Config

Role

Interactively edit one run's configuration. Two surfaces: (1) run-level settings that state.json is authoritative for (profile.*, parallelism, thresholds, naming) - written through bin/state.sh and mirrored into config.yaml as a human record; (2) the bucket role display labels and other record-only knobs that live in config.yaml; (3) the active SOURCE/SINK connector contract (source-connector.md / sink-connector.md), which you can swap to a shipped template or re-author via interview. This skill applies changes FORWARD only and refuses edits that would corrupt work already committed (a synthesize/apply already in flight).

Invocation

/claude-migrate:config [<run-name>] [<area>]

• <run-name> optional. If omitted, auto-detect: list .planning/claude-migrate/ and if exactly one run dir exists, use it; otherwise print available runs and exit.
• <area> optional shortcut: one of tier | parallelism | thresholds | naming | buckets | source | sink. If omitted, present the full menu (Step 2).

Protocol

Step 1: Locate the run + snapshot current config

• Parse <run-name> and the optional <area> from $ARGUMENTS.
• If <run-name> absent: ls .planning/claude-migrate/ - if exactly one dir, use it; else print available runs and exit.
• RUN_PATH=".planning/claude-migrate/<run-name>".
• If $RUN_PATH/state.json is missing, STOP: "No such run. Initialize with /claude-migrate:init <run-name> first."
• Read the current values to show the user what they are changing FROM:

CUR_STEP=$(bash ${CLAUDE_PLUGIN_ROOT}/bin/state.sh get "$RUN_PATH" .current_step)
jq '{profile, decisions:{naming_convention:.decisions.naming_convention}, input:.input.mode, output:.output.mode}' "$RUN_PATH/state.json"
sed -n '1,200p' "$RUN_PATH/config.yaml"

Step 2: Pick an area (AskUserQuestion, unless <area> was given)

Use AskUserQuestion with header "Which setting do you want to change?" and options:

• Profile tier (model choice + /ultra gate rigor + suggested parallelism)
• Parallelism (worker fan-out; seed pacing)
• Thresholds (cost guardrails + max brief size + copy-page inline limits)
• Naming (how migrated chats are titled)
• Bucket labels (display names for the four roles)
• Source connector (swap or re-author the export/live contract)
• Sink connector (swap or re-author the copy-page/browser contract)

Step 3: Forward-only safety check (refuse unsafe edits)

Before writing, gate on CUR_STEP:

• Tier / models: if CUR_STEP == "synthesize" or later (build-page, verify-gate, ready, pre-apply-gate, apply, finalize, done), REFUSE a tier change: "synth_model and the gate rigor are already committed for this run; a new tier applies forward only and there is no forward work left to apply it to. Start a new run if you need a different tier."
• Parallelism: safe to change any time the relevant queue has NOT fully drained; it only affects future launch-worker.sh terminals. Already-running workers keep the model and count they started with - say so.
• Thresholds: cost guardrails are read at the G-COST gate; refuse to lower hard_stop_usd retroactively if decisions.cost_acknowledged == true and CUR_STEP is past filter-gate (the cost was already approved). max_brief_tokens / inline limits are safe to change before distill / build-page respectively; warn if changed after, because existing briefs/pages will not be regenerated unless re-run.
• Naming: safe before distill; after distill the briefs already hold derived titles, so warn that a change only affects re-distilled units.
• Bucket labels: always safe (display-only).
• Connector swap: REFUSE if CUR_STEP is past split (units are already extracted under the current source contract; changing the source mid-run would desync the corpus). For sink: refuse a swap once apply has started. Before split, a swap is free. If a change is refused, explain why and stop - do NOT partially apply.

Step 4: Apply the change

Tier

Validate against the closed enum {small, medium, large, xl}; reject others with the valid list. Set the whole profile sub-object so downstream skills read it without re-deriving. Per-tier values:

tier	preflight_model	distill_model	synth_model	validator_model	ultra_gate_tier	parallelism	seed_parallelism
small	haiku	haiku	sonnet	sonnet	--small	1-2	1
medium	haiku	sonnet	sonnet	opus	--medium	2-3	1
large (default)	haiku	sonnet	opus	opus	--large	4	1
xl	haiku	opus	opus	sonnet	--xl	5-10	1

Cross-model rule (runtime-enforced by verify): validator_model != distill_model. The table satisfies it; if you ever edit the table re-verify it. seed_parallelism stays 1 in v0.1.0 regardless of tier (in-session serial apply). Write:

bash ${CLAUDE_PLUGIN_ROOT}/bin/state.sh set "$RUN_PATH" .profile '<full-profile-json-object>'

Parallelism

• profile.parallelism (integer >= 1): worker terminals for preflight/distill.
• profile.seed_delay_ms (integer >= 0): pacing between seed submissions in browser apply.
• profile.seed_parallelism: MUST remain 1 in v0.1.0. If the user asks for more, refuse and explain it is reserved for the future CDP-library path.

bash ${CLAUDE_PLUGIN_ROOT}/bin/state.sh set "$RUN_PATH" .profile.parallelism <N>
bash ${CLAUDE_PLUGIN_ROOT}/bin/state.sh set "$RUN_PATH" .profile.seed_delay_ms <MS>

Thresholds

State.json profile fields: max_brief_tokens, inline_card_limit, inline_byte_limit, ok_wait_ms, breaker_threshold, capture_screenshots. Cost guardrails (warn_usd, hard_stop_usd, warn_chat_count, warn_chat_tokens) live in config.yaml under cost: and are read at the G-COST gate. If the user turns capture_screenshots ON, WARN: per-attempt screenshots may capture PII; they are excluded from git by the run-dir .gitignore, and a banner warning fires in apply.

bash ${CLAUDE_PLUGIN_ROOT}/bin/state.sh set "$RUN_PATH" .profile.max_brief_tokens <N>

Then mirror the cost block into config.yaml (Step 5).

Naming

• decisions.naming_convention: keep (default; derive a title only when the original name is empty/generic) or custom:<scheme>. For custom, drive a structured pick with a prefilled worked example (Name DD.MM tag) via AskUserQuestion, then store the resolved scheme string.

bash ${CLAUDE_PLUGIN_ROOT}/bin/state.sh set "$RUN_PATH" .decisions.naming_convention '<keep|custom:SCHEME>'

Bucket labels (display-only)

The four ROLES are a closed enum (GROUPED | STANDALONE | REFERENCE | DROP); you may relabel how each is shown but NEVER add or remove a role. Edit the bucket_labels: map in config.yaml only (the core switches on the role enum, never on the label). Read the current labels, ask for new strings, and write config.yaml.

Source / Sink connector (swap or re-author)

A connector is a markdown CONTRACT (not code) copied into the run dir. Two paths via AskUserQuestion ("Swap to a shipped template, or re-author the current one?"):

1. SWAP to a shipped template - copy it over the run's connector (overwrites the user-editable copy in the run dir, never the shipped template):

   # SOURCE options:  export-file (default) | browser (live)
   cp ${CLAUDE_PLUGIN_ROOT}/templates/sources/export-file.md "$RUN_PATH/source-connector.md"
   # SINK options:    copy-page (the reliable floor) | browser (accelerator)
   cp ${CLAUDE_PLUGIN_ROOT}/templates/sinks/browser.md "$RUN_PATH/sink-connector.md"
   

   If the swap changes input.mode (export <-> live) or output.mode (copy-page <-> auto), update the matching state.json field too:

   bash ${CLAUDE_PLUGIN_ROOT}/bin/state.sh set "$RUN_PATH" .input.mode '<export|live>'
   bash ${CLAUDE_PLUGIN_ROOT}/bin/state.sh set "$RUN_PATH" .output.mode '<copy-page|auto>'
   

2. RE-AUTHOR via interview - a connector.md MUST implement all 7 ops for its side. Interview the user for each section and Write the file. Required SOURCE sections (§5.1): enumerate, extract_unit, extract_projects, unit_project_ref, account_check, citation_anchor, forbidden_fields. Required SINK sections (§5.2): prepare, dedupe_probe, create_project, seed_unit, finalize_unit, finalize_run, rate_limit_check. Open the file with the copy-first banner and, for browser/live sources, the mandatory secret-strip note. Never hardcode a secret in a connector - reference an env var. After re-authoring, advise re-running /claude-migrate:run so /ultra Gate 1 re-validates the new contract.

Step 5: Mirror record-only fields + confirm

For any state.json field that also appears in config.yaml (tier, parallelism, thresholds, naming, cost block, bucket_labels), update the config.yaml copy so the human record stays consistent (state.json remains authoritative for the fields it tracks). Then print a confirmation:

Config updated for run <run>:
  <field>: <old> -> <new>
  ...
Effect: applies on the NEXT step. Already-running workers keep their started settings.
Next action: /claude-migrate:run <run>   (or /claude-migrate:progress <run> to review)

Hard rules

• Forward-only. NEVER alter completed work or apply a tier/model change once synthesize has committed it. Refuse with a clear reason rather than corrupting state.
• All state.json mutations go through bin/state.sh (locked, atomic, injection-defended). NEVER hand-edit state.json.
• config.yaml is a human record for the fields state.json also tracks; on conflict state.json wins. Keep them consistent after every change.
• The four bucket ROLES are a closed enum. You may relabel them in config.yaml; you may NEVER add or remove a role.
• seed_parallelism stays 1 in v0.1.0. Refuse any higher value.
• A re-authored connector MUST implement all 7 ops for its side and MUST NOT hardcode a secret (reference env vars). Browser/live sources MUST keep the mandatory secret-strip note.
• Never reference any specific domain in defaults, examples, or labels; use neutral placeholders (Project Alpha, topic-1).