v5.0.0: GenerateTelosSummary.ts has 3 parser bugs that drop content from auto-generated PRINCIPAL_TELOS.md

[jdrolls]

Summary

Three independent parser bugs in PAI/TOOLS/GenerateTelosSummary.ts cause the auto-generated PRINCIPAL_TELOS.md (loaded into the DA's context at every session start via CLAUDE.md @-import) to silently miss most of the user's TELOS data.

The DA boots without seeing the user's missions, active goals, narratives, challenges, or formative experiences — even when those source files are fully populated.

---

Bug 1 — parseItems() only matches bullet form; standard PAI source files use header form

File: PAI/TOOLS/GenerateTelosSummary.ts, function parseItems (lines 47-59)

Symptom: Generated PRINCIPAL_TELOS.md has empty ## Missions, ## Active Goals (2026), ## Active Narratives, ## Personal Challenges, and ## Core Models sections — even when MISSION.md, GOALS.md, NARRATIVES.md, CHALLENGES.md, and MODELS.md are fully populated.

Root cause: parseItems matches only this single regex:

const match = line.match(/^-\s+\*?\*?(\w+)\*?\*?:\s*(.+)/);

That matches bullet form: - **M0**: Family Presence. But the canonical PAI source files (after a real /interview run, after /migrate, or after the user fills them in by hand) use header form:

### M0: Family Presence

Be fully present for family — attend every kids' event...

parseItems returns zero items for these files, so all five callers (parseMissions, parseGoals, parseNarratives, parseChallenges, parseModels) yield empty arrays. Note: parseProblems and parseStrategies already handle the header form correctly via their own regex — parseItems is the inconsistent one.

Proposed fix: Extend parseItems to try header form when bullet form misses:

function parseItems(content: string): ParsedItem[] {
  const items: ParsedItem[] = [];
  for (const line of content.split('\n')) {
    // Bullet: "- **M0**: text"
    const bullet = line.match(/^-\s+\*?\*?(\w+)\*?\*?:\s*(.+)/);
    if (bullet) {
      items.push({ id: bullet[1], text: bullet[2].trim() });
      continue;
    }
    // Header: "## M0: text" / "### G0: text" / "#### MO0: text"
    const header = line.match(/^#{2,4}\s+([A-Z]+\d+):\s*(.+?)(?:\s*\(.*\))?\s*$/);
    if (header) {
      items.push({ id: header[1], text: header[2].trim() });
    }
  }
  return items;
}

---

Bug 2 — Hardcoded {{PRINCIPAL_FULL_NAME}} placeholder in generated output

File: PAI/TOOLS/GenerateTelosSummary.ts, line 220

Symptom: Generated PRINCIPAL_TELOS.md starts with the literal string:

# Principal TELOS — {{PRINCIPAL_FULL_NAME}}

The {{PRINCIPAL_FULL_NAME}} token is never substituted because there is no substitution mechanism in this generator (it's a write, not a render). Different problem from #1135 (which is about the installer not substituting in PAI_SYSTEM_PROMPT.md); this is a generator that hardcodes a placeholder it can't resolve.

Proposed fix: Read the principal name from settings.json daidentity (or PRINCIPAL_IDENTITY.md frontmatter) at generation time:

import { readFileSync } from 'fs';
const settings = JSON.parse(readFileSync(join(process.env.HOME!, '.claude/settings.json'), 'utf-8'));
const principalName = settings.principal?.fullName || settings.principal?.name || 'Principal';
// then:
'# Principal TELOS — ' + principalName,

Or accept it as an environment variable / config injection.

---

Bug 3 — parseTraumas() requires ID-prefixed items but the bootstrap TRAUMAS.md uses plain titles

File: PAI/TOOLS/GenerateTelosSummary.ts, function parseTraumas (lines 189-193)

Symptom: Generated ## Formative Experiences (Traumas) section is empty even when TRAUMAS.md is fully populated.

Root cause: parseTraumas calls parseItems (which expects TR0, TR1, etc. prefixed items), but the canonical/bootstrap TRAUMAS.md uses plain titled headers:

## Major Transformations

### Religious Worldview Shift
**Age:** ~28 (after 7-year marriage)
**What happened:** ...

### Divorce
...

No TR# prefix, so parseItems returns zero, and the section in the generated summary is empty.

Proposed fix: Fall back to extracting ### Title headers and assigning synthetic TR0..TRn IDs:

function parseTraumas(): string[] {
  const content = readTelosFile('TRAUMAS.md');
  if (!content) return [];

  // Try ID-prefixed format first
  const items = parseItems(content);
  if (items.length > 0) {
    return items.map(i => `- **${i.id}**: ${truncate(i.text, 90)}`);
  }

  // Fallback: ### Title headers, assign synthetic TR0..TRn
  const headers = [...content.matchAll(/^###\s+(.+?)\s*$/gm)];
  return headers.map((m, i) => `- **TR${i}**: ${truncate(m[1].trim(), 90)}`);
}

---

Reproduce

# Use the bootstrap-default TRAUMAS.md and populate MISSION.md / GOALS.md with `### M0: Title` form
bun ~/.claude/PAI/TOOLS/GenerateTelosSummary.ts
cat ~/.claude/PAI/USER/TELOS/PRINCIPAL_TELOS.md
# Expect: Missions, Goals, Narratives, Challenges, Formative Experiences sections all empty
# Expect: header reads "# Principal TELOS — {{PRINCIPAL_FULL_NAME}}"

After applying the three patches above:

# Principal TELOS — <user's actual name>
## Missions
- **M0**: Family Presence
- **M1**: Financial Independence
...
## Formative Experiences (Traumas)
- **TR0**: Religious Worldview Shift
- **TR1**: Divorce
...

---

Impact

PRINCIPAL_TELOS.md is @-imported into the DA's system context on every session start (per CLAUDE.md). When these bugs fire, the DA boots blind to the user's missions, goals, narratives, challenges, and formative experiences — the entire point of the TELOS system. Users may not realize until they ask the DA something like "what's my current mission?" and get an unsteered answer.

---

Environment

• PAI v5.0.0
• Bun 1.3.8

Related

• v5.0.0: installer never substitutes {{VAR}} placeholders in PAI_SYSTEM_PROMPT.md #1135 — same anti-pattern ({{PLACEHOLDER}} literal that never gets resolved) but in a different file (PAI_SYSTEM_PROMPT.md) and from a different mechanism (installer skipping the file). Both fixes should land together as a "no literal {{}} tokens in shipped/generated PAI files" hygiene rule.

[wojteksbt]

Hi, the first bug duplicates #1113

[MHoroszowski]

Heads up — I implemented bug 1 (header-form support in parseItems) and bug 2 (placeholder substitution from PRINCIPAL_IDENTITY.md) on a private fork and they cleared 5 of the 9 empty sections in our generated PRINCIPAL_TELOS.md. Bug 3 (parseTraumas robustness) wasn't blocking us so we deferred it.

Implementation notes that may be useful upstream:

1. Bug 1 fix: added a second regex pass after the bullet match. Header regex: /^#{2,4}\s+([A-Z]+\d+)(?::|\s+[—-])\s*(.+?)(?:\s*\(.*\))?\s*$/ — supports ## ID:, ### ID:, and also accepts em-dash / hyphen as separators (real-world authoring tends toward em-dashes for readability).

2. Bug 2 fix: read **Name:** line from PRINCIPAL_IDENTITY.md Quick Reference. Falls back to Principal if not found. ~10 LOC.

Happy to PR these back if useful.

[danielmiessler]

Two of your three are fixed in the current parsers — header form and un-ID'd traumas both parse now, with an ID-less prose fallback on top. Your bug 2 was the sleeper: the placeholder survives because the installer's substitution pass never walks the TOOLS directory, so we're fixing it your suggested way — resolving the name at generation time instead of shipping a token the generator can't substitute. Genuinely useful report, all three were real.