Ticket graphs

Rewrite ONE unique forge module as entirely new, behavior-preserving, installable, control-config'd Python, per the FORGE-wide refactor standard #200 (#std). This ticket covers ONE module — the one identified below — and nothing else. It was filed against the module's DEDUPLICATED identity, so if that module was copy-pasted across the forge, every copy is covered by this single ticket. MODULE KEY : forge-rw:4b850fc5c8793fd5 (the ticketer's idempotency marker — it lives in this ticket's TITLE. Do not edit the title, or the module will be re-filed as a duplicate.) IDENTITY : AST-NORMALIZED — codebean's `_ast_norm` (#183), i.e. ast.dump(ast.parse(ast.unparse(ast.parse(src)))). Copies of this module that differ only in comments, blank lines, quoting or spacing carry the SAME module-sha and are covered by THIS ticket. A differing docstring is real content, so a module whose docstring differs is a DIFFERENT module with its own ticket. SIZE : 6102 bytes EXTENSION : .py CANONICAL SOURCE (the location this ticket is named for): repo : agent_tools_suite path : validate.py ref : HEAD (HEAD) blob : 69b50b41e700038739a5c7eace132b1b4b4162ca (git blob sha1) REFERENCE ONLY — you do NOT need this command. This module's bytes are embedded COMPLETE in THE ORIGINAL SOURCE block below, and that is what the rewrite is based on. For a reader who does have forge access: git --git-dir=/Users/stevenpeterson/code/installed/jazz-project/config/forge/remotes/agent_tools_suite.git cat-file blob 69b50b41e700038739a5c7eace132b1b4b4162ca THE SAME MODULE ALSO LIVES AT THESE FORGE TIPS: none — this module is unique to agent_tools_suite among the repos scanned so far. HISTORICAL OCCURRENCES of this exact blob — every OTHER DISTINCT {repo, path} the omni-git index (`occ`, joined on the git blob sha1) records this byte-identical file at, with the index's own `ts` for the first row of each. Any place this run enumerated LIVE is listed under the forge tips above, not here. Context for the rewrite, not extra work: testmonkeyalpha-group/jazz-forge/agent_tools_suite:validate.py [occ.ts gitlab] THE ORIGINAL SOURCE, EMBEDDED — YOU NEED NO FORGE ACCESS TO DO THIS TICKET. The module's own bytes are reproduced below, inside this ticket. THEY are the authoritative copy and the behavior-preserving basis: rewrite from them. The `git --git-dir=... cat-file blob` command under CANONICAL SOURCE above is kept only as a REFERENCE for a reader who happens to have forge access — running it is NOT part of this ticket. (Why the bytes are here at all: the pair_loop worker that executes this ticket is sandboxed to the ticket's own congruency worktree and cannot read the forge's bare repos. Ticket #346 is the fix that put them on the ticket.) The block BEGINS at the line ===ORIGINAL SOURCE (verbatim, the behavior-preserving basis)=== and ENDS at the first line ===END SOURCE=== Everything strictly between those two lines is the module — 6102 byte(s), byte for byte, with no re-indentation, no re-wrapping and nothing elided. ===ORIGINAL SOURCE (verbatim, the behavior-preserving basis)=== #!/usr/bin/env python3 """validate.py — does the agent's tooling actually work, and is it safe? Two levels, because "validating an agent" is two different questions: LEVEL 1 — the TOOLS are correct + safe (deterministic; this file). · each crankimg verb does what it says (build/verify/disks/flash/base) · the destructive verb (flash) REFUSES without an exact confirm, and refuses the system disk / internals / a backup volume · the SURFACE is minimal — the agent's whole action space is crankimg (5) + the bash decoy + gate meta, and nothing else (no file-edit, no git, no real shell) · the GATE enforces intent — every routed tool gets a REQUIRED `intent` field, and a call is recorded {timestamp, tool, intent} before it runs LEVEL 2 — the AGENT behaved (a replay audit, not this file). The gate's intent trail IS the validation surface: after a run, `gate__query` / recall replays the {intent, tool, time} sequence — you assert it matches the expected crank flow (build → verify → disks → flash) and that the decoy logged no shell reach. Deterministic tools + a mandatory-intent trail = you can prove BOTH "the tools work" and "the agent used them for the reason it claimed." Run: python3 validate.py """ import json import os import subprocess import sys import time HERE = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, HERE) sys.path.insert(0, os.path.join(HERE, "crankimg-mcp")) import mcplib # noqa: E402 import server as crank # noqa: E402 — crankimg-mcp/server.py P, F = 0, 0 def ok(msg): globals().__setitem__("P", P + 1); print(" \033[32m✓\033[0m " + msg) def bad(msg): globals().__setitem__("F", F + 1); print(" \033[31m✗ FAIL\033[0m " + msg) def check(cond, msg): ok(msg) if cond else bad(msg) def text(res): return "".join(c.get("text", "") for c in res.get("content", [])) def call(tool, **args): return crank.call_tool(tool, args) def level1_tools(): print("\n\033[1mLEVEL 1 — tools correct + safe\033[0m") # surface: exactly the five verbs, no more names = {t["name"] for t in crank.list_tools()} check(names == {"build", "verify", "disks", "flash", "base"}, "surface is exactly {build,verify,disks,flash,base} (got %s)" % sorted(names)) # disks: system disk refused, at least one candidate visible d = text(call("disks", intent="recon")) check("/dev/disk0" in d and "REFUSE" in d.split("/dev/disk0")[1].split("\n")[0], "disks REFUSES the system disk /dev/disk0") check("candidate" in d, "disks shows at least one candidate (external)") # base: reports the (already-present) image b = text(call("base", intent="check base")) check("present" in b or "MISSING" in b, "base reports present/missing cleanly") # verify: on the real baked seed (built earlier) → verified v = text(call("verify", intent="verify image")) check("verified" in v or "verification failed" in v, "verify runs a read-only image check") # flash SAFETY — the whole point check(text(call("flash", intent="x")).startswith("device required") or "device required" in text(call("flash", intent="x")), "flash refuses with no device") check("must echo the device" in text(call("flash", intent="x", device="/dev/disk9", confirm="/dev/disk8")), "flash refuses when confirm != device (fat-finger guard)") check("refused" in text(call("flash", intent="x", device="/dev/disk0", confirm="/dev/disk0")).lower(), "flash REFUSES the system disk even with a matching confirm") # build precondition: a missing wallet secret → clear guidance, no crash bd = text(call("build", intent="build", wifi_secret="no-such-secret-xyz")) check("wallet" in bd.lower(), "build handles a missing wallet secret with guidance (no crash)") def level1_end_to_end(): print("\n\033[1mLEVEL 1 — end-to-end through gate → coupler → crankimg\033[0m") proc = subprocess.Popen(["python3", "gate.py"], cwd=HERE, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True, bufsize=1) def rpc(mid, method, params): proc.stdin.write(json.dumps({"jsonrpc": "2.0", "id": mid, "method": method, "params": params}) + "\n") proc.stdin.flush() for _ in range(400): line = proc.stdout.readline() if not line: return None m = json.loads(line) if m.get("id") == mid: return m try: rpc(1, "initialize", {"protocolVersion": "2024-11-05", "capabilities": {}}) tl = rpc(2, "tools/list", {}) tools = {t["name"]: t for t in tl["result"]["tools"]} check("crankimg__disks" in tools, "crankimg tools surface through the gate (namespaced)") check("intent" in tools["crankimg__disks"]["inputSchema"].get("required", []), "the gate marked crankimg__disks' `intent` REQUIRED through the whole stack") check("bash__run" in tools, "the bash DECOY is present (catches a shell reach)") # a real routed call WITH intent r = rpc(3, "tools/call", {"name": "crankimg__disks", "arguments": {"intent": "validate: recon flashable disks"}}) check(r and not r["result"].get("isError") and "/dev/disk" in text(r["result"]), "routed crankimg__disks (with intent) returns disk recon") # the intent was recorded time.sleep(0.2) logp = os.path.join(HERE, "gate-memories.json") logged = os.path.exists(logp) and "validate: recon flashable disks" in open(logp).read() check(logged, "the intent was RECORDED to the gate memory log before routing") finally: proc.stdin.close(); proc.terminate() if __name__ == "__main__": level1_tools() level1_end_to_end() print("\n\033[1mresult: %d passed, %d failed\033[0m" % (P, F)) print(" LEVEL 2 (agent behaved) = replay the gate intent trail after a run: `gate__query`") sys.exit(1 if F else 0) ===END SOURCE=== THE CALLABLE SURFACE — every callable in the original, lowered with codebean's `lower_node` (#183). The rewrite must present THIS surface (same names, same parameters, same defaults) so callers of the original keep working: ok(msg) bad(msg) check(cond, msg) text(res) call(tool, **args) level1_tools() level1_end_to_end() CAPTURED BEHAVIORAL EXEMPLARS: none. no function in this module is minable — criteria_miner mines only self-contained pure functions (it refuses anything that mentions a world-touching or non-deterministic name, plus methods and no-arg functions), and nothing here qualified This does NOT weaken the ticket's bar — it MOVES the work: you must write the equivalence exemplars by hand from the original's behavior before rewriting, and they are what settles this ticket. THE STANDARD — #200's four requirements, which this ticket exists to enforce. All four are the standard's own words; none is optional: 1. BEHAVIOR-PRESERVING, VERIFIED NOT ASSERTED — the deliverable is ENTIRELY NEW source code (rewritten, not copied). It must reproduce the ORIGINAL's behavior, and the proof is the exemplars above (captured via codebean lower_* #183 + criteria_miner IO-exemplars, exactly as #200 specifies). 2. A PYTHON MODULE — regardless of the source language. This module's source is `.py`; the rewrite is Python either way. 3. INSTALLABLE — a proper package: pyproject.toml, `pip install -e .` works, console entry points where applicable. WRITE ONLY YOUR OWN MODULE FILE at forge2/<repo>/<module>.py. Do NOT hand-author pyproject.toml or __init__.py: the harness DERIVES them ADDITIVELY from the modules present (congruency/tools/ forge_package.py, #410 — the analogue of the tag-registry fix #281), so your unit never conflicts with, or clobbers, a sibling module of the same forge2/<repo>/ package. A package-level pyproject includes every .py in the directory, so your module installs and imports as `<repo>.<module>` the moment it lands. 4. CONTROL-CONFIG'D — NO hardcoded paths or params. A JSON/registry config object drives it (the install.json / registry.json idiom of this repo), and per note-for-claude a tool that cannot see its config THROWS rather than guessing a default. Add ONLY your own `forge2.<repo>.<module>` section to the package's registry.json (create the file if it is absent); NEVER read, edit or drop a sibling module's section — the sections are disjoint keys and stay additive that way. ACCEPTANCE (#200's, per rewrite ticket) — this ticket settles when ALL of: a. the new source reproduces every captured exemplar above, plus the edge cases you add for what sampling cannot reach (paste the runs as evidence); b. it presents the callable surface above, so existing callers keep working; c. `pip install -e .` installs it clean from its own pyproject.toml; d. it runs off a config object with ZERO hardcoded settings, and throws without one. Kind is `build` (a pair: coder + adversarial tester). This ticket hangs off #200 by a `spawned` edge, so a stop ticket aimed at #200 brakes every rewrite in this programme. Filed mechanically by checkouts/current/congruency/tools/forge_refactor_ticketer.py under ticket #201, from the forge bare repos at /Users/stevenpeterson/code/installed/jazz-project/config/forge/remotes. Its original source is embedded above (#346), so this ticket is complete on its own.

— browse congruency source (forge-parked tree)