Build-script mode¶

agedum --build-script [--check] <config.json> [output.sh]

--build-script turns a provider config JSON into a self-contained bash wrapper. The wrapper, when run, sources a .env, validates the required variables, exports the provider/model/auth environment for the harness, and finally execs agedum --wrapper <harness> -- <harness> "$@" — composing this codegen step with the wrapper mode that injects your skills and instructions.

This is how a content repo (e.g. an agentsconf) builds its provider launchers: providers/<name>.json → bin/<name>.sh. agedum stays a pure translator — it never reads a token at run time. All secret handling lives in the generated shell.

Config shape¶

The config is the condash-style agent envelope:

{
  "harness": "claude",
  "name": "Claude Deepseek Auto",
  "slug": "claude-deepseek-auto",
  "secretEnv": "DEEPSEEK_API_KEY",
  "requiredEnv": ["DEEPSEEK_API_KEY"],
  "config": { "...": "per-harness options" }
}

Field	Meaning
`harness`	`claude`, `kimi`, or `opencode`. Selects the translation; this is the build input, so it is read from the file (unlike wrapper mode, where the flag is authoritative).
`secretEnv`	The env var holding the API token. For `claude` it is mapped to `ANTHROPIC_AUTH_TOKEN` / `ANTHROPIC_API_KEY`; for `kimi` / `opencode` it is passed through under its own name.
`requiredEnv`	Vars the generated wrapper validates and exports. `secretEnv` is always appended if not already listed. Declare a provider's API-key var here so `kimi` / `opencode` (which read it from the environment) see it.
`config`	The per-harness option block — see below.
`name` / `slug`	Labels; `slug` (else `name`, else the harness) names the provider in the generated comment + error messages.

What the generated wrapper does¶

#!/usr/bin/env bash
set -euo pipefail

env_file="${AGENTS_ENV_FILE:-$HOME/.config/agents/.env}"   # override with AGENTS_ENV_FILE
if [[ ! -f "$env_file" ]]; then
  echo 'claude-deepseek-auto: env file not found:' "$env_file" >&2
  exit 1
fi
# shellcheck source=/dev/null
source "$env_file"
export DEEPSEEK_API_KEY="${DEEPSEEK_API_KEY:?DEEPSEEK_API_KEY is required by provider claude-deepseek-auto but is not set}"

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN="$DEEPSEEK_API_KEY"
# … the rest of the per-harness mapping …

exec agedum --wrapper claude -- claude "$@"

.env location — ${AGENTS_ENV_FILE:-$HOME/.config/agents/.env}. Set AGENTS_ENV_FILE to point elsewhere; no rebuild needed.
Validation — each requiredEnv var is checked with ${VAR:?…} (fails fast with a clear message if unset or empty) and exported.
Passthrough — "$@" forwards any args you give the wrapper to the harness.
No env block is emitted when a provider declares no secrets (e.g. native Claude), so the wrapper just execs the harness.

Per-harness `config` mapping¶

claude¶

When baseUrl is empty the wrapper runs bare (exec agedum --wrapper claude -- claude). Otherwise:

`config` key	Exported as
`baseUrl`	`ANTHROPIC_BASE_URL`
`authStyle` + `secretEnv`	`ANTHROPIC_AUTH_TOKEN` (bearer) or `ANTHROPIC_API_KEY` (apikey); the other is `unset`
`model`	`ANTHROPIC_MODEL`
`smallFastModel`	`ANTHROPIC_SMALL_FAST_MODEL`
`haikuAlias` / `sonnetAlias` / `opusAlias`	`ANTHROPIC_DEFAULT_{HAIKU,SONNET,OPUS}_MODEL`
`subagentModel`	`CLAUDE_CODE_SUBAGENT_MODEL`
`maxContextTokens` (> 0)	`CLAUDE_CODE_MAX_CONTEXT_TOKENS`
`effortLevel`	`CLAUDE_CODE_EFFORT_LEVEL`
`disableCaching`	`DISABLE_PROMPT_CACHING=1`
`disable1M`	`CLAUDE_CODE_DISABLE_1M_CONTEXT=1`
`disableAdaptiveThinking`	`CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1`
`disableTelemetry`	`DISABLE_TELEMETRY=1`
`disableErrorReporting`	`DISABLE_ERROR_REPORTING=1`
`disableClaudeApiSkill`	`CLAUDE_CODE_DISABLE_CLAUDE_API_SKILL=1`

CLAUDE_CODE_USE_{BEDROCK,VERTEX,FOUNDRY,MANTLE} are always unset defensively. Empty strings, zero, and false are omitted.

kimi¶

kimi's knobs become appended CLI flags on the exec line:

`config` key	Appended
`model`	`--model <model>`
`thinking`	`--thinking` (true) / `--no-thinking` (false)
`plan`	`--plan` (true)
`configInline`	`--config <value>`

opencode¶

`config` key	Effect
`model`	`model` field of `OPENCODE_CONFIG_CONTENT`
`disableExternalSkills`	`OPENCODE_DISABLE_EXTERNAL_SKILLS=1`
`defaultOptions.{reasoningEffort,textVerbosity,reasoningSummary}`	the default model's `provider.<id>.models.<model>.options`
`effortLevel` (flat alias)	the default model's `reasoningEffort` (explicit `defaultOptions.reasoningEffort` wins)
`agentOptions[]`	per-agent `agent.<name>` model + options; `primary: true` sets `mode: "primary"` for custom (non-built-in) agents
`extraConfigJson`	parsed and deep-merged into the document

The document is emitted as a single OPENCODE_CONFIG_CONTENT env var (no file written).

Drift guard — `--check`¶

--check regenerates the script in memory and compares it to the committed output.sh, exiting non-zero if they differ. Run it in CI so a stale bin/<name>.sh (someone edited the JSON but forgot to regenerate) fails the build:

agedum --build-script --check providers/claude-deepseek-auto.json bin/claude-deepseek-auto.sh

Generation is deterministic, so --check is byte-exact.