Cline¶

Cline, like opencode, is pure path-discovery in wrapper mode — it reads its instructions and skills from fixed locations and needs no flags. Cline reads AGENTS.md as a cross-tool rules file, which is exactly the agent-neutral source, so the project instructions stay in place.

Wrapper resolution¶

Source	Injected at
project `AGENTS.md`	(not injected — read natively at `./AGENTS.md`)
project `.agents/skills/`	`<root>/.cline/skills/`
global `~/.config/agents/AGENTS.md` (+ optional `AGENTS.cline.md` overlay)	`~/.agents/AGENTS.md`
global `~/.config/agents/skills/`	`$CLINE_DATA_DIR/skills/` (default `~/.cline/skills/`)

Project instructions — Cline reads the project-root AGENTS.md as a cross-tool rules file. That is the agent-neutral source, already in place, so agedum injects nothing for it — and never could, since the root AGENTS.md is git-tracked.
Global instructions — Cline reads the cross-tool global path ~/.agents/AGENTS.md, so the global AGENTS.md is bound there — base merged with an optional AGENTS.cline.md overlay. Note this is not under $CLINE_DATA_DIR: Cline reads global skills from ~/.cline/skills/ but global cross-tool instructions from ~/.agents/AGENTS.md — the asymmetry is Cline's.
Skills — each skill is a SKILL.md folder (the shape Cline already expects), compiled with the SKILL.cline.md overlay and bound to ./.cline/skills/ (project) and $CLINE_DATA_DIR/skills/ (global, default ~/.cline/skills/).
extra_args: none — Cline discovers everything from disk, like Claude and opencode.

agedum --wrapper cline -- cline task "review this change"
agedum --wrapper cline --dry-run -- cline        # show what would be injected

Provider config¶

Cline's knobs become appended CLI flags, like kimi:

{
  "harness": "cline",
  "slug": "cline-deepseek",
  "secretEnv": "DEEPSEEK_API_KEY",
  "config": {
    "provider": "deepseek",
    "model": "deepseek-v4-pro",
    "effortLevel": "xhigh"
  }
}

`config` key	Appended
`model`	`--model <model>`
`provider`	`--provider <id>` (Cline provider id; default `cline`)
`effortLevel`	`--thinking <none\\|low\\|medium\\|high\\|xhigh>`
`plan`	`--plan` (true)
`secretEnv` value	`--key <token>`

The config above launches cline --model deepseek-v4-pro --provider deepseek --thinking xhigh --key ***.

The token lands in argv. Cline takes its API token as a per-run flag (--key), so — unlike every other harness — the secret is in the launched command, visible in the process list while Cline runs. That is Cline's documented mechanism, not agedum's choice; agedum masks the token in --dry-run (the command line shows --key ***), and it still rides the requiredEnv export so the environment block masks it too.

Custom baseUrl. Cline has no run-time base-URL flag, and a --provider/--model flag set rebuilds the provider from flags (dropping any stored base URL and posting to the OpenAI default). So for a custom OpenAI-compatible endpoint (a Kimi coding subscription, OpenCode-Go, …) agedum generates a single-provider providers.json — the generic openai-compatible provider with lastUsedProvider set so Cline selects it with no flag — under an isolated CLINE_DATA_DIR (~/.cache/agedum/cline/<endpoint-slug>), and passes the key via --key. Set model (the upstream id served there) and secretEnv; optional contextWindow / maxTokens become a one-entry models array so Cline learns the window (its X/N meter + compaction trigger) and output cap. The API key is never written to disk. That providers.json is seeded writable into CLINE_DATA_DIR (not read-only bound), so under a sandbox launch Cline can rewrite it to persist its own selection rather than hitting EROFS on a read-only filesystem; agedum re-seeds the correct endpoint config on every launch. baseUrl and a named provider are mutually exclusive.

Named provider. To use a provider Cline already knows instead, set provider (no baseUrl) to a built-in (e.g. deepseek) or one set up via cline auth (stored in ~/.cline/data/settings/providers.json). Under a sandbox launch agedum makes ~/.cline writable so that store (and Cline's task state) persists too.

No subagent-model tiering. Cline runs a single model per session, so there is no agedum equivalent of opencode's agentOptions[] or reasonix's subagentModel, and a tiered cline-*-flash (strong main model + cheap-flash subagents) config is not expressible. Cline's subagents are fixed read-only research agents — their tools are limited to read_file, list_files, search_files, list_code_definition_names, read-only execute_command, and use_skill; they can't write, browse, reach MCP, or nest, and they inherit the session model. cline --help (v3.0.15) confirms a single -m/--model with no --subagent-model, so the only meaningful flash config is the flat cline-flash (flash as the whole-session model).

--prompt/--run. Cline's prompt is a positional argument; --tui opens the interactive TUI seeded with it, and a bare positional runs the task once in act mode and exits. So agedum cline-<name> --prompt "<text>" maps to cline --tui "<text>" and --run "<text>" maps to cline "<text>" — see the prompt-seeding table.