Add the any-coding-agent guide; route non-named harnesses explicitly

[pengfei-threemoonslab]

Why

Agent-adoption review finding: the agent-native surface is real but gated on being one of three named harnesses. Cline, Windsurf, Devin, Aider, OpenHands get no auto agent-mode (only CLAUDECODE=1 / CURSOR_TRACE_ID auto-enable), the AGENTS_SHIPGATE_AGENT_MODE=1 override lived only deep in AGENTS.md §errors, llms.txt never mentioned it, and docs/agents/ listed exactly codex/claude-code/cursor. An agent outside the big three that fetched llms.txt had no path to structured errors or a named entry point.

What

• docs/agents/any-coding-agent.md (new): the harness-agnostic control guide — force agent mode, the three prominent commands with the agent-handoff-first read order, exit-code table, the MCP-server path for shell-less harnesses, forbidden shortcuts, and the redacted feedback export loop with the agent_feedback.yml template link. Generic profile documented as --agent codex.
• docs/agents/README.md + llms.txt: explicit "any other agent" routing (llms.txt gains its first mention of the agent-mode env var).
• AGENTS.md: header names the wider harness set; a two-line router atop Run (canonical) sends in-flight capability changes to the control loop first (setup only when no shipgate.yaml) — closing the last first-command divergence from README's check-first ordering (the read-order half was Align read-first prose with the contract's agent_read_order #251).
• llms-full.txt regenerated (AGENTS.md is a build source).

Discipline

Docs-only — no new CLI surface, schema, adapter, or check; deliberately did not guess env vars for auto-detecting other harnesses (that would be invented facts; the explicit override is the honest path until verified detection signals exist).

Verification

test_public_surface_contract (120), test_trigger_command, test_packaging green; ruff clean; shipgate check --agent claude-code on this diff (touches the agent_instructions trust-root surface): decision: allow, completion_allowed: true.

🤖 Generated with Claude Code

[pengfei-threemoonslab]

P2 addressed in bdcf618: agent_onramps.any_coding_agent now points to docs/agents/any-coding-agent.md, and the public-surface test's expected_onramps pins it (with a comment explaining why the machine-readable map must route non-named harnesses). Valid point that this belonged in the same PR — I had deferred the .well-known change on surface-discipline grounds, but leaving the primary discovery path stale for exactly the audience the guide unblocks was the worse inconsistency; this is an extension of an existing discovery map, not a new parallel surface. JSON validates; test_public_surface_contract.py green (120 tests); ruff clean.

One operational note: the site vendors .well-known from release tags (sync-well-known.sh), so this on-ramp reaches threemoonslab.com at the next tagged release — raw.githubusercontent consumers see it as soon as this merges.

🤖 Generated with Claude Code

[pengfei-threemoonslab]

P2 fixed: the exit-code table's 0 row now reads "Command completed — not a verdict. Inspect the JSON fields (decision, merge_verdict, release_decision.decision): in advisory mode, verify exits 0 even for review_required / human_review_required." Matching the .well-known exit_code_policy semantics. (The README's human-facing table uses the shorter "Pass (advisory mode…)" wording — same ambiguity in principle; left out of this PR's scope but worth a follow-up sweep.)

🤖 Generated with Claude Code