Files
zero/internal/agent/prune.go
Vasanth T b2bb6c4b41 feat: five capability features from the reference-agent comparison (#276)
* feat(commands): user-defined slash commands from markdown files

Drop a markdown file at `.zero/commands/<name>.md` (project, shared with the
team) or `<userConfigDir>/zero/commands/<name>.md` (personal); typing
`/<name> args` expands its body template and submits it as a normal prompt.
Optional YAML frontmatter (`description:`, `model:`, `agent:`) feeds help text
and routing. This turns Zero's model-pulled skills into user-invokable,
repo-checked-in team workflows (`/release v1.2`, `/fix-flaky <test>`).

New internal/usercommands package: Load() scans the project + user dirs
(project shadows user on name collision, mirroring specialist scope
precedence), validates names to a safe slash-command shape, and parses
frontmatter with a self-contained parser. Expand() substitutes $ARGUMENTS,
$1..$9, and $$ (literal $); a placeholder-free template gets the raw args
appended so a trivial command still receives input.

Wiring: model loads commands at construction; the commandUnknown dispatch
checks the user set before reporting "unknown"; autocomplete lists user
commands alongside builtins. Builtins win on collision (listed first, shadow at
dispatch). Tests cover load/precedence/name-validation, the placeholder
substitution table, and the TUI dispatch + autocomplete + still-unknown paths.

* feat(init): guided /init that generates AGENTS.md seeded with repo facts

`/init` (TUI) and `zero init` (headless) investigate the repo and write a
concise AGENTS.md so future runs start with project context. Zero already has
the AGENTS.md loader and the repoinfo scanner, but the file only existed if a
user hand-authored it — this activates the machinery in one command.

New internal/agentinit.BuildPrompt seeds the bootstrap prompt with
repoinfo.Collect() output (primary language, build/test tools, CI, workspace
layout, git) formatted as a fact block, then instructs the agent to investigate
the gaps and write a tight AGENTS.md (build/test commands, conventions, things
to avoid). Collection failure is non-fatal — the agent investigates from
scratch. This is one better than a pure-prompt /init: the agent starts from
known facts instead of re-deriving them.

TUI: /init command + handleInitCommand reuse launchPrompt (standard run, tools,
AGENTS.md write path). Headless: `zero init` builds the prompt and forwards it
through the existing exec machinery, so all exec flags (--model, -w, --add-dir)
work unchanged. Tests cover FormatFacts (rendered/omitted fields), BuildPrompt,
and the TUI dispatch.

* feat(tools): lsp_navigate — semantic code navigation via the language server

A read-only, model-callable tool exposing LSP navigation: jump-to-definition,
find-all-references, find-implementations, and workspace_symbol search. The
agent can answer "where is X defined / who calls X / what implements this
interface" precisely, where grep collides on common names or misses indirect
references — cutting wasted read/grep turns and wrong-edit risk on large repos.

Zero already runs language servers for diagnostics, so this adds the
request/response navigation methods on top: Manager.Navigate syncs the file
(didOpen) then issues textDocument/definition|references|implementation or
workspace/symbol via the existing client.Call, decoding all three result shapes
(Location, []Location, []LocationLink). 1-based file:line:col in and out
(converting LSP's 0-based positions); the tool has its own lazily-started
manager (servers start on first use, reused across calls).

Opportunistic: a file type with no installed server, or a server lacking the
capability, degrades to a clear "unavailable — fall back to grep" message, not
an error. Registered in CoreReadOnlyToolsScoped. Tests cover the response-shape
decoding (incl. LocationLink + null), symbol decoding/kinds, unknown-op and
unsupported-extension degradation, and the tool's arg validation + read-only
safety.

* feat(agent): prune stale tool output before paid summarization

Long, dump-heavy sessions accumulate large read_file/grep/glob/bash tool
results the model has already acted on; their bodies are dead weight the loop
otherwise keeps paying an LLM summarizer to compress. Add a zero-cost prune
stage inside maybeCompact, before the summarizer: when over threshold, replace
the bodies of OLD, large tool results with a compact placeholder, recompute
size, and only fall through to the paid Compact() if still over.

pruneStaleToolOutput walks newest-first, protects the last preserveLast
messages plus a trailing 40k-token window of tool output (most likely still in
use), and prunes only tool results older than that and larger than 200 tokens —
gated so the whole pass is a no-op unless reclaimable output exceeds 20k tokens.
It keeps the tool message and its ToolCallID (provider replay stays valid; the
model knows what was there and can re-run the tool), never touches non-tool
messages, never re-prunes a placeholder (idempotent), and copies-on-write so it
never mutates the caller's slice. Recent turns are preserved verbatim — no
paraphrase loss, no token or latency cost.

Tests cover the reclaim gate, recent-window protection, idempotence,
non-tool-message safety, ToolCallID/role preservation, and no-input-mutation.

* feat(agent): reconnect on a transient model disconnect instead of ending the run

A long autonomous task (a big refactor, a swarm member, a headless/cron run)
should survive a single transient upstream hiccup rather than dying and
re-burning every token on a restart. When the initial StreamCompletion connect
fails with a disconnect-shaped error, streamWithReconnect re-issues the SAME
request with exponential backoff up to 2 times.

It retries only the connect — before any content is forwarded — so no
already-streamed OnText is ever duplicated. A context-limit error (the
compactor's job), an image rejection, a non-disconnect error, or a cancelled
context is returned immediately without retry, so the reconnect never fights
the existing handlers. Disconnect classification matches EOF / connection
reset|refused|closed / broken pipe / timeout / 502 / 503 / server-closed.
Reconnect attempts surface through OnReasoning ("connection lost —
reconnecting N/max"), a non-content channel that never corrupts the answer
text; silent when there's no reasoning sink.

Tests: recover-from-transient, give-up-after-max, no-retry-on-context-limit,
stop-on-context-cancel, the disconnect-vs-not classifier, backoff growth, and
the notice routing.

* fix(specialist): add lsp_navigate to knownToolNames

The new lsp_navigate tool was added to the core read-only registry but not to
specialist.knownToolNames, which mirrors the core tool set — so
TestKnownToolNamesMatchCoreRegistry failed (the list drifted from the registry).
Add it to the read-only group to match.

* fix(tools): confine lsp_navigate to the workspace (close arbitrary-file-read)

Review caught that lsp_navigate dropped the PathScope every sibling read-only
tool (read_file/glob/grep) carries, and its readFile joined or used the
model-supplied path verbatim — so `path:"/etc/passwd"` (absolute) or
`path:"../../etc/..."` (escaping) read arbitrary files off disk AND opened them
into the language-server subprocess, reachable via indirect prompt injection.

Add NewScopedLSPNavigateTool(workspaceRoot, scope) (NewLSPNavigateTool now
delegates with nil scope), resolve the path through resolveScopedReadPath BEFORE
reading or building the LSP request — the same confinement the neighbouring
tools enforce — and hand the LSP manager the resolved absolute path so it never
opens an out-of-workspace file. Errors echo only the relative path, never an
absolute one. Registered via the scoped constructor in CoreReadOnlyToolsScoped.

Regression test asserts `/etc/passwd` and `../../` are rejected for both the
position and workspace_symbol ops; verified by mutation (bypassing the scope
resolution makes the test fail).
2026-06-20 23:12:15 +05:30

124 lines
4.2 KiB
Go

package agent
import (
"fmt"
"strings"
"github.com/Gitlawb/zero/internal/zeroruntime"
)
// Stale tool-output pruning reclaims context at ZERO token/latency cost before
// the loop falls back to the paid LLM summarizer. A long, dump-heavy session
// accumulates large read_file/grep/glob/bash tool results that the model has
// long since acted on; their bodies are dead weight. Pruning replaces those
// older bodies with a compact placeholder (keeping the tool message + its
// ToolCallID so provider replay stays valid and the model knows what was
// there), preserving recent turns exactly — no paraphrase loss.
const (
// pruneProtectRecentTokens is the trailing window of tool output kept
// verbatim — the model is most likely still using it.
pruneProtectRecentTokens = 40000
// pruneMinReclaimTokens gates the whole pass: only prune when the reclaimable
// (older, large) tool output exceeds this, so a short session is untouched.
pruneMinReclaimTokens = 20000
// pruneMinBodyTokens skips small tool results — replacing a tiny body saves
// nothing and just loses information.
pruneMinBodyTokens = 200
)
// prunedPlaceholder is the body a pruned tool result is replaced with. It names
// the tool and the original size so the model can re-fetch if it needs to.
func prunedPlaceholder(toolName string, originalTokens int) string {
if toolName == "" {
toolName = "tool"
}
return fmt.Sprintf("[pruned %s output (~%d tokens) to reclaim context — re-run the tool if you need it again]", toolName, originalTokens)
}
// pruneStaleToolOutput walks messages newest-first, protects the last
// preserveLast messages and a trailing pruneProtectRecentTokens of tool output,
// then replaces the body of older large tool results with a placeholder. It
// returns the (possibly) rewritten slice and the number of tokens reclaimed.
// When nothing meets the bar it returns the input unchanged with reclaimed=0.
//
// It never drops a message, never touches non-tool messages, and never prunes
// an already-pruned body — so it is idempotent and safe to run every turn.
func pruneStaleToolOutput(messages []zeroruntime.Message, preserveLast int) ([]zeroruntime.Message, int) {
if len(messages) == 0 {
return messages, 0
}
if preserveLast < 0 {
preserveLast = 0
}
// Index → original token cost of each prunable tool body, scanning newest
// first and skipping the protected suffix.
type candidate struct {
index int
tokens int
}
var candidates []candidate
recentToolTokens := 0
reclaimable := 0
protectUntil := len(messages) - preserveLast // indices >= this are protected
for i := len(messages) - 1; i >= 0; i-- {
msg := messages[i]
if msg.Role != zeroruntime.MessageRoleTool {
continue
}
bodyTokens := ApproxTextTokens(msg.Content)
if i >= protectUntil {
recentToolTokens += bodyTokens
continue
}
// Still within the trailing-output protection window?
if recentToolTokens < pruneProtectRecentTokens {
recentToolTokens += bodyTokens
continue
}
if bodyTokens < pruneMinBodyTokens || isPrunedPlaceholder(msg.Content) {
continue
}
candidates = append(candidates, candidate{index: i, tokens: bodyTokens})
reclaimable += bodyTokens
}
if reclaimable < pruneMinReclaimTokens || len(candidates) == 0 {
return messages, 0
}
// Copy-on-write so we never mutate the caller's slice in place.
out := make([]zeroruntime.Message, len(messages))
copy(out, messages)
reclaimed := 0
for _, c := range candidates {
toolName := toolNameForResult(out, c.index)
newBody := prunedPlaceholder(toolName, c.tokens)
reclaimed += c.tokens - ApproxTextTokens(newBody)
out[c.index].Content = newBody
}
return out, reclaimed
}
// toolNameForResult finds the tool name for the tool-result message at index by
// matching its ToolCallID against the assistant tool_call that produced it.
func toolNameForResult(messages []zeroruntime.Message, index int) string {
id := messages[index].ToolCallID
if id == "" {
return ""
}
for j := index - 1; j >= 0; j-- {
for _, call := range messages[j].ToolCalls {
if call.ID == id {
return call.Name
}
}
}
return ""
}
func isPrunedPlaceholder(content string) bool {
return strings.HasPrefix(strings.TrimSpace(content), "[pruned ")
}