Compare commits

...

2 Commits

Author SHA1 Message Date
Tim Lingo 62af5649fe feat(engine): port Anthropic server-side web_search into the agentic loop
Neuron Soul CI / build (pull_request) Failing after 10m45s
Neuron Soul CI / deploy (pull_request) Has been skipped
Re-authors soul-webfix-20260711.patch in El (the patch is a diff against
generated C at month-old offsets, so nothing was applied as a patch). Its
last two hunks — an unrelated /api/safety-contact implementation — were
deliberately not ported; that route already exists and is safety-critical.

Activation restores existing design, it does not invent a mechanism:
commit 8eea1d9 (2026-06-09, Tim-approved) made native web_search built-in
with no user-facing toggle, and tests/test_agentic_tools.el section 2 still
asserts agentic_tools_all() contains it — an assertion main currently fails.
The call site was lost when agentic_tools_all() (connector tools, PR #19)
replaced agentic_tools_with_web(). Attaching in agentic_tools_all() covers
handle_chat_agentic, handle_dharma_room_turn_agentic and agentic_resume.

pause_turn handling is included and was genuinely missing: the shipped
binary has zero occurrences of it. Without it a paused server-side search
returns only the text written so far and the loop treats it as final —
a silently truncated answer. final_text now accumulates across resume
cycles rather than overwriting (overwriting would discard everything
written before the pause).

Default tool version is web_search_20250305, NOT the newer _20260209, and
that is a measured choice: _20260209's dynamic filtering uses server-side
programmatic tool calling, which the API refuses to combine with ADR 0005's
stopgap —

  HTTP 400 invalid_request_error
  tool_choice.disable_parallel_tool_use: true cannot be used with
  programmatic tool calling

Dropping the stopgap would resurrect neuron#78 bug b (killed runs 3/3 in
ADR 0005's own A/B). The basic variant is compatible with the stopgap and
returns real results, so neither feature is dropped. Version lives in state
key web_search_tool_version; flipping it once the stopgap retires is a
config write, no recompile. The fallback also fires on "programmatic tool
calling" so a premature flip self-heals loudly instead of dying.

Also fixes a real bug this port exposed: json_get is a first-match scanner
and a cited text block serialises citations FIRST, so json_get(block,"type")
returned the nested citation's type and every citation-bearing block — the
ones carrying the searched facts — was silently dropped from the reply.
Reply length on the same question: 79 -> 360 chars.

Other loop changes: server_tool_use accounting into tools_used, iteration
cap 8->12 for pause/resume cycles, max_tokens 4096->16384, container-id
carry-forward, API error head logged instead of swallowed, tools_used gated
on is_tool_turn so a truncated tool block is not reported as work done.

dist/soul.c is deliberately NOT regenerated — the local elc predates Will's
last regen (e610a41) and regenerating with an older compiler risks unrelated
codegen drift. Source only; regen is Will's.

E2E-VERIFIED on a sandbox soul (scratch HOME/engram, dead axon+ISE, explicit
NEURON_PORT): live Bentonville weather with tools_used ["web_search"]; the
27-route contract gate PASSes; the parallel-tool stopgap still completes a
3-file mission with no 400; a 3-search 3,486-char answer kept its end
sentinel. Honest gap: pause_turn is compiled in but not exercised by a live
pause (max_uses:5 caps the server loop below the pause threshold).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-04 22:57:59 -05:00
Tim Lingo 710761e2d5 fix(engine): send tool_choice.disable_parallel_tool_use on agentic loop (STOPGAP)
The agentic loop keeps only the FIRST tool_use block per round (chat.el:2281,
"Capture first tool_use block only"). Anthropic lets a model emit several
tool_use blocks in one message and requires a tool_result for every one, so a
parallel-tool turn is answered once, the rest are dropped, and the next request
dies with:

  tool_use ids were found without tool_result blocks immediately after

(neuron#78 quotes this as "tool_use ids found without tool_result"; the above is
the API's actual wording - recorded so the next person's grep matches.)

This constrains the wire to match what the loop can assemble:
  "tool_choice":{"type":"auto","disable_parallel_tool_use":true}

STOPGAP - AND THE DURABLE FIX ALREADY EXISTS. A correct multi-tool loop is
already in Will's EL runtime, in C, and the soul does not call it. Verified on
el:origin/main lang/el-compiler/runtime/el_runtime.c: llm_register_tool:9616,
llm_build_tool_results:9743 - which walks EVERY content block, emits one
tool_result per tool_use, and sets is_error for an unregistered tool -
llm_call_agentic:9817 calling it at :9918, iteration cap 10 at :9847. Will's
commit 12d5e77 (2026-04-30). grep for llm_call_agentic/llm_register_tool across
every neuron/*.el returns nothing; dist/soul.c has zero references. chat.el
hand-rolls its own single-tool loop instead, and that is the one that breaks.

The durable fix is therefore to register the soul's tools via llm_register_tool
and call llm_call_agentic - deleting a loop, not writing one. See ADR 0005.

Our own approved spec called this seven weeks ago:
docs/research/agentic-tool-approval-design.md (2026-06-12, "Approved for build"),
line 20 on the defect, line 30 on the goal ("Execute all tool_use blocks in a
turn (one result per block)").

Two edits, because dist/soul.c cannot be regenerated here (Will's gated elc/elb
toolchain is not on this machine):
  (a) chat.el:2255 - source of truth, so a later regen carries the fix.
      One edit covers all three routes: agentic_loop is called from chat.el:2152
      (/api/chat agentic), :2676 (dharma room) and :2496 (agentic_resume).
  (b) dist/soul.c:28173 - generated form, hand-spliced. Line 27624 is the
      non-agentic/OpenAI-compat req_body (no tools) and was left untouched.
Prior art reused rather than reinvented: soul-narrated-runs-20260713.patch
(27,824 bytes) line 78 spliced this same string into the same concat chain on
2026-07-13.

Deliberately NOT ported from that patch, having read it: max_tokens is not
changed by it (16384 sits on both sides of the hunk; our main's 4096 is a
separate output-truncation concern), and its pause_turn pairing fix - same
defect class - is unreachable today because no server-side web_search is wired
(agentic_tools_with_web at chat.el:1418 is never called), so it is untestable
and logged instead.

Proven E2E on a scratch profile and port 7791, never the live chain. A/B against
a pristine origin/main control built from the same vendored runtime: fixed
completed the mission (tools_used read_file x3, 4 iterations, correct answer);
control failed 3/3. Direct API probe confirmed the mechanism - without the field
the model emits 3 parallel tool_use blocks and replaying the unfixed loop's next
turn returns HTTP 400; with it, exactly 1 block.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-08-03 17:07:04 -05:00
2 changed files with 265 additions and 23 deletions
+264 -22
View File
@@ -1410,6 +1410,71 @@ fn agentic_tools_literal() -> String {
"]"
}
// web_search_tool_json the ONE place the server-side web_search tool block is built.
//
// Anthropic executes this tool on their side, inside the same /v1/messages call: no
// third-party search key, no new account, and no anthropic-beta header (the plain
// "anthropic-version: 2023-06-01" this soul already sends is sufficient).
//
// FUTURE-PROOF (design carried from soul-webfix-20260711.patch, Tim-approved): the tool
// VERSION lives in exactly one place state key "web_search_tool_version" so a future
// bump is a config write, not a recompile.
//
// DEFAULT IS THE BASIC VARIANT, AND THAT IS A DELIBERATE, MEASURED CHOICE.
// The July patch defaulted to web_search_20260209 (dynamic filtering). That variant is
// HARD-INCOMPATIBLE with the parallel-tool stopgap this engine currently depends on
// (ADR 0005). Dynamic filtering is implemented with server-side programmatic tool
// calling, and the API rejects the pair outright verified live 2026-08-04, verbatim:
//
// HTTP 400 invalid_request_error
// "tool_choice.disable_parallel_tool_use: true cannot be used with programmatic
// tool calling"
//
// So the two cannot both ship today. Dropping the stopgap would resurrect neuron#78
// bug b, which killed agentic runs 3/3 in the ADR's own A/B a functional regression.
// Dropping web search entirely would leave the product's promise unmet. The basic
// variant is compatible with the stopgap AND returns real, current results (proven
// E2E), so it is what we default to. What we give up is dynamic filtering a
// token-efficiency and result-quality nicety, not the capability itself.
//
// REMOVAL TRIGGER: this default should move to web_search_20260209 the moment ADR
// 0005's stopgap is retired (i.e. when the soul calls el_runtime's llm_call_agentic
// and no longer needs disable_parallel_tool_use). At that point it is a one-line
// config write state_set("web_search_tool_version", "web_search_20260209") with
// no recompile. Flagged for Will's ratification; see the PR body.
fn web_search_tool_json() -> String {
let ver: String = state_get("web_search_tool_version")
let eff: String = if str_eq(ver, "") { "web_search_20250305" } else { ver }
return "{\"type\":\"" + eff + "\",\"name\":\"web_search\",\"max_uses\":5}"
}
// strip_client_web_search remove any CLIENT-side tool literally named "web_search"
// from a tools-array interior, so attaching Anthropic's server-side tool cannot produce
// two tools sharing one name (the API rejects that with "Tool names must be unique",
// and before it did, the model picked between them nondeterministically).
//
// The soul's own literal set carries "web_get", not "web_search", so today this is a
// no-op there it exists for the MCP connector tools merged in by agentic_tools_all(),
// which are third-party and may well ship a "web_search".
//
// LIMITATION (inherited from the July patch, stated rather than hidden): the scan keys
// on the object terminator "]}}," so it only strips a matching tool that is followed by
// another tool. A client web_search in the LAST array position is left in place.
fn strip_client_web_search(tools_inner: String) -> String {
let ws_start: Int = str_index_of(tools_inner, "{\"name\":\"web_search\"")
if ws_start < 0 {
return tools_inner
}
let ws_rest: String = str_slice(tools_inner, ws_start, str_len(tools_inner))
let ws_end: Int = str_index_of(ws_rest, "]}},")
if ws_end <= 0 {
return tools_inner
}
let head: String = str_slice(tools_inner, 0, ws_start)
let tail: String = str_slice(ws_rest, ws_end + 4, str_len(ws_rest))
return head + tail
}
// agentic_tools_with_web the standard tool set, always plus Anthropic's NATIVE
// server-side web_search tool. Web search is BUILT IN: the model invokes it only when a
// query needs fresh info (max_uses caps it), so there is no user-facing toggle. The native
@@ -1417,8 +1482,8 @@ fn agentic_tools_literal() -> String {
// and needs no local runtime it sidesteps the soul's lack of executable tools entirely.
fn agentic_tools_with_web() -> String {
let base: String = agentic_tools_literal()
let inner: String = str_slice(base, 1, str_len(base) - 1)
return "[" + inner + ",{\"type\":\"web_search_20250305\",\"name\":\"web_search\",\"max_uses\":5}]"
let inner: String = strip_client_web_search(str_slice(base, 1, str_len(base) - 1))
return "[" + inner + "," + web_search_tool_json() + "]"
}
// ---------------------------------------------------------------------------
@@ -1443,20 +1508,35 @@ fn connector_tools_json() -> String {
return arr
}
// Built-in tools + every connector tool, as one tools array.
// Uses agentic_tools_literal (not agentic_tools_with_web) to avoid a duplicate
// "web_search" name the literal already includes a custom web_search handler,
// and adding the Anthropic server-side web_search_20250305 (same name) causes
// Anthropic to reject with "Tool names must be unique."
// agentic_tools_all built-in tools + every connector tool + Anthropic's NATIVE
// server-side web_search, as one tools array. This is the tool set EVERY agentic route
// uses (handle_chat_agentic and handle_dharma_room_turn_agentic both call it; a
// bridge-suspended run carries the same array through agentic_resume), so attaching the
// web tool here is what actually turns web search on for the product.
//
// ACTIVATION restoring an existing design, not inventing a mechanism:
// * Commit 8eea1d9 (2026-06-09, Tim-approved) made native web_search BUILT IN with no
// user-facing toggle "the model invokes it only when a query needs fresh info
// (max_uses:5 caps it)" and the desktop app removed its web-search toggle to pair
// with that. The call site was lost when agentic_tools_all() (connector tools, PR #19)
// replaced agentic_tools_with_web() in handle_chat_agentic.
// * tests/test_agentic_tools.el §2 still asserts agentic_tools_all() contains the
// native web_search tool. main currently FAILS that assertion; this restores it.
// The old comment here claimed "the literal already includes a custom web_search handler"
// that is stale: agentic_tools_literal() ships web_get, and is_builtin_tool() has no
// web_search. The duplicate-name hazard now lives only in third-party connector tools,
// which is exactly what strip_client_web_search() handles.
fn agentic_tools_all() -> String {
let base: String = agentic_tools_literal()
let conn: String = connector_tools_json()
let base_inner: String = str_slice(base, 1, str_len(base) - 1)
let conn_inner: String = str_slice(conn, 1, str_len(conn) - 1)
if str_eq(conn_inner, "") {
return base
let merged: String = if str_eq(conn_inner, "") {
base_inner
} else {
base_inner + "," + conn_inner
}
let base_open: String = str_slice(base, 0, str_len(base) - 1)
return base_open + "," + conn_inner + "]"
return "[" + strip_client_web_search(merged) + "," + web_search_tool_json() + "]"
}
// Proxy one tool call to the bridge. The model-supplied input is written to a
@@ -2223,6 +2303,20 @@ fn agentic_loop(session_id: String, model: String, safe_sys: String, tools_json:
let iteration: Int = 0
let keep_going: Bool = true
// Server-side web_search state, all three carried across iterations.
//
// tools_eff: a local copy of the caller's tools array, because the DRIFT branch below
// may rewrite the web_search version in place after an API rejection. (Parameters are
// not rebindable across loop iterations; a top-level local is.)
// container_id: web_search's dynamic filtering runs server-side code execution on
// Anthropic's side, which hands back a container id that MUST be echoed on every
// follow-up request in the same turn chain or the API 400s with "container_id is
// required". Captured from each response, replayed on the next.
// ws_drift: set when we fell back, so we only ever fall back once per run.
let tools_eff: String = tools_json
let container_id: String = ""
let ws_drift: Bool = false
// Suspension state captured at top level so it escapes the while body.
let pending: Bool = false
let pend_tool_id: String = ""
@@ -2242,11 +2336,39 @@ fn agentic_loop(session_id: String, model: String, safe_sys: String, tools_json:
state_set("run_progress_" + session_id, "")
}
while keep_going && iteration < 8 {
// Cap raised 8 -> 12: a server-side web_search turn can pause and resume several
// times (see is_pause below), and each resume consumes an iteration. At 8 a
// multi-search answer could exhaust the loop before the model finished writing.
while keep_going && iteration < 12 {
// Carry the server-side code-execution container forward when we have one.
let cont_frag: String = if str_eq(container_id, "") {
""
} else {
",\"container\":\"" + container_id + "\""
}
// STOPGAP (2026-08-03, neuron#78 bug b / ADR 0005): the block walk below keeps
// only the FIRST tool_use block per round, so a PARALLEL tool_use response leaves
// the other ids without tool_result and the next turn 400s with
// "tool_use ids found without tool_result" killing the run. Sending Anthropic's
// documented tool_choice.disable_parallel_tool_use makes the wire contract match
// what this loop can actually assemble. The real fix a multi-tool_result loop
// that answers every block in a round is Will's; see neuron#78.
//
// The stopgap and server-side web_search coexist: disable_parallel_tool_use
// constrains how many CLIENT tool_use blocks the model may emit per response, and
// Anthropic runs web_search itself (it comes back as server_tool_use +
// web_search_tool_result, never as a client tool_use the loop has to answer).
// Verified live, not assumed see the README's probe transcript.
//
// max_tokens raised 4096 -> 16384: web_search injects whole result documents into
// the response, and a multi-search answer at 4096 truncated mid-sentence. This is
// the second half of the anti-truncation fix; pause_turn handling is the first.
let req_body: String = "{\"model\":\"" + model + "\""
+ ",\"max_tokens\":4096"
+ cont_frag
+ ",\"max_tokens\":16384"
+ ",\"tool_choice\":{\"type\":\"auto\",\"disable_parallel_tool_use\":true}"
+ ",\"system\":\"" + safe_sys + "\""
+ ",\"tools\":" + tools_json
+ ",\"tools\":" + tools_eff
+ ",\"messages\":" + messages
+ "}"
@@ -2255,11 +2377,60 @@ fn agentic_loop(session_id: String, model: String, safe_sys: String, tools_json:
let is_error: Bool = str_starts_with(raw_resp, "{\"error\"")
|| str_starts_with(raw_resp, "{\"type\":\"error\"")
|| str_contains(raw_resp, "authentication_error")
if is_error {
// DRIFT / FALLBACK: if the API rejected our configured web_search variant (the
// usual cause is a model too old for it chat_default_model() is sonnet-4-5,
// which predates web_search_20260209), swap to the known-old basic variant,
// record the drift for the drift-watch, and retry this iteration.
//
// Everything here is computed as top-level if-expressions rather than inside an
// if-BLOCK, because in El a mutation inside an if-block does not escape its scope
// (see the block-walk note below). There is deliberately no `continue`: the
// fallback simply leaves `messages` untouched and keeps `keep_going` true, so the
// next iteration re-sends the same turn with the corrected tools array.
let ws_pos: Int = if is_error { str_index_of(tools_eff, "\"type\":\"web_search_") } else { 0 - 1 }
let ws_rest: String = if ws_pos >= 0 { str_slice(tools_eff, ws_pos + 8, str_len(tools_eff)) } else { "" }
// '"type":"' is 8 chars, so the version starts at ws_pos+8 and ends at the next quote.
let ws_vend: Int = if ws_pos >= 0 { str_index_of(ws_rest, "\"") } else { 0 - 1 }
let ws_cur: String = if ws_vend > 0 { str_slice(ws_rest, 0, ws_vend) } else { "" }
// Fall back on either of two signals, and ONLY these two — a loose prefix guess
// here once matched errors that had nothing to do with the tool:
// 1. the API names OUR configured version in its complaint (version drift, e.g.
// a model too old for the configured variant); or
// 2. the API rejects the request for "programmatic tool calling" — the verified
// signature of the dynamic-filtering variant colliding with ADR 0005's
// disable_parallel_tool_use. An operator who sets web_search_tool_version to
// web_search_20260209 while the stopgap still stands would otherwise get a
// dead chat; this downgrades them automatically and says so in the log.
let ws_conflict: Bool = str_contains(raw_resp, "programmatic tool calling")
let can_fallback: Bool = is_error && !ws_drift && ws_pos >= 0 && ws_vend > 0
&& !str_eq(ws_cur, "") && !str_eq(ws_cur, "web_search_20250305")
&& (str_contains(raw_resp, ws_cur) || ws_conflict)
let tools_eff = if can_fallback {
str_slice(tools_eff, 0, ws_pos + 8) + "web_search_20250305" + str_slice(ws_rest, ws_vend, str_len(ws_rest))
} else { tools_eff }
let ws_drift = if can_fallback { true } else { ws_drift }
if can_fallback {
println("[soul] DRIFT: web_search variant '" + ws_cur + "' rejected by API - fell back to web_search_20250305")
state_set("web_search_version_drift", ws_cur)
}
if is_error && !can_fallback {
// Log the actual API error head — the old code swallowed it entirely, which
// made every upstream failure look identical from the outside.
let err_head: String = if str_len(raw_resp) > 220 { str_slice(raw_resp, 0, 220) } else { raw_resp }
println("[soul] llm error: " + err_head)
return "{\"error\":\"llm unavailable\",\"reply\":\"\"}"
}
let stop_reason: String = json_get(raw_resp, "stop_reason")
// Capture/refresh the server-side container id when the response carries one.
let cont_raw: String = json_get_raw(raw_resp, "container")
let cont_id_new: String = if !str_eq(cont_raw, "") && !str_eq(cont_raw, "null") {
json_get(cont_raw, "id")
} else { "" }
let container_id = if !str_eq(cont_id_new, "") { cont_id_new } else { container_id }
// json_get_raw needed — content is an array, json_get returns "" for non-strings
let content_arr: String = json_get_raw(raw_resp, "content")
let eff_content: String = if str_eq(content_arr, "") { "[]" } else { content_arr }
@@ -2271,13 +2442,39 @@ fn agentic_loop(session_id: String, model: String, safe_sys: String, tools_json:
let tool_id: String = ""
let tool_name: String = ""
let tool_input: String = ""
// Server-executed tool names seen this round, quoted and comma-joined. Kept
// separate from tools_log so this inner walk has exactly one mutation site per
// variable (the El scope rule below), then merged in at the outer level.
let srv_log: String = ""
let ci: Int = 0
let c_total: Int = json_array_len(eff_content)
while ci < c_total {
let block: String = json_array_get(eff_content, ci)
let btype: String = json_get(block, "type")
// CITATION-BLOCK FIX (2026-08-04, found by E2E once web_search was live).
// json_get is a first-match scanner, and a cited text block serialises as
// {"citations":[{"type":"web_search_result_location",...}],"type":"text",...}
// — citations FIRST. So json_get(block,"type") returns the nested citation's
// type, "text" never matches, and the block is silently dropped. Those are
// exactly the blocks carrying the searched facts, so the user got an answer
// with the data punched out of it ("The current temperature is , with .").
// Only text blocks carry a citations array, so its presence identifies the
// block type unambiguously without needing a real JSON parser.
let cit_raw: String = json_get_raw(block, "citations")
let has_cit: Bool = !str_eq(cit_raw, "") && !str_eq(cit_raw, "null")
let btype_scan: String = json_get(block, "type")
let btype: String = if has_cit { "text" } else { btype_scan }
// Accumulate text at top level using if-expression
let text_out = if str_eq(btype, "text") { text_out + json_get(block, "text") } else { text_out }
// FUTURE-PROOF: tools Anthropic runs on our behalf (web_search today, whatever
// ships tomorrow) arrive as server_tool_use blocks, never as client tool_use.
// Count the CATEGORY by the block's own name so a new server tool appears in
// tools_used with zero code changes — honest accounting for work we didn't run.
let is_srv: Bool = str_eq(btype, "server_tool_use")
let srv_name_raw: String = if is_srv { json_get(block, "name") } else { "" }
let srv_name: String = if is_srv && str_eq(srv_name_raw, "") { "server_tool" } else { srv_name_raw }
let srv_log = if is_srv {
if str_eq(srv_log, "") { "\"" + srv_name + "\"" } else { srv_log + ",\"" + srv_name + "\"" }
} else { srv_log }
// Capture first tool_use block only
let is_new_tool: Bool = str_eq(btype, "tool_use") && !has_tool
let has_tool = if is_new_tool { true } else { has_tool }
@@ -2291,6 +2488,24 @@ fn agentic_loop(session_id: String, model: String, safe_sys: String, tools_json:
// A real tool turn that targets a tool the soul cannot run in-process is a
// CLIENT bridge: suspend the loop and hand the tool to the client.
let is_tool_turn: Bool = str_eq(stop_reason, "tool_use") && has_tool
// pause_turn — REQUIRED for server-side web_search, not optional.
// Anthropic's server-side tool loop has its own iteration limit. When it hits it
// mid-answer the turn comes back with stop_reason "pause_turn" and only the text
// written SO FAR. Per Anthropic's contract the caller must re-send with the
// assistant content appended and NO tool_result, and the server resumes where it
// left off. Skip this and the user silently gets a truncated answer — no error,
// no warning, just a sentence that stops. (The bug that ate Tim's report,
// 2026-07-11.) Nothing else in this engine handles it: "pause_turn" appears
// nowhere else in the .el sources or in the shipped binary.
let is_pause: Bool = str_eq(stop_reason, "pause_turn")
// Unknown stop reasons (future API drift): finish honestly and log loudly rather
// than treating an unrecognised terminal state as a completed answer.
if !str_eq(stop_reason, "end_turn") && !str_eq(stop_reason, "tool_use") && !is_pause
&& !str_eq(stop_reason, "max_tokens") && !str_eq(stop_reason, "refusal")
&& !str_eq(stop_reason, "") {
println("[soul] DRIFT: unknown stop_reason from API: " + stop_reason)
}
// If the user previously chose "always allow" for this tool in this session,
// treat it like a builtin — run server-side via dispatch_tool and skip the
// bridge suspension entirely so the approval UI is never shown again.
@@ -2318,10 +2533,19 @@ fn agentic_loop(session_id: String, model: String, safe_sys: String, tools_json:
let tool_msg: String = "{\"type\":\"tool_result\",\"tool_use_id\":\"" + tool_id + "\",\"content\":\"" + tool_result + "\"}"
// Accumulate tool names for the tools_used log surfaced in the response.
// Gated on is_tool_turn, not has_tool: a round truncated by max_tokens can carry a
// half-written tool_use block that will never run, and logging it would claim work
// the soul did not do.
let tool_quoted: String = "\"" + tool_name + "\""
let tools_log = if has_tool {
let tools_log = if is_tool_turn {
if str_eq(tools_log, "") { tool_quoted } else { tools_log + "," + tool_quoted }
} else { tools_log }
// Merge in the server-executed tools (web_search) seen in this round's blocks.
let tools_log = if str_eq(srv_log, "") {
tools_log
} else {
if str_eq(tools_log, "") { srv_log } else { tools_log + "," + srv_log }
}
// The assistant turn that requested the tool — needed verbatim on resume so the
// tool_use/tool_result pairing stays valid when the client posts its result.
@@ -2332,15 +2556,22 @@ fn agentic_loop(session_id: String, model: String, safe_sys: String, tools_json:
// Local built-in tool turn: append assistant + tool_result and keep looping.
let local_continue: Bool = is_tool_turn && !needs_bridge
// Pause resume: the assistant content goes back VERBATIM with NO tool_result —
// that is the whole contract. Anthropic's server resumes its own tool loop from
// there; sending a tool_result (or a "continue" user turn) breaks it.
let is_pause_resume: Bool = is_pause && !needs_bridge
let messages = if local_continue {
let inner2: String = str_slice(messages_with_assistant, 1, str_len(messages_with_assistant) - 1)
"[" + inner2 + ",{\"role\":\"user\",\"content\":[" + tool_msg + "]}]"
} else if is_pause_resume {
messages_with_assistant
} else { messages }
// Live progress ledger: one entry per round — the model's own narration
// (its pre-tool prose, previously discarded here) plus the tool it reached
// for. Clients poll /api/run-progress/<sid> to render these live.
if !str_eq(session_id, "") {
// Skipped on a version-fallback round: nothing happened that a poller should see.
if !str_eq(session_id, "") && !can_fallback {
let prog_key: String = "run_progress_" + session_id
let prog_prev: String = state_get(prog_key)
let prog_snip: String = if str_len(text_out) > 280 { str_slice(text_out, 0, 280) } else { text_out }
@@ -2365,8 +2596,19 @@ fn agentic_loop(session_id: String, model: String, safe_sys: String, tools_json:
bridge_save(session_id, model, safe_sys, tools_json, messages_with_assistant, tools_log, pend_tool_id)
}
let final_text = if !is_tool_turn { text_out } else { final_text }
let keep_going = if local_continue { keep_going } else { false }
// ACCUMULATE across pause/resume cycles instead of overwriting. A resumed turn
// CONTINUES the answer, it does not repeat it — overwriting here would throw away
// everything the model wrote before the pause, which is the same truncation the
// pause handling exists to prevent. A version-fallback round contributes nothing.
let final_text = if !is_tool_turn && !can_fallback { final_text + text_out } else { final_text }
// Output cap hit mid-action: the tool block is truncated and will NOT run. Say so
// instead of ending on silent almost-work.
let final_text = if str_eq(stop_reason, "max_tokens") && has_tool {
final_text + "\n\n[Output limit reached mid-action - the last planned action did not run. Ask me to continue to finish it.]"
} else { final_text }
// Keep looping for a local tool round, a server-side pause resume, or a one-shot
// web_search version fallback.
let keep_going = if local_continue || is_pause_resume || can_fallback { keep_going } else { false }
let iteration = iteration + 1
}
@@ -2390,9 +2632,9 @@ fn agentic_loop(session_id: String, model: String, safe_sys: String, tools_json:
// means the task was too complex for the agentic loop depth — surface it clearly
// so the caller/operator knows to increase the cap or break the task apart.
if str_eq(final_text, "") {
let hit_cap: Bool = iteration >= 8
let hit_cap: Bool = iteration >= 12
let err_msg: String = if hit_cap {
"agentic loop hit the 8-iteration cap without producing a final reply - task may be too complex or a tool call is looping"
"agentic loop hit the 12-iteration cap without producing a final reply - task may be too complex or a tool call is looping"
} else {
"no response"
}
Generated Vendored
+1 -1
View File
@@ -28170,7 +28170,7 @@ el_val_t agentic_loop(el_val_t session_id, el_val_t model, el_val_t safe_sys, el
state_set(el_str_concat(EL_STR("run_progress_"), session_id), EL_STR(""));
}
while (keep_going && (iteration < 8)) {
el_val_t req_body = el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(EL_STR("{\"model\":\""), model), EL_STR("\"")), EL_STR(",\"max_tokens\":4096")), EL_STR(",\"system\":\"")), safe_sys), EL_STR("\"")), EL_STR(",\"tools\":")), tools_json), EL_STR(",\"messages\":")), messages), EL_STR("}"));
el_val_t req_body = el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(el_str_concat(EL_STR("{\"model\":\""), model), EL_STR("\"")), EL_STR(",\"max_tokens\":4096,\"tool_choice\":{\"type\":\"auto\",\"disable_parallel_tool_use\":true}")), EL_STR(",\"system\":\"")), safe_sys), EL_STR("\"")), EL_STR(",\"tools\":")), tools_json), EL_STR(",\"messages\":")), messages), EL_STR("}"));
el_val_t raw_resp = http_post_with_headers(api_url, req_body, h);
el_val_t is_error = ((str_starts_with(raw_resp, EL_STR("{\"error\"")) || str_starts_with(raw_resp, EL_STR("{\"type\":\"error\""))) || str_contains(raw_resp, EL_STR("authentication_error")));
if (is_error) {