tandem-mode 0.2.0 → 0.5.0

package/README.md

@@ -1,204 +1,249 @@
-# Tandem Mode
-
-[![npm version](https://img.shields.io/npm/v/tandem-mode.svg)](https://www.npmjs.com/package/tandem-mode)
-[![license](https://img.shields.io/npm/l/tandem-mode.svg)](./LICENSE)
-[![node](https://img.shields.io/node/v/tandem-mode.svg)](./package.json)
-[![good first issues](https://img.shields.io/github/issues/zdrave08/Tandem-Mode/good%20first%20issue)](https://github.com/zdrave08/Tandem-Mode/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22)
-
-**English** | [Srpski](./README.sr.md)
-
-A terminal coding agent built specifically for the DeepSeek API — not a generic
-LLM harness with DeepSeek bolted on. Interactive REPL, planner/worker/reviewer
-orchestration, cache-aware prompt architecture, and first-class support for
-Windows, macOS, and Linux.
-
-> **Independent open-source project. Not affiliated with or endorsed by
-> DeepSeek.**
-
-## Status
-
-M0–M4 of the [development plan](./deepseek-cli-development-plan.md) are done.
-Published on npm, usable today for real coding tasks against a local repo.
-
-First benchmark run is in — see [full results](./docs/benchmark-results.md).
-Headline: all three configurations (Pro-only, Flash-only, orchestration)
-fixed 4/4 injected bugs correctly in a real WooCommerce plugin, with a clean
-`php -l` and zero unnecessary file changes. But for this class of task
-(single-file, well-scoped fixes), orchestration cost ~9.5× more and took
-~19× longer than Pro-only for the same correctness — the opposite of what
-you'd want orchestration for. Read as: orchestration's overhead needs a
-bigger, more decomposable task to pay for itself, which this first run didn't
-test. Reported as-is, not adjusted to fit expectations.
-
-## Features
-
-- **Agentic loop** — `deepseek-v4-pro`/`deepseek-v4-flash` with tool calling:
-  `read_file`, `list_dir`, `search`, `edit`, `git_diff`, `shell`, `view_image`,
-  `web_search`.
-- **Two entry points** — interactive REPL for normal work, single-shot mode
-  (`tandem "task" --yes`) for scripting and CI.
-- **Orchestration** (`/plan`) — a planner produces a JSON task list, workers
-  execute each task with a fresh minimal context, a reviewer checks the git
-  diff against the original task and either approves or requests corrections
-  (capped by `max_review_loops`).
-- **Cache-aware prompts** — the worker's system prompt + repo map + plan stay
-  byte-identical across calls in one orchestration run, so DeepSeek's context
-  cache actually hits. Measured live: cache-hit tokens grew call over call
-  within a single `/plan` run.
-- **Vision** — `view_image` lets the agent look at a screenshot you point it
-  to; `/paste` grabs an image straight from the clipboard (Windows, macOS,
-  and Linux via X11/`xclip` or Wayland/`wl-clipboard`).
-- **Web search** — isolated calls to DeepSeek's Responses API when the agent
-  needs something outside the repo (an error message, a library's docs).
-- **Session persistence** — every session is a JSONL file; `/resume` continues
-  the latest (or a named) one, `/fork` branches off without touching the
-  original.
-- **Safety** — destructive actions (file edits, shell commands not on a small
-  read-only allowlist) always ask for confirmation; `y`/`n`/`a` where `a`
-  remembers the choice for that tool for the rest of the session. `--yes`
-  bypasses this for CI but always prints what it auto-approved.
-- **Cost control** — usage is read only from the API's own `usage` field,
-  never estimated. A budget (`/budget`, `--budget`) hard-stops the loop before
-  the next API call if exceeded, without losing the session. `/usage` also
-  shows what the session would have cost off-peak, and the REPL warns you at
-  startup if you're currently in a peak-pricing window.
-- **ESC to interrupt** — cancels the in-flight call (and any running shell
-  command) cleanly, keeping whatever was generated so far.
-- **Cross-platform** — the shell tool runs PowerShell natively on Windows and
-  `sh` on macOS/Linux, file edits preserve the original CRLF/LF line ending
-  style, paths are handled with `path.resolve` throughout.
-
-## Quick start
-
-```bash
-npm install -g tandem-mode
-tandem            # interactive REPL
-tandem "explain what this repo does" --yes   # one-shot
-```
-
-Or from source (useful for contributing, or to track `main` instead of the
-latest npm release):
-
-```bash
-git clone https://github.com/zdrave08/Tandem-Mode.git
-cd Tandem-Mode
-pnpm install
-pnpm dev            # interactive REPL
-pnpm dev "explain what this repo does" --yes   # one-shot
-```
-
-On first run you'll be asked for a DeepSeek API key (get one at
-[platform.deepseek.com](https://platform.deepseek.com)) and a default
-model/reasoning effort. The key is stored in your OS's own secure credential
-store — never in a config file or log:
-
-- **Windows** — DPAPI-encrypted (tied to your OS user account) in
-  `%LOCALAPPDATA%\TandemMode\credentials.dat`.
-- **macOS** — the login Keychain, via the built-in `security` CLI.
-- **Linux** — the Secret Service (gnome-keyring, KWallet, or similar), via
-  `secret-tool` (install `libsecret-tools` if it's missing).
-
-`DEEPSEEK_API_KEY` works everywhere and always takes priority, for CI use.
-
-## Usage
-
-```bash
-tandem                              # interactive REPL
-tandem "fix the off-by-one in foo.ts"
-tandem --plan "add dark mode to the settings page"
-tandem "task" --model deepseek-v4-flash --effort low --budget 0.50 --yes
-```
-
-| Flag | Effect |
-|---|---|
-| `--resume` / `-r` | Continue the latest session in this directory |
-| `--yes` / `-y` | Auto-approve destructive actions (prints what it approved) |
-| `--plan` | Run planner/worker/reviewer orchestration instead of a single turn |
-| `--model` | `deepseek-v4-pro` or `deepseek-v4-flash` |
-| `--effort` | `low`, `high`, or `max` reasoning effort |
-| `--budget` | Hard USD cap for the run |
-| `--max-review-loops` | Cap on orchestration correction cycles (default 3) |
-
-### REPL commands
-
-| Command | Effect |
-|---|---|
-| `/plan <task>` | Planner/worker/reviewer orchestration |
-| `/model [name]` | Show or switch model |
-| `/thinking [on\|off]` | Toggle thinking mode |
-| `/effort [low\|high\|max]` | Show or set reasoning effort |
-| `/status` | Current session/settings |
-| `/usage` | Tokens, cost, off-peak projection |
-| `/budget [amount]` | Show or set the session budget |
-| `/new` | Start a fresh session |
-| `/fork` | New session, carrying over the current history |
-| `/resume [id]` | Continue the latest (or a named) session |
-| `/paste [message]` | Send a clipboard image to the agent |
-| `/clear` | Clear the screen |
-| `/help` | This list |
-| `/exit` | Quit |
-
-Press **ESC** at any time during a turn to interrupt it.
-
-## Architecture
-
-```
-tandem (no args)  ─┬─> interactive REPL ─┬─> single turn: agent loop (tools)
-                    │                     └─> /plan: planner → worker(s) → reviewer
-tandem "task"     ──┘   (single-shot, same agent loop, for scripts/CI)
-```
-
-- `src/deepseek/` — thin client for DeepSeek's OpenAI-compatible API
-  (streaming + non-streaming chat completions, vision, web search via
-  Responses API). See [`docs/api-notes.en.md`](./docs/api-notes.en.md)
-  (Serbian original: [`docs/api-notes.md`](./docs/api-notes.md)) for the
-  API behavior this is built against, verified live rather than assumed.
-- `src/agent/` — the tool-calling loop, tool implementations, session
-  persistence, usage/cost tracking.
-- `src/orchestrator/` — planner/worker/reviewer, repo-map generation, the
-  cache-aware stable-prefix construction.
-- `src/repl/` — slash commands, runtime config, the first-run wizard.
-- `src/cli.tsx` — Ink-based terminal UI and the REPL/single-shot entry points.
-
-## Configuration
-
-Layered: **session** (in-REPL commands) > **project** (`.tandem/config.json`
-in the current directory) > **global** (`~/.tandem/config.json`).
-
-```json
-{
-  "defaultModel": "deepseek-v4-pro",
-  "defaultReasoningEffort": "high",
-  "budgetUsd": 5,
-  "maxReviewLoops": 3,
-  "verifyCommand": "pnpm typecheck && pnpm test",
-  "language": "en"
-}
-```
-
-`verifyCommand` runs before each `/plan` review round and its result goes to
-the reviewer. `language` is the UI language (`en` or `sr`, default `en`); the
-`--lang` flag and the `TANDEM_LANG` environment variable override it, in that
-order. The first-run wizard asks for it.
-
-The API key is never stored in these files — it lives in the OS credential
-store (or `DEEPSEEK_API_KEY` env var, which takes priority, for CI use).
-
-## Development
-
-See [CONTRIBUTING.md](./CONTRIBUTING.md).
-
-```bash
-pnpm install
-pnpm typecheck
-pnpm dev "some task"
-```
-
-## Disclaimer
-
-This is an independent, community-run project. It is not affiliated with,
-endorsed by, or sponsored by DeepSeek.
-
-## License
-
-[MIT](./LICENSE)
+# Tandem Mode
+
+[![npm version](https://img.shields.io/npm/v/tandem-mode.svg)](https://www.npmjs.com/package/tandem-mode)
+[![license](https://img.shields.io/npm/l/tandem-mode.svg)](./LICENSE)
+[![node](https://img.shields.io/node/v/tandem-mode.svg)](./package.json)
+[![good first issues](https://img.shields.io/github/issues/zdrave08/Tandem-Mode/good%20first%20issue)](https://github.com/zdrave08/Tandem-Mode/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22)
+
+**English** | [Srpski](./README.sr.md)
+
+A terminal coding agent built specifically for the DeepSeek API — not a generic
+LLM harness with DeepSeek bolted on. Interactive REPL, planner/worker/reviewer
+orchestration, cache-aware prompt architecture, and first-class support for
+Windows, macOS, and Linux.
+
+> **Independent open-source project. Not affiliated with or endorsed by
+> DeepSeek.**
+
+## Status
+
+M0–M4 of the [development plan](./deepseek-cli-development-plan.md) are done.
+Published on npm, usable today for real coding tasks against a local repo.
+
+First benchmark run is in — see [full results](./docs/benchmark-results.md).
+Headline: all three configurations (Pro-only, Flash-only, orchestration)
+fixed 4/4 injected bugs correctly in a real WooCommerce plugin, with a clean
+`php -l` and zero unnecessary file changes. But for this class of task
+(single-file, well-scoped fixes), orchestration cost ~9.5× more and took
+~19× longer than Pro-only for the same correctness — the opposite of what
+you'd want orchestration for. Read as: orchestration's overhead needs a
+bigger, more decomposable task to pay for itself, which this first run didn't
+test. Reported as-is, not adjusted to fit expectations.
+
+## Features
+
+- **Agentic loop** — `deepseek-v4-pro`/`deepseek-v4-flash` with tool calling:
+  `read_file`, `list_dir`, `search`, `edit`, `git_diff`, `shell`, `view_image`,
+  `web_search`.
+- **Two entry points** — interactive REPL for normal work, single-shot mode
+  (`tandem "task" --yes`) for scripting and CI.
+- **Orchestration** (`/plan`) — a planner produces a JSON task list, workers
+  execute each task with a fresh minimal context, a reviewer checks the git
+  diff against the original task and either approves or requests corrections
+  (capped by `max_review_loops`).
+- **Cache-aware prompts** — the worker's system prompt + repo map + plan stay
+  byte-identical across calls in one orchestration run, so DeepSeek's context
+  cache actually hits. Measured live: cache-hit tokens grew call over call
+  within a single `/plan` run.
+- **Vision** — `view_image` lets the agent look at a screenshot you point it
+  to; `/paste` grabs an image straight from the clipboard (Windows, macOS,
+  and Linux via X11/`xclip` or Wayland/`wl-clipboard`).
+- **Web search** — isolated calls to DeepSeek's Responses API when the agent
+  needs something outside the repo (an error message, a library's docs).
+- **Session persistence** — every session is a JSONL file; `/resume` continues
+  the latest (or a named) one, `/fork` branches off without touching the
+  original.
+- **Safety** — destructive actions (file edits, shell commands not on a small
+  read-only allowlist) always ask for confirmation; `y`/`n`/`a` where `a`
+  remembers the choice for that tool for the rest of the session. `--yes`
+  bypasses this for CI but always prints what it auto-approved.
+- **Cost control** — usage is read only from the API's own `usage` field,
+  never estimated. A budget (`/budget`, `--budget`) hard-stops the loop before
+  the next API call if exceeded, without losing the session. `/usage` also
+  shows what the session would have cost off-peak, and the REPL warns you at
+  startup if you're currently in a peak-pricing window.
+- **Status line** — the last line of every turn shows model, mode (chat/plan),
+  cost (and budget), and the current call out of the iteration limit; it turns
+  yellow, then red, as you approach the budget or the limit.
+- **Readable tool output** — `edit` shows a red/green diff, `read_file`, `search`
+  and `list_dir` collapse to a one-line summary (`120 lines read`, `no matches`).
+- **ESC to interrupt** — cancels the in-flight call (and any running shell
+  command) cleanly, keeping whatever was generated so far.
+- **Cross-platform** — the shell tool runs PowerShell natively on Windows and
+  `sh` on macOS/Linux, file edits preserve the original CRLF/LF line ending
+  style, paths are handled with `path.resolve` throughout.
+
+## Quick start
+
+```bash
+npm install -g tandem-mode
+tandem            # interactive REPL
+tandem "explain what this repo does" --yes   # one-shot
+```
+
+Or from source (useful for contributing, or to track `main` instead of the
+latest npm release):
+
+```bash
+git clone https://github.com/zdrave08/Tandem-Mode.git
+cd Tandem-Mode
+pnpm install
+pnpm dev            # interactive REPL
+pnpm dev "explain what this repo does" --yes   # one-shot
+```
+
+On first run you'll be asked for a DeepSeek API key (get one at
+[platform.deepseek.com](https://platform.deepseek.com)) and a default
+model/reasoning effort. The key is stored in your OS's own secure credential
+store — never in a config file or log:
+
+- **Windows** — DPAPI-encrypted (tied to your OS user account) in
+  `%LOCALAPPDATA%\TandemMode\credentials.dat`.
+- **macOS** — the login Keychain, via the built-in `security` CLI.
+- **Linux** — the Secret Service (gnome-keyring, KWallet, or similar), via
+  `secret-tool` (install `libsecret-tools` if it's missing).
+
+`DEEPSEEK_API_KEY` works everywhere and always takes priority, for CI use.
+
+## Usage
+
+```bash
+tandem                              # interactive REPL
+tandem "fix the off-by-one in foo.ts"
+tandem --plan "add dark mode to the settings page"
+tandem "task" --model deepseek-v4-flash --effort low --budget 0.50 --yes
+```
+
+| Flag | Effect |
+|---|---|
+| `--resume` / `-r` | Continue the latest session in this directory |
+| `--yes` / `-y` | Auto-approve destructive actions (prints what it approved) |
+| `--plan` | Run planner/worker/reviewer orchestration instead of a single turn |
+| `--model` | `deepseek-v4-pro` or `deepseek-v4-flash` |
+| `--effort` | `low`, `high`, or `max` reasoning effort |
+| `--budget` | Hard USD cap for the run |
+| `--max-review-loops` | Cap on orchestration correction cycles (default 3) |
+| `--max-iterations` | Model calls per chunk of a turn (default 25); at the limit the agent asks whether to continue |
+| `--auto-continue` | Continue at the iteration limit without asking — only together with a budget (`--budget`) |
+
+### REPL commands
+
+| Command | Effect |
+|---|---|
+| `/plan <task>` | Planner/worker/reviewer orchestration |
+| `/model [name]` | Show or switch model |
+| `/thinking [on\|off]` | Toggle thinking mode |
+| `/effort [low\|high\|max]` | Show or set reasoning effort |
+| `/status` | Current session/settings |
+| `/diff [--staged] [path]` | Working-tree diff, no API call |
+| `/usage` | Tokens, cost, off-peak projection |
+| `/budget [amount]` | Show or set the session budget |
+| `/new` | Start a fresh session |
+| `/save-profile <name>` | Save model, effort, budget and review loops as a named profile |
+| `/profile [name]` | Load a saved profile, or list them |
+| `/compact` | Summarize the history into a new session to free up context; the old session stays on disk |
+| `/fork` | New session, carrying over the current history |
+| `/resume [id]` | Continue the latest (or a named) session |
+| `/paste [message]` | Send a clipboard image to the agent |
+| `/clear` | Clear the screen |
+| `/help` | This list |
+| `/exit` | Quit |
+
+Press **ESC** at any time during a turn to interrupt it.
+
+## Architecture
+
+```
+tandem (no args)  ─┬─> interactive REPL ─┬─> single turn: agent loop (tools)
+                    │                     └─> /plan: planner → worker(s) → reviewer
+tandem "task"     ──┘   (single-shot, same agent loop, for scripts/CI)
+```
+
+- `src/deepseek/` — thin client for DeepSeek's OpenAI-compatible API
+  (streaming + non-streaming chat completions, vision, web search via
+  Responses API). See [`docs/api-notes.en.md`](./docs/api-notes.en.md)
+  (Serbian original: [`docs/api-notes.md`](./docs/api-notes.md)) for the
+  API behavior this is built against, verified live rather than assumed.
+- `src/agent/` — the tool-calling loop, tool implementations, session
+  persistence, usage/cost tracking.
+- `src/orchestrator/` — planner/worker/reviewer, repo-map generation, the
+  cache-aware stable-prefix construction.
+- `src/repl/` — slash commands, runtime config, the first-run wizard.
+- `src/cli.tsx` — Ink-based terminal UI and the REPL/single-shot entry points.
+
+## Configuration
+
+Layered: **session** (in-REPL commands) > **project** (`.tandem/config.json`
+in the current directory) > **global** (`~/.tandem/config.json`).
+
+```json
+{
+  "defaultModel": "deepseek-v4-pro",
+  "defaultReasoningEffort": "high",
+  "budgetUsd": 5,
+  "maxReviewLoops": 3,
+  "maxIterations": 25,
+  "autoContinue": false,
+  "verifyCommand": "pnpm typecheck && pnpm test",
+  "notifyCommand": "curl -s -d \"Tandem: $TANDEM_OUTCOME\" https://ntfy.sh/my-topic",
+  "language": "en"
+}
+```
+
+`maxIterations` is how many model calls one turn may make before Tandem stops and
+asks "continue for N more calls? (y/n/a)"; `a` remembers the answer for the rest of
+the session. With `autoContinue` (or `--auto-continue`) it continues on its own, but
+only when a budget is set, because the budget is the hard stop before the next API
+call. In non-interactive runs (`--yes`) without `--auto-continue` it stops instead of
+waiting for an answer, and the session is saved so you can send "continue".
+
+`verifyCommand` runs before each `/plan` review round and its result goes to
+the reviewer. `language` is the UI language (`en` or `sr`, default `en`); the
+`--lang` flag and the `TANDEM_LANG` environment variable override it, in that
+order. The first-run wizard asks for it.
+
+`notifyCommand` runs when a `/plan` run ends: reviewer approved, `max_review_loops`
+reached, or an error (not when you interrupt it with ESC). The details come in as
+environment variables, never spliced into the command text: `TANDEM_OUTCOME`
+(`approved`, `max_review_loops` or `error`), `TANDEM_TASK`, `TANDEM_NOTES` (the
+reviewer's notes, or the error message), `TANDEM_REVIEW_LOOPS` and `TANDEM_COST_USD`.
+The command runs in your shell (PowerShell on Windows, `sh` elsewhere) with a 15 s
+limit, and a failure only prints a warning. Any shell command works, so Slack,
+ntfy, a desktop notification or a webhook are all just examples:
+
+```json
+{ "notifyCommand": "curl -s -d \"Tandem: $TANDEM_OUTCOME\" https://ntfy.sh/my-topic" }
+```
+
+On Windows (PowerShell) the variables are written `$env:TANDEM_OUTCOME`.
+
+Like `verifyCommand`, `notifyCommand` runs without asking, so only put it in a
+project's `.tandem/config.json` if you trust that project.
+
+**Profiles:** `/save-profile <name>` stores the current model, reasoning effort,
+budget and review loops in `~/.tandem/profiles/<name>.json`, and `/profile <name>`
+loads them back into the running session (only the fields present in the file are
+applied). `/profile` alone lists the saved ones. Names use `a-z`, `0-9`, `-` and `_`.
+Profiles never include `verifyCommand` or the API key.
+
+The API key is never stored in these files — it lives in the OS credential
+store (or `DEEPSEEK_API_KEY` env var, which takes priority, for CI use).
+
+## Development
+
+See [CONTRIBUTING.md](./CONTRIBUTING.md).
+
+```bash
+pnpm install
+pnpm typecheck
+pnpm dev "some task"
+```
+
+## Disclaimer
+
+This is an independent, community-run project. It is not affiliated with,
+endorsed by, or sponsored by DeepSeek.
+
+## License
+
+[MIT](./LICENSE)

package/README.sr.md

@@ -130,9 +130,13 @@ tandem "task" --model deepseek-v4-flash --effort low --budget 0.50 --yes
 | `/thinking [on\|off]` | Uključi/isključi thinking mode |
 | `/effort [low\|high\|max]` | Prikaži ili postavi reasoning effort |
 | `/status` | Trenutna sesija/podešavanja |
+| `/diff [--staged] [putanja]` | Diff radnog stabla, bez API poziva |
 | `/usage` | Tokeni, cena, off-peak projekcija |
 | `/budget [iznos]` | Prikaži ili postavi budžet sesije |
 | `/new` | Nova sesija |
+| `/save-profile <ime>` | Sačuvaj model, effort, budžet i review krugove kao imenovan profil |
+| `/profile [ime]` | Učitaj sačuvan profil, ili prikaži listu |
+| `/compact` | Sažmi istoriju u novu sesiju da se oslobodi kontekst; stara sesija ostaje na disku |
 | `/fork` | Nova sesija sa kopijom trenutne istorije |
 | `/resume [id]` | Nastavi poslednju (ili navedenu) sesiju |
 | `/paste [poruka]` | Pošalji sliku iz clipboard-a agentu |
@@ -175,6 +179,7 @@ Slojevito: **sesija** (komande u REPL-u) > **projekat**
   "budgetUsd": 5,
   "maxReviewLoops": 3,
   "verifyCommand": "pnpm typecheck && pnpm test",
+  "notifyCommand": "curl -s -d \"Tandem: $TANDEM_OUTCOME\" https://ntfy.sh/my-topic",
   "language": "sr"
 }
 ```
@@ -184,6 +189,30 @@ ide reviewer-u. `language` je jezik interfejsa (`en` ili `sr`, podrazumevano
 `en`); `--lang` flag i `TANDEM_LANG` env promenljiva ga nadjačavaju, tim
 redom. First-run wizard ga pita.
 
+`notifyCommand` se pokreće kad se `/plan` završi: reviewer odobrio, dostignut
+`max_review_loops`, ili greška (ne i kad ga prekineš sa ESC). Detalji dolaze kao
+promenljive okruženja, nikad ubačeni u tekst komande: `TANDEM_OUTCOME`
+(`approved`, `max_review_loops` ili `error`), `TANDEM_TASK`, `TANDEM_NOTES`
+(beleške reviewer-a, ili poruka greške), `TANDEM_REVIEW_LOOPS` i `TANDEM_COST_USD`.
+Komanda radi u tvojoj školjki (PowerShell na Windows-u, `sh` drugde) sa limitom
+od 15 s, a neuspeh samo ispiše upozorenje. Može bilo koja shell komanda, pa su
+Slack, ntfy, desktop obaveštenje ili webhook samo primeri:
+
+```json
+{ "notifyCommand": "curl -s -d \"Tandem: $TANDEM_OUTCOME\" https://ntfy.sh/my-topic" }
+```
+
+Na Windows-u (PowerShell) promenljive se pišu `$env:TANDEM_OUTCOME`.
+
+Kao i `verifyCommand`, `notifyCommand` se pokreće bez pitanja, pa ga stavljaj u
+`.tandem/config.json` projekta samo ako tom projektu veruješ.
+
+**Profili:** `/save-profile <ime>` čuva trenutni model, reasoning effort, budžet i
+review krugove u `~/.tandem/profiles/<ime>.json`, a `/profile <ime>` ih vraća u
+sesiju koja radi (primenjuju se samo polja koja postoje u fajlu). `/profile` bez
+argumenta prikazuje sačuvane. Imena koriste `a-z`, `0-9`, `-` i `_`.
+Profili nikad ne sadrže `verifyCommand` ni API ključ.
+
 API ključ se nikad ne čuva u ovim fajlovima — živi u OS-ovom skladištu
 kredencijala (ili u `DEEPSEEK_API_KEY` env promenljivoj, koja ima
 prioritet, za CI upotrebu).

package/dist/agent/compact.js

@@ -0,0 +1,80 @@
+import { chatCompletionOnce } from "../deepseek/client.js";
+// Tool output is where long sessions grow (file dumps, shell output), so it is
+// cut hard in the transcript the summarizer sees. The summary only needs the gist.
+const MAX_TOOL_OUTPUT_CHARS = 1_500;
+const MAX_TOOL_ARGS_CHARS = 300;
+const MAX_TRANSCRIPT_CHARS = 300_000;
+export const SUMMARY_MODEL = "deepseek-v4-flash";
+const SUMMARIZER_SYSTEM_PROMPT = `You compact the history of a coding-agent session so work can continue from your
+summary alone. Write a dense summary with these parts, skipping any that are empty:
+- Goal: what the user is trying to achieve, plus constraints and preferences they stated.
+- Done: files read or changed (exact paths), commands run and their results.
+- Decisions: what was decided and why.
+- State: where things stand right now, including failing tests or errors still open.
+- Next: what was about to happen or is still to do.
+Keep exact file paths, identifiers and error messages that matter. Do not invent anything
+that is not in the conversation. No preamble. Reply in the language the user wrote in.`;
+export const SUMMARY_PREFIX = "Summary of our conversation so far (the earlier messages were compacted to save context):";
+const SUMMARY_ACK = "Understood. I'll continue from this summary.";
+function truncate(text, max) {
+    return text.length > max ? `${text.slice(0, max)} … (truncated, ${text.length} chars)` : text;
+}
+/** Plain-text rendering of the history. Text avoids tool_call/tool pairing rules in the summarizer request. */
+export function buildTranscript(messages) {
+    const toolNames = new Map();
+    for (const m of messages) {
+        for (const call of m.toolCalls ?? [])
+            toolNames.set(call.id, call.function.name);
+    }
+    const parts = [];
+    for (const m of messages) {
+        if (m.role === "system")
+            continue;
+        if (m.role === "user") {
+            parts.push(`USER:\n${m.content ?? ""}`);
+        }
+        else if (m.role === "assistant") {
+            const calls = (m.toolCalls ?? [])
+                .map((c) => `[called ${c.function.name}(${truncate(c.function.arguments, MAX_TOOL_ARGS_CHARS)})]`)
+                .join("\n");
+            const text = [m.content ?? "", calls].filter(Boolean).join("\n");
+            if (text)
+                parts.push(`ASSISTANT:\n${text}`);
+        }
+        else {
+            const name = (m.toolCallId && toolNames.get(m.toolCallId)) || "tool";
+            parts.push(`TOOL RESULT (${name}):\n${truncate(m.content ?? "", MAX_TOOL_OUTPUT_CHARS)}`);
+        }
+    }
+    const transcript = parts.join("\n\n");
+    if (transcript.length <= MAX_TRANSCRIPT_CHARS)
+        return transcript;
+    // Too long even after trimming tool output: keep the most recent part.
+    return `[earlier part of the conversation omitted]\n\n${transcript.slice(-MAX_TRANSCRIPT_CHARS)}`;
+}
+/** One model call. Throws on API errors or an empty summary; the caller's history is never touched here. */
+export async function compactConversation(env, messages, systemPrompt, signal) {
+    const response = await chatCompletionOnce(env, {
+        model: SUMMARY_MODEL,
+        messages: [
+            { role: "system", content: SUMMARIZER_SYSTEM_PROMPT },
+            { role: "user", content: `Conversation to summarize:\n\n${buildTranscript(messages)}` },
+        ],
+        ...(signal ? { signal } : {}),
+    });
+    const summary = response.content.trim();
+    if (!summary)
+        throw new Error("The model returned an empty summary.");
+    // The acknowledgement keeps user/assistant alternation, so the next user turn
+    // never follows another user message.
+    return {
+        messages: [
+            { role: "system", content: systemPrompt },
+            { role: "user", content: `${SUMMARY_PREFIX}\n\n${summary}` },
+            { role: "assistant", content: SUMMARY_ACK },
+        ],
+        summary,
+        usage: response.usage,
+    };
+}
+//# sourceMappingURL=compact.js.map

package/dist/agent/compact.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"compact.js","sourceRoot":"","sources":["../../src/agent/compact.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,kBAAkB,EAAE,MAAM,uBAAuB,CAAC;AAG3D,+EAA+E;AAC/E,mFAAmF;AACnF,MAAM,qBAAqB,GAAG,KAAK,CAAC;AACpC,MAAM,mBAAmB,GAAG,GAAG,CAAC;AAChC,MAAM,oBAAoB,GAAG,OAAO,CAAC;AAErC,MAAM,CAAC,MAAM,aAAa,GAAG,mBAAmB,CAAC;AAEjD,MAAM,wBAAwB,GAAG;;;;;;;;uFAQsD,CAAC;AAExF,MAAM,CAAC,MAAM,cAAc,GACzB,2FAA2F,CAAC;AAE9F,MAAM,WAAW,GAAG,8CAA8C,CAAC;AAEnE,SAAS,QAAQ,CAAC,IAAY,EAAE,GAAW;IACzC,OAAO,IAAI,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,kBAAkB,IAAI,CAAC,MAAM,SAAS,CAAC,CAAC,CAAC,IAAI,CAAC;AAChG,CAAC;AAED,+GAA+G;AAC/G,MAAM,UAAU,eAAe,CAAC,QAAuB;IACrD,MAAM,SAAS,GAAG,IAAI,GAAG,EAAkB,CAAC;IAC5C,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;QACzB,KAAK,MAAM,IAAI,IAAI,CAAC,CAAC,SAAS,IAAI,EAAE;YAAE,SAAS,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC;IACnF,CAAC;IAED,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;QACzB,IAAI,CAAC,CAAC,IAAI,KAAK,QAAQ;YAAE,SAAS;QAClC,IAAI,CAAC,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YACtB,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC,CAAC,OAAO,IAAI,EAAE,EAAE,CAAC,CAAC;QAC1C,CAAC;aAAM,IAAI,CAAC,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAClC,MAAM,KAAK,GAAG,CAAC,CAAC,CAAC,SAAS,IAAI,EAAE,CAAC;iBAC9B,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,WAAW,CAAC,CAAC,QAAQ,CAAC,IAAI,IAAI,QAAQ,CAAC,CAAC,CAAC,QAAQ,CAAC,SAAS,EAAE,mBAAmB,CAAC,IAAI,CAAC;iBACjG,IAAI,CAAC,IAAI,CAAC,CAAC;YACd,MAAM,IAAI,GAAG,CAAC,CAAC,CAAC,OAAO,IAAI,EAAE,EAAE,KAAK,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YACjE,IAAI,IAAI;gBAAE,KAAK,CAAC,IAAI,CAAC,eAAe,IAAI,EAAE,CAAC,CAAC;QAC9C,CAAC;aAAM,CAAC;YACN,MAAM,IAAI,GAAG,CAAC,CAAC,CAAC,UAAU,IAAI,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,IAAI,MAAM,CAAC;YACrE,KAAK,CAAC,IAAI,CAAC,gBAAgB,IAAI,OAAO,QAAQ,CAAC,CAAC,CAAC,OAAO,IAAI,EAAE,EAAE,qBAAqB,CAAC,EAAE,CAAC,CAAC;QAC5F,CAAC;IACH,CAAC;IAED,MAAM,UAAU,GAAG,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IACtC,IAAI,UAAU,CAAC,MAAM,IAAI,oBAAoB;QAAE,OAAO,UAAU,CAAC;IACjE,uEAAuE;IACvE,OAAO,iDAAiD,UAAU,CAAC,KAAK,CAAC,CAAC,oBAAoB,CAAC,EAAE,CAAC;AACpG,CAAC;AASD,4GAA4G;AAC5G,MAAM,CAAC,KAAK,UAAU,mBAAmB,CACvC,GAAc,EACd,QAAuB,EACvB,YAAoB,EACpB,MAAoB;IAEpB,MAAM,QAAQ,GAAG,MAAM,kBAAkB,CAAC,GAAG,EAAE;QAC7C,KAAK,EAAE,aAAa;QACpB,QAAQ,EAAE;YACR,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,wBAAwB,EAAE;YACrD,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,iCAAiC,eAAe,CAAC,QAAQ,CAAC,EAAE,EAAE;SACxF;QACD,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC9B,CAAC,CAAC;IAEH,MAAM,OAAO,GAAG,QAAQ,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC;IACxC,IAAI,CAAC,OAAO;QAAE,MAAM,IAAI,KAAK,CAAC,sCAAsC,CAAC,CAAC;IAEtE,8EAA8E;IAC9E,sCAAsC;IACtC,OAAO;QACL,QAAQ,EAAE;YACR,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,YAAY,EAAE;YACzC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,GAAG,cAAc,OAAO,OAAO,EAAE,EAAE;YAC5D,EAAE,IAAI,EAAE,WAAW,EAAE,OAAO,EAAE,WAAW,EAAE;SAC5C;QACD,OAAO;QACP,KAAK,EAAE,QAAQ,CAAC,KAAK;KACtB,CAAC;AACJ,CAAC"}

package/dist/agent/loop.js

@@ -17,7 +17,17 @@ const MAX_REASONING_CHARS = 200_000;
 export async function* runAgentLoop(messages, opts) {
     const ctx = { cwd: opts.cwd, env: opts.env, usage: opts.usage, ...(opts.signal ? { signal: opts.signal } : {}) };
     const maxIterations = opts.maxIterations ?? 25;
-    for (let iteration = 0; iteration < maxIterations; iteration++) {
+    // Granica se pomera za još jedan "komad" svaki put kad korisnik (ili autoContinue) dozvoli nastavak.
+    let limit = maxIterations;
+    for (let iteration = 0;; iteration++) {
+        if (iteration >= limit) {
+            const proceed = opts.onIterationLimit ? await opts.onIterationLimit({ iterations: iteration }) : false;
+            if (!proceed) {
+                yield { type: "max_iterations_reached" };
+                return;
+            }
+            limit += maxIterations;
+        }
         if (opts.signal?.aborted) {
             yield { type: "interrupted" };
             return;
@@ -29,6 +39,7 @@ export async function* runAgentLoop(messages, opts) {
                 return;
             }
         }
+        yield { type: "iteration", n: iteration + 1, max: limit };
         let reasoning = "";
         let content = "";
         const toolCallBuf = new Map();
@@ -159,6 +170,5 @@ export async function* runAgentLoop(messages, opts) {
             await appendSessionMessage(opts.session, toolMsg);
         }
     }
-    yield { type: "max_iterations_reached" };
 }
 //# sourceMappingURL=loop.js.map