| # PR 7: finally good code |
|
|
| You are the OpenCode coding subagent, the sole code writer for this task. Apply |
| the attached Ponytail skill (full). Read it first and report its SHA-256 in your |
| final evidence. You are not alone in this workspace: preserve existing/concurrent |
| changes and never revert others' work. Do not commit, push, merge, install globally, |
| alter live ~/.hai, ~/.hermes, agent profiles, or unrelated repos. |
|
|
| ## Ownership |
|
|
| Implement only in `/home/smlflg/Projekte/CodingAgent/hai-mcp-pr7`, branch |
| `feat/coding-worker-lanes`. Own code, tests and matching project documentation |
| for the worker lane slice. Root agent orchestrates and reviews; it does not code. |
|
|
| ## User requirements (mandatory) |
|
|
| HAI-MCP becomes the common entry point for existing main agents. There are two |
| worker lanes: `coding` and `sonstiges`. Main agent chooses the lane autonomously |
| and can split mixed tasks; no owner approval for routine lane choice/delegation. |
| ALL coding via this entry point must delegate to an OpenCode subprocess with |
| Ponytail actually supplied as instructions. No main-agent coding or alternate |
| coding harness path, including fixes. Sonstiges is ordinary non-code work; main |
| can do it itself or use a worker. Distinct technical review must precede coding |
| acceptance; process exit 0 is not acceptance. Keep commit/push/destructive/global |
| configuration human gates. No built-in auto commits/pushes. |
|
|
| ## Architecture and scope |
|
|
| Read AGENTS.md, GOAL.md, docs/TOOL_CONTRACT.md and all touched code first. Current |
| main is the single-writer contract kernel. Open PR #6 is a separate owner-gate |
| change: do NOT merge or copy it. Reuse existing contracts, session validation, |
| path confinement, storage, locking and audit. Update slice documentation to |
| reflect user-authorized worker execution; retain no direct LLM SDK calls inside |
| server (external bounded OpenCode process performs model calls). |
| |
| Provide a minimal coherent real implementation, not just proposals or prompts: |
| - MCP entry to dispatch bounded coding job against a valid lease, criterion, |
| approved project scope; lane metadata explicit, allowed values only. |
| - Existing main session decides lane without extra owner gate; noncoding path |
| must not silently launch coding. Reject known code mutation requests routed as |
| noncoding. Clearly document that MCP cannot remove an external agent's shell |
| rights, nor semantically classify deceptive prose; do NOT claim OS enforcement. |
| Provide an actionable main-agent client contract for closing direct-write paths. |
| - Coding launch is hardwired to OpenCode (not agent-supplied executable/shell). |
| Full Ponytail instructions must be packaged with license attribution and sent |
| to every builder/fix invocation; record hash/provenance. Do not just accept |
| `ponytail=true` supplied by a caller. No large new framework/dependencies. |
| - Async start returns job ID quickly. Status/results and stop work on that job. |
| Bound duration to lease/mission expiry and stop on revoked/stale contracts, |
| explicit stop, timeout; no automatic retry/provider escalation. Concurrent |
| starts cannot bypass allowed capacity. Failed/missing process after restart |
| must never look active or completed forever. Don't signal stale/reused PIDs. |
| - Capture stdout/stderr without leaking credentials, exit/result status, actual |
| diff and evidence provenance tied to code snapshot; edits invalidate review. |
| Use subprocess argv, never shell=True or interpolation. Treat agent output as |
| untrusted and don't auto-execute suggested commands. Test commands and writable |
| project scopes must come from authorized task/contract, not generated output. |
| - Separate review records need a valid different session from builder/requester |
| as appropriate, matching mission/criterion/contract and exact diff/snapshot. |
| Reject stale or self-review. No `accepted` merely for successful process exit. |
| Do not fabricate semantic verification: tests/review are evidence with limits. |
| - Work in an isolated workspace for coding and return changes for inspection; |
| avoid changing source repo automatically. Capture untracked as well as tracked |
| files, avoid including secrets/.git/runtime dirs. If clean Git prerequisite is |
| necessary, enforce and document it honestly. Scope is meaningful, not only a |
| prompt; detect out-of-scope changes and refuse acceptance. |
| - Respect declared non-goals and all existing gates; old non-worker workflows |
| should remain functional. No dashboard, registry, cron, mass client install. |
| |
| If above needs a specific simplification, do not silently drop a requirement: |
| report it as unresolved with evidence. Keep solution small but complete. |
| |
| ## Verification |
| |
| Use isolated HAI_HOME and dependency cache under /tmp, never live state. Existing |
| interpreter `/home/smlflg/Projekte/hai-mcp-installed/.venv/bin/python` has pytest |
| and mcp installed; run with PYTHONPATH pointing at this checkout's src so no |
| installation is needed. Tests must cover actual subprocess boundary with a |
| deterministic fake opencode executable, not only mocked success dictionaries. |
| Add focused tests for real lifecycle, coding forcing/Ponytail content, lane choice, |
| invalid args, no shell injection, expired/revoked leases, capacity, stop/timeout, |
| subprocess errors, worktree scope escapes/symlinks, untracked diff, and distinct |
| review snapshot binding. Also expose tools through real MCP schema/stdio test. |
| Run whole pytest suite once focused tests pass. Existing failures must be reported. |
|
|
| Output final diff summary, exact test command/result, Ponytail hash, unresolved |
| limitations. No live paid worker tests unless root explicitly instructs one. |
|
|
