v0.4.0

First public release.

v0.3.1

Changelog

Excise v0.3.0 — opt-in local-Ollama LLM rerank

[0.3.0] — 2026-05-23

Added

• Opt-in LLM rerank via local Ollama (--llm). Layers a local model on
  top of the v0.2 heuristic shortlist: heuristic stays the cheap pre-filter,
  LLM only judges that shortlist and writes a one-line reason per turn.
• New flags --llm, --llm-model, --llm-host on both suggest and pick.
• New optional config file excise.toml (./excise.toml → $XDG_CONFIG_HOME/excise/
  → ~/.config/excise/) with [llm] block: host, model, top_n, timeout_sec.
• internal/llm/ package (tiny Ollama HTTP client + rerank logic + hard-coded
  prompt template) and internal/config/ package (TOML loader with sane defaults).
• New testdata claude_session_llm_rerank.jsonl + cmd-level tests for the
  happy path, fallback path, and JSON-parse-failure path against a stubbed
  Reranker. Live-Ollama integration test gated behind EXCISE_LIVE_OLLAMA=1.

Changed

• suggest and pick route the heuristic shortlist through the LLM reranker
  when --llm is set. Exit codes unchanged: 0 on success (incl. graceful
  fallback), 2 only on usage error.
• TUI shows per-turn LLMReason next to pre-marked turns when present.

Fixed

• Test helper NewStubReranker was originally placed in _test.go (invisible
  across packages); moved to internal/llm/stub.go so cmd-level tests resolve.
• go.sum regenerated to match real BurntSushi/toml@v1.4.0 checksum.

Trust contract (unchanged from v0.1/v0.2)

• No outbound network beyond the user-configured Ollama host.
• No autocut. --llm only changes ranking; user still confirms every excision.
• No telemetry, no acceptance log, no cross-session learning.

Still out of scope

• Remote LLM API key backends (OpenAI / Anthropic / OpenRouter) — deferred to v0.4.
• Streaming LLM output / partial-rerank UI.
• LLM + autocut combination.
• User-editable prompt template.

Excise v0.2.0 — heuristic suggestions

v0.2.0 adds a pure-stdlib heuristic suggestion engine that pre-computes the top-K candidate turns to excise from a polluted session. Zero new dependencies, zero network, no LLM. Same single binary, same trust premise.

New: excise suggest

$ excise suggest ~/.claude/projects/.../session.jsonl

 #   role        tokens   heuristic                                        preview
---  ---------   ------   ----------------------------------------------   --------------------
 17  assistant     2840   high_cost + user_correction_follows_up           "Let me try refactoring …"
 19  tool_use       420   tool_use_error + correction_follows              Edit(path=foo, …) ERROR
 32  assistant     3100   3rd file-edit on same file                       "Actually let me revert …"
 33  assistant     1820   user_correction_follows_up                       "I'll switch to using …"
 47  assistant     2200   long_drift_no_tool_calls                         "To summarize what we …"

5 candidates totalling ~10,380 tokens. Run `excise pick` to review interactively.

Output formats: table (default), --json for piping, --top=N for K, --min-score=X to filter.

Five heuristics (pure stdlib)

Trigger	Rule
high_token_cost	turn ≥ 2000 estimated tokens
repeated_file_edit	same file path edited 3+ times in a 3-turn window
user_correction_follows_up	next user turn matches correction lexicon (no / wrong / actually / let me try / forget that / instead / 不对 / 重来 / 换个思路 / …)
tool_use_error_then_correction	tool_use returned error AND user replied with correction lexicon
long_drift_no_tool_calls	5+ consecutive assistant turns with no tool_use

Bilingual correction lexicon (en + zh-CN) lives at internal/suggest/correction_lexicon.go.

TUI pre-marking

excise pick (default) now seeds the bubbletea picker with the top-K suggested turns highlighted. --no-suggest restores v0.1 behavior. Suggested turns render in a distinct color with a footer legend (◆ suggested — press space to uncheck).

What didn't change

• v0.1's session loading (Claude Code JSONL + Cursor SQLite), dependency-aware writer, snapshot+rollback — all untouched.
• Zero network. No LLM call of any kind (local Ollama / user API key — both deferred to v0.3 as opt-in plugins).
• No auto-cut. Scorer suggests; user confirms.
• No cross-session learning / acceptance log.

Tests

34 PASS / 0 FAIL across 5 packages (v0.1's 13 + 21 new for the suggest package). New test asserts the scorer's top-5 on testdata/claude_session_polluted.jsonl matches the planted failure indices.

Trust premise unchanged

Excise never talks to anything other than your local filesystem.

🤖 Auto-shipped from ai-radar pipeline scan-2026-05-18-1216 amendment v0.2.0.

Excise v0.1.0 — surgical context editing

First runnable release of Excise — a per-turn surgical editor for coding-agent session context.

The problem

Anthropic's own docs admit it: "If you've corrected Claude more than twice on the same issue in one session, the context is cluttered with failed approaches." Claude Code / Codex / Cursor give you /clear (nuke everything) and /compact (summarize everything) — but nothing in between. When a 4-hour session goes off the rails on a bad approach, your only escape is start over.

Excise gives you the scalpel that was missing.

What it does

• excise list <session> — show every turn in a Claude Code or Cursor session with role + token estimate + content preview.
• excise pick — interactive bubbletea TUI; mark turns for removal with vim keys, hit enter to commit.
• excise cut 5-7,9 — non-interactive surgical removal by turn range.
• excise rollback — every edit auto-snapshots before/after; restore in one command. Snapshots prune after 30 days.

Dependency-aware: when you excise an assistant turn that issued tool_use, Excise also removes the matching tool_result so the next turn doesn't reference a phantom call. Two invariants checked on every write: closure (no orphans) and round-trip (replaying the edited session produces the same sha as the writer output).

Supported sessions

• Claude Code (m1) — ~/.claude/projects/<encoded-cwd>/<uuid>.jsonl. Schema verified against a real session on the maintainer's machine; tests use fixtures matching the actual shape.
• Cursor (m2) — ~/Library/Application Support/Cursor/User/globalStorage/state.vscdb (cursorDiskKV table, bubbleId:<composer>:<bubble> keys). Read-only against Cursor's live db; emits a side-car .excised.jsonl so an open Cursor session is never corrupted.
• Rollback (m3) — snapshots at ~/.excise/snapshots/<session-id>/<iso>.jsonl.gz.

Stack

• Go 1.24, single binary, zero network.
• bubbletea + lipgloss TUI; cobra CLI; pure stdlib for jsonl; sqlite3 CLI for Cursor (no CGO).

Tests

14 PASS / 0 FAIL across 3 packages. Covers: dependency closure invariant, orphaned-result warning, round-trip after excise, snapshot+rollback, TUI state machine + token math, Cursor auto-detect, Claude loader against the real-machine sample.

What's next

v0.2 will tighten the Cursor write path (currently emits side-car JSONL; goal is in-place vscdb update behind a --unsafe-inplace flag). Codex / Aider support is explicitly out of scope until v0.1 ships traction signal — Excise stays focused.

🤖 Auto-shipped from ai-radar pipeline scan-2026-05-18-1216 (winner: need_t3c7gg07, surgical context editing).