← Bot runs

adr-rechallenge (ReArchi) — bilans

Human-in-the-loop ADR re-challenger. Loads one ADR + the current code, frames fresh arguments (changed assumptions, matured alternatives, dependency/code drift, triggered consequences), and asks the operator: keep / change / addendum. change files a backlog ticket; addendum appends a dated note to the ADR, then a second commit-or-skip gate (the note is optional). Uses interaction: human (not the heavier interaction: review); claude_code only (no forfait dependency).

2026-09-17 — profile 2: a full re-challenge on a one-ADR repository, the addendum committed (run 01a0af95-64fb)

• Status: validated end to end — load_adr → survey_code → frame_arguments → human_decision (addendum) → write_addendum → human_commit_gate (commit) → commit_changes → done.
• Versions: bot adr-rechallenge 0.1.1 (dsl: 2, wave 3b of #1344) · iterion 5fa790d82 for the bots; the engine the branch binary v3.157.0+c31559517 (built on the wave-3a branch from main at v3.157.0; main is at v3.159.0 as this is written, one PR of engine work ahead), served through the host's Anthropic-compatible facade (z.ai) as the runs' provenance records.
• Method: CLI iterion run <bundle>/main.bot launched FROM a scratch greeter repository (seven files, one ADR forbidding third-party imports, a requirements.txt pinning requests that nothing consumes), --var adr_path=docs/adr/0001-standard-library-only.md plus a one-line scope_notes, --store-dir the operator's workspace store, --sandbox none, --merge-into none, caps --max-cost-usd 1.5 --max-duration 10m, every LLM node on claude_code/claude-opus-5 (the GPT forfait is closed until 2026-09-20). The two gates answered with iterion resume --answer decision=addendum … then --answer commit=true ….
• Result: three LLM nodes, $1.10, 9 min 54 s of wall including the two pauses. survey_code found the one real signal (the pin is attributed to "optional tooling scripts" that do not exist in the tree), frame_arguments argued keep / change / addendum without pre-deciding ("no signal worth a change proposal"), write_addendum produced a dated eight-line note that reconfirms the decision and names the record error, committed as 9c65ec1 on the run worktree (detached HEAD, one file changed).
• Value: the ten prompts' paragraph breaks reach the models — the run's events.jsonl carries the authors' blank lines as \n\n for seven of the eight prompts visited nodes rendered (addendum_user has no paragraph break to witness; ticket_* belong to the change route, not taken). The bot's contract holds: the human owns the decision, the addendum corrects the record instead of re-litigating the ADR.
• Findings / misses: none bot-side. A resume does not carry --merge-into none — finalize reported "merge into main pending UI confirmation" after the second gate; nothing merged.
• Engine hardening: none needed.
• Lessons for next run: a repository with one ADR and one commit exercises every node; the addendum commit lives on the run worktree — read it there, or launch with --merge-into.

2026-06-14 — first dogfood, re-challenge ADR-008 (run 019ec5bc)

• Status: validated — full human-in-the-loop cycle exercised end-to-end; one commit-message bug found + fixed.
• Versions: bot 0.1.0 · iterion @ 92ccd136 (main)
• Method: iterion run bots/adr-rechallenge/main.bot --var adr_path=docs/adr/008-bot-golden-replay-framework.md --var workspace_dir=<worktree> (contained to the dogfood worktree, not main, since ReArchi has no worktree: auto and edits the live tree). claude_code (host OAuth) only — no claw/forfait nodes, so no token risk. Driven interactively: launch → pause → iterion resume --answers-file.
• Result: full cycle, all branches reachable. load_adr → survey_code → frame_arguments → pause at human_decision (operator chose addendum) → derive_decision → write_addendum → pause at human_commit_gate (operator chose commit) → derive_commit → commit_changes → done. The two human gates rendered + routed correctly; the launch→pause→resume→pause→resume flow worked via --answers-file.
• Value: high — honest, evidence-based re-challenge. Survey verified the ADR's claims against the current code with file:line evidence: the api.APIClient injection seam still doesn't exist (executor.go:377-524), so rejected Alternative 1 stays infeasible exactly as the ADR said; the assignee-coupling consequence hasn't fired (set_bot is only the MCP routing tool). It found no change driver and correctly refused to manufacture a change case (the argument-framing skill's anti-fabrication contract held). It surfaced two real record-keeping divergences (stale pre-rename bot names c9996d98/8784d677; four golden fixtures still original 2026-05-29 seeds never regenerated) and wrote a crisp, correctly-formatted addendum. The addendum was landed on main's ADR-008 (a9e65e50).

Findings

#	Finding	Severity	State
1	commit_changes committed the addendum with an unresolved/malformed message (docs(adr): addendum to ADR-{{outputs.load_adr.num}} re-challenge\n\n…) — embedded refs in the with{} edge string don't resolve (only PURE refs do; files: "{{outputs.load_adr.path_list}}" worked), and DSL \n stayed literal	real	FIXED 3a639f27 — write_addendum now emits a complete full_message, the edge passes it as a pure ref (matches the docs-refresh/feature-dev pattern). Live-validation deferred (static validate OK).

Lessons for next run

• Never build a multi-part value with embedded refs inside an edge with{} string — iterion resolves a pure ref ("{{x}}") there but not refs embedded in a longer string, and DSL \n is literal. Build the value in a node (agent/compute) and pass it as a pure ref. (Tool command: templates DO resolve embedded refs — that's why git -C {{input.workspace_dir}} works — so the asymmetry is edge-mappings vs command bodies.)
• Run contained (--var workspace_dir=<scratch> or a throwaway clone) since ReArchi has no worktree: auto and commits to the live tree on the addendum path — otherwise it edits + commits the operator's checked-out ADRs directly.
• The change branch (file a backlog ticket) was not exercised this run — ReArchi correctly judged ADR-008 had no change driver. A future dogfood on an ADR with a genuine maturing alternative would cover it.