codex-cissor — edit your Codex session context

cissor edits the context of Codex CLI sessions: trim bloated tool outputs, summarize stale turns in place, evict spans to files with recovery pointers, transplant or inject content — then resume the leaner session with codex resume <id>. It also inspects what is eating the window, exports markdown archives, and answers questions about oversized material without loading it into your context.

Edits happen on fork drafts; publishing (gated on offline validation) is how a finished fork returns to the canonical sessions tree.

Sessions live in ~/.codex/sessions/YYYY/MM/DD/rollout-<ts>-<uuid>.jsonl. cissor parses that format directly, including compacted records.

Install

uv tool install codex-cissor        # from PyPI (or: pipx install codex-cissor)

Requires OPENROUTER_API_KEY in the environment for LLM-backed commands (summarize / index / export chronicle / ask / replace / evict --with-summary). Inspection, forking, editing, validation, and transcript export work offline. Token estimates use tiktoken (o200k_base by default).

Config

cissor config path                 # where config/manifest/logs live (~/.codex/cxcm/)
cissor config show
cissor config set model z-ai/glm-5.3-flash          # summarizer
cissor config set ask_max_input_tokens 120000        # above this, ask() switches to map-reduce
cissor config set encoding o200k_base

Any OpenRouter model id works; env overrides via CISSOR_MODEL, CISSOR_ASK_MAX_INPUT_TOKENS, etc.

• ask_max_input_tokens (120k): the switch-point where ask stops sending the whole material in one call and switches to question-keyed map-reduce.
• context_window_tokens (0 = unknown): the model's context limit, purely advisory — inspect reports % of window. Codex defaults to 272k (expandable to 1M); auto-compaction triggers around the ~682k hard max.

Compaction-aware scope

A compacted record contains {message, replacement_history}. On resume, Codex replays only the last compaction's replacement_history + everything after it. cissor therefore:

• numbers turns/items over the effective context (post-last-compaction) by default,
• warns when you touch pre-compaction items (they are not in the live window),
• supports --scope full to address the entire file including pre-compaction history, and --scope post-compact|auto|full everywhere selection applies,
• edit uncompact removes the compacted record to restore the full pre-compaction history (all of it is still in the file),
• items inside replacement_history are addressable and editable like any other item,
• history lists every compaction point with what it hides.

Commands

Inspection

cissor list [--limit N] [--json]                    # sessions, newest first (fork names shown)
cissor inspect <session> [--scope full] [--json]    # per-turn + per-type token breakdown
cissor inspect <session> --turns 3-7                # breakdown of a selection
cissor show <session> --turns 3-4 [--max-chars N]   # actual content of a selection
cissor history <session> [--json]                   # compaction points: line, summary, hidden tokens
cissor validate <session> [--json]                  # structural check (pairs, reasoning, meta)
cissor forks [--channel X] [--json]                 # registered forks: channels, lineage, op counts
cissor log [--session X] [--channel C] [--project P] [--stats] [--path]  # activity trail

Address a session by: full/prefix id, manifest name, latest, or file path.

Two kinds of placeholders: <session> = any of the above, for read-only commands. <fork-id> = the draft id returned by cissor fork — required by every mutating command (edit, transplant, inject, extract --evict); latest is refused there and non-fork originals are always refused.

Selection (shared by inspect/show/summarize/index/export/extract/ask/edit ops)

--scope auto|post-compact|full
--turns "2-10"          turn ranges/lists
--items "12,15-20"      item numbers
--from-text "needle"    from first item containing text to end (inclusive)
--to-text "needle"      from start to last item containing text (inclusive)
--turn-from-text / --turn-to-text
--all                   everything in scope

--from-text + --to-text together bound a single inclusive span (both anchor items included). One-sided use opens a range to the start/end of scope.

Summarize (read-only)

cissor summarize <session> --turns 1-40 --style index
cissor summarize <session> --turns 2-10 --style key-points --out s.md
cissor summarize <session> --turns 2-10 --prompt "focus only on pricing decisions"

Styles are editable markdown files (see Styles below): chronicle, index, short, key-points, export-chronicle, plus your own — or pass a one-off --prompt/--prompt-file inline. Long selections are map-reduced in ~48k-token chunks (chunk_tokens).

Event index

cissor index <session>                                  # generate + save to cissor state
cissor index <session> --locate "pricing"               # search existing index, else generate
cissor index <session> --out D:\indexes\voice.md        # custom location

Produces Items A-B (turn T): title — description entries for fast navigation; --locate answers "where was X discussed" in item/turn ranges.

Export (archival)

cissor export <session>                                  # markdown transcript to stdout
cissor export <session> --turns 3-9 --out archive.md
cissor export <session> --strategy chronicle             # LLM chronicle segment
cissor export <session> --strategy chronicle --chronicle-dir docs\chronicles
cissor export <session> --strategy index-only
cissor export <session> --no-outputs --include-reasoning --max-item-chars 2000

Transcripts use **User** (item 4): / **Assistant** / **Tool call: name** labels under ## Turn N headings. Developer/system messages and synthetic user context (<environment_context> etc.) are excluded by default — pass --include-developer / --include-synthetic to keep them (the header records what was excluded).

--chronicle-dir maintains the conversation-chronicle skill layout: numbered segments (NNNN-YYYY-MM-DD-HHMM-title.md) + master INDEX.md with event index.

Extract / evict to file

cissor extract <session> --turns 5-11 --out span.md              # read-only dump (md|json|raw)
cissor extract <fork> --items 17-56 --out span.md --evict        # + pointer message in fork
cissor extract <fork> --from-text A --to-text B --evict --with-summary --out span.md

--evict removes the span from the fork and leaves a <context_evicted> user message with the archive path (plus an LLM summary with --with-summary). Text anchors are inclusive (both anchor items are part of the span); evictions covering >60% of scope or >150k tokens are refused without --force (over-broad anchor protection). --format raw writes re-importable JSONL lines.

Ask — query oversized material with another LLM (stateful)

cissor ask D:\data\huge.json "What are the top-level keys and their types?"
cissor ask <session> --scope full "What did we decide about X early on?"
cissor ask <ask-id> "follow-up question"              # continues the session
cissor asks list / cissor asks show <ask-id>

The target (any text file, or a session slice) is sent WHOLE to the ask model — your agent's context never holds it. Above ask_max_input_tokens it falls back to question-keyed map-reduce: the material is split into ~48k-token chunks, and each chunk is filtered through the QUESTION ("extract every fact relevant to this question, with exact values and anchors") — that's the map phase, keyed to the question rather than generic summarization. A final reduce call merges the extracts into a grounded answer citing which fragments support each claim. Note: each new question on an oversized target re-scans the chunks. Answers cite item/turn anchors when the material provides them. PDFs/binary are rejected with a conversion hint — convert to text/markdown first, then ask the converted file.

Fork (draft) → edit → publish

Forks are drafts stored in ~/.codex/cxcm/forks/ — invisible to Codex until published, so the canonical session list stays clean while editing is in progress.

cissor fork <session> --name my-fork        # or --suffix trimmed, or neither (auto -fork)
cissor fork <session> --channel diet-auth   # stable work-stream id (default: source session id,
                                          # inherited by fork-of-fork so a chain stays grouped)
cissor edit <fork-id> ...                   # work on the draft
cissor publish <fork-id>                    # offline validation gate, then copies into sessions tree
codex resume <published-id>               # user resumes
cissor forks [--channel diet-auth]          # registered forks: status, lineage, op counts

Fork discipline: the fork is agent-invoked and returns an id; every editing action must pass that id explicitly. latest is refused for edits. Multiple forks are fine when each is deliberate; channels group related forks, and a repeat fork in the same channel within 30 minutes prints a warning pointing at the existing one. Fork of a fork inherits the parent channel and extends the lineage chain.

Originals are never edited in place — no override exists. Editing requires a registered fork; --force only exists for the over-broad eviction guard.

cissor publish <fork-id> runs the offline validator first (no Codex invocation, no tokens) and refuses to publish a structurally broken fork; --open then launches codex resume <id> in the fork's project directory. Re-publishing an already published fork is refused unless --republish.

Note: forks keep the source session's cwd (provenance). Codex's resume lookup filters by cwd, so resume from the same project works directly; from a different directory use codex resume --all <id> / codex exec resume --all <id>.

Edit (fork drafts only — originals are immutable; requires the explicit fork id)

cissor edit <fork> drop --turns 3-7
cissor edit <fork> replace --turns 2-40 --style index          # summarize + splice summary
cissor edit <fork> replace --turns 2-40 --summary-file s.md    # splice a hand-written summary
cissor edit <fork> trim-outputs --turns 2-10 --keep-chars 250 --min-tokens 200
cissor edit <fork> trim-outputs --all --keep-chars 400
cissor edit <fork> trim-inputs --all --keep-chars 200
cissor edit <fork> uncompact [--which last|first|all]
cissor edit <fork> transplant --from <other-session> --turns 3-9   # copy a span in (pairs kept intact)
cissor edit <fork> inject --file notes.md --label project_notes    # add a whole document as a turn

inject appends ONE message at the end — user role by default (--role developer|assistant available) — wrapped in <label> tags; no fake assistant turn is added. transplant requires complete call/output pairs in the span and renames colliding call_ids automatically.

Safety model:

• originals are immutable — edits only ever touch fork drafts, never --force-able
• never touches session_meta, turn_context, event_msg, world_state records
• dropping/replacing a tool call auto-includes its paired output (and vice versa), plus any reasoning items directly attached to them — a broken call/output pair or a dangling reasoning item makes resume fail
• refuses to remove the last user message
• every save: pre-edit .cxcm-bak backup, post-edit validation; pre-existing structural errors (some of Codex's own compactions leave orphan calls) are reported but don't block edits — only new errors do
• all writes are atomic (temp file + rename)
• reports/log lines compress item lists to ranges (17-56) to save agent context

Styles — editable prompts

cissor styles list                # builtin | file | custom
cissor styles show index
cissor styles edit index          # materializes the default, prints the path; edit at will
cissor styles reset index         # back to builtin

Style files live in ~/.codex/cxcm/styles/<name>.md; a file overrides the builtin. Create <your-style>.md for custom compaction flavours and use --style your-style.

Traceability

Every mutating or LLM-backed action appends one line to ~/.codex/cxcm/activity.jsonl (ts, kind, session, channel, project cwd, op, items, tokens before/after, model, outputs). Forks additionally keep per-fork op history in ~/.codex/cxcm/manifest.json.

cissor log                       # last 30 events (item lists range-compressed)
cissor log --stats               # compact aggregate: events by kind, total tokens saved, channels
cissor log --channel diet-auth   # everything for one work stream
cissor log --project ai-landscape  # everything for one project cwd
cissor log --path                # just print the file paths — grep them with bash for deep dives

Never dump the whole log into an agent context; use --path + shell search, or --stats for a summary.

Agent self-service workflow

cissor fork latest --suffix context-diet
cissor index <fork> --locate "the pricing bug"        # find what matters
cissor extract <fork> --turns 3-8 --out dump.md --evict --with-summary
cissor edit <fork> trim-outputs --all --keep-chars 300 --min-tokens 300
cissor edit <fork> replace --turns 9-40 --style index
cissor validate <fork>
codex resume <new-id>

Notes

• Token estimates cover conversation items only — not the system prompt, developer instructions stored in session_meta, or tool schemas. Real API usage (from token_count events) is shown by inspect for comparison.
• .cxcm-bak files sit next to edited sessions; Codex ignores them.
• Roadmap ideas live in ROADMAP.md.