- Upgrade menu now shows status buffer instead of running binary upgrade directly
- Distinct behavior for E vs a: E always prompts for workspace, a reuses existing session
- Clearer menu descriptions for ECA commands
- Removed y-or-n-p confirmation from package upgrade (user already clicked upgrade button)
- Added backend check to ai-code-eca--add-menu-group
- Fixed tests to handle batch mode limitations

- Behavior injection into gptel-plan/gptel-agent buffers via transform
- Project-scoped behavior state management
- Auto-classify with preset suggestion and confidence threshold
- Completion for #behavior and @preset in gptel buffers
- C-c P to show last processed prompt
- Only process latest user message (not entire buffer)
- Pending preset system for preset-only prompts
- Sync checking with upstream ai-behaviors repo

Priority order for gptel-agent:
1. hashtags → auto-classify (with confidence) → pending → session → none

Priority order for regular:
1. hashtags → pending → session → auto-classify → none

- ai-code--behaviors-show-last-prompt now gets project root from
  the source buffer (via gptel--fsm-last FSM) in gptel-agent buffers
- Shows project root in the output buffer for debugging
- Added defvar for gptel--fsm-last to silence byte-compiler

- Add ai-code--extract-clean-user-prompt to strip behavior blocks
- Extract content within <user-prompt> tags if present
- Strip AdditionalContext blocks from classification input
- Add tests for clean prompt extraction
- Classification now ignores injected behavior instructions

- Add ai-code--behaviors-extract-project-from-buffer-name fallback
- Try to extract project from buffer name (e.g., *gptel-agent:.emacs.d*)
- Show all available project roots if current one has no data
- Use maphash instead of hash-table-keys for compatibility

- Try multiple candidate roots in priority order
- If only one stored entry exists, use it automatically
- Show available roots when no match found
- Fixed variable reference (found-root vs project-root)

- ai-code--fontify-behavior-keyword now normalizes matched strings
- #review now matches =review in ai-code--behavior-operating-modes
- Fixes: #=code highlighted but #review was not

- Remove redundant member check in face form
- Font-lock function already validates match with = prefix normalization
- Fixes: #review now highlights correctly after completion

- mode-line only shows in gptel-mode or ai-code-prompt-mode buffers
- Added project root fallback from gptel-agent buffer name
- Removed global mode-line-enable calls from enable-auto-presets
- Mode-line enabled via hooks instead of top-level calls
- Cleaner UX: no noise in unrelated buffers

- Add ai-code--behaviors-wrap-with-instruction helper
- Add ai-code--behaviors-meets-confidence-threshold-p for confidence check
- Add ai-code--behaviors-apply-and-format to reduce duplication
- Refactor ai-code--process-behaviors to use helpers
- Refactor ai-code--gptel-agent-process-behaviors to use helpers
- Fixes double-calling of classify-prompt-intent in gptel-agent
- Reduces code duplication by ~80 lines

- Fix void-variable 'it' bug in ai-code--gptel-agent-process-behaviors
  by extracting classification to let* bindings
- Fix mode-line to always show [○] when no behaviors active
- Update test to expect [○] instead of nil after clear

Features:
- Define readonly modes (=review, =research, =spec, etc.) and modify modes (=code, =debug)
- Auto-switch to agent mode when modify mode/preset used in gptel-plan
- Syntax-aware completion: #= prefix for modes, # prefix for modifiers
- Filter completion based on gptel-plan context
- Show notification when switching modes

Technical changes:
- Add ai-code--behavior-readonly-modes and ai-code--behavior-modify-modes constants
- Add ai-code--behaviors-mode-readonly-p and ai-code--behaviors-preset-readonly-p helpers
- Modify ai-code--extract-and-remove-hashtags to accept context-preset and return switch flag
- Modify ai-code--gptel-agent-process-behaviors to handle mode switching
- Rewrite ai-code--behavior-hashtag-capf for syntax-aware completion
- Update tests for new return format

The previous code was expanding the project name relative to default-directory,
causing doubled paths like /Users/davidwu/.emacs.d/.emacs.d

For gptel-agent buffers, default-directory is already correctly set by gptel-agent
when the buffer is created, so we can just return it directly.

The transform function had signature (callback fsm) but gptel calls
transforms with just (fsm). Changed to:
- Take only one argument (fsm)
- Return t if buffer was modified, nil otherwise
- Remove callback calls (gptel transforms don't use callbacks)

The transform runs in a prompt-copy buffer where the gptel text property
covers the entire buffer including AI responses. The old extraction logic
used prop-match-end which equals point-max, returning empty string.

Fix: Search for the user prompt marker ### instead of relying on
gptel text properties. This works in both gptel chat buffers and
prompt-copy buffers where properties are duplicated.

The regex ^### \(.*\) only captured the first line because . does
not match newlines in Emacs regex. Multi-line prompts would be
truncated.

Fix: Use buffer-substring-no-properties from the ### marker to
point-max instead of match-string. This correctly captures the
entire prompt including multi-line content.

Previously we registered behavior presets (deep-review, tdd-dev, etc.)
with gptel--known-presets with :system "". This caused gptel's
gptel--transform-apply-preset to remove @preset from prompts BEFORE
our transform could see it.

Fix: Don't register with gptel. Our transform handles @preset and
we provide completion via ai-code--behavior-preset-gptel-capf.

If prompt-text already contains <user-prompt> tags, extract the
content before wrapping to avoid nested/duplicate tags.

Also add original-prompt variable to store prompt before processing
(preparation for showing original with @preset in C-c P).

The transform was being called twice because:
1. Our transform was in the default-value of gptel-prompt-transform-functions
2. We added 't' to the local list telling gptel to merge with default
3. After merge, the transform appeared twice in the effective list

Fix: Don't add 't' to local list if our transform is already in the default value.
This prevents gptel from merging and duplicating the transform.

Also remove debug messages added for diagnosis.

…ation

- Add 20 constraint bundles for common tech stacks (@rust-stack, @react-stack, etc.)
- Expand constraint catalog to 50+ constraints (testing, safety, errors, performance, API, logging)
- Auto-detect constraints from project config files (tsconfig.json, pyproject.toml, Cargo.toml)
- Add project-scoped persistence via .ai-behaviors/constraints file
- Update mode-line to show bundle name and filter presets by gptel-plan mode
- Add completion for @bundle-name in prompts
- Fix glob pattern matching for *.csproj patterns in auto-detect
- Fix bundle loading from persistence file
- Fix mode-line update on bundle apply/clear
- Add 2 new tests for glob expansion

- Preserve session state (custom-suffix) in constraint functions
  - ai-code-constraints-clear: use plist-put to preserve other state keys
  - ai-code-constraints-apply-bundle: use plist-put to preserve other state keys
  - ai-code-constraints-auto-detect-and-apply: use plist-put to preserve other state keys

- Fix confidence threshold when GPTel classifier returns no confidence
  - Default to 'high when :confidence key is missing from classification result

- Fix path matching for directory patterns like .github/workflows
  - Pass project-root to ai-code--constraints-detect-from-file
  - Match against both basename and relative path

- Fix bundle side effect without project root context
  - Return bundle name as 4th element from ai-code--extract-and-remove-hashtags
  - Callers now set bundle using correct project root