> ## Documentation Index
> Fetch the complete documentation index at: https://prism.ntecdev.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# ADR-15: E3 closed — PID resolution rule: exact → case-insensitive → ask, never fuzzy

<Info>**Status:** `accepted` · **ADR-15** · Filed 2026-04-17</Info>

## Decision

*(none)*

## Rationale

E3 was the third of my Plan #3 Wave E open questions. Plan #3 got superseded but E3's text lived on in orphaned TODOs, and Donna's cold-start read it as blocking Wave B drift detection work. The rule itself is straightforward and deserves codification: exact match is the 95% path, case-insensitive catches filesystem-casing variance between OSes and careless typing, and escalation prevents the one catastrophic failure mode (silently attaching a session's records to the wrong project). This rule is the operational safety net for cross-agent, cross-machine continuity. The asymmetry — favor asking over guessing — is the right posture for a methodology tool whose entire value proposition is accurate shared memory.

## Alternatives Considered

Fuzzy matching (Levenshtein, prefix) — rejected; silent wrong-PID resolution is the worst failure mode in a shared-memory system because it corrupts one project's records with another's. Every ambiguous case must ask, never guess. Hardcode PID in a per-project file — rejected per BIOS v2 design; PID in CLAUDE.md was the exact anti-pattern we just fixed. Require explicit --pid flag on every session start — rejected; high friction, violates Context Is Currency, defeats the point of dir-based resolution. Dual keys (exact + project metadata) — already the model; codifying the fallback order is the addition.

## Decision

PID resolution rule for `prism_start` when called without explicit PID: (1) Exact match — compare current project dir basename to each registered project's name and to `basename(project_dir)`. First match wins. (2) Case-insensitive match — if step 1 returns zero hits, retry with case-folded comparison. First match wins. (3) Escalate — if step 2 returns zero or multiple hits, do NOT guess. Return an error payload listing candidate PIDs and prompting the user to call `prism_start` with an explicit PID. Specifically: never match on partial strings, never Levenshtein, never "closest fit." The "PrismGR vs Prism" case that triggered this question is resolved at the data layer, not the algorithm layer — PrismGR is not a project (legacy rename, see correction memory 2026-04-17), so there is no ambiguity to resolve.