@caveman-ai/cli 1.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/LICENSE ADDED
@@ -0,0 +1,21 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2026 Julius Brussee
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
package/README.md ADDED
@@ -0,0 +1,210 @@
1
+ # @caveman-ai/cli
2
+
3
+ The `caveman` (alias `cave`) CLI. Wrap supported coding agents so their LLM
4
+ traffic can use local metering and recoverable context compression:
5
+
6
+ ```sh
7
+ caveman claude # shorthand for `caveman wrap claude`; bootstraps signed runtime on first TTY run
8
+ caveman learn # interactive local Setup Score + grouped top moves
9
+ caveman setup # show which companion binaries are installed
10
+ caveman stats # local spend, savings labeled `inferred`
11
+ ```
12
+
13
+ `caveman learn` shows animated progress, compact result cards, and a keyboard
14
+ menu for implementation, full detail, or the visual report. Use `--plain` for
15
+ stable compact text, `--all` for every sink and detector id, `--json` for automation, or
16
+ `caveman learn implement [claude|codex] --prompt "<focus>"` to open an agent
17
+ that reviews fixes one by one. Agent path installs the existing learn guide when
18
+ missing, never edits load-bearing findings, and asks before every edit.
19
+
20
+ Claude Code configured for Bedrock has an explicit native Runtime lane:
21
+
22
+ ```sh
23
+ CAVEMAN_WRAP_PROVIDER=bedrock \
24
+ AWS_REGION=us-east-1 \
25
+ AWS_BEARER_TOKEN_BEDROCK=… \
26
+ caveman wrap claude
27
+ ```
28
+
29
+ IAM credentials (`AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`, and optional
30
+ `AWS_SESSION_TOKEN`) work too. Mantle remains opt-in with
31
+ `CAVEMAN_BEDROCK_ENDPOINT=mantle`; it is a distinct Bedrock endpoint contract,
32
+ not an alias for normal Anthropic traffic. Runtime uses the attributed
33
+ `/bedrock` base; Mantle uses `/bedrock/anthropic` because Claude Code appends
34
+ `/v1/messages` to its Mantle override.
35
+
36
+ When the gateway URL is managed, wrap merges exactly one
37
+ `x-cave-api-key: <CAVE_API_KEY>` into Claude Code's documented
38
+ `ANTHROPIC_CUSTOM_HEADERS`, preserving unrelated custom headers and replacing
39
+ any stale case variant. If `AWS_BEARER_TOKEN_BEDROCK` is present, wrap also
40
+ merges it as `x-cave-upstream-key`; otherwise a complete IAM environment is
41
+ encoded as
42
+ `AWS_ACCESS_KEY_ID:AWS_SECRET_ACCESS_KEY[:AWS_SESSION_TOKEN]`. Bearer wins when
43
+ both forms are present. Runtime still uses Claude Code's documented AWS
44
+ credential chain; there is no supported Runtime equivalent of
45
+ `CLAUDE_CODE_SKIP_MANTLE_AUTH`. With neither explicit environment form, no
46
+ upstream header is added, but Claude Code must still resolve a local AWS
47
+ credential/profile before the managed gateway can substitute the project's
48
+ stored credential. For stored-only, server-injected Claude Code auth, select the
49
+ opt-in Mantle lane, whose gateway auth bypass is documented. Stale
50
+ `x-cave-upstream-key` variants are removed so they cannot override stored
51
+ credentials. Newlines and incomplete IAM pairs fail before launch.
52
+
53
+ Local Bedrock wrap never adds Caveman gateway headers and keeps the inherited
54
+ AWS environment. This header-backed seam is repository-tested; a live Claude
55
+ Code → AWS smoke remains a release gate. See the official [custom-header reference](https://code.claude.com/docs/en/env-vars)
56
+ and [Bedrock gateway setup](https://code.claude.com/docs/en/bedrock-vertex-proxies).
57
+
58
+ ## What this package is (and is not)
59
+
60
+ This npm package ships **only the JavaScript front-end** — one CLI entrypoint
61
+ plus a lazy 25 KB interactive-learn chunk, with zero runtime dependencies.
62
+ The heavy lifting (compression, metering, streaming
63
+ recovery, browsing) runs in caveman's companion **Go binaries**:
64
+
65
+ | Binary | Powers |
66
+ |---|---|
67
+ | `caveman-proxy` | `start` · `wrap` · `stats` — local compression + truthful metering |
68
+ | `caveman-engine` | `compress` · `shrink` · `retrieve` · `toon` · `evals` |
69
+ | `caveman-mcp` | agent-side recovery so streaming requests can compress |
70
+ | `cavemem` | local remember · recall · learned-context offload |
71
+ | `caveman-browse` | compressed page snapshots (optional) |
72
+ | `caveman-shrink` | `tools compress catalog` — tool-schema compression, lint, and recovery (optional) |
73
+
74
+ Two shrink surfaces stay explicit: `caveman tools shrink -- <command>` compresses
75
+ command output; `caveman tools compress catalog [lint|recover]` delegates to
76
+ `caveman-shrink` for MCP/OpenAI tool catalogs.
77
+
78
+ On first interactive local wrap, `caveman <agent>` installs signed runtime bundle
79
+ and continues same command. Explicit `CAVEMAN_*_BIN` overrides and non-interactive
80
+ runs stay unchanged. Without binaries CLI degrades affected commands to explicit
81
+ pass-through: nothing is compressed, savings report 0, and warning says why. Run
82
+ `caveman setup` to see what works and how to repair missing pieces.
83
+
84
+ `wrap` is the one exception to "pass-through": it points the agent's
85
+ `ANTHROPIC_BASE_URL`/`OPENAI_BASE_URL` at the proxy, so with no proxy listening
86
+ the agent's requests **fail to route** rather than passing through. An
87
+ interactive wrap catches this and offers to launch the agent directly instead;
88
+ a non-interactive wrap prints the warning and still launches, so scripted/CI
89
+ callers must ensure the proxy is up (or use `--no-proxy`).
90
+
91
+ ## Getting the binaries
92
+
93
+ Normal install:
94
+
95
+ ```sh
96
+ npm i -g @caveman-ai/cli
97
+ caveman claude
98
+ ```
99
+
100
+ First interactive wrap downloads key-signed release artifacts, verifies the
101
+ checksum manifest plus every SHA-256, installs atomically into
102
+ `~/.caveman/bin`, then launches the agent. Manual install/repair is
103
+ `caveman setup --install`.
104
+
105
+ Contributors may still build from source with
106
+ `./scripts/install-local-cli.sh` on macOS/Linux or
107
+ `pwsh -File scripts/install-local-cli.ps1` on Windows. Lookup order:
108
+ `CAVEMAN_*_BIN` env override → `PATH` → `~/.caveman/bin`.
109
+
110
+ ## Honesty rules
111
+
112
+ - Local savings are always labeled `inferred`; `verified_savings` stays 0 until
113
+ an optimizer runs in active mode on real traffic.
114
+ - Recoverable compression: lossy transforms store original before replacement.
115
+ Missing safe recovery path, parse failure, or non-smaller output passes through
116
+ unchanged and claims nothing. `record` mode is byte-preserving.
117
+
118
+ Connected verbs (`login`, `plan`, `score`, `costs`, …) talk to Caveman Cloud
119
+ over HTTP and need no binaries.
120
+
121
+ Connected telemetry imports stay under existing `cloud audit` governance verb:
122
+
123
+ ```sh
124
+ caveman cloud audit import --format langfuse traces.json
125
+ caveman cloud audit eval-import evidence.jsonl --dry-run
126
+ caveman cloud audit eval-import evidence.jsonl
127
+ ```
128
+
129
+ Eval JSONL accepts one canonical external-evidence object per line, at most
130
+ 1,000 records / 4 MiB per batch. Server stamps org/project, `observed` basis,
131
+ and `external_observed` authority; imported evidence cannot approve rollout or
132
+ mint verified savings.
133
+
134
+ ## Agent-native setup
135
+
136
+ Install complete native Caveman integration into Codex or Claude Code: local
137
+ routing, lifecycle hooks, architecture-first Core, recovery MCP, project-scoped
138
+ read tools, and setup/discovery/review/optimization/management skills:
139
+
140
+ ```sh
141
+ caveman setup --agent-native codex
142
+ # or
143
+ caveman setup --agent-native claude
144
+ ```
145
+
146
+ This writes an MCP command, not an access token, into the agent config. Runtime
147
+ requests use the existing `caveman login` credential store and still pass through
148
+ control-api authentication, project scope, RBAC, audit, and tenant isolation.
149
+ Read tools expose structured reports, Cave Plan, trace metadata/spans, and
150
+ experiment evidence. Agent MCP intentionally exposes no lifecycle mutation:
151
+ control-api needs server-authoritative transition and evidence gates first.
152
+ Agent-facing CLI experiment commands are read-only for same reason.
153
+
154
+ Core changes coding behavior but remains independently controllable. Compression,
155
+ recovery, and telemetry continue when Core is off:
156
+
157
+ ```sh
158
+ caveman tools config get think.core
159
+ caveman tools config set think.core off
160
+ caveman tools config set think.core on
161
+ ```
162
+
163
+ Run `caveman doctor codex` or `caveman doctor claude` to see active Core state,
164
+ source, profile, and pack version. Changes affect later hook events; start a new
165
+ agent session to clear Core already delivered into model context. Aider's shallow,
166
+ static Core read cannot apply `think.core`; `caveman disable aider` removes it.
167
+ `caveman disable <agent>` removes native routing/hooks while preserving unrelated
168
+ host edits.
169
+
170
+ Agent-native setup preflights host/runtime requirements before writes, journals
171
+ bundle ownership, validates installed components, and rolls back failed installs.
172
+ Rerunning converges journal-owned drift. Remove only bundle-owned pieces and
173
+ restore prior skill/MCP bytes with:
174
+
175
+ ```sh
176
+ caveman setup --agent-native codex --remove
177
+ ```
178
+
179
+ Inspect before install:
180
+
181
+ ```sh
182
+ caveman tools skills list --json
183
+ caveman tools skills preview caveman-evidence-review
184
+ caveman tools skills install --suite agent-native --agent codex --no-pixel
185
+ caveman tools mcp install codex --server caveman-cloud
186
+ ```
187
+
188
+ Install any third-party skill through the official Skills CLI, then store its
189
+ instruction body as profitable pixel pages automatically:
190
+
191
+ ```sh
192
+ caveman tools skills add mattpocock/skills --skill tdd --agent codex -y
193
+ caveman tools skills add https://github.com/vercel-labs/agent-skills --agent claude-code --global -y
194
+ ```
195
+
196
+ GitHub shorthand, Git/GitLab URLs, direct skill URLs, local paths, and upstream
197
+ `--skill`/`--agent`/`--global`/`--list`/`--all` flags pass through unchanged.
198
+ Caveman forces upstream copy mode when pixelizing, converts only new or changed
199
+ Claude Code/Codex skills, keeps `SKILL.orig.md` for byte-exact revert, and
200
+ reports measured estimated-token reduction as `inferred` per invocation.
201
+ `--no-pixel` keeps upstream plain-text behavior. Third-party instructions and
202
+ resources are installed as supplied; Caveman does not security-review them.
203
+ Official Skills CLI telemetry behavior also applies; set
204
+ `DISABLE_TELEMETRY=1` to opt out upstream.
205
+
206
+ Dated local smoke on Matt Pocock's 35-skill repository converted 27 profitable
207
+ skills and left eight safely as text. Sum across one invocation of each
208
+ converted skill: 32,314 → 9,537 estimated tokens (−70% `inferred`). This is
209
+ estimator output, not provider-reported billing or a skill-quality benchmark;
210
+ see [`docs/research/skill-pixel-install-2026-08-09.md`](../../docs/research/skill-pixel-install-2026-08-09.md).