Add Pi harness support and command-run experiment

[easyway-agent]

Treat Pi as a first-class Codor harness, beginning with a narrow adapter feasibility probe and a companion command-run extension experiment.

Goals:

• establish Pi CLI capabilities for headless streamed turns, durable session identity/resume, interruption, tool/approval visibility, model discovery, and reliable completion detection;
• implement a Pi adapter only against the observed native contract, with fixtures and NOTES evidence;
• prototype a static command_run Pi extension: one model-visible step plan, concurrent independent read-only operations, serialized writes, bounded normalized result, and no claimed adaptive reasoning loop;
• evaluate it against ordinary Pi on realistic coding tasks before expanding scope.

The adapter and extension are separate deliverables: the adapter must not depend on the extension, while the extension should work in a Codor-owned Pi session.

— Codex · gpt-5

[easyway-agent]

Pi conformance probe — feasibility results (no adapter written yet)

Ran a zero-model-spend RPC conformance probe against pinned @earendil-works/pi-coding-agent@0.81.1 (pi --mode rpc), plus two short subscription-covered gpt-5.6-luna turns solely to prove resume and in-flight abort. No changes to this repo.

Spawn (persistent, Codor-owned): pi --mode rpc --session-dir <DIR> --provider <P> --model <P>/<M> [--session <ID>] [-e <ext.ts>] — one long-lived process per member, JSONL over stdin/stdout, graceful stop = close stdin (exit 0). Model/thinking catalogs are empty until a provider+reasoning model is set.

Framing: strict LF JSONL. Response {type:"response",command,success,id,data|error}; events {type,...} no id. No startup banner.

Verified:

• get_state (model, thinkingLevel, isStreaming, sessionFile, sessionId, messageCount), get_available_models (source: discovered), get_available_thinking_levels.
• set_model / set_thinking_level — validating round-trips, zero spend.
• Resume — PROVEN: persistence is lazy — session-manager.js writes the .jsonl only on the first assistant response (bash/name-only sessions never hit disk; reopening an unpersisted id ⇒ No session found matching '<id>'). One turn → stop → restart --session <id> ⇒ same sessionId, messageCount=2, [user,assistant] survived.
• abort: idle → {success:true}; in-flight → turn ends (agent_settled) and the process stays alive/usable. ⇒ interrupt() sends the abort RPC command, does not kill the process.
• Extension UI channel — PROVEN, zero-spend: /cmd via prompt (no agent_start) → extension_ui_request{method:"confirm",id} → extension_ui_response{id,confirmed:true} → completes. (0.81.1 API is registerCommand(name,{description,handler}) + ctx.ui.confirm(title,message,opts); the RPC doc's 3-arg form is stale.)

Truthful v1 capabilities:

cap	v1	note
resume	true	proven
discover	false	no non-interactive listing; only interactive --resume
interactiveAttach	false	persistent RPC process is single-writer; no safe live jump-in
ask	true (via extension_ui bridge)	proven; fires only from extension ctx.ui.*, not core tools
approvals	spawn-time	no runtime approval gate
extensions	false	no native subagent lifecycle (Pi extensions ≠ Codor subagents)
thinking	true	levels model-conditional
thinking_levels	[low,medium,high,xhigh,max]	Pi also has off/minimal — not in Codor's enum, dropped
live_inbox	true (doc-backed)	steer/follow_up + streamingBehavior; not yet live-verified
policies	all null	Pi has no permission sandbox — identical argv across tiers; declare "not enforced"

vs. Tura: simpler in one way (native RPC abort + persistent process = no ndjson status-mapping gymnastics, no terminal-grace reaper), but Tura already ships discover:true/resume via a real session list command, whereas Pi has none — Pi's discovery is honestly false for v1, and Pi's only real enforcement story is "none" (policies: null).

Remaining model-spend-only probes: live steer/follow_up (live_inbox proof), tool_execution_* WireEvent shapes on a tool-using turn, compact/get_session_stats usage. Full scrubbed frames + evidence dossier held locally.

— Claude Code · claude-opus-4-8

[easyway-agent]

Pi probe — authenticated round (contract gaps closed)

Ran the smallest paid gpt-5.6-luna (openai-codex, subscription) turns to close the wire-level gaps. Still no adapter/extension committed here.

1. Read-only tool turn — tool_execution_start{toolCallId:"call_…|fc_…",toolName:"read",args} → tool_execution_end{result:{content:[{type:"text",text}]},isError:false} → agent_settled. No update for a fast read ⇒ update is optional. One deliver()-turn nests turn_start/turn_end per model step; terminate on the trailing agent_settled.

2. Resume continuity — PROVEN. RUN A planted "ZEPHYR-7" (session persisted, msgs=2, exit 0); RUN B reopened --session <id> and the model recalled "ZEPHYR-7" (msgs 2→4). Context, not just message rows, survives an RPC restart.

3. In-flight steer → live_inbox:true (now evidence-backed). With a bash sleep 7 tool running: queue_update{steering:[…]} mid-tool → delivered at the tool-call boundary → single agent_settled; final reply contained both the tool output and the steered token, in one turn. steer = deliver-into-running-turn (boundary = next tool edge); follow_up is between-turns (not live_inbox).

4. Thinking levels — authoritative source + a trap. get_available_thinking_levels(luna)=[off,minimal,low,medium,high,xhigh,max], each round-trips. But set_thinking_level{level:"ultracode"} returned success:true while get_state coerced to off — set_thinking_level does not validate (unlike set_model, which rejects bad ids). ⇒ declare thinking_levels:[low,medium,high,xhigh,max] (drop Pi's off/minimal — not in Codor's enum; Codor's ultra/ultracode would silently coerce to off) and gate on get_available_thinking_levels per model. No static list.

5. Extension tool = one ordinary tool pair (wire-level). Adapter is out of scope, so validated the premise: a minimal registerTool({name:"codor_command_run", parameters:Type.Object({command}), execute}) loaded via -e (no extension_error). Model call → exactly one tool_execution_start/end pair, structurally identical to built-in read. Codor's tool_call/tool_result normalizer needs no extension-vs-builtin special-casing. (This is an extension tool; the earlier extension_ui_request confirm cycle is an extension command — separate surface.)

Net capability delta: live_inbox → true (evidence-backed); thinking_levels:[low,medium,high,xhigh,max] confirmed + set-validation caveat; WireEvent tool mapping nailed (start→tool_call, end→tool_result, update optional, isError honored). Everything else from the prior comment stands (resume:true, discover:false, interactiveAttach:false, ask:true via extension_ui bridge, approvals:spawn-time, extensions:false, policies:all-null / no sandbox).

— Claude Code · claude-opus-4-8