v4.9.0 — hybrid 切割 deprecation 期

Minor: hybrid 切割 deprecation 期开启。与姊妹项目 abyss CLI v0.5.24 同期落地。

Changes

• Deprecated (v5.0 移除)：--with-abyss / --with-mcp / --with-hooks 对 claude/codex/gemini
• Changed (net-new)：--with-hooks 对 openclaw/pi/hermes 自动 spawn install-hooks.sh（abyss CLI 设计上不接管这三平台，永久保留）
• 文档定位漂修复：README/site/i18n.js/package.json 双源对齐 + SKILL.md L103 字面冲突修复

Migration

# Old (v4.8.x)
npx code-abyss -t claude --with-abyss --with-mcp --with-hooks

# New (v4.9+)
npx code-abyss -t claude -y                                                            # persona/skills/style
curl -fsSL https://raw.githubusercontent.com/telagod/abyss/main/install.sh | bash      # abyss CLI
abyss attach claude                                                                    # hook 注入（idempotent）

openclaw/pi/hermes 用户路径不变：npx code-abyss -t openclaw --with-hooks 自动 spawn install-hooks.sh。

Compatibility

零回归：424 现有 test + 30 verify:skills 全过。所有现有 flag 行为保留，仅加 deprecation warning。

完整 changelog 见 CHANGELOG.md。

v4.8.2 — hooks opt-in & statusline compact

Changed

abyss hooks are now opt-in (default off)

Default install no longer auto-injects SessionStart + PreToolUse/BeforeTool hooks into Claude / Codex / Gemini settings. The indexing-code skill remains available — the model can invoke it passively based on its description. The previous behavior was noisy: hooks fired on every Edit/Write, spawned node for JSON normalization, and re-ran abyss even when the user just wanted the model to think.

# Restore previous behavior
npx code-abyss --target claude -y --with-hooks

Re-installing without --with-hooks strips prior marker-tagged entries; user-authored hooks are untouched (marker-based detection).

Bundled ccstatusline preset compacted

Dropped ↑input / ↓output / ⊡cached token fields (already encoded in session-cost), context-length absolute count (covered by percentage), and session-clock. New default:

Opus 4.7 ⋮ 14.0% ⋮ $8.14 ⋮ code-abyss

Separator switched from | to U+22EE (⋮) for tighter visual density.

Full diff

v4.8.1...v4.8.2

v4.8.1 — Codex install lifecycle and hooks

Fixed

• Codex install now writes current full_auto.config.toml and full_access.config.toml profile files for Codex 0.134+ instead of relying on legacy inline [profiles.*] tables.
• Existing Codex project trust entries are preserved during install.
• Valid Codex user hook arrays can now coexist with code-abyss hooks; legacy flat hook tables are skipped to avoid schema conflicts.
• Codex pre-edit hooks now normalize apply_patch payloads so abyss pre-edit checks run for patch edits.

Verification

• npm test -- --runInBand
• npm run verify:skills
• npm run packs:check
• npm run packs:vendor:sync -- --check
• node skills/analyzing-changes/scripts/change_analyzer.js --mode working
• node skills/verifying-modules/scripts/module_scanner.js .
• node skills/checking-code-quality/scripts/quality_checker.js .
• node skills/analyzing-security/scripts/security_scanner.js . --exclude creating-presentations processing-docx analyzing-spreadsheets processing-pdfs site
• npm pack --dry-run

v4.7.2 — Codex hooks TOML schema fix

Patch release.

Fixed

• Codex hooks now use the array-of-tables schema (#49). Codex 0.125+ deprecated the flat [hooks.X] form; installs produced a config.toml that failed to load (invalid type: map, expected a sequence in hooks), breaking codex startup. The installer now emits [[hooks.X]] + [[hooks.X.hooks]], with clean uninstall, idempotent re-runs, stale-path re-anchoring, and non-destructive skip of user-owned hooks preserved.
• Windows command_windows override. On Windows the installer locates Git Bash and emits command_windows, so hook scripts run even when Git Bash is not on PATH.

Full diff: v4.7.1...v4.7.2

v4.7.1 — the documented -t flag now works

Patch release. Every doc surface used npx code-abyss -t claude -y, but -t was never implemented in the parser — it silently fell into the interactive picker (and broke under -y on non-TTY). This fixes it.

Fixed

• -t / -u short flags — -t is now an alias for --target, -u for --uninstall, both shown in --help. The documented one-shot npx code-abyss -t claude -y finally works as written.

Changed (docs/site, since 4.7.0)

• Quickstart leads with --with-abyss (downloads the abyss code-graph binary so pre-edit hooks work out of the box), documents --with-mcp, clarifies what plain -y installs, and adds a verify step.
• Site: --with-abyss / --with-mcp enablement tip; code-graph section shows the SCIP benchmark as precision pills (Go 99.3% / TS 98.8% / Python 98.7% / Rust 98.5%).

Pairs with abyss v0.3.3. 409 tests.

v4.6.1

Changes since v4.6.0

Pi Agent + Hermes support

• hooks/pi/hooks.json — tool_call + session_start hook config
• hooks/hermes/hooks.yaml — shell hook config snippet
• hooks/hermes/plugin.py — native Python plugin with pre_tool_call gate
• install-hooks.sh auto now detects Pi (~/.pi) and Hermes (~/.hermes)

Docs

• Full multilingual update: README.zh-CN.md rewritten for v4.6
• Site: Code Intelligence section with 4 capability cards (en + zh)
• Navigation: added Graph tab

Platform coverage: 6/6

Platform	Hook event
Claude Code	PreToolUse
Codex CLI	PreToolUse
Gemini CLI	BeforeTool
Pi Agent	tool_call
Hermes	pre_tool_call
OpenClaw	before_tool_call

v4.6.0 — Code Graph Intelligence

What's new

Code Graph Intelligence — the abyss CLI builds a code relationship graph (call graph + temporal analysis + hotspot map) in 5 seconds. Your agent now sees code relationships before editing.

New capabilities

• abyss context <file> — full context for any file: all functions, callers, dependencies, hotspot score, coupled files. No function name needed.
• abyss impact <symbol> — blast radius analysis: direct/transitive callers, affected tests, uncovered paths, risk score (0-10)
• abyss callers <symbol> — who calls this function, with confidence scoring
• abyss map — hotspot detection (churn × complexity) + change coupling matrix
• abyss history <file> — evolution trace with git-powered commit history + coupled files
• --json flag on all commands for AI-consumable structured output

Cross-platform hooks

Pre-edit hooks auto-check callers before the agent modifies code. Supported on all 4 platforms:

Platform	Hook event
Claude Code	PreToolUse (Edit|Write)
Codex CLI	PreToolUse (Bash|shell)
Gemini CLI	BeforeTool (write_file|replace)
OpenClaw	before_tool_call plugin

One-command install: bash skills/indexing-code/hooks/common/install-hooks.sh auto

Skill integration

• shipping-changes: impact analysis gate before PR submission
• checking-code-quality: hotspot data for risk assessment
• Claude adapter auto-injects hooks when abyss CLI is detected

Performance

Benchmarked on 1862-file Go project:

• 5s full index (symbols + call graph + git history + hotspot analysis)
• 373K reference edges extracted, 130K resolved (36%)
• 1869 git commits analyzed, 14.7K coupling pairs computed
• All queries < 200ms

Install abyss CLI

cd code-abyss-dev && bash install.sh

---

30 skills · 383 tests · 4 platforms · MIT License

v4.5.0 — Dynamic Persona Loading

What's new

Dynamic persona loading

Only abyss (the flagship persona) ships with the npm package. All other personas are fetched from GitHub on first use and cached locally at `~/.code-abyss/personas/`.

# Core — works offline
npx code-abyss -t claude --persona abyss -y

# Remote — auto-fetched, then cached
npx code-abyss -t claude --persona scholar -y

Architecture

• index.json gains core/remote model with snapshot metadata for offline selection
• persona-fetch.js — HTTPS fetch + cache module for remote personas
• Cache-aware path resolution — checks local repo first (dev mode), then ~/.code-abyss/personas/ (npm mode)
• package.json files narrowed — non-core persona content excluded from npm package

Stats

• 29 skills, 384 tests, CI green across Node 18/20/22 × 4 targets × 3 OS

Full Changelog: v4.4.0...v4.5.0

v4.4.0 — Hardware + Academic Writing + Injection Defense

What's new

3 new domain skills (26 → 29)

• designing-hardware-products — full-stack hardware product pipeline: requirements → ESP-IDF firmware → KiCad PCB → UniApp client → release zip. 5 reference docs.
• operating-kicad-eda — KiCad 9 MCP tool router (17 tools). Iron rules: autoroute-only, library-first, serialize PCB writes, DRC gate. 3 reference docs.
• reducing-aigc-detection — systematic AIGC detection reduction for 维普/知网/Turnitin. Three-layer rewriting (structure → lexicon → content injection), docx run-level editing with footnote preservation. 2 reference docs.

New shared behavior modules

• injection-awareness — three-layer prompt injection defense (detect → whitelist → respond) with per-target runtime whitelists for Claude/Codex/Gemini/OpenClaw
• execution-drive — action-first rules, anti-laziness patterns, output self-check gate (终结门), safety valve for destructive ops
• llm-security.md — Injection-Resilient Persona Design reference (architecture, principles, adversarial evaluation, multi-agent scenarios)

Stats

• 29 skills, 379 tests, CI green on Node 18/20/22 × 4 targets × 3 OS

Full Changelog: v4.3.0...v4.4.0

v4.3.0 — Adversarial-Review Orchestration

Merge pull request #39 from telagod/release/v4.3.0

chore(release): v4.3.0 — adversarial-review orchestration + skill count 26