Phase 2: Orient
 │  └─ Skimmer agent (codebase context via skim)
 │
-├─ Phase 3: Explore Requirements (Agent Teams)
+├─ Phase 3: Load Project Knowledge
+│  └─ Read decisions.md + pitfalls.md
+│
+├─ Phase 4: Explore Requirements (Agent Teams)
 │  ├─ User Perspective Explorer (teammate)
 │  ├─ Similar Features Explorer (teammate)
 │  ├─ Constraints Explorer (teammate)
 │  ├─ Failure Mode Explorer (teammate)
 │  └─ Debate → consensus requirements findings
 │
-├─ Phase 4: Synthesize Exploration
+├─ Phase 5: Synthesize Exploration
 │  └─ Synthesizer agent (mode: exploration)
 │
-├─ Phase 5: Plan Scope (Agent Teams)
+├─ Phase 6: Plan Scope (Agent Teams)
 │  ├─ User Stories Planner (teammate)
 │  ├─ Scope Boundaries Planner (teammate)
 │  ├─ Acceptance Criteria Planner (teammate)
 │  └─ Debate → consensus scope plan
 │
-├─ Phase 6: Synthesize Planning
+├─ Phase 7: Synthesize Planning
 │  └─ Synthesizer agent (mode: planning)
 │
-├─ Phase 7: GATE 1 - Validate Scope ⛔ MANDATORY
+├─ Phase 8: GATE 1 - Validate Scope ⛔ MANDATORY
 │  └─ AskUserQuestion: Confirm scope and priorities
 │
-├─ Phase 8: GATE 2 - Confirm Criteria ⛔ MANDATORY
+├─ Phase 9: GATE 2 - Confirm Criteria ⛔ MANDATORY
 │  └─ AskUserQuestion: Final spec approval
 │
-├─ Phase 9: Create Issue
+├─ Phase 10: Create Issue
 │  └─ gh issue create
 │
 └─ Report: Issue number, URL, /implement command

package/plugins/devflow-specify/commands/specify.md

@@ -57,20 +57,24 @@ Find: project structure, similar features, patterns, integration points
 Return: codebase context for requirements (not implementation details)"
 ```
 
-### Phase 3: Explore Requirements (Parallel)
+### Phase 3: Load Project Knowledge
 
-Spawn 4 Explore agents **in a single message**, each with Skimmer context:
+Read `.memory/knowledge/decisions.md` and `.memory/knowledge/pitfalls.md`. Pass their content as context to the Explore agents below — prior decisions constrain requirements, known pitfalls inform failure modes.
+
+### Phase 4: Explore Requirements (Parallel)
+
+Spawn 4 Explore agents **in a single message**, each with Skimmer context and project knowledge (if loaded):
 
 | Focus | Thoroughness | Find |
 |-------|-------------|------|
 | User perspective | medium | Target users, goals, pain points, user journeys |
 | Similar features | medium | Comparable features, scope patterns, edge cases |
-| Constraints | quick | Dependencies, business rules, security, performance |
-| Failure modes | quick | Error states, edge cases, validation needs |
+| Constraints | quick | Dependencies, business rules, security, performance, **prior architectural decisions** |
+| Failure modes | quick | Error states, edge cases, validation needs, **known pitfalls** |
 
-### Phase 4: Synthesize Exploration
+### Phase 5: Synthesize Exploration
 
-**WAIT** for Phase 3, then spawn Synthesizer:
+**WAIT** for Phase 4, then spawn Synthesizer:
 
 ```
 Task(subagent_type="Synthesizer"):
@@ -79,7 +83,7 @@ Mode: exploration
 Combine into: user needs, similar features, constraints, failure modes"
 ```
 
-### Phase 5: Plan Scope (Parallel)
+### Phase 6: Plan Scope (Parallel)
 
 Spawn 3 Plan agents **in a single message**, each with exploration synthesis:
 
@@ -89,9 +93,9 @@ Spawn 3 Plan agents **in a single message**, each with exploration synthesis:
 | Scope boundaries | v1 MVP, v2 deferred, out of scope, dependencies |
 | Acceptance criteria | Success/failure/edge case criteria (testable) |
 
-### Phase 6: Synthesize Planning
+### Phase 7: Synthesize Planning
 
-**WAIT** for Phase 5, then spawn Synthesizer:
+**WAIT** for Phase 6, then spawn Synthesizer:
 
 ```
 Task(subagent_type="Synthesizer"):
@@ -100,7 +104,7 @@ Mode: planning
 Combine into: user stories, scope breakdown, acceptance criteria, open questions"
 ```
 
-### Phase 7: Gate 1 - Validate Scope
+### Phase 8: Gate 1 - Validate Scope
 
 Use AskUserQuestion to validate:
 - Primary problem being solved
@@ -108,14 +112,14 @@ Use AskUserQuestion to validate:
 - v1 scope selection (minimal/medium/full)
 - Explicit exclusions
 
-### Phase 8: Gate 2 - Confirm Criteria
+### Phase 9: Gate 2 - Confirm Criteria
 
 Present specification summary, then use AskUserQuestion for final confirmation:
 - Success UX pattern
 - Error handling approach
 - Ready to create issue / Needs changes / Cancel
 
-### Phase 9: Create Issue
+### Phase 10: Create Issue
 
 Create GitHub issue with `gh issue create`:
 - Title: feature name
@@ -135,30 +139,33 @@ Report issue number and URL.
 ├─ Phase 2: Orient
 │  └─ Skimmer agent (codebase context via skim)
 │
-├─ Phase 3: Explore Requirements (PARALLEL)
+├─ Phase 3: Load Project Knowledge
+│  └─ Read decisions.md + pitfalls.md
+│
+├─ Phase 4: Explore Requirements (PARALLEL)
 │  ├─ Explore: User perspective
 │  ├─ Explore: Similar features
 │  ├─ Explore: Constraints
 │  └─ Explore: Failure modes
 │
-├─ Phase 4: Synthesize Exploration
+├─ Phase 5: Synthesize Exploration
 │  └─ Synthesizer agent (mode: exploration)
 │
-├─ Phase 5: Plan Scope (PARALLEL)
+├─ Phase 6: Plan Scope (PARALLEL)
 │  ├─ Plan: User stories
 │  ├─ Plan: Scope boundaries
 │  └─ Plan: Acceptance criteria
 │
-├─ Phase 6: Synthesize Planning
+├─ Phase 7: Synthesize Planning
 │  └─ Synthesizer agent (mode: planning)
 │
-├─ Phase 7: GATE 1 - Validate Scope ⛔ MANDATORY
+├─ Phase 8: GATE 1 - Validate Scope ⛔ MANDATORY
 │  └─ AskUserQuestion: Confirm scope and priorities
 │
-├─ Phase 8: GATE 2 - Confirm Criteria ⛔ MANDATORY
+├─ Phase 9: GATE 2 - Confirm Criteria ⛔ MANDATORY
 │  └─ AskUserQuestion: Final spec approval
 │
-├─ Phase 9: Create Issue
+├─ Phase 10: Create Issue
 │  └─ gh issue create
 │
 └─ Report: Issue number, URL, /implement command

package/plugins/devflow-typescript/.claude-plugin/plugin.json

@@ -4,7 +4,7 @@
   "author": {
     "name": "Dean0x"
   },
-  "version": "1.5.0",
+  "version": "1.6.1",
   "homepage": "https://github.com/dean0x/devflow",
   "repository": "https://github.com/dean0x/devflow",
   "license": "MIT",

package/scripts/hooks/ambient-prompt

@@ -25,20 +25,22 @@ if [[ "$PROMPT" == /* ]]; then
 fi
 
 # Skip single-word confirmations (< 2 words)
-WORD_COUNT=$(echo "$PROMPT" | wc -w | tr -d ' ')
+WORD_COUNT=$(printf '%s' "$PROMPT" | wc -w | tr -d ' ')
 if [ "$WORD_COUNT" -lt 2 ]; then
   exit 0
 fi
 
-# Inject classification preamble
-PREAMBLE="AMBIENT MODE ACTIVE: Before responding, silently classify this prompt:
-Intent: BUILD | DEBUG | REVIEW | PLAN | EXPLORE | CHAT
-Depth: QUICK (no overhead) | GUIDED (load skills) | ELEVATE (suggest /command)
+# Normalize to lowercase for matching
+PROMPT_LOWER=$(printf '%s' "$PROMPT" | tr '[:upper:]' '[:lower:]')
 
-If GUIDED+: Use the ambient-router skill already in your session context for classification details and skill selection matrix. Then Read the selected skills from ~/.claude/skills/<name>/SKILL.md and apply their patterns. For BUILD tasks, also load test-driven-development skill and enforce RED-GREEN-REFACTOR.
+# Fast-path: git operations are always QUICK — skip preamble
+if printf '%s' "$PROMPT_LOWER" | grep -qE '^(commit|push|pull|merge|rebase|cherry-pick|squash|tag|stash)|create (a )?pr|open (a )?pr'; then
+  exit 0
+fi
 
-If QUICK: Respond normally without stating classification.
-Only state classification aloud for GUIDED/ELEVATE."
+# Inject classification preamble
+# SYNC: must match tests/integration/helpers.ts AMBIENT_PREAMBLE
+PREAMBLE="AMBIENT MODE ACTIVE: Before responding, silently classify this prompt using the ambient-router skill already in your session context. If QUICK, respond normally without stating classification. If GUIDED or ORCHESTRATED, you MUST load the selected skills using the Skill tool before proceeding."
 
 jq -n --arg ctx "$PREAMBLE" '{
   "hookSpecificOutput": {

package/scripts/hooks/background-memory-update

@@ -2,8 +2,9 @@
 
 # Background Working Memory Updater
 # Called by stop-update-memory as a detached background process.
-# Resumes the parent session headlessly to update .memory/WORKING-MEMORY.md.
-# On failure: logs error, does nothing (no fallback).
+# Reads the last turn from the session transcript, then uses a fresh `claude -p`
+# invocation to update .memory/WORKING-MEMORY.md.
+# On failure: logs error, does nothing (stale memory is better than fake data).
 
 set -e
 
@@ -29,7 +30,7 @@ rotate_log() {
 
 # --- Stale Lock Recovery ---
 
-# Portable mtime in epoch seconds (same pattern as stop-update-memory:35-39)
+# Portable mtime in epoch seconds
 get_mtime() {
   if stat --version &>/dev/null 2>&1; then
     stat -c %Y "$1"
@@ -72,11 +73,65 @@ cleanup() {
 }
 trap cleanup EXIT
 
+# --- Transcript Extraction ---
+
+extract_last_turn() {
+  # Compute transcript path: Claude Code stores transcripts at
+  # ~/.claude/projects/{cwd-with-slashes-replaced-by-hyphens}/{session_id}.jsonl
+  local encoded_cwd
+  encoded_cwd=$(echo "$CWD" | sed 's|^/||' | tr '/' '-')
+  local transcript="$HOME/.claude/projects/-${encoded_cwd}/${SESSION_ID}.jsonl"
+
+  if [ ! -f "$transcript" ]; then
+    log "Transcript not found at $transcript"
+    return 1
+  fi
+
+  # Extract last user and assistant text from JSONL
+  # Each line is a JSON object with "type" field
+  local last_user last_assistant
+
+  last_user=$(grep '"type":"user"' "$transcript" 2>/dev/null \
+    | tail -3 \
+    | jq -r '
+      if .message.content then
+        [.message.content[] | select(.type == "text") | .text] | join("\n")
+      else ""
+      end
+    ' 2>/dev/null \
+    | tail -1)
+
+  last_assistant=$(grep '"type":"assistant"' "$transcript" 2>/dev/null \
+    | tail -3 \
+    | jq -r '
+      if .message.content then
+        [.message.content[] | select(.type == "text") | .text] | join("\n")
+      else ""
+      end
+    ' 2>/dev/null \
+    | tail -1)
+
+  # Truncate to ~4000 chars total to keep token cost low
+  if [ ${#last_user} -gt 2000 ]; then
+    last_user="${last_user:0:2000}... [truncated]"
+  fi
+  if [ ${#last_assistant} -gt 2000 ]; then
+    last_assistant="${last_assistant:0:2000}... [truncated]"
+  fi
+
+  if [ -z "$last_user" ] && [ -z "$last_assistant" ]; then
+    log "No text content found in transcript"
+    return 1
+  fi
+
+  LAST_USER_TEXT="$last_user"
+  LAST_ASSISTANT_TEXT="$last_assistant"
+  return 0
+}
+
 # --- Main ---
 
-# Wait for parent session to flush transcript.
-# 3s provides ~6-10x margin over typical flush times.
-# If --resume shows stale transcripts, bump to 5s.
+# Wait for parent session to flush transcript
 sleep 3
 
 log "Starting update for session $SESSION_ID"
@@ -87,7 +142,6 @@ break_stale_lock
 # Acquire lock (other sessions may be updating concurrently)
 if ! acquire_lock; then
   log "Lock timeout after 90s — skipping update for session $SESSION_ID"
-  # Don't clean up lock we don't own
   trap - EXIT
   exit 0
 fi
@@ -102,97 +156,72 @@ if [ -f "$MEMORY_FILE" ]; then
   PRE_UPDATE_MTIME=$(get_mtime "$MEMORY_FILE")
 fi
 
-# Build instruction
-if [ -n "$EXISTING_MEMORY" ]; then
-  PATTERNS_INSTRUCTION=""
-PATTERNS_FILE="$CWD/.memory/PROJECT-PATTERNS.md"
-EXISTING_PATTERNS=""
-if [ -f "$PATTERNS_FILE" ]; then
-  EXISTING_PATTERNS=$(cat "$PATTERNS_FILE")
-  PATTERNS_INSTRUCTION="
-
-Also update $PATTERNS_FILE by APPENDING any new recurring patterns discovered during this session. Do NOT overwrite existing entries — only add new ones. Skip if no new patterns were observed. Format each entry as: - **Pattern name**: Brief description (discovered: YYYY-MM-DD). Keep patterns.md under 40 entries. When approaching the limit, consolidate related patterns into broader entries rather than adding duplicates.
-
-Existing patterns:
-$EXISTING_PATTERNS"
-else
-  PATTERNS_INSTRUCTION="
-
-If recurring patterns were observed during this session (coding conventions, architectural decisions, team preferences, tooling quirks), create $PATTERNS_FILE with entries formatted as: - **Pattern name**: Brief description (discovered: YYYY-MM-DD). Only create this file if genuine patterns were observed — do not fabricate entries."
+# Gather git state (always available, used as fallback too)
+GIT_STATE=""
+if cd "$CWD" 2>/dev/null && git rev-parse --git-dir >/dev/null 2>&1; then
+  GIT_STATUS=$(git status --short 2>/dev/null | head -20)
+  GIT_LOG=$(git log --oneline -5 2>/dev/null)
+  GIT_DIFF=$(git diff --stat HEAD 2>/dev/null | tail -10)
+  GIT_STATE="Branch: $(git branch --show-current 2>/dev/null || echo 'unknown')
+Recent commits:
+${GIT_LOG}
+Changed files:
+${GIT_STATUS}
+Diff summary:
+${GIT_DIFF}"
 fi
 
-INSTRUCTION="First, Read the file $MEMORY_FILE to satisfy Claude Code's read-before-write requirement. Then update it with working memory from this session. The file already has content — possibly from a concurrent session that just wrote it moments ago. Merge this session's context with the existing content to produce a single unified working memory snapshot. Both this session and the existing content represent fresh, concurrent work — integrate both fully. Working memory captures what's active now, not a changelog. Deduplicate overlapping information. Keep under 120 lines total. Use the same structure: ## Now, ## Progress, ## Decisions, ## Modified Files, ## Context, ## Session Log.
-
-## Progress tracks Done (completed items), Remaining (next steps), and Blockers (if any). Keep each sub-list to 1-3 items. This section reflects current work state, not historical logs.
-
-## Decisions entries must include date and status. Format: - **[Decision]** — [rationale] (YYYY-MM-DD) [ACTIVE|SUPERSEDED]. Mark superseded decisions rather than deleting them.${PATTERNS_INSTRUCTION}
-
-Existing content:
-$EXISTING_MEMORY"
+# Extract last turn from transcript (or fall back to git-only)
+LAST_USER_TEXT=""
+LAST_ASSISTANT_TEXT=""
+EXCHANGE_SECTION=""
+
+if extract_last_turn; then
+  log "--- Extracted user text (${#LAST_USER_TEXT} chars) ---"
+  log "$LAST_USER_TEXT"
+  log "--- Extracted assistant text (${#LAST_ASSISTANT_TEXT} chars) ---"
+  log "$LAST_ASSISTANT_TEXT"
+  log "--- End transcript extraction ---"
+  EXCHANGE_SECTION="Last exchange:
+User: ${LAST_USER_TEXT}
+Assistant: ${LAST_ASSISTANT_TEXT}"
 else
-  PATTERNS_INSTRUCTION=""
-  PATTERNS_FILE="$CWD/.memory/PROJECT-PATTERNS.md"
-  if [ -f "$PATTERNS_FILE" ]; then
-    EXISTING_PATTERNS=$(cat "$PATTERNS_FILE")
-    PATTERNS_INSTRUCTION="
-
-Also update $PATTERNS_FILE by APPENDING any new recurring patterns discovered during this session. Do NOT overwrite existing entries — only add new ones. Skip if no new patterns were observed. Format each entry as: - **Pattern name**: Brief description (discovered: YYYY-MM-DD). Keep patterns.md under 40 entries. When approaching the limit, consolidate related patterns into broader entries rather than adding duplicates.
-
-Existing patterns:
-$EXISTING_PATTERNS"
-  else
-    PATTERNS_INSTRUCTION="
-
-If recurring patterns were observed during this session (coding conventions, architectural decisions, team preferences, tooling quirks), create $PATTERNS_FILE with entries formatted as: - **Pattern name**: Brief description (discovered: YYYY-MM-DD). Only create this file if genuine patterns were observed — do not fabricate entries."
-  fi
-
-  INSTRUCTION="First, Read the file $MEMORY_FILE if it exists (to satisfy Claude Code's read-before-write requirement). Then create it with working memory from this session. Keep under 120 lines. Use this structure:
-
-# Working Memory
-
-## Now
-<!-- Current focus, status, blockers (1-3 bullets) -->
-
-## Progress
-<!-- Done: completed items (1-3). Remaining: next steps (1-3). Blockers: if any. -->
+  log "Falling back to git-state-only context"
+  EXCHANGE_SECTION="(Session transcript not available — using git state only)"
+fi
 
-## Decisions
-<!-- Format: - **[Decision]** — [rationale] (YYYY-MM-DD) [ACTIVE|SUPERSEDED] -->
+# Build prompt for fresh claude -p invocation
+PROMPT="You are a working memory updater. Your ONLY job is to update the file at ${MEMORY_FILE} using the Write tool. Do it immediately — do not ask questions or explain.
 
-## Modified Files
-<!-- File paths only, most recent first -->
+Current working memory:
+${EXISTING_MEMORY:-"(no existing content)"}
 
-## Context
-<!-- Branch, PR, architectural context, open questions -->
+${EXCHANGE_SECTION}
 
-## Session Log
+Git state:
+${GIT_STATE:-"(not a git repo)"}
 
-### Today
-<!-- Chronological summary of work done today (2-5 bullets) -->
+Instructions:
+- Use the Write tool to update ${MEMORY_FILE} immediately
+- Keep under 120 lines
+- Use sections: ## Now, ## Progress, ## Decisions, ## Modified Files, ## Context, ## Session Log
+- Integrate new information with existing content
+- Deduplicate overlapping information
+- ## Progress tracks Done (completed), Remaining (next steps), Blockers (if any)
+- ## Decisions entries: format as - **[Decision]** — [rationale] (YYYY-MM-DD) [ACTIVE|SUPERSEDED]"
 
-### This Week
-<!-- Broader multi-day context if relevant -->${PATTERNS_INSTRUCTION}"
-fi
+log "--- Full prompt being passed to claude -p ---"
+log "$PROMPT"
+log "--- End prompt ---"
 
-# Resume session headlessly to perform the update
-TIMEOUT=120  # Normal runtime 30-60s; 2x margin
+# Run fresh claude -p (no --resume, no conversation confusion)
+TIMEOUT=120
 
-DEVFLOW_BG_UPDATER=1 env -u CLAUDECODE "$CLAUDE_BIN" -p \
-  --resume "$SESSION_ID" \
+DEVFLOW_BG_UPDATER=1 "$CLAUDE_BIN" -p \
   --model haiku \
-  --tools "Read,Write,Bash" \
-  --allowedTools \
-    "Read($CWD/.memory/WORKING-MEMORY.md)" \
-    "Read($CWD/.memory/PROJECT-PATTERNS.md)" \
-    "Write($CWD/.memory/WORKING-MEMORY.md)" \
-    "Write($CWD/.memory/PROJECT-PATTERNS.md)" \
-    "Bash(git status:*)" \
-    "Bash(git log:*)" \
-    "Bash(git diff:*)" \
-    "Bash(git branch:*)" \
-  --no-session-persistence \
+  --dangerously-skip-permissions \
   --output-format text \
-  "$INSTRUCTION" \
+  "$PROMPT" \
   >> "$LOG_FILE" 2>&1 &
 CLAUDE_PID=$!