soma-run

Skill: soma-run

Version: 1.1.0 v1.0.0: initial implementation — live-verified 2026-05-16 v1.1.0 changes (2026-05-16): FIX 1 — Winners-log written immediately after HW (not end-of-pipeline) to prevent hook text loss on session compaction FIX 2 — Intermediate output persistence: each agent output saved to temp vault note; cleaned up on success FIX 3 — TI drift detection: original_input vs ti_trend comparison; logged in evo-log FIX 4 — Score cross-validation: HW scores vs CR scores; degradation flagged in report FIX 5 — Retry logic: 1 retry per agent on timeout or abort sentinel (+30s timeout on retry) FIX 6 — CR quality violations gate: if CR detects violations, amend winners-log with warning

---

Purpose

Runs the full SOMA content pipeline in a single skill invocation:

[User input] → VALIDATE → TI → HW → [Winners-log] → CR → [Evo-logs] → [Report]

Key change from v1.0.0: winners-log is written immediately after HW completes, while hook texts are still in context. Evo-logs are written after CR completes.

---

Hard rules — zero hallucination

• Never fabricate agent outputs, scores, hook text, or log entries
• Every evo-log entry must contain only data extracted from actual agent responses
• If an agent times out or returns an error → log FAILED — never invent a plausible output
• All Obsidian paths are fixed (confirmed 2026-05-16) — do NOT invent new paths
• Retry is limited to 1 attempt per agent — never retry more than once

---

Confirmed constants (live-verified 2026-05-16)

TIMEOUTS (primary attempt):
  TI  → timeout_seconds: 180
  HW  → timeout_seconds: 120
  CR  → timeout_seconds: 120

TIMEOUTS (retry — +30s each):
  TI retry  → timeout_seconds: 210
  HW retry  → timeout_seconds: 150
  CR retry  → timeout_seconds: 150

EVO-LOG PATHS:
  TI  → agents/trend-intelligence/evo-log.md
  HW  → agents/hook-writer/evo-log.md
  CR  → agents/content-repurposer/evo-log.md

WINNERS-LOG PATH:
  → agents/hook-writer/winners-log.md
  → Threshold: score ≥ 17/20

TEMP NOTE PATHS (deleted after successful run):
  TI output  → temp/soma-run-{run_id}-ti.md
  HW output  → temp/soma-run-{run_id}-hw.md
  CR output  → temp/soma-run-{run_id}-cr.md

EVO-LOG FORMATS:
  TI: date | original_input (≤60 chars) | trend_found [DRIFT if detected] | confidence | angle_suggested | hook_writer_triggered
  HW: date | trend | platforms | scores | winner_platform | winner_score | flags
  CR: date | trend | platforms_completed | scores | flag | notes

WINNERS-LOG FORMAT:
  date | trend | platform | hook_text | score | pattern

HW SCORE PATTERN (regex): LI:\d+ X:\d+ YT:\d+ IG:\d+ TT:\d+

DRIFT THRESHOLD:
  If fewer than 3 words from original_input appear in ti_trend → flag as DRIFT

SCORE DEGRADATION THRESHOLD:
  If any platform score in CR is more than 2 points below HW score → flag as DEGRADED

ABORT SENTINELS (case-insensitive):
  - empty string or len < 50 characters
  - starts with "I cannot"
  - starts with "I don't have"
  - starts with "I'm unable"
  - starts with "I'm sorry, I"
  - contains "As an AI, I"

---

STEP 0 — Task List

Create tasks before starting:

• "VALIDATE — input gate"
• "TI — Trend Intelligence run"
• "HW — Hook Writer run + winners-log"
• "CR — Content Repurposer run"
• "LOG — Write evo-logs"
• "CLEANUP — Remove temp notes"
• "REPORT — Final summary"

Mark each in_progress before starting, completed when done.

---

STEP 1 — Determine Pipeline Scope

Full pipeline (default)

Run all three agents: TI → HW → CR.

Partial pipeline (scope override)

If user explicitly specifies a shorter scope:

User says	Scope
"samo TI" / "run TI only"	Run only TI. Stop after TI log.
"TI i HW" / "TI and HW" / "stop before CR"	Run TI → HW. Stop after HW log + winners-log.
(anything else)	Full pipeline: TI → HW → CR

Store as pipeline_scope: "TI" / "TI+HW" / "FULL".

---

STEP 2 — VALIDATE: Input Gate

2a — Validator recommendation

Before running, check if the user has already run pipeline-input-validator on this input. If they have not, recommend it:

"💡 Preporučujem da prvo pokreneš pipeline-input-validator na ovom inputu. Ako je status PASS ili WARN+, nastavi sa soma-run. Nastavljamo svejedno?"

If user confirms (or if they already have a PASS/WARN+ result) → proceed to 2b. If user has a WARN- or FAIL result → warn but allow override: "Input ima slab score. Sigurno želiš da ga pustiš kroz pipeline?"

2b — Minimum input check

Extract the raw trend input from the user's message. Apply:

Check	Abort condition
Empty	Input is empty or whitespace only → ABORT
Too short	Input is < 20 characters → ABORT: "Input je prekratak. Opiši trend konkretno."
Abort sentinel in input	Input contains an abort sentinel string → ABORT: "Input sadrži nevalidan sadržaj."

If input passes → store as {trend_input}. Store also as {original_input} (immutable copy, used for drift detection in Step 4). Proceed to Step 3.

---

STEP 3 — Generate run_id

Generate run ID using current date and time:

run_id = YYYY-MM-DD-HHMMSS   (e.g. 2026-05-16-143022)

Use today's actual date. Do not guess or fabricate. If unsure of current time, use YYYY-MM-DD only as the run_id.

Store as {run_id}. This ID appears in evo-log entries and temp note paths.

Initialize retry counters: {ti_retries} = 0, {hw_retries} = 0, {cr_retries} = 0.

---

STEP 4 — TI: Run Trend Intelligence

4a — Mark task in_progress

4b — Build TI message

Construct the message as follows (date injection is mandatory):

Today is {YYYY-MM-DD}. {trend_input}

Example:

Today is 2026-05-16. Anthropic released Claude Sonnet 4 — 40% SWE-bench improvement.

CRITICAL: The Today is {date} prefix MUST be included. Without it, TI runs without date context and may misclassify freshness. Confirmed bug on 2026-05-15.

4c — Call TI (with retry)

Primary attempt:

as_chat_with_agent(
  agent_name:      "Trend Intelligence",
  message:         "Today is {YYYY-MM-DD}. {trend_input}",
  timeout_seconds: 180
)

On ABORT or timeout (retry — max 1): If {ti_retries} == 0:

• Set {ti_retries} = 1
• Wait 5 seconds
• Retry with timeout_seconds: 210
• Log: "⚠️ TI retry 1/1 — original attempt failed."

If retry also fails → mark TI as FAILED. Report ABORT. Stop pipeline. Do NOT attempt a 3rd call.

4d — Capture output

Store the full reply text as {ti_output}.

4e — Save TI output to temp note

Call obsidian_create_note:

path:    "temp/soma-run-{run_id}-ti.md"
content: "# TI Output — soma-run {run_id}\n\n{ti_output}"

If this call fails → log warning but continue. Temp note is a safety net, not a blocker.

4f — Validate TI output

Check {ti_output} against abort sentinels:

• If ABORT condition matched → mark TI as FAILED. Stop pipeline. Report: "⛔ TI vrati prazan ili nevalidan output. Pipeline abortiran."
• If OK → proceed to 4g.

4g — Extract TI data

From {ti_output}, extract:

• {ti_trend}: the trend name/title TI identified (first headline or sentence)
• {ti_confidence}: confidence rating (⭐⭐⭐ = HIGH, ⭐⭐ = MED, ⭐ = LOW/EVERGREEN)
• {ti_angle}: the content angle TI suggested
• Set {ti_status} = "yes" (hook_writer_triggered)

4h — Drift detection (FIX 3)

Compare {original_input} to {ti_trend}:

1. Tokenize both strings (split on spaces, lowercase, strip punctuation)
2. Count how many tokens from {original_input} appear in {ti_trend}

Drift rule:

• If < 3 tokens overlap → set {drift_flag} = "DRIFT", {drift_detected} = true
• Otherwise → set {drift_flag} = "none", {drift_detected} = false

If drift detected, log to console:

⚠️ DRIFT DETECTED
Original input : {original_input (first 60 chars)}
TI trend found : {ti_trend}
Overlap tokens : {N}

The pipeline does NOT abort on drift — it continues but records it in the evo-log and the final report.

---

STEP 5 — HW: Run Hook Writer + Write Winners-log

Skip if pipeline_scope == "TI".

5a — Mark task in_progress

5b — Call HW (with retry)

Pass the full TI output as the message. Do not summarize or truncate.

Primary attempt:

as_chat_with_agent(
  agent_name:      "Hook Writer",
  message:         {ti_output},
  timeout_seconds: 120
)

On ABORT or timeout (retry — max 1): If {hw_retries} == 0:

• Set {hw_retries} = 1
• Wait 5 seconds
• Retry with same message, timeout_seconds: 150
• Log: "⚠️ HW retry 1/1 — original attempt failed."

If retry also fails → mark HW as FAILED. Log TI evo-log only. Stop pipeline. Report: "⛔ HW vrati nevalidan output. TI je logiran. Pipeline abortiran."

5c — Capture output

Store the full reply text as {hw_output}.

5d — Save HW output to temp note (FIX 2)

Call obsidian_create_note:

path:    "temp/soma-run-{run_id}-hw.md"
content: "# HW Output — soma-run {run_id}\n\n{hw_output}"

If this call fails → log warning and continue. Hook texts are still in {hw_output} in active context — proceed. Temp note is a safety net only.

5e — Extract scores from HW output

Scan {hw_output} for the score pattern LI:\d+ X:\d+ YT:\d+ IG:\d+ TT:\d+.

• If pattern found → extract individual platform scores:

  {hw_scores_raw} = "LI:19 X:18 YT:17 IG:17 TT:18"   (example)
  {hw_scores} = { LI: 19, X: 18, YT: 17, IG: 17, TT: 18 }
  

• If pattern NOT found → set {hw_scores_raw} = "UNSCORED", {hw_scores} = null.

Determine winner platform:

• If {hw_scores} is not null → find platform with highest score. Ties: prefer LinkedIn > X > YouTube > Instagram > TikTok. Store as {hw_winner_platform} and {hw_winner_score}.
• If {hw_scores} is null → {hw_winner_platform} = "n/a", {hw_winner_score} = "n/a".

Determine flags:

• If {hw_scores} is null → flag = "UNSCORED"
• If all platform scores identical → flag = "SINGLE_HOOK_BUG" (same hook on all platforms)
• If any q

---

Content truncated.