返回 CodeWhale
LIVE_SMOKE.md
根目录 / docs / LIVE_SMOKE.md
1 # Opt-in live smoke runs
2
3 This page is **manual, opt-in, and never automated.** Nothing in CI, no test,
4 no build script, and no skill runs these commands. The repository's automated
5 suite is provider-free by design; see
6 [`crates/tui/assets/skills-catalog-matrix.json`](../crates/tui/assets/skills-catalog-matrix.json)
7 and the catalog-matrix tests for what is actually asserted without a provider.
8
9 Run this only when you want to answer one narrow question: *does a real route
10 to a real model return a well-formed receipt on this machine?*
11
12 ## What a live smoke run does and does not prove
13
14 | Question | Answered here? |
15 | --- | --- |
16 | Does the receipt record the provider/model I asked for? | Yes. This does not by itself prove which network endpoint handled the request. |
17 | Does the run emit an inspectable route/usage receipt? | Yes, when the harness reaches that stage. |
18 | Does one response prove my account's entitlement state? | **No.** Provider configuration, authentication/entitlement, and harness behavior remain candidate causes until corroborated. |
19 | Does the model semantically pick the right skill? | **No.** Not measured. |
20 | Is skill registry/catalog/alias behavior correct? | **No** — that is the provider-free suite's job. |
21
22 Treat these as investigation starting points, not proven failure classes:
23
24 - **Provider error response** — HTTP 401/403, unknown-model, quota, or region
25 errors can reflect the configured provider/endpoint, credential
26 authentication or entitlement, provider availability, or a harness
27 routing/request defect. The response alone does not distinguish them.
28 - **Receipt or process anomaly** — wrong `provider`/`model` in the receipt,
29 missing receipt fields, a crash, or failure to use the isolated state
30 directory is evidence to investigate the harness, but still needs a minimal
31 reproduction or other corroboration before assigning the cause.
32
33 ## Isolation rules these snippets follow
34
35 1. `env -i` clears the inherited environment, so your ambient `HOME`,
36 `CODEWHALE_HOME`, and `*_API_KEY` values are not forwarded. Only variables
37 listed explicitly on the `env` line survive.
38 2. Only `CODEWHALE_HOME` points at the task-specific throwaway directory, so
39 Codewhale config, sessions, and the bundled skill install land in scratch
40 state. `HOME` is intentionally left unset; the smoke run never repurposes it.
41 3. You name the credential variable yourself (`CW_SMOKE_CRED_VAR`). Nothing is
42 guessed from the provider.
43 4. The isolated child reads the secret with echo disabled, restores the prior
44 terminal state on `EXIT`, `INT`, `HUP`, or `TERM`, and exports it only in
45 that child. The value is not persisted to disk or placed in a command
46 argument or shell history.
47 5. `PATH` is forwarded explicitly, and is the only host variable carried over.
48
49 Portable `sh` is used throughout; `stty` and `mktemp -d` are the only non-POSIX
50 niceties and both exist on macOS and mainstream Linux.
51
52 ## Step 1 — create the throwaway state (both runs)
53
54 ```sh
55 CW_SMOKE_CODEWHALE_HOME="$(mktemp -d)" || exit 1
56 mkdir -p "$CW_SMOKE_CODEWHALE_HOME/tmp"
57 echo "scratch Codewhale state: $CW_SMOKE_CODEWHALE_HOME"
58 ```
59
60 ## Step 2 — name the credential variable
61
62 `CW_SMOKE_CRED_VAR` must be the variable name the provider expects. Codewhale
63 reads `MOONSHOT_API_KEY` (or `KIMI_API_KEY`) for the Moonshot/Kimi route and
64 `DEEPSEEK_API_KEY` for the DeepSeek route.
65
66 ```sh
67 CW_SMOKE_CRED_VAR="MOONSHOT_API_KEY" # you choose this; nothing is inferred
68 ```
69
70 The run command prompts for the value inside its isolated child process. It
71 does not create a credential file.
72
73 ## Step 3a — run A: Kimi K3
74
75 `kimi-k3` is a model id this build knows about. The configured provider and its
76 resolved endpoint determine the route: `--provider moonshot` selects the
77 configured Moonshot route; selecting `opencode_go` would select that separately
78 configured route. The account does not choose between them, and the harness
79 does not switch between them based on a response. Set
80 `CW_SMOKE_PROVIDER` / `CW_SMOKE_MODEL` for the route you intend to exercise. A
81 model-not-found response is an unclassified result until the provider/endpoint
82 configuration, credential access, and harness request are corroborated.
83
84 ```sh
85 CW_SMOKE_PROVIDER="moonshot"
86 CW_SMOKE_MODEL="kimi-k3"
87 CW_SMOKE_EFFORT="medium"
88 CW_SMOKE_PROMPT="Reply with exactly: SMOKE OK"
89
90 env -i \
91 PATH="$PATH" \
92 TMPDIR="$CW_SMOKE_CODEWHALE_HOME/tmp" \
93 CODEWHALE_HOME="$CW_SMOKE_CODEWHALE_HOME" \
94 CW_SMOKE_CRED_VAR="$CW_SMOKE_CRED_VAR" \
95 sh -c '
96 CW_SMOKE_STTY_STATE="$(stty -g)" || exit 1
97 restore_terminal() {
98 stty "$CW_SMOKE_STTY_STATE" 2>/dev/null || :
99 }
100 trap "restore_terminal" EXIT
101 trap "restore_terminal; exit 129" HUP
102 trap "restore_terminal; exit 130" INT
103 trap "restore_terminal; exit 143" TERM
104
105 printf "Paste value for %s (input hidden): " "$CW_SMOKE_CRED_VAR" >&2
106 stty -echo || exit 1
107 if ! IFS= read -r CW_SMOKE_CRED; then
108 printf "\nCredential input failed.\n" >&2
109 exit 1
110 fi
111 restore_terminal
112 trap - EXIT HUP INT TERM
113 unset CW_SMOKE_STTY_STATE
114 printf "\n" >&2
115
116 export "$CW_SMOKE_CRED_VAR=$CW_SMOKE_CRED"
117 unset CW_SMOKE_CRED
118 exec codewhale-tui exec \
119 --provider "$1" --model "$2" --reasoning-effort "$3" --json "$4"
120 ' sh "$CW_SMOKE_PROVIDER" "$CW_SMOKE_MODEL" "$CW_SMOKE_EFFORT" "$CW_SMOKE_PROMPT"
121 ```
122
123 ## Step 3b — run B: a second provider/model (DeepSeek)
124
125 Set `CW_SMOKE_CRED_VAR="DEEPSEEK_API_KEY"`, then:
126
127 ```sh
128 CW_SMOKE_PROVIDER="deepseek"
129 CW_SMOKE_MODEL="deepseek-v4-pro"
130 ```
131
132 …and re-run the identical `env -i …` block from step 3a; it prompts for a fresh
133 credential value. Running the *same* command shape against two providers is the
134 point: a difference in outcome is an observation to investigate, not proof of
135 route, entitlement, or harness correctness. Provider/endpoint configuration,
136 credentials, provider health, and the generated request all remain possible
137 explanations.
138
139 ## Step 4 — optional: tool-and-reasoning receipt
140
141 The `--json` one-shot above records the resolved route claimed by the harness;
142 it does not independently prove which endpoint handled the request. To also see
143 tool-catalog and reasoning receipts, use the streaming form (still inside the
144 same `env -i` wrapper, substituting the `exec` line):
145
146 ```sh
147 exec codewhale-tui exec --auto --max-turns 3 \
148 --output-format stream-json \
149 --provider "$1" --model "$2" --reasoning-effort "$3" "$4"
150 ```
151
152 ## Step 5 — what to record
153
154 From the `--json` one-shot receipt:
155
156 | Field | Expectation |
157 | --- | --- |
158 | `mode` | `one-shot` |
159 | `provider` | exactly the `--provider` you passed |
160 | `model` | exactly the `--model` you passed |
161 | `success` | `true` |
162 | `output` | the model's text; content is *not* a pass/fail criterion |
163
164 From the `stream-json` metadata receipt:
165
166 | Field | Expectation |
167 | --- | --- |
168 | `provider`, `model` | match the flags you passed |
169 | `route_source` | records *why* that route was chosen |
170 | `reasoning_tokens` | present when the receipt reports reasoning; absence can reflect model/provider behavior, configuration, or a harness omission and needs corroboration |
171 | `tool_catalog_sha256` | present when a tool surface was offered |
172 | `approval_posture`, `sandbox_posture` | match the flags you passed |
173 | `duration_ms`, `input_tokens`, `output_tokens` | present for a completed run |
174
175 Report the receipt fields. Do **not** paste the credential, the key file, or
176 raw provider error bodies (they can echo request headers).
177
178 ## Step 6 — clean up
179
180 ```sh
181 rm -rf "$CW_SMOKE_CODEWHALE_HOME"
182 unset CW_SMOKE_CODEWHALE_HOME CW_SMOKE_CRED_VAR \
183 CW_SMOKE_PROVIDER CW_SMOKE_MODEL CW_SMOKE_EFFORT CW_SMOKE_PROMPT
184 ```
185
186 ## Scope note
187
188 A green live smoke run is evidence that the configured live attempt completed
189 today. By itself it does not prove endpoint identity, durable account
190 entitlement, or the absence of a harness defect; corroborate those claims
191 separately. It says nothing about skill selection, alias resolution, locale
192 routing, or prompt budget — all of which are covered deterministically and
193 provider-free in `crates/tui/src/skills/catalog_matrix.rs`.
194
194 lines MARKDOWN