Changelog

ed3d-research-agents 1.1.0

Stable repo caching for the remote-code-researcher agent.

Changed:

Remote code researcher now caches cloned repos at a stable path ($TMPDIR/claude-code-repos/<host>/<org>/<repo>) instead of using mktemp each time
Subsequent invocations against the same repo fetch-and-reset instead of re-cloning
Removed cleanup step — cached repos persist across agent invocations

ed3d-plan-and-execute 1.11.0

Remove SessionStart hook that injected the entire using-plan-and-execute skill into context on every session start.

Fixed:

Removed SessionStart hook that wasted context tokens by injecting the full SKILL.md content as additionalContext on every conversation

ed3d-session-reflection 0.2.0

New:

/export-session-as-markdown skill for exporting sessions as GitHub-flavored Markdown
reduce-transcript.py now supports --markdown flag for full Markdown export with metadata header, collapsible tool results, and thinking blocks
Conversation reviewer now detects agent-discovered multi-step workflows as automation candidates (not just user-directed ones)
Conversation reviewer suggests build system improvements using the project's existing tooling

ed3d-session-reflection 0.1.0

EXPERIMENTAL. Session awareness and conversation review tooling.

New:

SessionStart hook that injects session ID and transcript path into Claude's context
reduce-transcript.py script that strips JSONL transcripts to token-efficient text (78-99% reduction)
conversation-reviewer Opus agent that analyzes sessions for prompting effectiveness, agent performance, and environment gaps
/review-session skill for single-session deep review
/review-recent-sessions skill for parallel multi-session review with cross-session synthesis

ed3d-hook-security-hardening 1.0.1

Expanded detection coverage to address common Claude-generated access patterns.

New:

Detect polyglot env readers (python3, node, ruby, perl, awk reading environment variables)
Detect declare -p on secret variables
Detect curl -d @.env and curl -F file=@.env file exfiltration
Detect while read loops on secret files
Detect git remote set-url and git config with embedded tokens
Detect file reading via sed, awk, strings, base64, xxd, od, dd, tee, perl
Detect grep '' .env (empty/wildcard pattern reads entire file)
197 tests for bash secrets hook, 37 tests for sensitive file hook

Fixed:

Echo check now catches all secret variables in multi-variable commands
Dot-source regex correctly catches . .env syntax
Handle non-string command input without crashing

ed3d-hook-security-hardening 1.0.0

Hook plugin that catches common secrets leakage patterns in Claude Code sessions.

New:

PreToolUse hook on Bash: detects echoing secret env vars, printenv, env|grep without -q, cat/source on .env files, grep on shell configs showing values, tokens in git clone URLs, tokens in curl URL parameters
PostToolUse hook on Write/Edit: reminds about gitignore and chmod 600 after writing to sensitive files (.env, .envrc, credentials, secrets, .pem, .key, .netrc, .npmrc)
Uses shlex tokenization for robust command parsing over fragile regex
High-confidence leaks (echo/printenv/length/substring) are denied; medium-confidence patterns (cat .env, source .env, env|grep) force user approval

ed3d-extending-claude 1.1.1

Add prompt-security-hardening skill and require it from writing-claude-directives.

New:

prompt-security-hardening skill: 7 rules covering secrets leakage into LLM context, env var existence checks, file permissions, gitignore verification, URL/process token exposure, input sanitization, and context contamination from files

Changed:

writing-claude-directives: now unconditionally requires prompt-security-hardening as a subskill

ed3d-house-style 1.0.3

Relax FCIS file classification to target only files with runtime behavior.

Changed:

FCIS skill: classification now mandatory only for files containing runtime logic (functions, classes with methods, orchestration)
FCIS skill: added exemptions for type-only files, constants, barrels, tests, and generated files
FCIS skill: added threshold note — exempt files that grow to include runtime logic must be classified
TypeScript skill: added clarifying note in FCIS Integration section about exemptions

ed3d-plan-and-execute 1.10.3

Add session isolation for parallel planning/execution to prevent file collisions.

New:

SCRATCHPAD_DIR parameter with unique session ID (e.g., /tmp/plan-2025-01-24-feature-a7f3b2/) ensures isolation when multiple planning or execution sessions run in parallel
Random session ID component prevents collisions on retry attempts

Changed:

writing-implementation-plans: creates and passes SCRATCHPAD_DIR to code-reviewer in Finalization step
executing-an-implementation-plan: creates SCRATCHPAD_DIR at startup, passes to all code review invocations
requesting-code-review: accepts and forwards SCRATCHPAD_DIR to code-reviewer subagent
code-reviewer agent: documents SCRATCHPAD_DIR usage for any scratch files

ed3d-extending-claude 1.1.0

Adds marketplace management skill for creating and maintaining Claude Code Plugin Marketplaces.

New:

maintaining-a-marketplace skill covering marketplace.json schema, version management, release checklists, changelog conventions, validation, and distribution — generalizable for any user publishing a marketplace

ed3d-plan-and-execute 1.10.2

Fix typo in planning handoff command.

Fixed:

starting-a-design-plan: Phase 6 handoff command had /ed3d-ed3d-plan-and-execute:start-implementation-plan instead of /ed3d-plan-and-execute:start-implementation-plan

ed3d-extending-claude 1.0.4

Add model-level testing guidance to testing-skills-with-subagents.

Changed:

testing-skills-with-subagents: RED phase should use production-level model (default: Sonnet); GREEN/REFACTOR phases should use one tier down (default: Haiku) to ensure skill clarity under weaker reasoning
Quick reference table now includes model column

ed3d-basic-agents 1.1.0

Add fan-out analysis skill for large corpus processing.

New:

doing-a-simple-two-stage-fanout skill: orchestrates parallel Worker subagents, Critic review subagents, and a Summarizer for analyzing corpora that exceed a single agent's context window
compute_layout.py helper script for computing segment counts, agent assignments, and context window budgets
diagram-templates.md reference with Mermaid and Graphviz templates for visualizing fan-out pipelines
First user-invocable: true skill in this plugin

[ed3d-hook-skill-reinforcement] 1.0.1, [ed3d-plan-and-execute] 1.10.1, [ed3d-basic-agents] 1.0.2, [ed3d-extending-claude] 1.0.3

Remove stale <available_skills> XML tag references that no longer match how Claude Code injects skill lists.

Fixed:

Replaced all references to <available_skills> with format-agnostic language ("your available skills shown in your system context") across hooks, skills, and agent prompts
Hook reminder now uses emphatic "MUST" / "Do NOT skip" phrasing for stronger compliance
Added warning comment in CLAUDE_MD_TESTING.md example to prevent re-introducing the stale tag

ed3d-plan-and-execute 1.10.0

Scoped acceptance criteria for cross-plan uniqueness.

New:

AC identifiers now use scoped format {slug}.AC{N}.{M} (e.g., oauth2-svc-authn.AC1.1) to prevent collisions across multiple plan-and-execute rounds
Design plan naming now prompts user explicitly via AskUserQuestion — supports ticket names (e.g., PROJ-1234) or descriptive slugs
Slug naming guidance: prefer terse unambiguous names (authn not authentication, but not auth since ambiguous with authz)

Changed:

starting-a-design-plan: Added Step 1 to get design plan name before file creation
starting-an-implementation-plan: Slug definition now documents its three uses (directory, worktree, AC scope)
writing-design-plans: AC structure uses scoped format with slug prefix
writing-implementation-plans: Task templates and AC coverage sections use scoped format
executing-an-implementation-plan: AC coverage check references scoped format
All examples updated to use terse slugs (e.g., oauth2-svc-authn instead of oauth2-service-auth)