task-start

name	task-start
version	1.2.0
category	workflow
description	MANDATORY first action for every task. Bundles the complete pre-task sequence into one reference. Load this BEFORE any other tool call — it prescribes the exact sequence that follows.
pinned	true
aliases	["session-start-discipline"]
related_skills	["agent-flow","survey-before-action","change-checklist","reflexion-check","loop-governance","task-persistence"]

Task Start — Mandatory First Action

Nothing comes before this skill. No tool calls, no file reads, no commands. This is the first and only thing you do when a new task arrives.

The Sequence

Execute these steps in order. Do not skip, batch, or defer any step.

Step 1: Cache search

mcp_loop_governance_cache_search(query="<the user's request>")

Learn from past cycles before starting work.

Step 2: Begin governance

mcp_loop_governance_begin_change(task_id="<short-name>", description="<what you are about to do>")

Mandatory before any code/config change. Write tools are blocked without this.

Step 3: Load always skills

Call skill_view(name) for every skill listed in the always section of .hermes-cortex/skills.yaml (or skills.yaml in the current working directory).

These skills define HOW you think and work — they are active context for the entire task, not one-time loads. Do not skip this step. (Formerly its own skill session-start-discipline — merged here 2026-08-20.)

Marker mechanics: the governance enforcer auto-creates a per-session marker at ~/.hermes-cortex/state/skills-loaded/<session-id> once all always-section skills are loaded via actual skill_view() calls. Never touch the marker — a bare file fails content verification. Concurrent sessions each own their marker file, so no session can stomp another's proof. If write tools are blocked with "session skills not fully loaded":

• The per-session loaded-set lives in the GATEWAY PROCESS's memory. If the set still holds all 7 (no gateway restart since you loaded them), ONE serial skill_view(name='<any-always-skill>') call re-triggers the auto-create (batching all loads in one turn does NOT trigger it — a serial extra call does).
• After a GATEWAY RESTART or a DEPLOY that changed a skill file (the marker pins a fingerprint of skill mtimes), the in-memory set is empty and the stored fingerprint is stale — you must re-load ALL 7 always-skills via skill_view() (they are read-only, so the gate does not block them); the 7th call regenerates the marker. This is the "7/7 loaded ✅ but still blocked" state — the message's ✅ reflects the current load, the block reflects the stale marker.

Step 4: Load domain skills (new Phase 0a)

After classification but BEFORE begin_change, run survey-before-action Phase 0a — the operation-type × file-extension domain-skill mapping table. Load ALL matching skills. Discovery fallback: skills_list(category=<domain>).

Step 5: Select reasoning pattern

Load agent-flow and use its embedded reasoning-pattern table to choose:

• Plan-Execute-Verify (default) — write a plan, execute steps, verify each
• ReAct — debugging, exploration
• Reflexion — add to any pattern when quality is critical
• Tree of Thoughts — design decisions with trade-offs

State your choice: "Using Plan-Execute-Verify with Reflexion check." (The former reasoning-patterns skill was folded into agent-flow 2026-08-20.)

Step 6: Classify with agent-flow

Match the request against the 12 workflow patterns. This determines toolset, output format, and checklist.

Step 7: Load on-task skills

After classification, load every skill matching your classification from the on_task section of .hermes-cortex/skills.yaml. Then call skills_list() for the relevant category to discover skills not in the manifest.

Step 8: Restore cross-session todos

~/.hermes-cortex/scripts/task-db.py pending

If items exist, restore them to the in-memory todo list:

todo(todos=<json_items>, merge=true)

(From the former session-start-discipline skill — merged here 2026-08-20.)

Step 9: Survey before action

Load survey-before-action and run its checklist BEFORE creating any file, writing any code, or running any command. Search for existing resources first. Its repo-specific pre-flight (git search, Hermes boundary, deploy verification — formerly cortex-preflight) is embedded in the same skill since 2026-08-20.

Step 10: Work

Execute the task using the loaded skills, following the chosen reasoning pattern and the classified workflow pattern's checklist.

Step 11: Reflexion check before delivery

After completing the work but BEFORE presenting results:

1. Load reflexion-check and run the seven-question audit
2. Score confidence (HIGH / MEDIUM / LOW / ZERO)
3. If LOW or ZERO: fix before delivering

Step 12: Change checklist before end_change

For code/config/cron changes: load change-checklist and run all phases before calling end_change(). Phase 6 (Reflexion) is mandatory.

Step 13: Score and close

mcp_loop_governance_cycle_query(task_id="<task-id>")
mcp_loop_governance_feedback_accept(id=<cycle-id>, note="<summary>")
mcp_loop_governance_end_change(task_id="<task-id>")

Quick Reference Table

Step	Action	Tool/Skill
1	Cache search	cache_search()
2	Begin governance	begin_change()
3	Load always skills	skill_view() per skill in always section
4	Choose reasoning	agent-flow embedded pattern table
5	Classify workflow	agent-flow skill
6	Load on-task skills	skill_view() + skills_list()
7	Restore todos	task-db.py pending → todo()
8	Survey + preflight	survey-before-action (cortex-preflight merged in)
9	Work	(task execution)
10	Reflexion check	reflexion-check skill
11	Change checklist	change-checklist skill
12	Score and close	cycle_query → feedback_accept → end_change

Known Gaps & Recovery

• Date-format daemon session IDs: the enforcer's daemon guard protects cron_/bg_ prefixes only. Sessions with HERMES_CRON_SESSION=1 but date-format IDs bypass the guard; concurrent sessions can then overwrite the marker. Symptom: "7/7 loaded" immediately followed by "not fully loaded". Fix: re-call skill_view('<any>') to re-trigger auto-create.
• Subagents overwrite markers: delegate_task subagents run with date-format session IDs and load skills themselves, which can invalidate the parent's marker. Include the marker workaround in subagent context (see references/recovery-from-violation.md).
• Mid-session correction: if the user asks "did you load all skills?" mid-task, you skipped Step 3. Stop, load the always skills, and only then continue. See references/recovery-from-violation.md and references/survey-before-creating.md for the pre-creation survey discipline.

Why This Exists

The pre-task sequence was previously 8+ independent steps documented across 3 files (AGENTS.md, SOUL.md, skills). Agents skipped them because the path from "user says do X" to "actually doing X" required too many remembered actions. This skill bundles all steps into one reference so the agent only has to remember one call: skill_view('task-start').

Every existing rule still applies. Nothing is weakened. The difference is that the sequence is now a single skill invocation rather than 8 separate things an agent must independently remember.

On 2026-08-20 session-start-discipline (todo-restore + marker mechanics) was merged into this skill and deleted as a standalone always-skill — one load, one sequence, no duplicated marker lore.