self-healing-context

name	self-healing-context
description	Reconciliation pass. Reads ground-truth.md + crm/companies.md + crm/opportunities.md + crm/outreach.md and surfaces discrepancies. Never silently overwrites — every finding is a proposed delta the user reviews. Writes to <WORKING_FOLDER>/reconcile/<YYYY-MM-DD>-findings.md.
disable-model-invocation	false
allowed-tools	Read, Write, Bash(mkdir *), Bash(test *), Bash(date *)

Self-Healing Context

Workspace state drifts: ground-truth says "currently working with X" but X's engagement closed two months ago; ground-truth says "looking for full-time" but recent outreach is 90% consulting. This skill surfaces the drift; the user owns the reconciliation.

Inputs

$ARGUMENTS:

• Optional: --strict — surface even minor signal mismatches.
• Optional: --apply — for findings the user pre-approves, apply changes (rare; default is propose-only).

Examples

/career:reconcile
/career:reconcile --strict

Procedure

1. Load state

• ground-truth.md (all sections).
• crm/companies.md, crm/opportunities.md, crm/outreach.md.
• Recent week-logs (last 4 weeks) from plans/.
• Recent recommendations (last 30 days).

2. Run discrepancy checks

Each check has a description, the rule, and the action.

a) Engagement state vs ground-truth

• ground-truth looking-for says "currently working with client X" but crm/opportunities.md shows that engagement at status closed-won 90+ days ago, or closed-lost, or paused → propose ground-truth update.
• ground-truth lists "currently working with X" but no row for X exists in opportunities → propose adding the row or removing the claim.

b) Direction vs activity

• ground-truth says "looking for full-time" but ≥75% of last 30 days outreach used consulting templates → flag drift.
• ground-truth says "freelance only" but recent outreach mentions employment-shaped roles → flag drift.

c) Constraints vs targets

• ground-truth hard-constraints excludes a region (e.g. US-only) but recent recommendations include US-only companies → flag the recommendation logic, not ground-truth.
• ground-truth salary floor exceeds the highest known band on companies in recommendations/ → flag mismatch.

d) Stale CRM rows

• crm/companies.md rows with status engaged but no outreach.md activity in 60+ days → propose status change to paused or passed.
• crm/outreach.md rows in status sent for >45 days with no follow-up → propose status stalled.

e) Domain coverage

• Recent week-logs show outreach concentrated in 1 domain while ground-truth domains-of-interest lists 4+ → flag narrowness.
• ground-truth domains-of-interest lists a domain with no domain-notes file → suggest running /career:discover by-domain for it.

f) Companion configuration

• companions.transcription set in config but companion not actually installed → flag.
• Outreach drafts using business-brand exist but freelance-brand in ground-truth is empty → flag.

3. Severity

Tag each finding:

• critical — internal contradiction (claim X is currently true, evidence X is false).
• drift — direction has shifted; ground-truth probably needs editing.
• staleness — row hasn't been touched, low cost to fix.
• coverage — gap (a domain not yet researched, a config not set).

--strict surfaces signal mismatches that would otherwise be filtered (small drifts, single-row stale rows).

4. Write the findings file

# Reconcile findings — {{date}}

> Generated: {{ISO timestamp}}
> Sources: ground-truth.md ({{mtime}}), crm/* ({{summary}}), recent recommendations + week-logs.

## Critical

### {{finding-id}}: {{title}}
- evidence: {{specific row + file}}
- rule: {{which check fired}}
- proposed action: {{e.g. "edit ground-truth section `looking-for`"}}
- to apply: `{{exact next command}}`

## Drift

(same shape)

## Staleness

(same shape)

## Coverage gaps

(same shape)

## No discrepancies found

(if a category is clean, list it here so the user sees the system checked)

5. Optionally apply

If --apply and a finding has a clearly-mechanical action (e.g. flip outreach row from sent → stalled after 45 days):

• Show diff.
• Ask per-finding: y/n.
• Apply only confirmed.

Never auto-apply ground-truth edits — those always remain proposals.

6. Print summary

findings: reconcile/<date>-findings.md
critical: <n>
drift: <n>
staleness: <n>
coverage: <n>

next:
  critical → /career:ground-truth edit-section <section>
  staleness → /career:reconcile --apply (mechanical updates only)
  coverage → /career:discover by-domain "<missing>"

Guardrails

• Never silently overwrite ground-truth. Critical and drift findings are always proposals.
• Every finding cites a specific file + row. If you can't cite a source, the finding is unfounded — drop it.
• Don't moralize. "You haven't followed up with X in 50 days" is staleness, not a failure.
• Severity is conservative. Don't escalate staleness to critical because the user "should" act on it.

Failure modes

• Empty workspace (no CRM, no recommendations) → return "nothing to reconcile yet — run /career:discover and /career:research-company first".
• Ground-truth missing → bail; reconciliation needs ground-truth as the anchor.