mini-coder 0.5.10 → 0.5.12

package/BENCHMARK.md

@@ -0,0 +1,408 @@
+# Benchmark workflow
+
+Use this as the default tuning loop for `mini-coder` on Terminal-Bench.
+
+The goal is not to rerun the whole benchmark after every change. The goal is to get fast enough feedback that small prompt / loop / tool-use changes can be judged quickly, then promote only the promising ones to bigger runs.
+
+Terminal-Bench is a signal, not the product. The target is a better coding agent, not a higher benchmark score from benchmark-shaped patches.
+
+## Principles
+
+- Keep changes small.
+- Change one thing at a time.
+- Compare against a fresh baseline, not an old leaderboard run.
+- Use fast suites for iteration, broad suites for promotion.
+- Keep structured `mc --json` logs in trial artifacts so behavior can be analyzed.
+- Optimize for general coding-agent behavior first.
+- Use benchmark failures to extract general behavior gaps, not to encode benchmark lore into the agent.
+- Prefer generic improvements over task-named patches, reminders, or stop-time nudges.
+
+## Suites
+
+### 1. Guardrail suite
+
+Use this to catch regressions on known-good behavior.
+
+Tasks:
+
+- `cancel-async-tasks`
+- `bn-fit-modify`
+- `sparql-university`
+- `extract-elf`
+
+Settings:
+
+- `1` attempt
+- `2` concurrent
+- `0` retries
+
+Estimated wall time:
+
+- about `8–10 minutes`
+
+### 2. Fast suite
+
+Use this after almost every small change.
+
+Tasks:
+
+- `polyglot-rust-c`
+- `overfull-hbox`
+- `gcode-to-text`
+- `mteb-retrieve`
+- `dna-insert`
+- `torch-tensor-parallelism`
+
+Settings:
+
+- `2` attempts
+- `2` concurrent
+- `0` retries
+
+Estimated wall time:
+
+- about `44 minutes`
+
+### 3. Focused suite
+
+Use this only if a change looks good on the fast suite.
+
+Tasks:
+
+- `polyglot-rust-c`
+- `overfull-hbox`
+- `gcode-to-text`
+- `mteb-retrieve`
+- `torch-tensor-parallelism`
+- `filter-js-from-html`
+- `configure-git-webserver`
+- `qemu-alpine-ssh`
+
+Settings:
+
+- `2` attempts
+- `2` concurrent
+- `0` retries
+
+Estimated wall time:
+
+- about `1.1 hours`
+
+### 4. Broad promotion suite
+
+Use this after 1–3 promising changes, or overnight.
+
+Options:
+
+- rerun recent failures
+- wider local failure suite
+- full leaderboard-style run
+
+Typical settings for rerunning recent failures:
+
+- `4` attempts
+- `2` concurrent
+- `0` retries
+
+## Baseline procedure
+
+Before changing code:
+
+1. run the guardrail suite
+2. run the fast suite
+3. run the fast suite again
+
+That gives a fresh local baseline for:
+
+- pass count
+- runtime
+- variance / noise
+
+Do not compare a new change only against an old run from days ago if a fresh same-HEAD baseline is available.
+
+## Experiment quality bar
+
+Before changing code, write the hypothesis in two layers:
+
+1. the benchmark symptom
+2. the general coding-agent behavior gap behind it
+
+Only run an experiment if you can answer all of these:
+
+- what general behavior is being improved?
+- why should that help outside Terminal-Bench?
+- what would make this change obviously overfit?
+
+Good experiment themes:
+
+- verification-equivalence before completion
+- artifact-grounded verification of final outputs
+- preferring a task-named local source of truth over approximations
+- reducing shell thrash before first meaningful verification
+
+Reject or redesign experiments that:
+
+- depend on benchmark-specific task names, file names, package names, or tool names in product logic
+- inject reminders or guards keyed to one benchmark noun unless that rule maps cleanly to a real product behavior
+- only make sense because a particular verifier is known
+- cannot be explained without citing a single task transcript
+
+Hard rule:
+
+- no task-specific nouns in agent logic unless they map to a real product feature
+
+## Iteration loop
+
+For each change:
+
+1. inspect the last fast / focused failures
+2. translate them into **one** general behavior gap
+3. reject benchmark-shaped ideas; if you cannot phrase the change without task-specific nouns, keep diagnosing
+4. if the change depends on a dynamic trigger, confirm that the trigger actually appears in the target failures
+5. form **one** narrow hypothesis
+6. make **one** small change
+7. run:
+   - guardrails
+   - fast suite
+8. compare to baseline
+9. decide:
+   - keep
+   - revert
+   - refine
+
+Only run the focused suite if:
+
+- the fast suite improved, and
+- the guardrails did not regress
+
+Only run the broad promotion suite if:
+
+- the focused suite also looks good, or
+- enough promising changes have accumulated to justify it
+
+## Decision rules
+
+Treat the fast suite as a noisy but useful signal.
+
+For the 12-trial fast suite:
+
+- `+2` or more passes: probably meaningful improvement
+- `-2` or more passes: probably meaningful regression
+- `±1`: likely noise unless it repeats
+
+Keep a change if:
+
+- the fast suite improves materially
+- guardrails do not regress
+- the result repeats on another fast-suite rerun
+- the mechanism is still a general agent-quality improvement, not just a benchmark-specific patch
+
+Revert a change if:
+
+- it clearly loses fast-suite passes, or
+- it regresses guardrails
+
+Call it inconclusive if:
+
+- the delta is tiny and does not repeat
+
+Also prefer:
+
+- a smaller, more general improvement over a larger but obviously benchmark-specific patch
+- a clearly exercised mechanism over a theory that never fired in the target runs
+
+## Failure buckets to optimize against
+
+Use these as the main buckets when inspecting logs:
+
+### Exact-contract / cleanup misses
+
+Examples:
+
+- right artifact, wrong path
+- right output plus extra junk
+- correct core work, but violates a strict file / in-place requirement
+
+### Verification-equivalence / premature completion
+
+Examples:
+
+- waited too long to run a meaningful check
+- never checked the exact contract
+- stopped after a weaker non-equivalent check
+- verified the intended design instead of the written artifact
+
+### Source-of-truth selection gaps
+
+Examples:
+
+- task names an exact local tool/package/interface, but the agent uses an approximation instead
+- lower-level library reasoning replaces package-local or task-local semantics
+
+### Over-exploration / shell thrash
+
+Examples:
+
+- too many shell / read steps before first write
+- repeated discovery commands without narrowing the problem
+
+### Persistent correctness gaps
+
+Examples:
+
+- `torch-tensor-parallelism`
+- `filter-js-from-html`
+- `configure-git-webserver`
+
+### Long-horizon / timeout-heavy tasks
+
+These are useful for promotion runs, not tight inner loops.
+Examples:
+
+- `gpt2-codegolf`
+- `query-optimize`
+- `winning-avg-corewars`
+- `train-fasttext`
+
+## Behavior analysis requirements
+
+Behavior analysis depends on structured agent logs.
+
+Keep wrappers on:
+
+- `mc --json -p ...`
+
+Per trial, keep:
+
+- result JSON
+- verifier output
+- exception type
+- agent stderr
+- structured `agent/mini-coder.ndjson`
+- timestamps
+
+With those artifacts, analyze things like:
+
+- time to first tool call
+- time to first edit
+- number of `shell` / `read` / `grep` / `edit` calls
+- whether the agent ran a verifier-like shell command
+- whether it used the named local source of truth when one was available
+- whether it verified the final artifact it actually wrote
+- whether a dynamic intervention visibly exercised
+- whether it had a local pass before ending
+- whether it kept changing files after a pass-worthy state
+- whether it left extra artifacts
+
+## Commands
+
+### Guardrail suite
+
+```bash
+GUARDRAIL_TASKS=(
+  cancel-async-tasks
+  bn-fit-modify
+  sparql-university
+  extract-elf
+)
+
+cmd=(
+  harbor run -y
+  --job-name local-guardrails-$(date +%F__%H-%M-%S)
+  --jobs-dir "$PWD/terminal-bench/jobs"
+  --agent-import-path mini_coder_local_agent:MiniCoderLocalAgent
+  --model openai-codex/gpt-5.4
+  --dataset terminal-bench@2.0
+  --n-attempts 1
+  --n-concurrent 2
+  --max-retries 0
+)
+
+for t in "${GUARDRAIL_TASKS[@]}"; do
+  cmd+=(--include-task-name "$t")
+done
+
+PYTHONPATH="$PWD/terminal-bench" "${cmd[@]}"
+```
+
+### Fast suite
+
+```bash
+FAST_TASKS=(
+  polyglot-rust-c
+  overfull-hbox
+  gcode-to-text
+  mteb-retrieve
+  dna-insert
+  torch-tensor-parallelism
+)
+
+cmd=(
+  harbor run -y
+  --job-name local-fast-$(date +%F__%H-%M-%S)
+  --jobs-dir "$PWD/terminal-bench/jobs"
+  --agent-import-path mini_coder_local_agent:MiniCoderLocalAgent
+  --model openai-codex/gpt-5.4
+  --dataset terminal-bench@2.0
+  --n-attempts 2
+  --n-concurrent 2
+  --max-retries 0
+)
+
+for t in "${FAST_TASKS[@]}"; do
+  cmd+=(--include-task-name "$t")
+done
+
+PYTHONPATH="$PWD/terminal-bench" "${cmd[@]}"
+```
+
+### Focused suite
+
+```bash
+FOCUS_TASKS=(
+  polyglot-rust-c
+  overfull-hbox
+  gcode-to-text
+  mteb-retrieve
+  torch-tensor-parallelism
+  filter-js-from-html
+  configure-git-webserver
+  qemu-alpine-ssh
+)
+
+cmd=(
+  harbor run -y
+  --job-name local-focus-$(date +%F__%H-%M-%S)
+  --jobs-dir "$PWD/terminal-bench/jobs"
+  --agent-import-path mini_coder_local_agent:MiniCoderLocalAgent
+  --model openai-codex/gpt-5.4
+  --dataset terminal-bench@2.0
+  --n-attempts 2
+  --n-concurrent 2
+  --max-retries 0
+)
+
+for t in "${FOCUS_TASKS[@]}"; do
+  cmd+=(--include-task-name "$t")
+done
+
+PYTHONPATH="$PWD/terminal-bench" "${cmd[@]}"
+```
+
+## Minimal experiment log format
+
+Keep this in `PROGRESS.md`, a final summary for each completed change:
+
+- benchmark symptom
+- general behavior gap
+- why this should help outside Terminal-Bench
+- overfitting risk / why this is still general
+- trigger evidence, if relevant
+- hypothesis
+- files changed
+- local verification
+- guardrail result
+- fast-suite result
+- focused-suite result, if run
+- mechanism exercised?
+- keep / revert / refine (Make the decision very visible in the file).

package/PROGRESS.md

@@ -0,0 +1,5 @@
+# CURRENT TASK PROGRESS
+
+- Benchmark loop stopped.
+- In-progress experiment changes reverted.
+- Progress reset for a fresh benchmark restart.

package/README.md

@@ -39,9 +39,11 @@ $ mc
 
 ## Tools
 
-Four built-in tools, plus a read-only image tool:
+Six built-in tools, plus a conditional read-only image tool:
 
 - **`shell`** — runs commands in the user's shell. Returns stdout, stderr, and exit code. Large output is truncated to protect model context.
+- **`read`** — reads UTF-8 text files from disk, optionally by line window.
+- **`grep`** — searches file contents with ripgrep-style options and returns structured matches.
 - **`edit`** — exact-text replacement in a single file. Fails deterministically if the target is missing or ambiguous. Creates new files when old text is empty.
 - **`todoWrite`** — creates or updates the session todo list incrementally and returns the full current snapshot.
 - **`todoRead`** — returns the full current session todo list snapshot.
@@ -123,6 +125,17 @@ bun run format
 bun run typecheck
 ```
 
+## Also makes LLMs smarter
+
+LLMs famously tell you to walk 50 meters to the car wash — forgetting the car needs to be there too. Not on our watch.
+
+<table align="center">
+  <tr>
+    <td><img src="assets/mc-claude-smart.png" alt="Claude correctly answering the car wash question" width="400" /></td>
+    <td><img src="assets/mc-gpt-smart.png" alt="GPT correctly answering the car wash question" width="400" /></td>
+  </tr>
+</table>
+
 ## License
 
 MIT

package/benchmark-loop.sh

@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+stop() {
+  echo "stopping loop"
+  exit 0
+}
+
+trap stop INT TERM
+
+while :; do
+  echo ""
+  echo "> Running step"
+  echo ""
+  mc -p "See BENCHMARK.md and PROGRESS.md first. Do the next step in the process. Once the step is complete update PROGRESS.md"
+
+  sleep 60
+done
+

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mini-coder",
-  "version": "0.5.10",
+  "version": "0.5.12",
   "description": "A small, fast CLI coding agent",
   "module": "src/index.ts",
   "type": "module",

package/src/agent.ts

@@ -623,6 +623,10 @@ function getTodoReminderSignature(
   );
 }
 
+function wrapSystemReminder(content: string): string {
+  return `<system_reminder>\n${content}\n</system_reminder>`;
+}
+
 function createTodoReminderMessage(
   messages: readonly Message[],
   remindedTodoSignatures: Set<string>,
@@ -651,7 +655,7 @@ function createTodoReminderMessage(
 
   return {
     role: "user",
-    content: lines.join("\n"),
+    content: wrapSystemReminder(lines.join("\n")),
     timestamp: Date.now(),
   };
 }

package/src/headless.ts

@@ -22,32 +22,64 @@ type HeadlessStopReason = "stop" | "length" | "error" | "aborted";
 /** Options for a headless NDJSON run. */
 export interface HeadlessRunOptions {
   /** Optional line writer for completed NDJSON event output. */
-  writeLine?: (line: string) => void;
+  writeLine?: (line: string) => void | Promise<void>;
 }
 
 /** Options for a headless final-text run. */
 export interface HeadlessTextRunOptions {
   /** Optional writer for the final assistant text output. */
-  writeText?: (text: string) => void;
+  writeText?: (text: string) => void | Promise<void>;
 }
 
 interface HeadlessOutputController {
-  /** Write text to stdout with broken-pipe handling. */
+  /** Queue text for stdout with broken-pipe handling. */
   write(text: string): void;
   /** Attach SIGINT/stdout error handlers for the active run. */
   attach(): void;
   /** Remove SIGINT/stdout error handlers after the run. */
   detach(): void;
-  /** Resolve the final stop reason, converting broken pipes into quiet shutdowns. */
-  finalize(stopReason: HeadlessStopReason): HeadlessStopReason;
+  /** Wait for queued writes and resolve the final stop reason. */
+  finalize(stopReason: HeadlessStopReason): Promise<HeadlessStopReason>;
 }
 
 // ---------------------------------------------------------------------------
 // Helpers
 // ---------------------------------------------------------------------------
 
-function defaultWrite(text: string): void {
-  process.stdout.write(text);
+function defaultWrite(text: string): Promise<void> {
+  return new Promise((resolve, reject) => {
+    let settled = false;
+
+    const cleanup = (): void => {
+      process.stdout.off("error", handleError);
+    };
+
+    const settle = (callback: () => void): void => {
+      if (settled) {
+        return;
+      }
+      settled = true;
+      cleanup();
+      callback();
+    };
+
+    const handleError = (error: unknown): void => {
+      settle(() => {
+        reject(error);
+      });
+    };
+
+    process.stdout.on("error", handleError);
+    try {
+      process.stdout.write(text, () => {
+        settle(resolve);
+      });
+    } catch (error) {
+      settle(() => {
+        reject(error);
+      });
+    }
+  });
 }
 
 function isBrokenPipeError(error: unknown): boolean {
@@ -126,10 +158,11 @@ function shouldWriteHeadlessJsonEvent(event: AgentEvent): boolean {
 
 function createHeadlessOutputController(
   state: AppState,
-  writeImpl: (text: string) => void,
+  writeImpl: (text: string) => void | Promise<void>,
 ): HeadlessOutputController {
   let brokenPipe = false;
   let outputError: unknown = null;
+  let pendingWrite = Promise.resolve();
   const sigintHandler = createSigintHandler(state);
 
   const stopForBrokenPipe = (): void => {
@@ -140,30 +173,37 @@ function createHeadlessOutputController(
     state.abortController?.abort();
   };
 
-  const stdoutErrorHandler = (error: unknown): void => {
+  const failOutput = (error: unknown): void => {
     if (isBrokenPipeError(error)) {
       stopForBrokenPipe();
       return;
     }
 
-    outputError = error;
+    outputError = outputError ?? error;
     state.abortController?.abort();
   };
 
+  const stdoutErrorHandler = (error: unknown): void => {
+    failOutput(error);
+  };
+
   return {
     write(text) {
-      if (brokenPipe) {
+      if (brokenPipe || outputError) {
         return;
       }
 
-      try {
-        writeImpl(text);
-      } catch (error) {
-        if (!isBrokenPipeError(error)) {
-          throw error;
+      pendingWrite = pendingWrite.then(async () => {
+        if (brokenPipe || outputError) {
+          return;
         }
-        stopForBrokenPipe();
-      }
+
+        try {
+          await writeImpl(text);
+        } catch (error) {
+          failOutput(error);
+        }
+      });
     },
     attach() {
       process.stdout.on("error", stdoutErrorHandler);
@@ -173,7 +213,8 @@ function createHeadlessOutputController(
       process.stdout.off("error", stdoutErrorHandler);
       process.off("SIGINT", sigintHandler);
     },
-    finalize(stopReason) {
+    async finalize(stopReason) {
+      await pendingWrite;
       if (outputError) {
         throw outputError;
       }
@@ -222,7 +263,7 @@ export async function runHeadlessPrompt(
       state,
       hooks,
     );
-    return output.finalize(stopReason);
+    return await output.finalize(stopReason);
   } finally {
     output.detach();
   }
@@ -271,7 +312,7 @@ export async function runHeadlessPromptText(
     if (finalText.length > 0) {
       output.write(finalText);
     }
-    return output.finalize(stopReason);
+    return await output.finalize(stopReason);
   } finally {
     output.detach();
   }