Synaxi predict
Wraps subagent dispatch with cost prediction, model selection, and automatic actuals recording. Use this INSTEAD OF calling the Agent tool directly whenever delegating a coding task to a subagent. Pass the full task description as args.From its SKILL.md
npx -y skills add BeadW/synaxi-predict --skill synaxi-predictAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 4 stars4 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.0 KB, as published. Nobody here has run it
Cost prediction
python3 -m predictor.predict "$ARGUMENTS" --models single --repo-path "$(pwd)" 2>/dev/null \
|| echo "WARNING: synaxi-predict not installed — restart Claude Code to trigger setup"
Parse the block above. Store internally:
- PRED_ID — 8-char hex from the
Prediction ID:line (noneif prediction failed) - PREDICT_BIN — path to the
predictbinary used (derive siblings fromdirname) - REPO_PATH —
$(pwd)used in the prediction command above
Write prediction state so the recording step can find it later:
mkdir -p /tmp/synaxi
printf '{"pred_id":"%s","repo_path":"%s","predict_bin":"%s"}\n' \
"PRED_ID" "REPO_PATH" "PREDICT_BIN" \
> /tmp/synaxi_${CLAUDE_SESSION_ID}.json
Model selection
If PRED_ID is none, skip to Execute using model haiku.
Otherwise use AskUserQuestion with one option per model row in the prediction table:
- label: model name (e.g.
single-haiku) - description:
Est. $X.XX · N turns · X% pass
Include a final option: label "Just predict — don't execute", description "Stop here".
If the user picks "Just predict", stop.
Execute
Dispatch the task agent with the chosen model:
Agent(
model = "haiku" or "sonnet" (from selection above),
prompt = "Working in REPO_PATH.\n\nTask: $ARGUMENTS\n\nComplete the task using all available tools. Summarise what you changed and whether it succeeded."
)
After Agent returns, extract AGENT_ID from agentId: <hex> in the result metadata.
Parse metrics
PARSE="$(dirname "PREDICT_BIN")/parse-session"
"$PARSE" AGENT_ID REPO_PATH
Extract from the JSON output:
- ACTUAL_TURNS from
turns - ACTUAL_COST from
cost_usd
Eval
Dispatch a second Agent to assess success independently:
Agent(
prompt = "Task attempted: $ARGUMENTS\nRepo: REPO_PATH\n\n1. Run: git diff HEAD\n2. Run tests if any exist (pytest -q, npm test, etc.)\n3. Assess whether the task requirements are met.\n\nRespond with ONLY valid JSON, no markdown:\n{\"passed\": true, \"reason\": \"one sentence\"}"
)
Parse PASSED and REASON from the JSON.
Record
RECORD="$(dirname "PREDICT_BIN")/record-actual"
"$RECORD" PRED_ID --turns ACTUAL_TURNS --cost ACTUAL_COST --passed PASSED
Report to the user:
Recorded PRED_ID:
Turns: ACTUAL_TURNS (predicted: X)
Cost: $ACTUAL_COST (predicted: $X.XX)
Passed: PASSED — REASON