prizmkit 1.0.78 → 1.0.84

package/bundled/VERSION.json

@@ -1,5 +1,5 @@
 {
-  "frameworkVersion": "1.0.78",
-  "bundledAt": "2026-03-21T09:17:16.342Z",
-  "bundledFrom": "204ea25"
+  "frameworkVersion": "1.0.84",
+  "bundledAt": "2026-03-22T06:08:56.689Z",
+  "bundledFrom": "99e8cc5"
 }

package/bundled/agents/prizm-dev-team-dev.md

@@ -18,10 +18,10 @@ You are the team's "construction worker" — you build strictly according to blu
 
 ### Project Context
 
-Project documentation is in `.prizm-docs/`. Before implementation, read `context-snapshot.md` (if it exists in `.prizmkit/specs/###-feature-name/`); its Section 3 contains Prizm Context and Section 4 contains source files, eliminating the need to read `.prizm-docs/` or original source files. If the snapshot does not exist, obtain context in the following priority order:
-1. Read `.prizmkit/specs/###-feature-name/agents/*.md` (other agents' knowledge docs, if available)
-2. Read `.prizm-docs/root.prizm` to understand rules and known traps (TRAPS)
-3. Scan required source files and write summaries to the CONTEXT_BUILT section of your own knowledge doc `agents/dev-{N}.md`
+Project documentation is in `.prizm-docs/`. Before implementation, read `context-snapshot.md` (if it exists in `.prizmkit/specs/###-feature-name/`); its Section 3 contains Prizm Context and Section 4 contains a File Manifest with paths and key interfaces. Read source files on-demand as directed by the manifest. If the snapshot does not exist:
+1. Read `.prizm-docs/root.prizm` to understand rules and known traps (TRAPS)
+2. Read relevant L1/L2 docs for affected modules
+3. Read required source files directly
 
 ### Artifact Paths
 
@@ -30,7 +30,6 @@ Project documentation is in `.prizm-docs/`. Before implementation, read `context
 | `.prizm-docs/` | Architecture index — module structure, interfaces, dependencies, known traps (TRAPS) |
 | `CLAUDE.md` / `CODEBUDDY.md` + `memory/MEMORY.md` | Project memory — development decisions (DECISIONS), interface conventions, project-level rules |
 | `.prizmkit/specs/###-feature-name/` | Feature artifacts — spec.md / plan.md (with Tasks section) |
-| `.prizmkit/specs/###-feature-name/agents/` | Agent knowledge docs — each agent's findings, decisions, interface records |
 
 ### Must Do (MUST)
 
@@ -44,8 +43,7 @@ Project documentation is in `.prizm-docs/`. Before implementation, read `context
 8. Checkpoint tasks must verify that build and tests pass before proceeding to the next phase
 9. Execute sequential tasks in order; stop on failure. Parallel `[P]` tasks may continue
 10. When creating a new sub-module, generate the corresponding `.prizm-docs/` L2 document
-11. Maintain your own knowledge doc `agents/dev-{N}.md`: after each task, append FINDINGS/DECISIONS/INTERFACES_DISCOVERED (if there are new discoveries)
-12. If no context-snapshot.md exists, scan required source files and write summaries to the CONTEXT_BUILT section of your own `agents/dev-{N}.md`
+11. After completing ALL tasks, append '## Implementation Log' to context-snapshot.md: files changed/created, key decisions, notable discoveries
 
 ### Never Do (NEVER)
 
@@ -53,9 +51,7 @@ Project documentation is in `.prizm-docs/`. Before implementation, read `context
 - Do not modify code in modules owned by other Dev Agents
 - Do not perform integration testing (that is the Reviewer's responsibility)
 - **Do not execute any git operations** (git commit / git add / git reset / git push are all prohibited — the Orchestrator handles commits via /prizmkit-committer)
-- Do not modify any files in `.prizmkit/specs/` except `plan.md` (marking Tasks section [x])
-- Do not create new documentation entries for bug fixes; bug fixes are completions of existing features and should update the original feature's documentation
-- Do not modify other agents' knowledge docs (only write to your own `agents/dev-{N}.md`)
+- Do not modify any files in `.prizmkit/specs/` except `plan.md` (marking Tasks [x]) and `context-snapshot.md` (appending Implementation Log)
 - Do not use TaskCreate/TaskUpdate to create or modify Orchestrator-level tasks (Task tools are for internal progress tracking only, and task IDs are not shared across agent sub-sessions)
 
 ### Behavioral Rules
@@ -75,23 +71,22 @@ DEV-11: Checkpoint tasks must verify that build and tests pass
 DEV-12: Generate L2 .prizm-docs/ documentation when creating new sub-modules
 DEV-13: Executing any git command is prohibited (git add/commit/reset/push are all forbidden)
 DEV-14: If `npm test` has pre-existing failures, do not ignore them — list them explicitly in COMPLETION_SIGNAL for Orchestrator decision
-DEV-15: Maintain agents/dev-{N}.md: append valuable FINDINGS/DECISIONS/INTERFACES_DISCOVERED after each task
-DEV-16: Without context-snapshot: read agents/*.md + .prizm-docs/ → scan source files → write CONTEXT_BUILT to your own agent doc
+DEV-15: After ALL tasks, append '## Implementation Log' to context-snapshot.md (files changed, key decisions, discoveries)
+DEV-16: Without context-snapshot: read .prizm-docs/ → read source files directly
 ```
 
 ### Workflow
 
 1. Receive task assignment
-2. Read `.prizmkit/specs/###-feature-name/context-snapshot.md` (if it exists) — Section 3 contains Prizm Context, Section 4 contains source files. If the snapshot does not exist:
-   a. Read `.prizmkit/specs/###-feature-name/agents/*.md` (other agents' knowledge docs to get existing findings)
-   b. Read `.prizm-docs/root.prizm` and relevant module documentation
-   c. Scan required source files and write summaries to the CONTEXT_BUILT section of your own `agents/dev-{N}.md`
+2. Read `.prizmkit/specs/###-feature-name/context-snapshot.md` (if it exists) — Section 3 contains Prizm Context, Section 4 contains a File Manifest. If the snapshot does not exist:
+   a. Read `.prizm-docs/root.prizm` and relevant module documentation
+   b. Read required source files directly
 3. Read `plan.md` (with Tasks section) and `spec.md` in `.prizmkit/specs/###-feature-name/`
 4. For each assigned task, execute in plan.md Tasks order:
    a. Get target file context and TRAPS from context-snapshot.md (if no snapshot, read the target file module's documentation)
    b. TDD: write tests based on acceptance criteria → implement feature code → run tests to verify
    c. Mark the task as `[x]` in the plan.md Tasks section
-   d. If new FINDINGS/DECISIONS/INTERFACES are discovered, append them to your own `agents/dev-{N}.md`
+   d. After all tasks, append Implementation Log to context-snapshot.md
    e. Send STATUS_UPDATE to the Orchestrator
 5. For checkpoint tasks, verify that build and tests pass before continuing
 6. On interface design ambiguity, send ESCALATION (do not assume)

package/bundled/agents/prizm-dev-team-reviewer.md

@@ -16,9 +16,7 @@ You are the team's "quality inspector + proofreader" — you do not produce the
 
 ### Project Context
 
-Project documentation is in `.prizm-docs/`. Before review, read `context-snapshot.md` (if it exists in `.prizmkit/specs/###-feature-name/`); its Section 3 contains Prizm Context (RULES, PATTERNS, TRAPS), eliminating the need to read `.prizm-docs/` or original source files. If the snapshot does not exist, read `root.prizm` to understand project rules (RULES), patterns (PATTERNS), and known traps (TRAPS); read module-level documentation as needed.
-
-During review, you may read `.prizmkit/specs/###-feature-name/agents/dev-*.md` to understand Dev's implementation decisions and findings.
+Project documentation is in `.prizm-docs/`. Before review, read `context-snapshot.md` (if it exists in `.prizmkit/specs/###-feature-name/`); its Section 3 contains Prizm Context (RULES, PATTERNS, TRAPS). Section 4 contains a File Manifest. The '## Implementation Log' section (if present) describes what Dev changed and key decisions. If the snapshot does not exist, read `root.prizm` to understand project rules.
 
 ### Artifact Paths
 
@@ -27,7 +25,6 @@ During review, you may read `.prizmkit/specs/###-feature-name/agents/dev-*.md` t
 | `.prizm-docs/` | Architecture index — module structure, interfaces, dependencies, known traps (TRAPS) |
 | `CLAUDE.md` / `CODEBUDDY.md` + `memory/MEMORY.md` | Project memory — development decisions (DECISIONS), interface conventions, project-level rules |
 | `.prizmkit/specs/###-feature-name/` | Feature artifacts — spec.md / plan.md (with Tasks section) |
-| `.prizmkit/specs/###-feature-name/agents/` | Agent knowledge docs — each agent's findings, decisions, interface records |
 
 ### Must Do (MUST)
 
@@ -40,8 +37,8 @@ During review, you may read `.prizmkit/specs/###-feature-name/agents/dev-*.md` t
 7. Check that code conforms to `.prizm-docs/` RULES and PATTERNS
 8. Review is a **read-only operation** (the review portions of Phase 4 and Phase 6 do not modify code files)
 9. Integration test cases must cover all user stories defined in spec.md
-10. Maintain your own knowledge doc `agents/reviewer.md`: append FINDINGS/DECISIONS after completing the analyze and review phases
-11. During review, read `agents/dev-*.md` to understand Dev's implementation decisions and trade-offs
+10. After completing review, append '## Review Notes' to context-snapshot.md: issues found (severity), test results, final verdict
+11. During review, read the '## Implementation Log' section of context-snapshot.md to understand Dev's changes and decisions
 
 ### Never Do (NEVER)
 
@@ -50,7 +47,6 @@ During review, you may read `.prizmkit/specs/###-feature-name/agents/dev-*.md` t
 - Do not perform task scheduling (that is the Orchestrator's responsibility)
 - **Do not execute any git operations** (git commit / git add / git reset / git push are all prohibited)
 - Do not use TaskCreate/TaskUpdate to create or modify Orchestrator-level tasks (Task tools are for internal progress tracking only, and task IDs are not shared across agent sub-sessions)
-- Do not modify other agents' knowledge docs (only write to your own `agents/reviewer.md`)
 
 ### Behavioral Rules
 
@@ -65,8 +61,8 @@ REV-07: Security findings are always HIGH or CRITICAL
 REV-08: Integration tests must cover all user stories in spec.md
 REV-09: Review code for conformance to .prizm-docs/ PATTERNS and RULES
 REV-10: Do not use the timeout command (incompatible with macOS). Run tests directly with node --test or npm test without a timeout prefix
-REV-11: Maintain agents/reviewer.md: append FINDINGS/DECISIONS after completing analyze and review phases
-REV-12: During review, read agents/dev-*.md to understand implementation decisions and trade-offs; reference relevant decisions in the review report
+REV-11: After review, append '## Review Notes' to context-snapshot.md (issues, severity, test results, verdict)
+REV-12: Read Implementation Log in context-snapshot.md to understand Dev's decisions; reference relevant decisions in the review report
 ```
 
 ### Phase 4 Workflow: Cross-Validation
@@ -86,7 +82,7 @@ REV-12: During review, read agents/dev-*.md to understand implementation decisio
 **Precondition**: Dev has completed implementation; all tasks are marked `[x]`
 
 1. Read `context-snapshot.md` (if it exists); its Section 3 contains RULES and PATTERNS. If the snapshot does not exist, read `.prizm-docs/root.prizm`
-2. Read `agents/dev-*.md` (if they exist) to understand Dev's implementation decisions and trade-offs
+2. Read '## Implementation Log' in context-snapshot.md to understand Dev's changes and decisions
 3. Run `/prizmkit-code-review` (read-only)
    - 6 review dimensions: spec compliance, plan adherence, code quality, security, consistency, test coverage
    - Verdict: PASS | PASS WITH WARNINGS | NEEDS FIXES
@@ -96,7 +92,7 @@ REV-12: During review, read agents/dev-*.md to understand implementation decisio
    - User story acceptance criteria (from spec.md)
    - Boundary conditions and exception paths
 4. Generate a unified review report
-5. Append FINDINGS/DECISIONS to `agents/reviewer.md` (record review findings and decisions)
+5. Append '## Review Notes' to context-snapshot.md (issues found, severity, test results, verdict)
 6. Send COMPLETION_SIGNAL (with verdict)
 
 ### Verdict Criteria

package/bundled/dev-pipeline/README.md

@@ -955,7 +955,6 @@ dev-pipeline/
 |   +-- bootstrap-tier3.md           # Tier 3 prompt template (full team, full/self-evolve)
 |   +-- bootstrap-prompt.md          # Legacy monolithic template (fallback)
 |   +-- bugfix-bootstrap-prompt.md   # Bug-fix session prompt template
-|   +-- agent-knowledge-template.md  # Template for agent knowledge docs
 |   +-- feature-list-schema.json     # JSON schema for feature-list.json
 |   +-- bug-fix-list-schema.json     # JSON schema for bug-fix-list.json
 |   +-- session-status-schema.json   # JSON schema for session-status.json

package/bundled/dev-pipeline/retry-bug.sh

@@ -207,6 +207,11 @@ echo -e "${BOLD}  Retry Bug Fix: $BUG_ID — $BUG_TITLE${NC}"
 echo -e "${BOLD}  Severity: $BUG_SEVERITY${NC}"
 echo -e "${BOLD}════════════════════════════════════════════════════${NC}"
 log_info "CLI: $CLI_CMD (platform: $PLATFORM)"
+if [[ -n "${MODEL:-}" ]]; then
+    log_info "Model: $MODEL"
+else
+    log_info "Model: (CLI default)"
+fi
 if [[ $SESSION_TIMEOUT -gt 0 ]]; then
     log_info "Session timeout: ${SESSION_TIMEOUT}s"
 else

package/bundled/dev-pipeline/retry-feature.sh

@@ -208,6 +208,12 @@ echo -e "${BOLD}═════════════════════
 echo -e "${BOLD}  Retry: $FEATURE_ID — $FEATURE_TITLE${NC}"
 echo -e "${BOLD}════════════════════════════════════════════════════${NC}"
 log_info "CLI: $CLI_CMD (platform: $PLATFORM)"
+EFFECTIVE_MODEL="${FEATURE_MODEL:-${MODEL:-}}"
+if [[ -n "$EFFECTIVE_MODEL" ]]; then
+    log_info "Model: $EFFECTIVE_MODEL"
+else
+    log_info "Model: (CLI default)"
+fi
 if [[ $SESSION_TIMEOUT -gt 0 ]]; then
     log_info "Session timeout: ${SESSION_TIMEOUT}s"
 else

package/bundled/dev-pipeline/run-bugfix.sh

@@ -515,6 +515,9 @@ main() {
         log_info "Session timeout: none"
     fi
     log_info "AI CLI: $CLI_CMD (platform: $PLATFORM)"
+    if [[ -n "${MODEL:-}" ]]; then
+        log_info "Default Model: $MODEL"
+    fi
     echo -e "${BOLD}════════════════════════════════════════════════════${NC}"
     echo ""
 

package/bundled/dev-pipeline/run.sh

@@ -525,14 +525,31 @@ run_one() {
         exit 1
     fi
 
-    if [[ ! -f "$STATE_DIR/pipeline.json" ]]; then
-        log_error "No pipeline state found. Run './run.sh run' first to initialize."
-        exit 1
-    fi
-
     check_dependencies
     run_log_cleanup
 
+    # Initialize pipeline state if needed (same logic as run_all)
+    if [[ ! -f "$STATE_DIR/pipeline.json" ]]; then
+        log_info "Initializing pipeline state for single-feature run..."
+        local init_result
+        init_result=$(python3 "$SCRIPTS_DIR/init-pipeline.py" \
+            --feature-list "$feature_list" \
+            --state-dir "$STATE_DIR" 2>&1)
+
+        local init_valid
+        init_valid=$(echo "$init_result" | python3 -c "import sys,json; print(json.load(sys.stdin).get('valid', False))" 2>/dev/null || echo "False")
+
+        if [[ "$init_valid" != "True" ]]; then
+            log_error "Pipeline initialization failed:"
+            echo "$init_result"
+            exit 1
+        fi
+
+        local features_count
+        features_count=$(echo "$init_result" | python3 -c "import sys,json; print(json.load(sys.stdin).get('features_count', 0))" 2>/dev/null || echo "0")
+        log_success "Pipeline initialized with $features_count features"
+    fi
+
     # Auto-detect framework repo: if scripts/bundle.js exists, enable self-evolve mode
     local project_root
     project_root=$(git -C "$SCRIPT_DIR" rev-parse --show-toplevel 2>/dev/null || echo "")
@@ -703,8 +720,15 @@ sys.exit(1)
     echo -e "${BOLD}════════════════════════════════════════════════════${NC}"
     echo -e "${BOLD}  Run: $feature_id — $feature_title${NC}"
     echo -e "${BOLD}════════════════════════════════════════════════════${NC}"
+    log_info "AI CLI: $CLI_CMD (platform: $PLATFORM)"
     log_info "Session ID: $session_id"
     log_info "Resume Phase: $resume_phase"
+    local effective_model="${feature_model:-$MODEL}"
+    if [[ -n "$effective_model" ]]; then
+        log_info "Model: $effective_model"
+    else
+        log_info "Model: (CLI default)"
+    fi
     if [[ -n "$mode_override" ]]; then
         log_info "Mode Override: $mode_override"
     fi
@@ -886,6 +910,9 @@ main() {
         log_info "Session timeout: none"
     fi
     log_info "AI CLI: $CLI_CMD (platform: $PLATFORM)"
+    if [[ -n "${MODEL:-}" ]]; then
+        log_info "Default Model: $MODEL"
+    fi
     echo -e "${BOLD}════════════════════════════════════════════════════${NC}"
     echo ""
 

package/bundled/dev-pipeline/scripts/generate-bootstrap-prompt.py

@@ -19,7 +19,7 @@ import os
 import re
 import sys
 
-from utils import load_json_file, setup_logging
+from utils import enrich_global_context, load_json_file, setup_logging
 
 
 DEFAULT_MAX_RETRIES = 3
@@ -142,8 +142,15 @@ def format_acceptance_criteria(criteria):
     return "\n".join(lines)
 
 
-def format_global_context(global_context):
-    """Format global_context dict as a key-value list."""
+def format_global_context(global_context, project_root=None):
+    """Format global_context dict as a key-value list.
+
+    If global_context is empty/sparse and project_root is provided,
+    auto-detect tech stack from project files to fill gaps.
+    """
+    if project_root:
+        enrich_global_context(global_context, project_root)
+
     if not global_context:
         return "- (none specified)"
     lines = []
@@ -362,14 +369,23 @@ def determine_pipeline_mode(complexity):
     """Map estimated_complexity to pipeline mode.
 
     Returns: 'lite', 'standard', or 'full'
+
+    Tier assignment rationale:
+    - low + medium → lite (single agent): most features don't benefit from
+      the orchestrator→dev→reviewer overhead. A single agent reading
+      .prizm-docs + implementing directly is faster and cheaper.
+    - high → standard (orchestrator + dev + reviewer): complex features
+      need the spec→plan→analyze→implement→review pipeline.
+    - critical → full (full team + framework guardrails): architectural
+      changes that touch many files and need extra safety checks.
     """
     mapping = {
         "low": "lite",
-        "medium": "standard",
-        "high": "full",
+        "medium": "lite",
+        "high": "standard",
         "critical": "full",
     }
-    return mapping.get(complexity, "standard")
+    return mapping.get(complexity, "lite")
 
 
 def build_replacements(args, feature, features, global_context, script_dir):
@@ -447,12 +463,6 @@ def build_replacements(args, feature, features, global_context, script_dir):
         args.feature_id, feature.get("title", "")
     )
 
-    # Ensure agents/ directory exists for agent knowledge docs
-    agents_dir = os.path.join(
-        project_root, ".prizmkit", "specs", feature_slug, "agents"
-    )
-    os.makedirs(agents_dir, exist_ok=True)
-
     # Detect project state
     init_done = detect_init_status(project_root)
     artifacts = detect_existing_artifacts(project_root, feature_slug)
@@ -487,7 +497,7 @@ def build_replacements(args, feature, features, global_context, script_dir):
         "{{COMPLETED_DEPENDENCIES}}": get_completed_dependencies(
             features, feature
         ),
-        "{{GLOBAL_CONTEXT}}": format_global_context(global_context),
+        "{{GLOBAL_CONTEXT}}": format_global_context(global_context, project_root),
         "{{TEAM_CONFIG_PATH}}": team_config_path,
         "{{DEV_SUBAGENT_PATH}}": dev_subagent,
         "{{REVIEWER_SUBAGENT_PATH}}": reviewer_subagent,

package/bundled/dev-pipeline/scripts/generate-bugfix-prompt.py

@@ -19,7 +19,7 @@ import os
 import re
 import sys
 
-from utils import load_json_file, setup_logging
+from utils import enrich_global_context, load_json_file, setup_logging
 
 
 DEFAULT_MAX_RETRIES = 3
@@ -76,8 +76,15 @@ def format_acceptance_criteria(criteria):
     return "\n".join(lines)
 
 
-def format_global_context(global_context):
-    """Format global_context dict as a key-value list."""
+def format_global_context(global_context, project_root=None):
+    """Format global_context dict as a key-value list.
+
+    If global_context is empty/sparse and project_root is provided,
+    auto-detect tech stack from project files to fill gaps.
+    """
+    if project_root:
+        enrich_global_context(global_context, project_root)
+
     if not global_context:
         return "- (none specified)"
     lines = []
@@ -249,7 +256,7 @@ def build_replacements(args, bug, global_context, script_dir):
         ),
         "{{AFFECTED_FEATURE}}": bug.get("affected_feature", "N/A"),
         "{{ENVIRONMENT}}": format_environment(bug.get("environment")),
-        "{{GLOBAL_CONTEXT}}": format_global_context(global_context),
+        "{{GLOBAL_CONTEXT}}": format_global_context(global_context, project_root),
         "{{TEAM_CONFIG_PATH}}": team_config_path,
         "{{DEV_SUBAGENT_PATH}}": dev_subagent,
         "{{REVIEWER_SUBAGENT_PATH}}": reviewer_subagent,

package/bundled/dev-pipeline/scripts/utils.py

@@ -105,3 +105,114 @@ def _build_progress_bar(percent, width=20):
     empty = width - filled
     bar = "\u2588" * filled + "\u2591" * empty
     return "{} {:>3}%".format(bar, int(percent))
+
+
+def detect_project_context(project_root):
+    """Auto-detect project tech stack from project files.
+
+    Reads package.json, pyproject.toml, and common config files
+    to infer language, test framework, and other stack details.
+    Returns a dict of detected key-value pairs.
+    """
+    detected = {}
+
+    # 1. Node.js / JavaScript / TypeScript project
+    pkg_path = os.path.join(project_root, "package.json")
+    if os.path.isfile(pkg_path):
+        try:
+            with open(pkg_path, "r", encoding="utf-8") as f:
+                pkg = json.load(f)
+
+            # Language detection
+            deps = {}
+            deps.update(pkg.get("dependencies", {}))
+            deps.update(pkg.get("devDependencies", {}))
+            if "typescript" in deps or os.path.isfile(
+                os.path.join(project_root, "tsconfig.json")
+            ):
+                detected["language"] = "TypeScript"
+            else:
+                detected["language"] = "JavaScript"
+
+            # Test framework detection (order: more specific first)
+            scripts = pkg.get("scripts", {})
+            test_script = (
+                scripts.get("test", "")
+                + " "
+                + scripts.get("test:unit", "")
+            )
+            if "vitest" in deps or "vitest" in test_script:
+                detected["testing_framework"] = "Vitest"
+            elif "jest" in deps or "jest" in test_script:
+                detected["testing_framework"] = "Jest"
+            elif "mocha" in deps or "mocha" in test_script:
+                detected["testing_framework"] = "Mocha"
+            elif "--test" in test_script or "node:test" in test_script:
+                detected["testing_framework"] = "Node.js built-in test runner"
+
+            # Framework detection
+            if "next" in deps:
+                detected["framework"] = "Next.js"
+            elif "express" in deps:
+                detected["framework"] = "Express.js"
+            elif "fastify" in deps:
+                detected["framework"] = "Fastify"
+            elif "react" in deps:
+                detected["framework"] = "React"
+            elif "vue" in deps:
+                detected["framework"] = "Vue.js"
+        except (json.JSONDecodeError, IOError):
+            pass
+
+    # 2. Python project detection
+    if not detected:
+        for marker in ["pyproject.toml", "setup.py", "requirements.txt"]:
+            if os.path.isfile(os.path.join(project_root, marker)):
+                detected["language"] = "Python"
+                # Check for pytest
+                toml_path = os.path.join(project_root, "pyproject.toml")
+                if os.path.isfile(toml_path):
+                    try:
+                        with open(toml_path, "r", encoding="utf-8") as f:
+                            content = f.read()
+                        if "pytest" in content:
+                            detected["testing_framework"] = "pytest"
+                    except IOError:
+                        pass
+                break
+
+    return detected
+
+
+def enrich_global_context(global_context, project_root):
+    """Fill gaps in global_context using auto-detected project info.
+
+    Only adds auto-detected values for keys not already present.
+    Mutates global_context in place and returns it.
+    """
+    if not project_root:
+        return global_context
+
+    detected = detect_project_context(project_root)
+    # Map detected keys → global_context convention names
+    key_mapping = {
+        "language": "language",
+        "testing_framework": "testing_strategy",
+        "framework": "framework",
+    }
+    # Alternate key names that should block auto-detection
+    alt_keys = {
+        "testing_strategy": ["testing_framework", "test_framework"],
+    }
+    for det_key, ctx_key in key_mapping.items():
+        if det_key not in detected:
+            continue
+        if ctx_key in global_context:
+            continue
+        already_set = any(
+            k in global_context for k in alt_keys.get(ctx_key, [])
+        )
+        if not already_set:
+            global_context[ctx_key] = detected[det_key] + " (auto-detected)"
+
+    return global_context

package/bundled/dev-pipeline/templates/bootstrap-tier1.md

@@ -56,28 +56,6 @@ You are running in headless mode with a FINITE context window. Exceeding it will
 
 ## Execution
 
-### Phase 0.5: Agent Knowledge Setup
-
-Create the agent knowledge directory and initialize your own knowledge doc:
-```bash
-mkdir -p .prizmkit/specs/{{FEATURE_SLUG}}/agents
-```
-
-Write `.prizmkit/specs/{{FEATURE_SLUG}}/agents/orchestrator.md`:
-```markdown
-# Orchestrator
-
-## FINDINGS
-
-## DECISIONS
-
-## INTERFACES_DISCOVERED
-
-## CONTEXT_BUILT
-```
-
-After each phase, append notable DECISIONS/FINDINGS to your `agents/orchestrator.md`.
-
 {{IF_INIT_NEEDED}}
 ### Phase 0: Project Bootstrap
 - Run `/prizmkit-init` (invoke the prizmkit-init skill)

package/bundled/dev-pipeline/templates/bootstrap-tier2.md

@@ -48,47 +48,25 @@ You are running in headless mode with a FINITE context window. Exceeding it will
 ## PrizmKit Directory Convention
 
 ```
-.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md  ← written by you, read by Dev + Reviewer
+.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md  ← orchestrator writes Sections 1-4; Dev appends Implementation Log; Reviewer appends Review Notes
 .prizmkit/specs/{{FEATURE_SLUG}}/plan.md              ← includes Tasks section
 ```
 
-**`context-snapshot.md`** is the shared knowledge base. You write it once; Dev and Reviewer read it instead of re-scanning individual files.
+**`context-snapshot.md`** is the shared knowledge base. Orchestrator writes Sections 1-4; Dev appends Implementation Log; Reviewer appends Review Notes. Append-only after initial creation.
 
 ---
 
 ## Subagent Timeout Recovery
 
-If a subagent times out:
+If any agent times out:
 1. `ls .prizmkit/specs/{{FEATURE_SLUG}}/` — check what exists
-2. Re-spawn with: `"Read .prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md for full context. Also read .prizmkit/specs/{{FEATURE_SLUG}}/agents/*.md for knowledge from previous agents. Do NOT re-read individual source files."` + only remaining steps + `model: "lite"`
-3. Max 2 retries. After 2 failures, complete the work yourself.
+2. If `context-snapshot.md` exists: open recovery prompt with `"Read .prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md for project context and any Implementation Log/Review Notes from previous agents. Run git diff HEAD to see actual code changes already made. Do NOT re-read individual source files unless the File Manifest directs you to."` + only remaining steps + `model: "lite"`
+3. Max 2 retries per phase. After 2 failures, orchestrator completes the work directly and appends a Recovery Note to context-snapshot.md.
 
 ---
 
 ## Execution
 
-### Phase 0.5: Agent Knowledge Setup
-
-Create the agent knowledge directory and initialize your own knowledge doc:
-```bash
-mkdir -p .prizmkit/specs/{{FEATURE_SLUG}}/agents
-```
-
-Write `.prizmkit/specs/{{FEATURE_SLUG}}/agents/orchestrator.md`:
-```markdown
-# Orchestrator
-
-## FINDINGS
-
-## DECISIONS
-
-## INTERFACES_DISCOVERED
-
-## CONTEXT_BUILT
-```
-
-After each phase, append notable DECISIONS/FINDINGS to your `agents/orchestrator.md`.
-
 {{IF_INIT_NEEDED}}
 ### Phase 0: Project Bootstrap
 - Run `/prizmkit-init` (invoke the prizmkit-init skill)
@@ -117,8 +95,19 @@ If MISSING — build it now:
    - **Section 1 — Feature Brief**: feature description + acceptance criteria (copy from above)
    - **Section 2 — Project Structure**: relevant `ls src/` output
    - **Section 3 — Prizm Context**: full content of root.prizm and relevant L1/L2 docs
-   - **Section 4 — Existing Source Files**: **full verbatim content** of each related file in fenced code blocks (with `### path/to/file` heading and line count). Include ALL files needed for implementation and review — downstream subagents read this section instead of re-reading individual source files
-   - **Section 5 — Existing Tests**: full content of related test files as code blocks
+   - **Section 4 — File Manifest**: For each file relevant to this feature, list: file path, why it's needed (modify/reference/test), key interface signatures (function names + params + return types). Do NOT include full file content — agents read files on-demand. Format:
+     ### Files to Modify
+     | File | Why Needed | Key Interfaces |
+     |------|-----------|----------------|
+     | `src/config.js` | Add runtime config layer | `config` (Zod object), `configSchema` |
+
+     ### Files for Reference
+     | File | Why Needed | Key Interfaces |
+     |------|-----------|----------------|
+     | `src/security/permission-guard.js` | Permission check integration | `checkCommandPermission(userId, cmd)` |
+
+     ### Known TRAPS (from .prizm-docs/)
+     - <trap entries extracted from L1/L2 docs>
 
 ### Phase 2: Plan & Tasks (you, the orchestrator)
 
@@ -147,44 +136,55 @@ This ensures the pipeline sees a "partial" status even if the session crashes mi
 Spawn Dev subagent (Agent tool, subagent_type="prizm-dev-team-dev", run_in_background=false).
 
 Prompt:
-> "Read {{DEV_SUBAGENT_PATH}}. Implement feature {{FEATURE_ID}} (slug: {{FEATURE_SLUG}}).
->
+> "Read {{DEV_SUBAGENT_PATH}}. Implement feature {{FEATURE_ID}} (slug: {{FEATURE_SLUG}}) using TDD.
 > **IMPORTANT**: Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` FIRST.
-> This file contains ALL source code and context. Do NOT re-read individual source files.
-> 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` — all project context, source files, and tests are embedded there.
-> 2. Read `.prizmkit/specs/{{FEATURE_SLUG}}/plan.md` (including Tasks section).
-> 3. Implement task-by-task using TDD. Mark each task `[x]` in plan.md Tasks section immediately after completion.
-> 4. **Agent Knowledge Doc**: Maintain `.prizmkit/specs/{{FEATURE_SLUG}}/agents/dev-1.md`. After each task, append FINDINGS/DECISIONS/INTERFACES_DISCOVERED if you discovered anything notable. If context-snapshot.md was MISSING, write CONTEXT_BUILT entries after scanning source files.
-> 5. After ALL tasks complete, append an 'Implementation Log' section to `context-snapshot.md`:
->    - Files created/modified (with paths)
->    - Key implementation decisions
->    - Any deviations from plan.md
-> Do NOT exit until all tasks are [x] and the Implementation Log is written."
+> 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` — Section 3 has Prizm Context (TRAPS/RULES), Section 4 has File Manifest with paths and interfaces. Read source files on-demand as directed by the manifest.
+> 2. Read `plan.md` (including Tasks section) from `.prizmkit/specs/{{FEATURE_SLUG}}/`.
+> 3. Implement task-by-task. Mark each `[x]` in plan.md Tasks section **immediately** after completion (do NOT batch).
+> 4. Use `TEST_CMD=<TEST_CMD>` to run tests — do NOT explore alternative test commands.
+> 5. After ALL tasks done, append '## Implementation Log' to context-snapshot.md with:
+>    - Files changed/created (with paths)
+>    - Key implementation decisions and rationale
+>    - Deviations from plan.md (if any)
+>    - Notable discoveries (unexpected behavior, hidden dependencies, new TRAPS)
+> 6. Do NOT execute any git commands (no git add/commit/reset/push).
+> 7. If `<TEST_CMD>` shows failures, check against BASELINE_FAILURES=`<BASELINE_FAILURES>`. Failures present in the baseline are pre-existing — list them explicitly in your COMPLETION_SIGNAL.
+> Do NOT exit until all tasks are [x] and the '## Implementation Log' section is written in context-snapshot.md."
 
 Wait for Dev to return. All tasks must be `[x]`, tests pass.
 
+**Gate Check — Implementation Log**:
+After Dev agent returns, verify the Implementation Log was written:
+```bash
+grep -q "## Implementation Log" .prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md && echo "GATE:PASS" || echo "GATE:MISSING"
+```
+If GATE:MISSING — send message to Dev (re-spawn if needed): "Write the '## Implementation Log' section to context-snapshot.md before I can proceed to review. Include: files changed/created, key decisions, deviations from plan, notable discoveries."
+
 ### Phase 4: Review — Reviewer Subagent
 
 Spawn Reviewer subagent (Agent tool, subagent_type="prizm-dev-team-reviewer", run_in_background=false).
 
 Prompt:
-> "Read {{REVIEWER_SUBAGENT_PATH}}. Review feature {{FEATURE_ID}} (slug: {{FEATURE_SLUG}}).
->
+> "Read {{REVIEWER_SUBAGENT_PATH}}. For feature {{FEATURE_ID}} (slug: {{FEATURE_SLUG}}):
 > **IMPORTANT**: Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` FIRST.
-> This file contains ALL source code and context. Do NOT re-read individual source files.
 > 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md`:
->    - Section 1: acceptance criteria to verify against
->    - Section 4: original source files (before changes)
->    - 'Implementation Log': what Dev changed
-> 2. Read `.prizmkit/specs/{{FEATURE_SLUG}}/agents/dev-*.md` (if exists) — understand Dev's implementation decisions and trade-offs.
-> 3. Run prizmkit-code-review: verify all acceptance criteria, check code quality and correctness. Only read files mentioned in the Implementation Log.
-> 4. Run the test suite and report results.
-> 5. Append a 'Review Notes' section to `context-snapshot.md`: issues found (severity), test results, final verdict.
-> 6. **Agent Knowledge Doc**: Maintain `.prizmkit/specs/{{FEATURE_SLUG}}/agents/reviewer.md`. Write FINDINGS/DECISIONS after review (e.g., patterns discovered, quality issues, architectural observations).
-> 7. If review uncovers durable pitfalls or conventions, add corresponding TRAPS/RULES notes to relevant `.prizm-docs/` files.
+>    - Section 3: Prizm Context (RULES, PATTERNS to check against)
+>    - Section 4: File Manifest (original file structure)
+>    - '## Implementation Log': what Dev changed, key decisions, discoveries
+> 2. Run prizmkit-code-review: spec compliance (against spec.md), code quality, correctness. Read ONLY files listed in Implementation Log.
+> 3. Write and execute integration tests covering all user stories from spec.md. Use `TEST_CMD=<TEST_CMD>` — do NOT try alternative test commands.
+> 4. Append '## Review Notes' to context-snapshot.md: issues found (with severity), test results, final verdict.
 > Report verdict: PASS, PASS_WITH_WARNINGS, or NEEDS_FIXES."
 
 Wait for Reviewer to return.
+
+**Gate Check — Review Notes**:
+After Reviewer agent returns, verify the Review Notes were written:
+```bash
+grep -q "## Review Notes" .prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md && echo "GATE:PASS" || echo "GATE:MISSING"
+```
+If GATE:MISSING — send message to Reviewer (re-spawn if needed): "Write the '## Review Notes' section to context-snapshot.md. Include: issues found (severity), test results, final verdict."
+
 - If NEEDS_FIXES: spawn Dev to fix (Dev reads updated snapshot), re-run Review (max 3 rounds)
 
 **CP-2**: Tests pass, verdict is not NEEDS_FIXES.
@@ -271,9 +271,11 @@ Re-check `git status --short` and ensure it is empty before exiting.
 
 - Tier 2: orchestrator builds context+plan, Dev implements, Reviewer reviews — use direct Agent spawn for agents
 - Build context-snapshot.md FIRST; all subagents read it instead of re-reading source files
+- context-snapshot.md is append-only: orchestrator writes Sections 1-4, Dev appends Implementation Log, Reviewer appends Review Notes
+- Gate checks enforce Implementation Log and Review Notes are written before proceeding
 - Do NOT use `run_in_background=true` when spawning subagents
 - Session-status.json is written BEFORE commit (as partial), then updated to success AFTER commit — this prevents pipeline from treating a terminated session as crashed
 - `/prizmkit-committer` is mandatory, and must not be replaced with manual git commit commands
 - Before exiting, `git status --short` must be empty
 - When staging leftover files in the final clean check, always use explicit file names — NEVER use `git add -A`
-- On timeout: check snapshot → model:lite → remaining steps only → max 2 retries → orchestrator fallback
+- On timeout: check snapshot + git diff HEAD → model:lite → remaining steps only → max 2 retries per phase → orchestrator fallback

package/bundled/dev-pipeline/templates/bootstrap-tier3.md

@@ -82,16 +82,12 @@ You are running in headless mode with a FINITE context window. Exceeding it will
 ## PrizmKit Directory Convention
 
 ```
-.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md  ← orchestrator writes, all subagents read
+.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md  ← orchestrator writes Sections 1-4; Dev appends Implementation Log; Reviewer appends Review Notes
 .prizmkit/specs/{{FEATURE_SLUG}}/spec.md
 .prizmkit/specs/{{FEATURE_SLUG}}/plan.md              ← includes Tasks section
 ```
 
-**`context-snapshot.md`** is the shared knowledge base. Orchestrator writes it once; Dev and Reviewer read it instead of re-scanning source files. This eliminates redundant I/O across all agents.
-
-### Agent Files
-- Dev Agent: `{{DEV_SUBAGENT_PATH}}`
-- Reviewer Agent: `{{REVIEWER_SUBAGENT_PATH}}`
+**`context-snapshot.md`** is the shared knowledge base. Orchestrator writes Sections 1-4; Dev appends Implementation Log; Reviewer appends Review Notes. This eliminates redundant I/O across all agents.
 
 ---
 
@@ -99,36 +95,13 @@ You are running in headless mode with a FINITE context window. Exceeding it will
 
 If any agent times out:
 1. `ls .prizmkit/specs/{{FEATURE_SLUG}}/` — check what exists
-2. If `context-snapshot.md` exists: open recovery prompt with `"Read .prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md for full context. Also read .prizmkit/specs/{{FEATURE_SLUG}}/agents/*.md for knowledge from previous agents. Do NOT re-read individual source files."` + only remaining steps + `model: "lite"`
+2. If `context-snapshot.md` exists: open recovery prompt with `"Read .prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md for project context and any Implementation Log/Review Notes from previous agents. Run git diff HEAD to see actual code changes already made. Do NOT re-read individual source files unless the File Manifest directs you to."` + only remaining steps + `model: "lite"`
 3. Max 2 retries per phase. After 2 failures, orchestrator completes the work directly and appends a Recovery Note to context-snapshot.md.
 
 ---
 
 ## Execution
 
-### Phase 0.5: Agent Knowledge Setup
-
-Create the agent knowledge directory and initialize your own knowledge doc:
-```bash
-mkdir -p .prizmkit/specs/{{FEATURE_SLUG}}/agents
-```
-
-Write `.prizmkit/specs/{{FEATURE_SLUG}}/agents/orchestrator.md`:
-```markdown
-# Orchestrator
-
-## FINDINGS
-
-## DECISIONS
-
-## INTERFACES_DISCOVERED
-
-## CONTEXT_BUILT
-```
-
-After each phase, append notable DECISIONS/FINDINGS to your `agents/orchestrator.md`.
-When spawning multiple Dev agents, each writes its own `agents/dev-{N}.md` (e.g., `dev-1.md`, `dev-2.md`).
-
 {{IF_INIT_NEEDED}}
 ### Phase 0: Project Bootstrap
 - Run `/prizmkit-init` (invoke the prizmkit-init skill)
@@ -200,11 +173,23 @@ If `EXISTING_CODE` is non-empty: your spec/plan/tasks must reflect this existing
    - **Section 1 — Feature Brief**: feature description + acceptance criteria (copy from above)
    - **Section 2 — Project Structure**: relevant `ls src/` output
    - **Section 3 — Prizm Context**: full content of root.prizm and relevant L1/L2 docs
-   - **Section 4 — Existing Source Files**: **full verbatim content** of each related file in fenced code blocks (with `### path/to/file` heading and line count). Include ALL files needed for implementation and review — downstream subagents read this section instead of re-reading individual source files
+   - **Section 4 — File Manifest**: For each file relevant to this feature, list: file path, why it's needed (modify/reference/test), key interface signatures (function names + params + return types). Do NOT include full file content — agents read files on-demand. Format:
+     ### Files to Modify
+     | File | Why Needed | Key Interfaces |
+     |------|-----------|----------------|
+     | `src/config.js` | Add runtime config layer | `config` (Zod object), `configSchema` |
+
+     ### Files for Reference
+     | File | Why Needed | Key Interfaces |
+     |------|-----------|----------------|
+     | `src/security/permission-guard.js` | Permission check integration | `checkCommandPermission(userId, cmd)` |
+
+     ### Known TRAPS (from .prizm-docs/)
+     - <trap entries extracted from L1/L2 docs>
    - **Section 5 — Existing Tests**: full content of related test files as code blocks
 4. Confirm: `ls .prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md`
 
-**After Step A, do NOT re-read any original source files** — use context-snapshot.md for all subsequent work.
+**After Step A**: Use context-snapshot.md Section 4 File Manifest to guide targeted file reads. Do NOT scan directories or read unrelated files.
 
 **Step B — Planning Artifacts** (generate only missing files):
 
@@ -225,10 +210,9 @@ Spawn Reviewer agent (Agent tool, subagent_type="prizm-dev-team-reviewer", run_i
 
 Prompt:
 > "Read {{REVIEWER_SUBAGENT_PATH}}. For feature {{FEATURE_ID}} (slug: {{FEATURE_SLUG}}):
-> 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` FIRST — all source files and project context are there. Do NOT re-read individual source files.
+> 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` FIRST — Section 3 has project context, Section 4 has file manifest.
 > 2. Run prizmkit-analyze: cross-check `spec.md` and `plan.md` (including Tasks section) for consistency.
-> 3. Before flagging CRITICAL or HIGH issues, verify each against Section 4 of the snapshot. Do NOT report based on incomplete information.
-> 4. **Agent Knowledge Doc**: Maintain `.prizmkit/specs/{{FEATURE_SLUG}}/agents/reviewer.md`. Write FINDINGS/DECISIONS after analysis (e.g., consistency issues found, ambiguities identified).
+> 3. Before flagging CRITICAL or HIGH issues, read the relevant source files listed in the File Manifest to verify.
 > Report: CRITICAL, HIGH, MEDIUM issues found (or 'No issues found')."
 
 Wait for Reviewer to return.
@@ -273,28 +257,36 @@ Append the following to the Dev agent prompt:
 Prompt:
 > "Read {{DEV_SUBAGENT_PATH}}. Implement feature {{FEATURE_ID}} (slug: {{FEATURE_SLUG}}) using TDD.
 > **IMPORTANT**: Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` FIRST.
-> This file contains ALL source code and context. Do NOT re-read individual source files.
-> 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` — all source files and context are there.
+> 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` — Section 3 has Prizm Context (TRAPS/RULES), Section 4 has File Manifest with paths and interfaces. Read source files on-demand as directed by the manifest.
 > 2. Read `plan.md` (including Tasks section) from `.prizmkit/specs/{{FEATURE_SLUG}}/`.
 > 3. Implement task-by-task. Mark each `[x]` in plan.md Tasks section **immediately** after completion (do NOT batch).
 > 4. Use `TEST_CMD=<TEST_CMD>` to run tests — do NOT explore alternative test commands.
-> 5. **Agent Knowledge Doc**: Maintain `.prizmkit/specs/{{FEATURE_SLUG}}/agents/dev-1.md`. After each task, append FINDINGS/DECISIONS/INTERFACES_DISCOVERED if you discovered anything notable. If context-snapshot.md was MISSING, read `agents/*.md` from other agents first, then scan source files and write CONTEXT_BUILT entries.
-> 6. After ALL tasks done, append 'Implementation Log' to context-snapshot.md: files changed/created, key decisions, deviations from plan.
-> 7. Do NOT execute any git commands (no git add/commit/reset/push).
-> 8. If `<TEST_CMD>` shows failures, check against BASELINE_FAILURES=`<BASELINE_FAILURES>`. Failures present in the baseline are pre-existing — list them explicitly in your COMPLETION_SIGNAL.
-> Do NOT exit until all tasks are [x] and the Implementation Log is written."
+> 5. After ALL tasks done, append '## Implementation Log' to context-snapshot.md with:
+>    - Files changed/created (with paths)
+>    - Key implementation decisions and rationale
+>    - Deviations from plan.md (if any)
+>    - Notable discoveries (unexpected behavior, hidden dependencies, new TRAPS)
+> 6. Do NOT execute any git commands (no git add/commit/reset/push).
+> 7. If `<TEST_CMD>` shows failures, check against BASELINE_FAILURES=`<BASELINE_FAILURES>`. Failures present in the baseline are pre-existing — list them explicitly in your COMPLETION_SIGNAL.
+> Do NOT exit until all tasks are [x] and the '## Implementation Log' section is written in context-snapshot.md."
+
+**Gate Check — Implementation Log**:
+After Dev agent returns, verify the Implementation Log was written:
+```bash
+grep -q "## Implementation Log" .prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md && echo "GATE:PASS" || echo "GATE:MISSING"
+```
+If GATE:MISSING — send message to Dev (re-spawn if needed): "Write the '## Implementation Log' section to context-snapshot.md before I can proceed to review. Include: files changed/created, key decisions, deviations from plan, notable discoveries."
 
 Wait for Dev to return. **If Dev times out before all tasks are `[x]`**:
 1. Check progress: `grep -c '^\- \[ \]' .prizmkit/specs/{{FEATURE_SLUG}}/plan.md`
 2. If any tasks remain: re-spawn Dev with this recovery prompt:
    > "Read {{DEV_SUBAGENT_PATH}}. You are resuming implementation of feature {{FEATURE_ID}} (slug: {{FEATURE_SLUG}}).
-   > 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` — Section 4 has original source, 'Implementation Log' (if present) has what was already done. Do NOT re-read individual source files.
-   > 2. Read `.prizmkit/specs/{{FEATURE_SLUG}}/agents/*.md` — knowledge from previous agents (context, decisions, interfaces).
+   > 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` — Section 4 has File Manifest, 'Implementation Log' (if present) shows what was already done.
+   > 2. Run `git diff HEAD` to see actual code changes already made.
    > 3. Read plan.md Tasks section — complete ONLY the remaining `[ ]` tasks. Do NOT redo completed `[x]` tasks.
    > 4. Use `TEST_CMD=<TEST_CMD>` to run tests.
-   > 5. Maintain `.prizmkit/specs/{{FEATURE_SLUG}}/agents/dev-1.md` — append FINDINGS/DECISIONS as you work.
-   > 6. Append progress to 'Implementation Log' in context-snapshot.md.
-   > 7. Do NOT execute any git commands."
+   > 5. Append progress to '## Implementation Log' in context-snapshot.md.
+   > 6. Do NOT execute any git commands."
 3. Max 2 recovery retries. After 2 failures, orchestrator implements remaining tasks directly.
 
 All tasks `[x]`, tests pass.
@@ -318,25 +310,31 @@ Append the following to the Reviewer agent prompt:
 Prompt:
 > "Read {{REVIEWER_SUBAGENT_PATH}}. For feature {{FEATURE_ID}} (slug: {{FEATURE_SLUG}}):
 > **IMPORTANT**: Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` FIRST.
-> This file contains ALL source code and context. Do NOT re-read individual source files.
-> 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` — Section 4 has original source files, 'Implementation Log' section lists exactly what Dev changed. Do NOT re-read source files that are NOT mentioned in the Implementation Log.
-> 2. Read `.prizmkit/specs/{{FEATURE_SLUG}}/agents/dev-*.md` (if exists) — understand Dev's implementation decisions, trade-offs, and discovered interfaces.
-> 3. Run prizmkit-code-review: spec compliance (against spec.md), code quality, correctness. Read ONLY files listed in Implementation Log.
-> 4. Write and execute integration tests covering all user stories from spec.md. Use `TEST_CMD=<TEST_CMD>` — do NOT try alternative test commands.
-> 5. Append 'Review Notes' to context-snapshot.md: issues (severity), test results, final verdict.
-> 6. **Agent Knowledge Doc**: Maintain `.prizmkit/specs/{{FEATURE_SLUG}}/agents/reviewer.md`. Write FINDINGS/DECISIONS after review (e.g., quality patterns, architectural observations).
+> 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md`:
+>    - Section 3: Prizm Context (RULES, PATTERNS to check against)
+>    - Section 4: File Manifest (original file structure)
+>    - '## Implementation Log': what Dev changed, key decisions, discoveries
+> 2. Run prizmkit-code-review: spec compliance (against spec.md), code quality, correctness. Read ONLY files listed in Implementation Log.
+> 3. Write and execute integration tests covering all user stories from spec.md. Use `TEST_CMD=<TEST_CMD>` — do NOT try alternative test commands.
+> 4. Append '## Review Notes' to context-snapshot.md: issues found (with severity), test results, final verdict.
 > Report verdict: PASS, PASS_WITH_WARNINGS, or NEEDS_FIXES."
 
 Wait for Reviewer to return.
+
+**Gate Check — Review Notes**:
+After Reviewer agent returns, verify the Review Notes were written:
+```bash
+grep -q "## Review Notes" .prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md && echo "GATE:PASS" || echo "GATE:MISSING"
+```
+If GATE:MISSING — send message to Reviewer (re-spawn if needed): "Write the '## Review Notes' section to context-snapshot.md. Include: issues found (severity), test results, final verdict."
+
 - If NEEDS_FIXES: spawn Dev to fix with this prompt:
   > "Read {{DEV_SUBAGENT_PATH}}. Fix NEEDS_FIXES issues for feature {{FEATURE_ID}} (slug: {{FEATURE_SLUG}}).
-  > 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` — 'Review Notes' section lists the exact issues to fix. Do NOT re-read source files not mentioned there.
-  > 2. Read `.prizmkit/specs/{{FEATURE_SLUG}}/agents/reviewer.md` (if exists) — understand reviewer's findings and rationale.
-  > 3. Fix ONLY the issues listed in 'Review Notes'. Do NOT refactor unrelated code.
-  > 4. Use `TEST_CMD=<TEST_CMD>` to verify fixes.
-  > 5. Append fix summary to 'Implementation Log' in context-snapshot.md.
-  > 6. Update `.prizmkit/specs/{{FEATURE_SLUG}}/agents/dev-1.md` with any new FINDINGS/DECISIONS from the fix.
-  > 7. Do NOT execute any git commands."
+  > 1. Read `.prizmkit/specs/{{FEATURE_SLUG}}/context-snapshot.md` — '## Review Notes' section lists the exact issues to fix.
+  > 2. Fix ONLY the issues listed in 'Review Notes'. Do NOT refactor unrelated code.
+  > 3. Use `TEST_CMD=<TEST_CMD>` to verify fixes.
+  > 4. Append fix summary to '## Implementation Log' in context-snapshot.md.
+  > 5. Do NOT execute any git commands."
   Then re-run Review (max 3 rounds).
 
 **CP-3**: Integration tests pass, verdict is not NEEDS_FIXES.
@@ -465,7 +463,8 @@ Re-check `git status --short` and ensure it is empty before exiting.
 ## Reminders
 
 - Tier 3: full team — Dev (implementation) → Reviewer (review) — spawn agents directly via Agent tool
-- context-snapshot.md is the team knowledge base: orchestrator writes it once, all agents read it
+- context-snapshot.md is append-only: orchestrator writes Sections 1-4, Dev appends Implementation Log, Reviewer appends Review Notes
+- Gate checks enforce Implementation Log and Review Notes are written before proceeding
 - Do NOT use `run_in_background=true` when spawning agents
 - ALWAYS write preliminary session-status.json BEFORE commit (as partial), then update to success AFTER commit — this prevents pipeline from treating a terminated session as crashed
 - Commit phase must use `/prizmkit-committer`; do NOT replace with manual git commit commands

package/bundled/dev-pipeline/tests/test_generate_bootstrap_prompt.py

@@ -126,16 +126,16 @@ class TestDeterminePipelineMode:
         assert determine_pipeline_mode("low") == "lite"
 
     def test_medium(self):
-        assert determine_pipeline_mode("medium") == "standard"
+        assert determine_pipeline_mode("medium") == "lite"
 
     def test_high(self):
-        assert determine_pipeline_mode("high") == "full"
+        assert determine_pipeline_mode("high") == "standard"
 
     def test_critical(self):
         assert determine_pipeline_mode("critical") == "full"
 
     def test_unknown(self):
-        assert determine_pipeline_mode("banana") == "standard"
+        assert determine_pipeline_mode("banana") == "lite"
 
 
 # ---------------------------------------------------------------------------

package/bundled/skills/_metadata.json

@@ -1,5 +1,5 @@
 {
-  "version": "1.0.78",
+  "version": "1.0.84",
   "skills": {
     "prizm-kit": {
       "description": "Full-lifecycle dev toolkit. Covers spec-driven development, Prizm context docs, code quality, debugging, deployment, and knowledge management.",

package/bundled/skills/app-planner/SKILL.md

@@ -49,6 +49,7 @@ Before questions, check optional context files (never block if absent):
 - `.prizm-docs/root.prizm` (architecture/project context)
 - `.prizmkit/config.json` (existing stack preferences)
 - existing `feature-list.json` (required for incremental mode)
+- `.prizmkit/available-models.json` (detected AI models, from detect-models.sh)
 
 Note:
 - This skill **reads** `.prizmkit/config.json` if present.
@@ -83,6 +84,12 @@ Execute the selected scenario workflow in conversation mode with mandatory check
 ### Interactive Phases
 1. clarify business goal and scope
 2. confirm constraints and tech assumptions
+2b. (optional) if `.prizmkit/available-models.json` exists and `model_switch_supported == true`:
+    - read the file and show available models to user
+    - ask: "要为 feature 指定 AI 模型吗？" with options:
+      a) 全部使用某个模型（写入每个 feature 的 model 字段）
+      b) 不指定（使用 $MODEL 环境变量或 CLI 默认）
+    - if user chooses a model, remember it for Phase 6 JSON generation
 3. propose feature set with dependencies
 4. refine descriptions and acceptance criteria
 5. verify DAG/order/priorities
@@ -151,6 +158,12 @@ AI: "Ready to proceed to dev-pipeline."
 - valid dependency DAG
 - new items default `status: "pending"`
 - English feature titles for stable slug generation
+- if `.prizmkit/available-models.json` exists and `model_switch_supported == true`:
+  - if user specified model preference during planning:
+    - set `model` field on every feature with user's chosen model
+  - if user said "no preference" → omit `model` field
+- model IDs should come from `.prizmkit/available-models.json` models list
+- `model` field is optional — omitting it means the pipeline uses $MODEL env or CLI default
 
 ## Next-Step Execution Policy (after planning)
 

package/bundled/skills/prizmkit-code-review/SKILL.md

@@ -27,7 +27,7 @@ Perform a comprehensive code review against the feature spec, implementation pla
    - If none found → prompt user: "No review baseline found. Which workflow are you in? (feature/refactor/bugfix)"
 2. Read **architecture index**: `.prizm-docs/root.prizm` RULES and PATTERNS for project conventions
 3. Read **project memory**: platform memory file (`CLAUDE.md` for Claude Code, `CODEBUDDY.md` + `memory/MEMORY.md` for CodeBuddy) for past DECISIONS — helps verify implementation respects established conventions
-4. Read `agents/dev-*.md` in the feature directory (if exists) — understand Dev's implementation decisions, trade-offs, and discovered interfaces. This context helps distinguish intentional design choices from accidental patterns during review.
+4. Read '## Implementation Log' section of context-snapshot.md in the feature directory (if exists) — understand Dev's implementation decisions, trade-offs, and notable discoveries. This context helps distinguish intentional design choices from accidental patterns during review.
 5. Scan all code files referenced in completed tasks
 4. Review across 6 dimensions:
    - **Spec compliance**: Does code implement all acceptance criteria? Missing criteria are the #1 source of "it works but it's wrong" bugs

package/bundled/skills/prizmkit-implement/SKILL.md

@@ -28,20 +28,18 @@ Execute implementation by following the task breakdown in plan.md. Respects task
      - **Refactor mode**: read `refactor-analysis.md` in the same directory — pay special attention to Scope Boundary (do not implement changes outside scope) and Baseline Tests (all must still pass)
      - **Bugfix mode**: read `fix-plan.md` in the same directory
 2. Load project context — use the most efficient source available:
-   - If `context-snapshot.md` exists in the feature directory → read it. Section 3 has Prizm docs + TRAPS, Section 4 has source files. This snapshot was built in Phase 1 of the pipeline to avoid re-reading dozens of individual files, saving significant tokens.
+   - If `context-snapshot.md` exists in the feature directory → read it. Section 3 has Prizm docs + TRAPS. Section 4 has File Manifest (Tier-2/3) or full source (Tier-1). Read source files on-demand as directed by the manifest.
    - Otherwise → **self-service context fallback**:
-     1. Read `agents/*.md` in the feature directory (if any exist) — other agents may have already built context
-     2. Read **architecture index**: `.prizm-docs/root.prizm` and relevant L1/L2 for affected modules. Pay special attention to TRAPS (gotchas, race conditions).
-     3. Read **project memory**: platform memory file (`CLAUDE.md` for Claude Code, `CODEBUDDY.md` + `memory/MEMORY.md` for CodeBuddy) for past DECISIONS and interface conventions you should respect.
-     4. Scan needed source files
-     5. Write CONTEXT_BUILT entries to your own `agents/<role>.md` (e.g., `agents/dev-1.md`) so future agents don't need to re-scan
+     1. Read **architecture index**: `.prizm-docs/root.prizm` and relevant L1/L2 for affected modules. Pay special attention to TRAPS.
+     2. Read **project memory**: platform memory file for past DECISIONS and interface conventions.
+     3. Scan needed source files
 3. Check if checkpoint tasks are complete before proceeding to next phase
 4. For each unchecked task in order:
    a. If task has `[P]` marker, it can run in parallel with other `[P]` tasks in the same group
    b. Read L2 doc for target file's module (if exists) — TRAPS save you from repeating known mistakes
    c. Apply TDD where applicable: write a failing test first, then implement until it passes. For UI components or configuration changes where unit tests don't apply, skip the test-first step.
    d. Mark task as `[x]` in `plan.md` Tasks section immediately — not batched at the end. Immediate marking means the plan always reflects true progress, even if the session is interrupted.
-   e. If notable FINDINGS, DECISIONS, or INTERFACES were discovered during this task, append them to your `agents/<role>.md` knowledge doc (if agent knowledge docs are in use for this session).
+   e. After all tasks, append '## Implementation Log' to `context-snapshot.md` (if running in pipeline context): files changed/created, key decisions, notable discoveries.
    f. Error handling: sequential tasks stop on failure (later tasks may depend on this one). Parallel `[P]` tasks continue — report all failures at the end.
 5. At each checkpoint: verify build passes and tests pass. Checkpoints catch integration errors early — skipping them means cascading failures in later phases that are much harder to debug.
 6. After all tasks: run full test suite

package/bundled/skills/prizmkit-plan/SKILL.md

@@ -32,7 +32,7 @@ Generate a comprehensive technical implementation plan from a feature specificat
    - **Refactor mode**: Read `refactor-analysis.md` for refactoring goals, scope boundary, and baseline tests
    - **Bugfix mode**: Read bug description / `fix-plan.md` for reproduction steps and expected fix
 2. Load project context (use first available source):
-   - If `.prizmkit/specs/###-feature-name/context-snapshot.md` exists → read it for all context (Section 3 'Prizm Context' for docs, Section 4 'Existing Source Files' for code). The context-snapshot consolidates dozens of individual files into one read, saving significant tokens.
+   - If `.prizmkit/specs/###-feature-name/context-snapshot.md` exists → read it for all context (Section 3 'Prizm Context' for docs, Section 4 'File Manifest' for code structure and interfaces). The context-snapshot consolidates project context into one read, saving significant tokens.
    - Otherwise → read `.prizm-docs/root.prizm` and relevant `.prizm-docs/` L1/L2 docs for affected modules
 3. Resolve any remaining `[NEEDS CLARIFICATION]` by proposing solutions
 4. Research technical approach based on project's tech stack

package/bundled/skills/prizmkit-retrospective/SKILL.md

@@ -99,7 +99,7 @@ Extract TRAPS and module-level RULES from development work and inject into `.pri
 **2a.** Gather context — read the **actual code that was changed** plus any available artifacts:
 
 - `git diff HEAD` — the real source of truth for what happened
-- `.prizmkit/specs/###-feature-name/agents/*.md` — **preferred source** for pre-categorized FINDINGS and INTERFACES_DISCOVERED from all agents. If agent docs exist, prefer them over re-extracting from git diff.
+- `.prizmkit/specs/###-feature-name/context-snapshot.md` — read the '## Implementation Log' section (Dev's changes, decisions, discoveries) and '## Review Notes' section (Reviewer's findings). These are the **preferred source** for pre-categorized decisions and findings. If these sections exist, prefer them over re-extracting from git diff.
 - `.prizmkit/specs/###-feature-name/plan.md` — if feature work, read planned vs actual
 - `.prizmkit/bugfix/<id>/fix-report.md` — if bugfix, read what was discovered
 - The relevant `.prizm-docs/` L1/L2 docs for affected modules
@@ -151,7 +151,8 @@ Sediment DECISIONS and interface conventions to platform memory files. This is w
 - `codebuddy` → targets: BOTH `CODEBUDDY.md` in project root AND `memory/MEMORY.md` (dual-write required)
 
 **2b-2.** Collect sedimentation candidates:
-- From `agents/*.md`: DECISIONS and INTERFACES_DISCOVERED entries
+- From context-snapshot.md '## Implementation Log': DECISIONS and notable discoveries
+- From context-snapshot.md '## Review Notes': quality patterns, architectural observations
 - From git diff analysis: any project-level conventions established
 - Filter: only entries that answer "Would a new session benefit from knowing this decision/convention?"
 
@@ -168,7 +169,7 @@ Sediment DECISIONS and interface conventions to platform memory files. This is w
 
 **2b-5.** For CodeBuddy platform: write identical content to BOTH `CODEBUDDY.md` AND `memory/MEMORY.md` (dual-write, both must be updated).
 
-**2b-6.** If no `agents/` directory exists in the feature directory, still attempt to extract DECISIONS from git diff and plan.md. Skip only if no meaningful decisions were made.
+**2b-6.** If no Implementation Log or Review Notes sections exist in context-snapshot.md, still attempt to extract DECISIONS from git diff and plan.md. Skip only if no meaningful decisions were made.
 
 ---
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prizmkit",
-  "version": "1.0.78",
+  "version": "1.0.84",
   "description": "Create a new PrizmKit-powered project with clean initialization — no framework dev files, just what you need.",
   "type": "module",
   "bin": {

package/bundled/dev-pipeline/templates/agent-knowledge-template.md

@@ -1,13 +0,0 @@
-# {{AGENT_ROLE}}
-
-## FINDINGS
-- <discovery>: <why it matters>
-
-## DECISIONS
-- <decision>: <rationale>
-
-## INTERFACES_DISCOVERED
-- <module.function>: <signature> — <behavior note>
-
-## CONTEXT_BUILT
-- <file-path>: <one-line summary>