Research agents (repo-research-analyst, git-history-analyzer,
best-practices-researcher, framework-docs-researcher) were using
shell commands like find, rg, cat, and chained pipelines for routine
codebase exploration. This triggers permission prompts in Claude Code
and degrades the user experience when these agents run as sub-agents.
Updated all research agents with platform-agnostic tool selection
guidance that prefers native file-search/glob, content-search/grep,
and file-read tools over shell equivalents. Shell is now reserved for
commands with no native equivalent (ast-grep, bundle show, git).
Git-history-analyzer additionally limits shell to one simple git
command per call with no chaining or piping.
Added tool selection rules to AGENTS.md so future agents follow
the same pattern by default.
The Phase 0 vibes-based heuristic warned users and forced a mode choice
even with 80%+ context remaining. Full mode now runs by default;
compact-safe mode stays available on explicit request.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The model was offering "commit to current branch" on main instead
of "create a branch and PR." Added explicit branch detection step
and "Do NOT commit directly to main" instruction.
Let the agent generate a reasonable branch name based on context
and repo conventions instead of prescribing a date-based format
that would collide on multiple runs per day.
Handles committing changes at the end of a refresh run so doc
maintenance doesn't sit uncommitted. Detects git context and adapts:
autonomous mode uses sensible defaults (branch + PR on main, separate
commit on feature branches), interactive mode presents options. Always
selectively stages only compound-refresh files to avoid mixing with
in-progress feature work.
The file-tools-over-bash instruction was in the orchestrator's
context but not passed to spawned subagents. Changed to an explicit
quoted instruction block that must be included in each subagent's
task prompt so it's visible to the subagent, not just the orchestrator.
Auto-archive now requires both the implementation AND the problem
domain to be gone. If referenced files are deleted but the application
still deals with the same problem (auth, payments, migrations), the
learning should be Replace'd not Archive'd — the knowledge gap needs
to be filled. Uses agent reasoning about concepts, not mechanical
keyword searches.
The model was generating findings internally then outputting a
one-line summary. Added explicit instructions that the full report
must be printed as text output — every file, every classification,
every action. In autonomous mode, the report is the sole deliverable
and must be self-contained and complete.
- Restructure Phase 3 with explicit autonomous skip section that says
"do not ask, do not present, do not wait" before any interactive
instructions
- Add autonomous caveats to Core Rules 4, 7, 8 which previously had
unconditional "ask the user" language
- Clarify that missing referenced files is unambiguous Archive evidence,
not a doubt case requiring user input
Instead of requiring write permissions, autonomous mode attempts
writes and gracefully falls back to recommendations when denied.
Report splits into Applied (succeeded) and Recommended (could not
write) sections. Read-only invocations produce a maintenance plan.
Support mode:autonomous argument for unattended/scheduled runs.
In autonomous mode: skip all user questions, apply safe actions
directly, mark ambiguous cases as stale with conservative confidence,
and generate a detailed report for after-the-fact human review.
Documents five skill design patterns discovered during testing:
platform-agnostic tool references, auto-archive consistency,
smart triage for broad scope, replacement subagents over
ce:compound handoff, and file tools over shell commands.
- Broad scope triage: inventory + impact clustering + spot-check drift
for 9+ docs, recommends highest-impact area instead of blind ask
- Drift classification: sharp boundary between Update (fix references
in-skill) and Replace (subagent writes successor learning)
- Replacement subagents: sequential subagents write new learnings using
ce:compound's document format with investigation evidence already
gathered, avoiding redundant research
- Stale fallback: when evidence is insufficient for a confident
replacement, mark as stale and recommend ce:compound later
- Use platform-agnostic interactive question tool phrasing with examples
for Claude Code and Codex instead of hardcoding AskUserQuestion
- Fix contradiction between Phase 2 auto-archive criteria and Phase 3
always-ask-before-archive rule so unambiguous archives proceed without
unnecessary user prompts
Adds a new skill that reviews existing docs/solutions/ learnings against the
current codebase and decides whether to keep, update, replace, or archive them.
Also enhances ce:compound with Phase 2.5 selective refresh checks.
Co-Authored-By: Claude <noreply@anthropic.com>
Name specific blocking question tools (AskUserQuestion, request_user_input,
ask_user) so agents actually invoke them instead of printing questions as
text output. Updates skill compliance checklist to match.
Add a highest-leverage-move question to the product pressure test,
a challenger option in approach exploration, and a low-cost change
check to the finalization checklist.
Prevents future agent resolution failures by requiring skills to use
compound-engineering:<category>:<agent-name> instead of short names.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Explains how to test local plugin changes across Claude Code (shell
alias with --plugin-dir) and Codex (local path install) without
disrupting production installs.
Rebased onto main after #251 restructured commands into skills/.
Applied the same namespace fix to the new skill file locations
(ce-brainstorm, ce-plan, ce-review SKILL.md files).
Short agent names like `repo-research-analyst` are replaced with
fully-qualified `compound-engineering:research:repo-research-analyst`
to ensure correct agent resolution across all plugin targets.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
