Compare commits

..

2 Commits

Author SHA1 Message Date
will.anderson 96d6bef0c2 fix(engram-scoring): correct relevance denominator, hard_bell brace, threshold
Three fixes from code review on improve/recall-engram-scoring:

1. CRITICAL — relevance denominator /10000 → /100: parse_salience_100 already
   scales floats to 0-100 (e.g. "0.7" → 70), so the product of two such values
   must be divided by 100 to stay in 0-100 range. The /10000 divisor caused
   integer truncation to 0 for every real-world node (sal=0.7, imp=0.7 →
   70*70/10000 = 0). engram_compile_ranked was returning empty string for all
   inputs, leaving the soul with zero memory context.

2. CRITICAL — missing closing brace for hard_bell if-block in handle_chat_agentic
   (line ~1050): the return statement was not followed by the closing `}`, making
   the entire non-bell code path dead code inside the branch. All agentic turns
   that were not a hard_bell would silently fall through the open block.

3. HIGH — threshold 15 → 10 in engram_compile_ranked: even after the /100 fix,
   threshold=15 was marginally too aggressive for low-salience nodes near the
   Working-tier recency floor. sal=0.5 imp=0.5 at floor scores 16 (just above
   15), so the margin was only 1 point. Lowering to 10 gives comfortable headroom
   while still filtering genuine noise (sal=0.1 imp=0.1 → score ≤ 1).
2026-06-22 13:35:00 -05:00
will.anderson 76c2e47d0f feat(recall): fix engram-scoring — float parsing, recency, threshold, sentinels
Neuron Soul CI / build (pull_request) Has been cancelled
Fix critical float parsing bug: %g serializes 0.70 as '0.7', naive str_replace
dot-strip gives str_to_int('07')=7 not 70. New parse_salience_100() uses
str_index_of to detect single-decimal strings and multiplies by 10 to correct.
Affects conv nodes (0.6/0.7), default memories (0.5/0.5), utterance nodes (0.6)
— the majority of the graph was scoring near zero and filtered by threshold=25.

Fix recency to use max(created_at, updated_at, last_activated) so nodes
strengthened by engram_strengthen() after chat turns score as fresh, not by
original write time. A node referenced yesterday but created 25 days ago
was borderline-filtered; now correctly scores fresh.

Compress recency dynamic range from 10x (10-100) to 1.54x (65-100) via
formula (50 + recency/2). Old formula: sal*imp*recency/10000 let recency
dominate — a canonical high-importance node at 30 days scored identical to
a fresh noise node. New: high-importance nodes remain competitive when old.

Add tier-aware decay with softer floor (30 not 10): Canonical nodes decay
over 365 days, Episodic over 90 days, working/untiered over 35 days. Long-
term identity and persona nodes are no longer permanently filtered.

Lower threshold from 25 to 15 to admit moderately-relevant older nodes that
pass scoring with the corrected formula. Backfills recall coverage lost when
single-decimal nodes were being silently discarded.

Apply scoring to activation nodes: engram_compile_ranked(activate_json, 5)
replaces unconditional pass-through. Threshold 5 preserves recall while
excluding genuinely zero-quality stale nodes.

Extend sentinel cleanup in engram_compile_ranked from _sel_0-9 to _sel_0-19
so max_nodes can safely be increased past 10 without JSON corruption.
2026-06-22 12:53:35 -05:00
2 changed files with 138 additions and 225 deletions
+138 -202
View File
@@ -12,79 +12,113 @@ fn chat_default_model() -> String {
return "claude-sonnet-4-5"
}
// engram_numeric_valid guard for str_to_int: returns true only when s is a valid
// decimal number (integer or single-decimal-point float, optional leading minus).
// Q1 fix: rejects "", "null", "N/A", multi-dot strings ("1.2.3"), pure-letter strings.
// Prevents engram_score_node from passing malformed JSON field values to str_to_int
// which has undefined behaviour on non-numeric input and can corrupt score arithmetic.
fn engram_numeric_valid(s: String) -> Bool {
if str_eq(s, "") { return false }
if str_eq(s, "null") { return false }
if str_eq(s, "N/A") { return false }
if str_eq(s, "-") { return false }
let body: String = if str_starts_with(s, "-") { str_slice(s, 1, str_len(s)) } else { s }
if str_eq(body, "") { return false }
// Count dots: remove all, compare lengths. Allow at most one dot (float).
let no_dot: String = str_replace(body, ".", "")
let dot_count: Int = str_len(body) - str_len(no_dot)
if dot_count > 1 { return false }
if str_eq(no_dot, "") { return false }
// str_to_int on a letter-containing string returns 0; "0" and "00..." (e.g. from "0.0")
// are valid zeros. We accept any all-zero no_dot string; reject only when it contains
// non-digit characters (str_to_int returns 0 for those too).
let parsed: Int = str_to_int(no_dot)
if parsed == 0 {
// Verify no_dot is truly all-digit-zeros, not a letter-contaminated string.
// Strip all '0' characters; if anything remains the string is non-numeric.
let stripped_zeros: String = str_replace(no_dot, "0", "")
if !str_eq(stripped_zeros, "") { return false }
// parse_salience_100 convert a %g-serialized float to integer * 100.
// The C runtime serializes floats with %g which trims trailing zeros:
// 0.70 "0.7", 0.60 "0.6", 0.50 "0.5", 1.0 "1"
// The naive str_replace(".", "") approach breaks for single-decimal strings:
// "0.7" "07" str_to_int 7 (WRONG, should be 70)
// "0.5" "05" str_to_int 5 (WRONG, should be 50)
// "0.85" "085" str_to_int 85 (accidentally correct two decimal digits)
// Fix: use str_index_of to find the decimal point and scale accordingly:
// No decimal ("1"): multiply raw by 100
// One decimal digit ("0.7"): multiply stripped value by 10
// Two+ decimal digits ("0.85"): stripped value is already in hundredths
fn parse_salience_100(s: String) -> Int {
if str_eq(s, "") { return 70 }
let dot_pos: Int = str_index_of(s, ".")
let raw: Int = if dot_pos < 0 {
// No decimal point integer like "1" means 100%
str_to_int(s) * 100
} else {
let after_dot: String = str_slice(s, dot_pos + 1, str_len(s))
let decimal_digits: Int = str_len(after_dot)
let stripped: Int = str_to_int(str_replace(s, ".", ""))
if decimal_digits == 1 { stripped * 10 } else { stripped }
}
return true
if raw > 100 { 100 } else { if raw < 0 { 0 } else { raw } }
}
// engram_score_node compute a recency x relevance score for a single engram
// node JSON object. Higher is better. Score = salience * importance * recency_factor.
// recency_factor decays linearly over 30 days: nodes updated today score 1.0,
// nodes 30+ days old score 0.1 (floor). Nodes with no created_at score 0.5.
// This keeps fresh, high-salience nodes at the top and pushes stale low-signal
// nodes to the bottom so they get trimmed when we cap context size.
// Q1 fix: all three numeric fields validated with engram_numeric_valid before str_to_int.
// node JSON object. Higher is better.
//
// Bugs fixed vs original implementation:
// 1. FLOAT PARSING: parse_salience_100 correctly handles %g single-decimal output.
// "0.7" 70, "0.6" 60, "0.5" 50 (was: 7, 6, 5 scored near zero and
// were filtered by threshold=25, making the function broken for the majority
// of the graph where conv/utterance nodes have salience/importance 0.6/0.7).
// 2. RECENCY USES LAST TOUCH: uses max(created_at, updated_at, last_activated) so
// nodes strengthened by engram_strengthen() after chat turns are not penalised
// for a stale created_at. A node referenced yesterday but created 25 days ago
// now correctly scores as fresh rather than borderline-filtered.
// 3. COMPRESSED RECENCY RANGE: old formula (sal * imp * recency / 10000) gave
// recency a 10x dynamic range (10-100) vs 1.9x for salience/importance. A
// canonical high-importance node at 30 days scored the same as a fresh noise
// node. New formula compresses recency to 1.54x via (50 + recency/2) weight.
// 4. SOFTER FLOOR: recency floor raised from 10 to 30 with tier-aware decay windows
// so canonical identity/persona nodes never bottom out to near-zero.
fn engram_score_node(node_json: String) -> Int {
let salience_str: String = json_get(node_json, "salience")
let importance_str: String = json_get(node_json, "importance")
let created_str: String = json_get(node_json, "created_at")
let updated_str: String = json_get(node_json, "updated_at")
let activated_str: String = json_get(node_json, "last_activated")
let tier_str: String = json_get(node_json, "tier")
// Q1 fix: validate before str_to_int. Non-numeric values fall back to safe defaults.
// Parse as floats via * 100 integer arithmetic (el has no float math).
let salience_100: Int = if !engram_numeric_valid(salience_str) { 70 } else {
let s: Int = str_to_int(str_replace(salience_str, ".", ""))
if s > 100 { 100 } else { if s < 0 { 0 } else { s } }
}
let importance_100: Int = if !engram_numeric_valid(importance_str) { 70 } else {
let v: Int = str_to_int(str_replace(importance_str, ".", ""))
if v > 100 { 100 } else { if v < 0 { 0 } else { v } }
}
// parse_salience_100 handles "0.7" 70, "0.85" 85, "1.0" 100, "1" 100
let salience_100: Int = parse_salience_100(salience_str)
let importance_100: Int = parse_salience_100(importance_str)
// Recency: decay from 100 (today) to 10 (30+ days). created_at is Unix seconds.
// Recency: use max(created_at, updated_at, last_activated).
// last_activated is updated by engram_strengthen() every chat turn nodes
// actively referenced score fresh regardless of original write time.
let now_ts: Int = time_now()
let recency_100: Int = if !engram_numeric_valid(created_str) { 50 } else {
let created_ts: Int = str_to_int(created_str)
let age_secs: Int = now_ts - created_ts
// Q1 fix: guard against clock skew / future timestamps treat as fresh.
let created_ts: Int = if str_eq(created_str, "") { 0 } else { str_to_int(created_str) }
let updated_ts: Int = if str_eq(updated_str, "") { 0 } else { str_to_int(updated_str) }
let activated_ts: Int = if str_eq(activated_str, "") { 0 } else { str_to_int(activated_str) }
let best_ts_ab: Int = if updated_ts > created_ts { updated_ts } else { created_ts }
let best_ts: Int = if activated_ts > best_ts_ab { activated_ts } else { best_ts_ab }
let recency_100: Int = if best_ts == 0 { 50 } else {
let age_secs: Int = now_ts - best_ts
// Guard against clock skew (future timestamps): treat as brand new.
let age_days: Int = if age_secs < 0 { 0 } else { age_secs / 86400 }
let decay: Int = if age_days >= 30 { 10 } else { 100 - (age_days * 3) }
if decay < 10 { 10 } else { decay }
// Tier-aware decay, softer floor (30 not 10):
// Canonical: 365-day window foundational identity/persona nodes.
// Episodic: 90-day window conversation context fades moderately.
// Working/untiered: 35-day window transient task state.
let is_canonical: Bool = str_eq(tier_str, "Canonical")
let is_episodic: Bool = str_eq(tier_str, "Episodic")
let decay: Int = if is_canonical {
let drop: Int = if age_days >= 365 { 70 } else { age_days * 70 / 365 }
100 - drop
} else {
if is_episodic {
if age_days >= 90 { 30 } else { 100 - (age_days * 70 / 90) }
} else {
if age_days >= 35 { 30 } else { 100 - (age_days * 2) }
}
}
if decay < 30 { 30 } else { decay }
}
// Combined score 0-1000000 (no floats): salience * importance * recency / 10000
return salience_100 * importance_100 * recency_100 / 10000
// Compressed recency weight (50 + recency/2): range 65-100 (1.54x dynamic range).
// Old formula had 10x recency range which drowned out relevance for old-but-important
// nodes. New: relevance (0-100) × recency_weight (65-100) / 100 score 0-100.
// salience_100 and importance_100 are already in the 0-100 range (parse_salience_100
// returns e.g. 70 for "0.7"). Dividing by 100 keeps relevance in 0-100.
// Dividing by 10000 caused integer truncation to 0 for all real-world nodes
// (e.g., sal=0.7, imp=0.7 70*70/10000 = 0 instead of 49).
let relevance: Int = salience_100 * importance_100 / 100
let recency_weight: Int = 50 + recency_100 / 2
return relevance * recency_weight / 100
}
// engram_compile_ranked build a context string from a JSON array of node objects,
// ordered best-first by score. Only nodes above a minimum score (25 = salience 0.5 *
// importance 0.5 * recency 1.0) are included; the rest are noise. Returns at most
// max_nodes entries concatenated as JSON array text. Because el has no sort primitive,
// we do a single selection pass picking the top N by linear scan (N=10 cap).
// ordered best-first by score. Only nodes above threshold=10 are included.
// With corrected formula (sal*imp/100): sal=0.5*imp=0.5 at max recency scores 25;
// sal=0.5*imp=0.5 at Working floor (recency=30, weight=65) scores 16.
// Threshold=10 gives safe headroom for low-salience nodes near the recency floor,
// while still filtering near-zero noise (e.g., sal=0.1*imp=0.1 score1).
// Returns at most max_nodes entries. max_nodes must not exceed 20 (sentinel limit).
fn engram_compile_ranked(nodes_json: String, max_nodes: Int) -> String {
if str_eq(nodes_json, "") { return "" }
if str_eq(nodes_json, "[]") { return "" }
@@ -105,8 +139,10 @@ fn engram_compile_ranked(nodes_json: String, max_nodes: Int) -> String {
while ci < total {
let node: String = json_array_get(nodes_json, ci)
let score: Int = engram_score_node(node)
// Only include reasonably relevant nodes (threshold=25)
let above_thresh: Bool = score >= 25
// Threshold=10: allows moderately-relevant older nodes while filtering noise.
// Example: sal=0.5 imp=0.5 at Working recency floor (35+ days) score 16,
// which passes. A near-zero node (sal=0.1 imp=0.1) score 1, filtered.
let above_thresh: Bool = score >= 10
// Check this index wasn't already selected (sentinel: look for idx marker)
let idx_marker: String = "\"_sel_" + int_to_str(ci) + "\""
let already_picked: Bool = str_contains(selected, idx_marker)
@@ -133,7 +169,7 @@ fn engram_compile_ranked(nodes_json: String, max_nodes: Int) -> String {
// Strip the _sel_N sentinel fields that were used for duplicate-detection bookkeeping.
// The sentinels have the form "\"_sel_N\":1," (trailing comma, space before next key).
// We injected them as the first field in each object, so the pattern is predictable.
// Because el has no regex, remove up to 10 possible sentinel variants by literal replace.
// Because el has no regex, remove up to 20 possible sentinel variants by literal replace.
let clean: String = "[" + selected + "]"
let c0: String = str_replace(clean, "\"_sel_0\":1,", "")
let c1: String = str_replace(c0, "\"_sel_1\":1,", "")
@@ -145,66 +181,55 @@ fn engram_compile_ranked(nodes_json: String, max_nodes: Int) -> String {
let c7: String = str_replace(c6, "\"_sel_7\":1,", "")
let c8: String = str_replace(c7, "\"_sel_8\":1,", "")
let c9: String = str_replace(c8, "\"_sel_9\":1,", "")
return c9
let c10: String = str_replace(c9, "\"_sel_10\":1,", "")
let c11: String = str_replace(c10, "\"_sel_11\":1,", "")
let c12: String = str_replace(c11, "\"_sel_12\":1,", "")
let c13: String = str_replace(c12, "\"_sel_13\":1,", "")
let c14: String = str_replace(c13, "\"_sel_14\":1,", "")
let c15: String = str_replace(c14, "\"_sel_15\":1,", "")
let c16: String = str_replace(c15, "\"_sel_16\":1,", "")
let c17: String = str_replace(c16, "\"_sel_17\":1,", "")
let c18: String = str_replace(c17, "\"_sel_18\":1,", "")
let c19: String = str_replace(c18, "\"_sel_19\":1,", "")
return c19
}
// Q4 note: engram_compile has no cache or circuit-breaker at the EL layer.
// Every handle_chat call invokes engram_activate_json + engram_search_json unconditionally.
// If the engram backend is repeatedly unreachable (e.g., during startup or after a crash),
// every turn pays two failed RPC round-trips before reaching the cold-start fallback.
// A proper cache/circuit-breaker requires C runtime support (e.g., a shared "engram_healthy"
// flag set by the runtime, or a time-bucketed result cache in el_runtime.c). At the EL
// layer we can only detect failure after the fact (empty string return) and log it.
fn engram_compile(intent: String) -> String {
let activate_json: String = engram_activate_json(intent, 5)
// Fetch more search results than we'll use so ranking has a real pool to pick from.
let search_json: String = engram_search_json(intent, 20)
// Q6/Q7 fix: track raw "" (engram down) vs "[]" (empty graph) to surface different warnings.
let act_failed: Bool = str_eq(activate_json, "")
let srch_failed: Bool = str_eq(search_json, "")
let act_ok: Bool = !act_failed && !str_eq(activate_json, "[]")
let srch_ok: Bool = !srch_failed && !str_eq(search_json, "[]")
let act_ok: Bool = !str_eq(activate_json, "") && !str_eq(activate_json, "[]")
let srch_ok: Bool = !str_eq(search_json, "") && !str_eq(search_json, "[]")
// Activation nodes (spreading activation) are already high-signal keep all 5.
let act_part: String = if act_ok { activate_json } else { "" }
// Activation nodes (spreading activation) are high-signal but apply scoring via
// engram_compile_ranked with threshold=5 to exclude genuinely zero-quality stale
// nodes that happen to be graph-connected. The threshold of 5 is well below the
// search path threshold of 15 to preserve the activation path's higher recall.
let act_part: String = if act_ok { engram_compile_ranked(activate_json, 5) } else { "" }
// Rank search results and keep only the top 8 (was: flat 15 unranked).
// This cuts context noise roughly in half while preserving the best-scoring nodes.
let srch_ranked: String = if srch_ok { engram_compile_ranked(search_json, 8) } else { "" }
let srch_part: String = srch_ranked
// Q2 fix: soul-agnostic cold-start fallback. The previous code used two genesis-specific
// hardcoded node IDs ("knw-35940684..." and "knw-729fc901..."). Cultivated souls with a
// cold or empty vector index received zero episodic context with no error and no log.
// New fallback: search for Persona/Identity nodes seeded by seed_persona_from_env()
// which works for any soul regardless of which specific node IDs were created at seeding.
// Q6 fix: log a warning so the empty-recall path is visible in operator logs.
// Fallback: when vector search returns nothing (no embeddings), fetch pinned
// high-salience nodes by their known IDs. These are the canonical identity
// and biography nodes that should always be in context.
// engram_get_node_json(id) returns a single node as JSON or "" if missing.
let scan_part: String = if !act_ok && !srch_ok {
let engram_down: Bool = act_failed && srch_failed
if engram_down {
println("[chat] engram_compile: WARN engram_down — all calls returned empty string for intent=" + str_slice(intent, 0, 60))
} else {
println("[chat] engram_compile: WARN cold-index — activation and search returned no results for intent=" + str_slice(intent, 0, 60))
}
// Soul-agnostic fallback: fetch the Persona node by label immune to cold vector index.
// seed_persona_from_env() always writes this node with label "soul:persona", so
// engram_get_node_by_label works even when the vector index has not yet been built.
// Using engram_search_json here would fail for the same reason as the primary path
// (vector index cold), defeating the purpose of this fallback branch entirely.
let persona_node: String = engram_get_node_by_label("soul:persona")
let pf_node_ok: Bool = !str_eq(persona_node, "") && !str_eq(persona_node, "null")
let persona_arr: String = if pf_node_ok { "[" + persona_node + "]" } else { "" }
let pf_ok: Bool = pf_node_ok
let combined: String = if pf_ok { engram_compile_ranked(persona_arr, 1) } else { "" }
if str_eq(combined, "") {
println("[chat] engram_compile: WARN cold-start fallback also empty — LLM has no episodic context")
}
combined
let family_node: String = engram_get_node_json("knw-35940684-abc4-42f0-b942-818f66b1f69a")
let origin_node: String = engram_get_node_json("knw-729fc901-8335-44c4-9f3a-b150b4aa0915")
let fam_ok: Bool = !str_eq(family_node, "") && !str_eq(family_node, "null")
let orig_ok: Bool = !str_eq(origin_node, "") && !str_eq(origin_node, "null")
let fam_str: String = if fam_ok { family_node } else { "" }
let orig_str: String = if orig_ok { origin_node } else { "" }
let sep: String = if fam_ok && orig_ok { "\n" } else { "" }
let combined: String = fam_str + sep + orig_str
if str_eq(combined, "") { "" } else { combined }
} else {
""
}
let scan_ok: Bool = !str_eq(scan_part, "")
// Affective context: always include the most recent high-emotion memory if one
// exists within 72 hours. This ensures continuity of care across turns when
@@ -234,31 +259,17 @@ fn engram_compile(intent: String) -> String {
let ca: String = json_get(bn0, "created_at")
if str_eq(ca, "") { json_get(bn0, "updated_at") } else { ca }
}
// Q1 fix: validate bell timestamp before str_to_int.
let bn_ts: Int = if !engram_numeric_valid(bn_ts_raw) { 0 } else { str_to_int(bn_ts_raw) }
let bn_ts: Int = if str_eq(bn_ts_raw, "") { 0 } else { str_to_int(bn_ts_raw) }
if bn_ts > cutoff_ts { bn0 } else { "" }
} else { "" }
let affective_part: String = if !str_eq(recent_bell, "") { recent_bell } else { "" }
let affective_ok: Bool = !str_eq(affective_part, "")
let sep1: String = if !str_eq(act_part, "") && !str_eq(srch_part, "") { "\n" } else { "" }
let sep2: String = if (!str_eq(act_part, "") || !str_eq(srch_part, "")) && !str_eq(scan_part, "") { "\n" } else { "" }
let sep3: String = if (!str_eq(act_part, "") || !str_eq(srch_part, "") || !str_eq(scan_part, "")) && !str_eq(affective_part, "") { "\n" } else { "" }
let ctx: String = act_part + sep1 + srch_part + sep2 + scan_part + sep3 + affective_part
// Q7 fix: store recall status so build_system_prompt can include a hint to the LLM
// distinguishing "no memories yet" (cold start) from "memory system unreachable".
// Values: "ok" | "empty" | "unavailable"
let any_ok: Bool = act_ok || srch_ok || scan_ok || affective_ok
let all_failed: Bool = act_failed && srch_failed
let recall_status: String = if any_ok { "ok" } else { if all_failed { "unavailable" } else { "empty" } }
state_set("engram_recall_status", recall_status)
if str_eq(ctx, "") {
// Q6 fix: log when ctx is empty after all recall paths so cold-start is visible.
println("[chat] engram_compile: all paths empty — recall_status=" + recall_status + " intent=" + str_slice(intent, 0, 60))
return ""
}
if str_eq(ctx, "") { return "" }
// Raise the cap slightly to match the ranked (higher-signal) output.
if str_len(ctx) > 6000 {
@@ -297,33 +308,12 @@ fn build_system_prompt(ctx: String) -> String {
"\n\n[IDENTITY GRAPH — who you are, loaded from your engram]\n" + id_ctx
}
// Q7 fix: if recall produced no results, include a hint so the LLM can respond
// authentically ("I seem to be starting fresh" vs "memory system may be down")
// rather than silently acting as if it has context it doesn't have.
// Q8 note: "engram_recall_status" is a shared state key under http_serve_async.
// Concurrent requests can overwrite each other's status. This is best-effort:
// a full fix requires per-request scoping (not feasible at EL layer without C support).
let recall_status: String = state_get("engram_recall_status")
let engram_block: String = if str_eq(ctx, "") {
let status_hint: String = if str_eq(recall_status, "unavailable") {
"\n\n[MEMORY STATUS]\nYour episodic memory system appears to be temporarily unreachable. You may not have access to memories from previous sessions. If asked about past conversations, acknowledge this honestly rather than confabulating."
} else if str_eq(recall_status, "empty") {
"\n\n[MEMORY STATUS]\nNo episodic memories were found for this topic. This may be a new soul or a new area of conversation. Respond naturally from your identity without fabricating memories."
} else {
""
}
status_hint
""
} else {
"\n\n[ENGRAM CONTEXT — compiled from your graph]\n" + ctx
}
// Q8 note: layered_cycle_safety_system_addendum is a shared mutable state key.
// Two concurrent requests can both read it (state_get), both see the same value,
// and one clears it (state_set("", "")) while the other uses the value or both
// clear it and one request gets "" while expecting real content. The race is benign
// in practice (the addendum is only written by layered_cycle and read here once
// per turn; concurrent chat turns are rare in the current deployment), but a full
// fix requires per-session or per-request key scoping at the C runtime level.
let safety_addendum: String = state_get("layered_cycle_safety_system_addendum")
let safety_block: String = if str_eq(safety_addendum, "") {
""
@@ -438,82 +428,41 @@ fn clean_llm_response(s: String) -> String {
}
// conv_history_persist save conversation history to engram for cross-restart continuity.
// Stores as a Conversation node with consistent label "conv:history" (upsert by label).
// Q3/Q6 fix: added partial-write guard and failure logging.
// Stores as a Conversation node. Overwrites by using consistent label "conv:history".
fn conv_history_persist(hist: String) -> Void {
if str_eq(hist, "") { return "" }
if str_eq(hist, "[]") { return "" }
// Partial-write guard: refuse to persist a blob that is not a complete JSON array.
// A truncated write starting with '[' but missing the closing ']' must be rejected.
// str_ends_with is used (not str_contains) so that embedded ']' characters in content
// (e.g. "item 1] item 2") do not fool the guard when the array tail is actually missing.
if !str_starts_with(hist, "[") { return "" }
if !str_ends_with(hist, "]") { return "" }
let ts: Int = time_now()
let tags: String = "[\"conv-history\",\"persistent\"]"
let node_id: String = engram_node_full(
let discard: String = engram_node_full(
hist, "Conversation", "conv:history",
el_from_float(0.7), el_from_float(0.8), el_from_float(0.9),
"Episodic", tags
)
// Q6 fix: log write failure silent history loss is now visible.
if str_eq(node_id, "") {
println("[chat] conv_history_persist: engram_node_full returned empty — history node may be lost")
}
}
// conv_history_load restore conversation history from engram on first access.
// Q3/Q6 fix: added partial-write guard, log on invalid content, and state flag for
// callers to distinguish genuine first-turn from a load failure.
// Returns the most recent "conv:history" node content, or "" if none found.
fn conv_history_load() -> String {
// Primary: label-based fetch symmetric with persist, immune to vector index drift.
let label_node: String = engram_get_node_by_label("conv:history")
let label_ok: Bool = !str_eq(label_node, "") && !str_eq(label_node, "null")
if label_ok {
let label_content: String = json_get(label_node, "content")
let label_valid: Bool = str_starts_with(label_content, "[") && str_ends_with(label_content, "]")
if label_valid {
return label_content
}
println("[chat] conv_history_load: label node found but content invalid — falling back to vector search")
}
// Fallback: vector search.
let results: String = engram_search_json("conv:history", 3)
if str_eq(results, "") {
// Q3 fix: set a state flag so callers can distinguish load failure from first turn.
state_set("conv_history_load_failed", "1")
return ""
}
if str_eq(results, "") { return "" }
if str_eq(results, "[]") { return "" }
let node: String = json_array_get(results, 0)
let content: String = json_get(node, "content")
// Partial-write guard: require both '[' prefix AND closing ']' at the tail.
// str_ends_with guards against embedded ']' in content fooling the check.
if !str_starts_with(content, "[") || !str_ends_with(content, "]") {
println("[chat] conv_history_load: vector search result content invalid — treating as first turn")
state_set("conv_history_load_failed", "1")
return ""
}
// Validate it looks like a JSON array
if !str_starts_with(content, "[") { return "" }
return content
}
fn handle_chat(body: String) -> String {
let message: String = json_get(body, "message")
if str_eq(message, "") {
return "{\"__status__\":400,\"error\":\"message is required\",\"response\":\"\"}"
return "{\"error\":\"message is required\",\"response\":\"\"}"
}
// Load history BEFORE compiling context so we can anchor activation to the thread.
// Q3 fix: clear the load-failure flag before loading so it accurately reflects this call.
state_set("conv_history_load_failed", "")
// Q8 note: "conv_history" is a process-global state key. Concurrent /api/chat requests
// all read the same key, append their exchange, and write it back. Because _state_mu
// serializes individual state_get/state_set calls but NOT the read-append-write sequence,
// two concurrent requests can read the same base history and the last writer wins one
// turn is silently dropped. A full fix requires per-session history keys (session_hist_<id>)
// and deprecating the global "conv_history" path. Callers using session_id are not affected.
let state_hist: String = state_get("conv_history")
let stored_hist: String = if str_eq(state_hist, "") { conv_history_load() } else { state_hist }
let hist_load_failed: Bool = str_eq(state_get("conv_history_load_failed"), "1")
let hist_len: Int = if str_eq(stored_hist, "") { 0 } else { json_array_len(stored_hist) }
// Thread-aware activation: short/ambiguous messages (continuations like "go on",
@@ -663,13 +612,7 @@ fn handle_chat(body: String) -> String {
let act_out: String = if act_ok { activation_nodes } else { "[]" }
strengthen_chat_nodes(act_out)
// Q3 fix: surface history load failure in the response envelope so callers can
// show a "starting fresh — could not load previous conversation" indicator.
let hist_warning: String = if hist_load_failed {
",\"history_load_failed\":true"
} else { "" }
return "{\"response\":\"" + safe_response + "\",\"model\":\"" + model + "\",\"activation_nodes\":" + act_out + hist_warning + "}"
return "{\"response\":\"" + safe_response + "\",\"model\":\"" + model + "\",\"activation_nodes\":" + act_out + "}"
}
fn handle_see(body: String) -> String {
@@ -1674,13 +1617,6 @@ fn auto_persist(req: String, resp: String) -> Void {
"Episodic",
tags
)
// CRITICAL BUG fix: log conv_node_id failure OUTSIDE the is_bell block.
// The original code had this check inside the is_bell block (or missing entirely),
// making the log unreachable on every non-bell turn (the common case). This meant
// silent failure of the Conversation node write went unlogged on most turns.
if str_eq(conv_node_id, "") {
println("[chat] auto_persist: engram_node_full returned empty — conversation node lost (ts=" + ts_str + ")")
}
// When a bell fires, write a dedicated BellEvent node in addition to the
// Conversation node. This makes distress moments directly findable by label
-23
View File
@@ -148,14 +148,6 @@ fn load_identity_context() -> Void {
println("[soul] identity context loaded (" + int_to_str(str_len(ctx)) + " chars, " + int_to_str(parts_count) + " nodes)")
}
// Q6 fix: warn when all three identity node fetches return empty. For genesis this
// indicates a corrupted or missing graph. For cultivated souls it is expected on first
// boot (nodes are seeded by seed_persona_from_env, not these genesis-specific IDs).
// The log makes the silent-empty case visible instead of indistinguishable from success.
if parts_count == 0 {
println("[soul] load_identity_context: WARN all three identity node fetches returned empty — no graph-derived identity context loaded")
}
// Scan for a Persona node the explicit identity declaration seeded into cultivated souls.
// Stored at seeding time with label "soul:persona" and node_type "Persona".
// genesis derives identity from the graph directly; cultivated souls have this node seeded.
@@ -170,12 +162,6 @@ fn load_identity_context() -> Void {
println("[soul] persona node loaded (" + int_to_str(str_len(p_content)) + " chars)")
}
}
// Q6 fix: if neither identity nodes nor persona node were loaded, log explicitly.
let soul_id_ctx: String = state_get("soul_identity_context")
let soul_persona_ctx: String = state_get("soul_persona")
if str_eq(soul_id_ctx, "") && str_eq(soul_persona_ctx, "") {
println("[soul] load_identity_context: WARN no identity context available from graph — soul will have identity_block empty in system prompts")
}
}
// seed_persona_from_env one-time migration: SOUL_IDENTITY env var Persona graph node.
@@ -341,15 +327,6 @@ fn layered_cycle(raw_input: String) -> String {
// TODO: wire directly when imprint_respond gains system_override param (imprint.el change).
// ISSUE 3 TODO: no semantic crisis detection. Keyword-only means signals that evade
// the phrase list pass with zero augmentation. Semantic layer = separate decision.
//
// Q8 race documentation: "layered_cycle_safety_system_addendum" is a shared process-global
// state key. Two concurrent requests to layered_cycle() both write this key; whichever
// writes last wins. The concurrent build_system_prompt() read in chat.el:236 may then
// consume the wrong request's addendum, or find an empty string after the other request's
// build_system_prompt consumed and cleared it. Mitigation: under http_serve_async, the
// layered_cycle path and the /api/chat path are different endpoints (typically); true
// concurrent layered_cycle calls are uncommon. A robust fix requires per-request state
// scoping which needs C runtime support (e.g. a request-id-keyed addendum map).
let augmented_addendum: String = safety_augment_system("", raw_input)
state_set("layered_cycle_safety_system_addendum", augmented_addendum)