A coding agent harness in Go. Point it at a repository, tell it to make the failing test pass, and measure how often it actually does.
The code is the byproduct. The deliverable is METRICS.md. PLAN.md is the build plan this follows.
go build -o harness ./cmd/harness
./harness -C /path/to/a/throwaway/clone # full-screen TUI
./harness -C /path/to/repo -p "why does TestFoo fail?" # one prompt
If you are logged into Claude Code, that is all the setup there is: the harness
borrows the OAuth credential from ~/.claude/.credentials.json when nothing
else supplied one. Otherwise, or to use a different account:
export ANTHROPIC_API_KEY=sk-ant-...
A different endpoint — a gateway, a proxy, a local server speaking the Messages API:
./harness -base-url http://127.0.0.1:4000 -api-key ... -C .
# or ANTHROPIC_BASE_URL / ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN
Flags beat the saved config, which beats the built-in defaults; the config is
what the setup wizard writes on a first run. -auth-token sends
Authorization: Bearer, which is what OAuth tokens and most gateways want;
-api-key sends x-api-key. A non-default endpoint announces itself on
stderr, because your prompt and every file the agent reads go wherever that URL
points.
The Claude Code login is the last resort, tried only when no flag and no
environment variable supplied a credential, and only against Anthropic's own
endpoint — a credential found on disk is not one you chose to send, so it is
never forwarded to a -base-url you pointed the harness at. -no-claude-code-auth
skips it entirely. Note that the credential is issued to Claude Code; using it
from another client is outside Anthropic's terms for it, which API keys and
gateway tokens do not carry.
On macOS Claude Code keeps this in the Keychain rather than a file, so there is nothing to borrow there. An expired access token is refreshed automatically; the credentials file is left alone unless the server rotates the refresh token, in which case the replacement is written back — leaving the dead one there is what would break your Claude Code login.
Useful flags: -model, -effort low|medium|high|xhigh|max, -max-turns,
-resume latest, -no-stream, -no-cache, -v, -yolo, -no-tui, -no-skills,
-no-agents, -no-claude-code-auth.
On a terminal, the interactive mode is a full-screen Bubble Tea app. It opens on a welcome block naming the model, the endpoint and the root, and settles into a transcript above a prompt:
⏺ Let me see what the test expects before changing anything.
⏺ ReadFile(path=internal/agent/loop_test.go)
⎿ 1: package agent
…
⏺ EditFile(path=internal/agent/compact.go)
4 additions, 1 removal
@@ -50,7 +50,9 @@
+ if window == 0 {
✢ Noodling… (14s · turn 3 · ↓ 1.3k tokens · esc to interrupt) ● high · /effort
╭──────────────────────────────────────────────────────────────────────────────╮
│ › _ │
╰──────────────────────────────────────────────────────────────────────────────╯
joni ✳ Opus 5 ◐ 62.2k (6%) ↑ 4.2k ⟳ 58.0k ↓ 1.3k $0.0835
⏵ approvals on (shift+tab to cycle) · /help for commands
Every event hangs off a dot in the left gutter, and everything belonging to it is indented under that dot, so a screen of tool output still reads as a sequence of things that happened. Diffs are shown in full and in colour with a count of what changed; other output is elided after six lines.
The status bar reports the context gauge — the last turn's input size against the window, which drops when the conversation is compacted — and separates fresh input tokens from cache reads, because the gap between the two is what METRICS.md is about. The prompt box grows with what you type, up to six lines.
| Key | Does |
|---|---|
enter | send |
/ | open the command menu — see below |
! | run a shell command yourself — see below |
$ | the same, through python3 |
ctrl+j | newline without sending |
shift+tab | cycle the approval mode |
ctrl+b | open the agents sidebar — see below |
ctrl+r | expand elided tool output |
pgup/pgdn | scroll the transcript |
esc | interrupt the turn |
ctrl+c | interrupt the turn; again when idle to exit |
Commands: /allowed, /clear, /effort, /exit, /help, /model,
/mouse, /setup, /skills, and one per discovered skill.
A slash at the start of an empty prompt opens the command menu, which filters as you type — a name that starts with what you typed before one that merely contains it:
╭──────────────────────────────────────────────────────────────────────────────╮
│ › /s │
╰──────────────────────────────────────────────────────────────────────────────╯
❯ /skills the instruction sets found at startup
/systematic-debugging find the cause before changing code (project)
/brainstorming explore an idea before building it (personal)
↑/↓ move, tab or enter completes the highlighted name into the prompt —
completing is not sending, so there is room to type arguments for a skill or a
level for /effort — and esc dismisses the menu without touching what you
typed. The menu offers exactly the commands the TUI answers, so nothing it
suggests can come back as "unknown command".
! on an empty prompt points the box at a shell instead of at the model; $
points it at python3. The marker replaces the ›, the frame changes colour,
and enter runs the line where the agent's own bash tool would have run it —
from the working root, with the same timeout and the same output cap.
╭──────────────────────────────────────────────────────────────────────────────╮
│ ! go test ./internal/agent/ │
╰──────────────────────────────────────────────────────────────────────────────╯
⏺ ! go test ./internal/agent/
⎿ exit code: 1 (2.3s)
--- stdout ---
--- FAIL: TestCompactKeepsTheLastExchange (0.00s)
compact_test.go:88: history has 2 messages, want 4
… +9 more lines (ctrl+r)
The agent is shown what you saw. The command and its output are appended to the conversation as a user message, so the next thing you type can be "fix that" and the model already has the failure. Nothing is sent at that moment — it waits in the history and goes up with your next real turn, so several commands can be run before saying anything about them.
What reaches the model is the capped report, not the raw firehose: a noisy
command cannot cost more context from the prompt box than it would through the
tool. The transcript collapses it further to six lines, and ctrl+r expands to
exactly what the model was given, so the two never drift.
esc leaves the mode, as does backspace over the marker. The mode is
per-command — after a run the prompt returns to normal, the way a shell prompt
comes back. One command runs at a time; while it does, esc interrupts it and
a message you type waits rather than jumping ahead of the output it is about.
A ! typed mid-sentence is an exclamation mark.
These bypass the approval policy by design: you typed the command, so there is nobody left to ask. They are also the one part of the TUI that runs something without the model's involvement — worth remembering when the root is not a throwaway clone.
/model asks the endpoint what it serves — GET /v1/models, paginated — and
opens a picker on the answer:
Select model
from anthropic · ↑↓ move · enter apply · esc cancel
1. Claude Fable 5.1 claude-fable-5-1 · 1.0M in · $10 / $50 per Mtok
2. Claude Opus 5 claude-opus-5 · 1.0M in · $5 / $25 per Mtok ✓
❯ 3. Claude Sonnet 5 claude-sonnet-5 · 1.0M in · $2 / $10 per Mtok
4. Claude Haiku 4.5 claude-haiku-4-5-20251001 · 200.0k in · $1 / $5 per Mtok
The cursor starts on the model in force, which is also the one marked. The id, the context window and the list price come from the endpoint and from the pricing table; a model neither knows about simply says less. The switch applies from the next turn — the conversation so far is replayed to the new model — and costs already incurred stay priced at the model that incurred them, because a session that switched models was not all billed at the model it ended on.
Not every endpoint serves that route. When the fetch fails the picker says so,
names the host and the reason, and closes; /model <id> and the -model flag
still switch without asking anyone's permission.
/effort opens a slider in place of the prompt, on the level in force — an
unset level is the API's own default, high, so the marker starts there rather
than at an end:
Faster Smarter
──────────────────────────────────────▲───────────────────────────────────
low medium high xhigh max
←→ move · 1-5 jump · enter apply · esc cancel
Moving the marker changes nothing until enter; esc leaves the level alone.
/effort xhigh still applies a level directly for anyone who can type it, and
either way it is the -effort flag reached from the keyboard. The level in
force shows above the prompt on the right. A terminal too narrow for five
labels gets ◀ high ▶ and the same keys rather than a track with the words
sliced in half.
/model and /effort change the session. /setup changes what the next
session starts with — see Setup.
shift+tab cycles between asking before every mutating call and accepting them
all, which is -yolo reached from the keyboard. The mode line says which is in
force at all times, and the switch is recorded in the transcript, because a
mode that silently lets the agent write to disk is not one to keep in a corner.
Approvals appear inline as a bordered prompt, answered by number or by the key
in brackets — y, n, a for this exact command, t for the whole tool. The
agent goroutine blocks until you answer, which is the point: the tool has not
run yet. Interrupting while that prompt is open denies the call and cancels the
turn rather than leaving the agent waiting on a channel nobody will read.
The TUI is skipped automatically when stdin or stdout is not a terminal, and by
-p and -no-tui, which use the line-based renderer instead. Both front ends
consume the same agent.Event stream, so neither can show something the other
cannot.
The agent can start other agents. spawn_agent(preset, task) returns
immediately with a short id — a1, a2, … — and that agent runs in the
background on its own goroutine while the main agent carries on with whatever
it was doing. agent_wait(id) blocks until that one finishes, for the times
the parent genuinely cannot proceed without the answer.
A subagent is a whole agent, not a subroutine: its own conversation, its own context window, its own model, its own turn loop. The parent never sees its turns — only the report it ends with. That is the point. Work that would have filled the parent's window with file listings fills the child's instead, and comes back a paragraph long.
Three presets, compiled in rather than configurable, because a preset the model can invent is a blast radius the model can invent:
| Preset | For | Turn cap |
|---|---|---|
explore | finding things, and reporting paths and line numbers | 20 |
plan | working out how a change should be made, without making it | 30 |
general | a self-contained piece of work, edits and tests included | 40 |
All three get the full tool set — read, glob, grep, edit, write, bash, skill —
and none of them gets spawn_agent, so subagents cannot spawn subagents.
A preset pins no model and no effort level. Both are arguments to
spawn_agent, because what to spend is a judgement about the job and the agent
holding the task is the one making it: model picks what runs it — a small
fast model for mechanical search, a strong one for work that has to be right —
and effort picks how hard it thinks, low through max. Omit either and it
inherits the session's. A pinned id here would be a guess about someone else's
endpoint as well as about the job: point the harness at a gateway that serves
neither, and every spawn of that preset fails before its first turn.
The models it may choose between are in its own system prompt: the harness asks
the endpoint what it serves at startup — GET /v1/models, the same route the
picker uses — and lists them newest first, with the rule that the newest is the
one to reach for unless you asked for another or it has already failed this
session. An endpoint that does not answer that route simply gets no listing,
and the agent is told that naming no model runs the subagent on whatever the
session is using. The list is read once, because it sits in the cached prefix
and a prompt that changes mid-session throws the cache away. Subagents do not
get this paragraph — they have no such choice to make.
output_config.effort is a per-model capability, not a universal parameter.
Haiku 4.5 rejects a request carrying it, and xhigh did not exist before Opus
4.7 — so a request for a level its model does not take goes out without it,
rather than failing with a 400 that reads like a bug in the harness. Where that
happens to a subagent the answer says so outright (claude-haiku-4-5 takes no effort level, so "low" was not applied) instead of letting the parent believe
it got what it asked for, and the effort pill in the status line reads not on Haiku rather than claiming a setting that is not in force.
A model this build has never heard of — a gateway serving something else, a
family released after it — is treated the same way: the level is left off. The
two mistakes are not symmetric. Sending effort to a model that lacks it is a
400 that ends the run before any work happens; omitting it costs the level and
says so in the pill. Gateway decoration is seen past first, so
anthropic.claude-opus-5, openrouter/anthropic/claude-opus-5 and
claude-opus-5@20260401 are all recognised as the model they name — for
pricing and for the thinking parameter too.
Every spawn goes through the ordinary approval box, which shows the preset, the model and the task. It is not rememberable per task — the task text differs every time, so "always this exact one" would never match twice — but "always allow this tool" still applies, for anyone who has decided to trust spawning as such. A subagent's own tool calls go through the same policy, and the box says which agent is asking.
Four run at once; further spawns queue. Two agents editing one file is not a
special case: edit_file already refuses a file whose mtime moved since that
agent read it, so the second one is told to read it again rather than
clobbering the first.
A report reaches the parent exactly once, whichever way it gets there. Both
paths — agent_wait returning it as a tool result, and the front end putting
it into the conversation as a message — claim it from the supervisor first, and
only one claim wins. Without that, an agent that finished before its
agent_wait call was even emitted was handed over twice: once as the result,
once as a message a turn later, which is the context cost the whole feature
exists to avoid. A wait that loses the claim answers with a line saying where
to find the report instead of repeating it.
The report comes back on its own. When a subagent finishes, its report is
appended to the main conversation as a user message and the main agent starts a
turn on it — you do not have to type anything. It is held back only while a
turn is already running, while an approval is on screen, while a ! or $
command is running, or while you have text in the prompt box: a turn starting
under your fingers would blur the input and eat what you were typing. It goes
up the moment that clears. A failed or killed agent is reported the same way,
so the parent is never left waiting on a result that is not coming.
ctrl+b opens a 30-column pane on the right and gives it the keyboard; the
transcript reflows into what is left rather than being covered by it.
Open and focused are separate things. The pane answers every key while it has
focus — that is what lets k be a kill rather than a letter — so click the
prompt to type again, and the pane stays where it is. Click a row to hand the
keys back. ctrl+b always moves towards the pane: it opens a closed one,
focuses an open one, and only closes a pane that already has the keys. Which
one is live is visible in two places: the heading carries the accent and the
prompt frame goes dim, or the other way round.
⏺ Two of those are independent, so I will farm │ Agents 2 ⏵
them out. │
│ ❯ ✳ a1 plan $0.23
⏺ SpawnAgent(preset=explore, task=…) │ Opus · 2m04s · 6 turns
⎿ started a2 (explore) on claude-haiku-4-5, s…│ ● a2 explore
│ Haiku · waiting · 1 turn
│
│ ↑↓ move · enter open
✢ Deliberating… (8s · turn 5 · esc to interrup…│ p pause · k kill · esc back
╭──────────────────────────────────────────────╮│
│ › ask for a change, /help, ! for a shell ││
╰──────────────────────────────────────────────╯│
joni ✳ Opus 5 ⏵ 2 agents │
⏵ approvals on (shift+tab to cycle) · ctrl+b a…│
Each agent is two lines: the status glyph, the id and the preset, with the cost
at the right, and under it the model, how long it has been going and how many
turns it has taken. The model is the family alone — thirty columns will not
hold claude-haiku-4-5 as well as the numbers — and the full id is in the
header of the thread, one keystroke away.
| Key | Does |
|---|---|
↑/↓ | move |
enter | open that agent's thread in the main pane |
esc | back to your own thread; again to close the pane |
p | pause or resume |
k | stop it |
| click the prompt | type again, leaving the pane open |
The statuses are queued, running, waiting — blocked on an approval of yours —
paused, done, failed and killed. Pause is checked between turns: the model call
and the tool calls in flight finish, and it holds before the next request, so
the conversation is never left with a tool_use whose tool_result never
came. Kill cancels immediately, and what the agent already did to the tree
stays done — this ends an agent, it does not undo one.
Mouse reporting is on, so clicking the agents pill in the status line toggles
the pane and clicking a row selects it. /mouse turns reporting off and on
again, because while it is on the terminal's own text selection is not.
A child's usage is priced at the child's model and rolled into the session
total, so a haiku agent spawned from an opus session is not billed at opus
rates. Every metrics event a child writes carries its agent id, so
.harness/events.jsonl can be split per agent afterwards.
Subagent transcripts last as long as the session does: -resume brings back
your conversation, not theirs.
Subagents belong to the TUI: their reports come back as messages the front end
puts into the conversation, and the line-based REPL has nowhere to show four
agents at once. -p and -no-tui runs are not offered the tools at all, and
-no-agents turns them off for a TUI run too.
The first run on a terminal opens a full-screen wizard instead of the chat: the mark assembles, the name lights up, and it asks four questions one at a time.
harness setup step 1 of 4
◆━━━━━━━━━━━━━━━◇───────────────◇───────────────◇
Which model should the agent use?
▌ ◆ Claude Opus 5 $5 / $25 per Mtok
◇ Claude Sonnet 5 $2 / $10 per Mtok
◇ Claude Haiku 4.5 $1 / $5 per Mtok
◇ custom…
claude-opus-5 · 1.0M token context
↑↓ move enter next esc back ctrl+c cancel
| Question | Answers |
|---|---|
| Model | whatever GET /v1/models returns, plus custom… for any id |
| Effort | the API's own llm.EffortLevels: low medium high xhigh max |
| Approvals | ask before every mutating call, or accept them all |
| Theme | ember mint iris mono |
It is the same catalogue and the same pricing table /model uses, through the
same GET /v1/models call, so the two cannot disagree about what a model is
called or what it costs. Where they differ is failure: /model reports the
error and closes, because there is a session to go back to. The wizard cannot
do that on a first run, so it falls back to the models this build can price and
says so. custom… is there because a gateway may serve something no build has
heard of.
The theme step previews by applying the palette for real — reading a swatch tells you nothing about what a transcript looks like in it — and it recolours the whole TUI, pickers and slider included. Cancelling puts the old one back.
Answers land in ~/.config/harness/config.json (XDG_CONFIG_HOME is
honoured):
{
"version": 1,
"model": "claude-opus-5",
"effort": "high",
"approvals": "ask",
"theme": "ember"
}
Nothing in that file is required, and none of it wins an argument with a flag:
precedence is flag, then config, then the built-in default, so
-model claude-sonnet-5 is still a one-run override. accept-all saved as a
default is honoured, but -yolo can only turn approvals off, never on — a file
should not be able to quietly grant the agent more than the command line did.
A value this build does not recognise is dropped rather than forwarded to the
API, and effort is checked against llm.EffortLevels rather than a second copy
of the same five strings.
/setup reopens the wizard at any time. Finishing applies the answers through
the same setters /model and /effort use, so a setting chosen in the wizard
and the same setting chosen from the prompt cannot come out different — the
wizard just reports them in one line instead of three. It is refused mid-turn:
changing the model while a turn is in flight would finish it on a different
model than it started. A config file that cannot be parsed is reported and then
ignored, and is never overwritten until you finish the wizard.
/model and /effort change the running session and say "from the next turn
on"; they do not write to the file. The wizard is the thing that decides what
the next session starts with, which is why it is a separate screen rather
than a fourth picker.
The whole thing is ↑↓ and enter, esc to go back. Animations collapse to
their final frame under NO_COLOR, a dumb terminal, or a window under 60×20,
and any key skips the one playing.
| Tool | Approval | Notes |
|---|---|---|
read_file | automatic | Numbered lines, offset/limit for paging |
glob | automatic | Newest first; *.go matches at any depth |
grep | automatic | ripgrep when present, pure Go otherwise, identical output either way |
edit_file | prompts | Requires a prior read; refuses ambiguous matches; returns a diff |
write_file | prompts | Whole files only; separate from edit_file so the metrics stay honest |
bash | prompts | Own process group, killed whole on timeout; output capped head and tail |
A skill is a directory of instructions the agent pulls in when it is relevant:
~/.agents/skills/commit-helper/SKILL.md every repo
./.agents/skills/commit-helper/SKILL.md this one, checked in
---
name: commit-helper
description: Use when writing commit messages for this repo
---
Read the last twenty commits for tone before writing anything.
Conventions are in references/conventions.md.
The directory name is the skill's name, and a name: that disagrees with it is
a load error. Only the description reaches the model until the skill is
invoked, so it is what decides whether the skill fires at all; the body can be
as long as you like.
Two ways in, one way to load. The agent calls the skill tool when a
description matches what it is doing; you can force one with /commit-helper fix the flaky test, which the model answers by making the same call. /skills
lists what was found, project shadowing personal on a name collision.
Skill directories are readable by read_file, glob and grep even though
they sit outside the working root, so a skill can ship reference files and
point at them. They are not writable by anything: edit_file, write_file and
bash keep the working root as their only root. A skill is input, not a work
product.
Relative paths still anchor to the working root, so a skill pointing at its own
files is read with the directory the skill tool prints joined on. If the model
forgets, the "no such file" it gets back names the skill directories, which is
usually enough to recover on the next turn. A shadowed skill's directory is not
readable — only the skills actually in play widen anything.
Discovery happens at startup. The name-and-description listing goes in the
cached prompt prefix, so re-reading the directory each turn would cost a cache
miss on every turn for a file that never changes. Restart to pick up a new
skill. -no-skills turns the whole thing off; the eval runner never discovers
skills at all, so METRICS.md is unaffected by what is in your home directory.
./evals/selftest.sh # no API calls: checks the tasks measure something
go run ./evals -n 3 # the real thing
Ten tasks, a fresh temp directory per run, full transcripts under
evals/runs/<timestamp>/. See METRICS.md.
Three layers, in increasing order of how much they actually protect you.
Path jail. Every path a tool is given is resolved with
filepath.EvalSymlinks and refused if it lands outside the working root,
including through a symlink that points out of the tree. Cheap, and it catches
a lot.
Approval prompts. Read-only tools run unattended. bash, edit_file and
write_file ask, with "yes / no / always this exact one / always this tool" and
session memory. /allowed lists what you have agreed to.
Shell commands are parsed with mvdan.cc/sh/syntax so the prompt can tell you
what a command will actually run — git status; rm -rf ~ is shown as running
both git and rm. This is not a security boundary, and the harness does
not pretend otherwise: a command containing $(...), backticks, or an
interpreter with -c cannot be analysed at all, so it is marked opaque and
"always allow this exact command" is refused for it. String-matching
allowlists for shell commands do not work; this exists to make the prompt
specific, not to make the decision for you.
A sandbox. This is the layer that actually contains a mistake, and the harness cannot provide it for you. The honest recipe on Linux:
podman run --rm -it \
-v "$PWD":/work:Z -w /work \
-e ANTHROPIC_API_KEY \
docker.io/library/golang:1.26 \
/work/harness -C /work -yolo
Decide about egress explicitly. The container above has network access,
because the harness needs to reach the API — which means the agent's bash
has it too. That is the line between "broke my repo" and "read my .env and
made an HTTP request". If that matters for the tree you are pointing it at:
bash in a
network-less namespace. That is not implemented here.-yolo skips every prompt. It is correct inside a container or against a
throwaway clone, and wrong everywhere else. The eval runner sets it, because it
is headless and operates on a temp directory it created.
cmd/harness flag parsing, wiring, REPL
internal/llm Messages API: types, HTTP, SSE, pricing
internal/agent turn loop, conversation state, compaction, sessions
internal/tools the tool registry and the tools
internal/skills discovery and parsing of SKILL.md
internal/config saved defaults: model, effort, approvals, theme
internal/approval allow / deny / ask, with shell analysis
internal/metrics JSONL event log
internal/ui line-based terminal rendering
internal/tui full-screen Bubble Tea front end, and the setup wizard
evals task suite and runner
Two rules hold the shape: agent never imports ui or tui, so the eval
runner can drive the loop headless; and every tool returns (string, error)
where the error is text the model reads, not an exception the harness swallows.
A failed edit is a normal event in the transcript.
The TUI obeys the first rule through two one-way hatches. Events reach it via
agent.Observe, forwarded into the Bubble Tea loop with Program.Send, which
is safe from the agent's goroutine. Approvals go out the same way and then
block the agent on a reply channel until a key arrives. That channel is
buffered and the wait honours the turn's context, so neither side can strand
the other.